An Introduction to Multivariate Statistical Analysis Third Edition
T. W. ANDERSON Stanford University Department of Statistics Stanford, CA
fXt.WILEY-
~INTERSCIENCE A JOHN WILEY & SONS, INC., PUBLICATION
Copyright © 2003 by John Wiley & Sons, Inc. All rights reserved. Published by John Wiley & Suns, Inc. Huboken, New Jersey Published sinll.lt'lIleolisly in CanOl"". No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or hy any me'IIlS, electronic, mechanical, photocopying, recording, scanning 0,' otherwise, except as pcnnitteu unuel' Section 107 or 11I1l of the 1'176 United States Copyright Act, without either the prior writ'en permission of the Publisher, or at thorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, '!7X-7S0-X400, fax 97R-7S0-4470, or on the web at www.copyright.com. Requests t'1 the .'ublisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, e-mail: permreq(alwiley.com. Limit of Liability IDisciaimer of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respe<;t to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales materials. The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Neither the publisher nor au'hor shall be liable for any loss of profit or any other commercial uamages, including but not limited to special, incidental, consequential, or other damages. For general inf()rmalion un our other proollcls and services please contact our Customer
Care Department within the U,S. at 877-762-2974, outside the U.S. at 317-572-3993 or fax .117-572-4002. Wiley also publishes its books in a variety of electronic formats, Some content that appears in print, however, may not he availahle in electronic format. Library oj Congress Cataloging-in-Publication Data
Anuerson,T. W. (The()uore Wilbur), 191~An introduction to multivariate statistical analysis I Theodore W. Anderson.-- 3rd ed. p. cm.-- (Wiley series in probability and mathematical statistics) Incluues bibliographical rdcrences anu inucx. ISBN 0-471-36091-0 (cloth: acid-free paper) 1. Multivariate analysis. I. Title. II. Series. QA278.A5162003 519.5'35--dc21 Printed in the United States of America 109il7654321
2002034317
To
DOROTHY
Contents I Preface to the Third Edition
i I I
I,
I I I I I I
I
I I I
I II
xv
Preface to the Second Edition
xvii
Preface to the First Edition
xix
1 Introduction
1
1.1. Multivariate Statistical Analysis, 1.2. The Multivariate Normal Distribution, 3 2
The Multivariate Normal Distribution
6
2.1. Introduction, 6 2.2. Notions of Multivariate Distributions, 7 2.3. The Multivariate Normal Distribution, 13 2.4. The Distribution of Linear Combinations of Normally Distributed Variates; Independence of Variates; Marginal Distributions, 23 2.5. Conditional Distributions and Multiple Correlation Coefficient, 33 2.6. The Characteristic Function; Moments, 41 2.7. Elliptically Contoured Distributions, 47 Problems, 56 3 Estimation of the Mean Vector and the Covariance Matrix
3.1.
66
Introduction, 66 vii
viii
CONTENTS
3.2.
Thl: Maximum Likdihoou Estimaturs uf thl: Mean Vector and the Covariance Matrix, 67 3.3. The Distribution of the Sample Mean Vector; Inference Concerning the Mean When the Covariance Matrix Is Known, 74 3.4. Theoretical Properties of Estimators of the Mean Vector, 83 3.5. Improved Estimation of the Mean, 91 3.6. Elliptically Contoured Distributions, 101 Problems, 108 4
The Distributions and Uses of Sample Correlation Coefficients
115
Introduction, 115 Curn.:lation Codficient of a Bivariate Sample, 116 Partial Correlation Coefficients; Conditional Distributions, 136 4.4. The Multiplc Corrclation Coefficient, 144 4.5. Elliptically Contoured Distributions, 158 Problems, 163
4.1. 4.2. 4.3.
5
The Generalized T 2 -Statistic
170
Introduction, 170 Derivation of the Generalized T 2-Statistic and Its Distribution, 171 5.3. Uses of the T 2-Statistic, 177 5.4. The Distribution of T2 under Alternative Hypotheses; The Power Function, 185 5.5. The Two-Sample Problem with Unequal Covariance Matrices, 187 5.6. Some Optimal Properties of the T 2-Test, 190 5.7. Elliptically Contoured Distributions, 199 Problems, 20 I
5.1. 5.2.
6
Classification of Observations 6.1. 6.2. (d.
The Problem of Classification, 207 Standards of Good Classification, 208 Procl:dures of C1assifil:atiun into Onc of Two Populations with Known Probability Distributions, 2]]
207
CONTENTS
ix
6.4.
Classification into One of Two Known Multivariate Normal Populations, 215 6.5. Classification into One of Two Multivariate Normal Populations When the Parameters Are Estimated, 219 6.6. Probabilities of Misclassification, 227 6.7. Classification into One of Several Populations, 233 6.8. Classification into One of Several Multivariate Normal Populations, 237 6.9. An Example of Classification into One of Several Multivariate Normal Populations, 240 6.10. Classification into One of Two Known Multivariate Normal Populations with Unequal Covariance Matrices, 242 Problems, 248 7 The Distribution of the Sample Covarir.nce Matrix and the Sample Generalized Variance
251
7.1. 7.2. 7.3. 7.4. 7.5. 7.6.
Introduction, 251 The Wishart Distribution, 252 Some Properties of the Wishart Distribution, 258 Cochran's Theorem, 262 The Generalized Variance, 264 Distribution of the Set of Correlation Coefficients When the Population Covariance Matrix Is Diagonal, 270 7.7. The Inverted Wishart Distribution and Bayes Estimation of the Covariance Matrix, 272 7.8. Improved Estimation of the Covariance Matrix, 276 7.9. Elliptically Contoured Distributions, 282 Problems, 285
8 Testing the General Linear Hypothesis; Multivariate Analysis of Variance
8.1. 8.2.
Introduction, 291 Estimators of Parameters in Multivariate Linear Regression, 292 8.3. Likelihood Ratio Criteria for Testing Linear Hypotheses about Regression Coefficients, 298 1),4. The Distribution of the Likelihood Ratio Criterion When the Hypothesis Is True, 304
291
x
CONTENTS ~.5.
An Asymptotic Expansion of the Distribution of the Likelihood Ratio Criterion, 316 8.6. Other Criteria for Testing the Linear Hypothesis, 326 8.7. Tests of Hypotheses about Matrices of Regression Coefficients and Confidence Regions, 337 8.8. Testing Equality of Means of Several Normal Distributions with Common Covariance Matrix, 342 8.9. Multivariate Analysis of Variance, 346 8.10. Some Optimal Properties of Tests, 353 8.11. Elliptically Contoured Distributions, 370 Problems, 3""4
9 Testing Independence of Sets of Variates
381
Introductiom, 381 The Likelihood Ratio Criterion for Testing Independence of Sets of Variates, 381 9.3. The Distribution of the Likelihood Ratio Criterion When the Null Hypothesis Is True, 386 Y.4. An Asymptotic Expansion of the Distribution of the Likelihood Ratio Criterion, 390 9.5. Other Criteria, 391 9.6. Step-Down Procedures, 393 9.7. An Example, 396 9.8. The Case of Two Sets of Variates, 397 9.9. Admissibility of the Likelihood Ratio Test, 401 9.10. Monotonicity of Power Functions of Tests of Independence of Sets, 402 9.1l. Elliptically Contoured Distributions, 404 Problems, 408 Y.!. Y.l.
10 Testing Hypotheses of Equality of Covariance Matrices and Equality of Mean Vectors and Covariance Matrices 10.1. 10.2.
10.3. 10.4.
Introduction, 411 Criteria for Testing Equality of Several Covariance Matrices, 412 Criteria for Testing That Several Normal Distributions Are Identical, 415 Distributions of the Criteria, 417
411
xi
CONTENTS
10.5. Asymptotic Expansions of the Distributions of the Criteria, 424 10.6. The Case of Two Populations, 427 10.7. Testing the Hypothesis That a Covariance Matrix Is Proportional to a Given Matrrix; The Sphericity Test, 431 10.8. . Testing the Hypothesis That a Covariance Matrix Is Equal to a Given Matrix, 438 10.9. Testing the Hypothesis That a Mean Vector and a Covariance Matrix Are Equal to a Given Vector and Matrix, 444 10.10. Admissibility of Tests, 446 10.11. Elliptically Contoured Distributions, 449 Problems, 454 11
Principal Components
459
11.1. Introduction, 459 11.2. Definition of Principal Components in the Populat;on, 460 11.3. Maximum Likelihood Estimators of the Principal Components and Their Variances, 467 11.4. Computation of the Maximum Likelihood Estimates of the Principal Components, 469 11.5. An Example, 471 11.6. Statistical Inference, 473 11.7. Testing Hypotheses about the Characteristic Roots of a Covariance Matrix, 478 11.S. Elliptically Contoured Distributions, 482 Problems, 483 12 Canonical Correlations and Canonical Variables 12.1. 12.2. 12.3. 12.4. 12.5. 12.6.
Introduction, 487 Canonical Correlations and Variates in the Population, 488 Estimation of Canonical Correlations and Variates, 498 Statistical Inference, 503 An Example, 505 Linearly Related Expected Values, 508
487
xii
CONTENTS
Reduced Rank Regression, 514 Simultaneous Equations Models, 515 Problems, 526 12.7. 12.8.
13 The Distributions of Characteristic Roots and Vectors
528
Introduction, 528 The Case of Two Wishart Matrices, 529 The Case of One Nonsingular Wishart Matrix, 538 Canonical Correlations, 543 Asymptotic Distributions in the Case of One Wishart Matrix, 545 13.6. Asymptotic Distributions in the Case of Two Wishart Matrices, 549 13.7. Asymptotic Distribution in a Regression Model, 555 B.S. Elliptically Contoured Distributions, 563 Problems, 567
13.1. 13.2. 13.3. 13.4. 13.5.
14
Factor Analysis
569
Introduction, 569 The Model, 570 Maximum Likelihood Estimators for Random Oithogonal Factors, 576 14.4. Estimation for Fixed Factors, 586 14.5. Factor Interpretation and Transformation, 587 14.6. Estimation for Identification by Specified Zeros, 590 14.7. Estimation of Factor Scores, 591 Problems, 593
14.1. 14.2. 14.3.
15 Patterns of Dependence; Graphical Models 15.1. 15.2. 15.3. 15.4. 15.5.
Introduction, 595 Undirected Graphs, 596 Directed Graphs, 604 Chain Graphs, 610 Statistical Inference, 613
Appendix A Matrix Theory A1. A2.
Definition of a Matrix and Operations on Matrices, 624 Characteristic Roots and Vectors, 631
595
xiii
CONTENTS
A.3. A.4. A.5.
Partitioned Vectors and Matrices, 635 Some Miscellaneous Results, 639 Gram-Schmidt Orthogonalization and the Solution of Linear Equations, 647
Appendix B Tables B.1. B.2. B.3.
t
B.4.
B.S. B.6. B.7.
651
Wilks'Likelihood Criterion: Factors C(p, m, M) to Adjust to X;.m' where M = n - p + 1, 651 Tables of Significance Points for the Lawley-Hotelling Trace Test, 657 Tables of Significance Points for the ·Bartlett-Nanda-Pillai Trace Test, 673 Tables of Significance Points for the Roy Maximum Root Test, 677 Significance Points for the Modified Likelihood Ratio Test of Equality of Covariance Matrices Based on Equal Sample Sizes, 681 Correction Factors for Significance Points for the Sphericity Test, 683 Significance Points for the Modified Likelihood Ratio Test I = Io, 685
References
687
Index
713
Preface to the Third Edition
For some forty years the first and second editions of this book have been used by students to acquire a basic knowledge of the theory and methods of multivariate statistical analysis. The book has also served a wider community of statisticians in furthering their understanding and proficiency in this field. Since the second edition was published, multivariate analysis has been developed and extended in many directions. Rather than attempting to cover, or even survey, the enlarged scope, I have elected to elucidate several aspects that are particularly interesting and useful for methodology and comprehension. Earlier editions included some methods that could be carried out on an adding machine! In the twenty-first century, however, computational techniques have become so highly developed and improvements come so rapidly that it is impossible to include all of the relevant methods in a volume on the general mathematical theory. Some aspects of statistics exploit computational power such as the resampling technologies; these are not covered here. The definition of multivariate statistics implies the treatment of variables that are interrelated. Several chapters are devoted to measures of correlation and tests of independence. A new chapter, "Patterns of Dependence; Graphical Models" has been added. A so-called graphical model is a set of vertices or nodes identifying observed variables together with a new set of edges suggesting dependences between variables. The algebra of such graphs is an outgrowth and development of path analysis and the study of causal chains. A graph may represent a sequence in time or logic and may suggest causation of one set of variables by another set. Another new topic systematically presented in the third edition is that of elliptically contoured distributions. The multivariate normal distribution, which is characterized by the mean vect.or and covariance matrix, has a limitation that the fourth-order moments of the variables are determined by the first- and second-order moments. The class .of elliptically contoured xv
xvi
PREFACE TO THE THIRD EDITION
distribution relaxes this restriction. A density in this class has contours of equal density which are ellipsoids as does a normal density, but the set of fourth-order moments has one. further degree of freedom. This topic is expounded by the addition of sections to appropriate chapters. Reduced rank regression developed in Chapters 12 and 13 provides a method of reducing the number of regression coefficients to be estimated in the regression of one set of variables to another. This approach includes the limited-information maximum-likelihood estimator of an equation in a simultaneous equations model. The preparation of the third edition has been benefited by advice and comments of readers of the first and second editions as well as by reviewers of the current revision. In addition to readers of the earlier editions listed in those prefaces I want to thank Michael Perlman and Kathy Richards for their assistance in getting this manuscript ready. T. W. Stanford, California February 2003
ANDERSON
Preface to the Second Edition
Twenty-six years have plssed since the first edition of this book was published. During that tim~ great advances have been made in multivariate statistical analysis-particularly in the areas treated in that volume. This new edition purports to bring the original edition up to date by substantial revision, rewriting, and additions. The basic approach has been maintained, namely, a mathematically rigorous development of statistical methods for observations consisting .of several measurements or characteristics of each subject and a study of their properties. The general outline of topics has been retained. The method of maximum likelihood has been augmented by other considerations. In point estimation of the m~an vector and covariance matrix alternatives to the maximum likelihood estimators that are better with respect to certain loss functions, such as Stein and Bayes estimators, have been introduced. In testing hypotheses likelihood ratio tests have been supplemented by other invariant procedures. New results on distributions and asymptotic distributions are given; some significant points are tabulated. Properties of these procedures, such as power functions, admissibility, unbiasedness, and monO tonicity of power functions, are studied. Simultaneous confidence intervals for means and covariances are developed. A chapter on factor analysis replaces the chapter sketching miscellaneous results in the first edition. Some new topics, including simultaneous equations models and linear functional relationships, are introduced. Additional problems present further results. It is impossible to cover all relevant material in this book; what seems most important has been included. For a comprehensive listing of papers until 1966 and books until 1970 the reader is referred to A Bibliography of Multivariate Statistical Analysis by Anderson, Das Gupta, and Styan (1972). Further references can be found in Multivariate Analysis: A Selected and xvii
xviii
PREFACE TO THE SECOND EDITION
Abstracted Bibliography, 1957-1972 by Subrahmaniam and Subrahmaniam (1973). I am in debt to many students, colleagues, and friends for their suggestions and assistance; they include Yasuo Amemiya, James Berger, Byoung-Seon Choi, Arthur Cohen, Margery Cruise, Somesh Das Gupta, Kai-Tai Fang, Gene Golub. Aaron Han, Takeshi Hayakawa, Jogi Henna, Huang Hsu, Fred Huffer, Mituaki Huzii, Jack Kiefer, Mark Knowles, Sue Leurgans, Alex McMillan, Masashi No, Ingram Olkin, Kartik Patel, Michael Perlman, Allen Sampson. Ashis Sen Gupta. Andrew Siegel, Charles Stein, Patrick Strout, Akimichi Takemura, Joe Verducci, Marlos Viana, and Y. Yajima. I was helped in preparing the manuscript by Dorothy Anderson, Alice Lundin, Amy Schwartz, and Pat Struse. Special thanks go to Johanne Thiffault and George P. H. Styan for their precise attention. Support was contributed by the Army Research Office, the National Science Foundation, the Office of Naval Research, and IBM Systems Research Institute. Seven tables of significance points are given in Appendix B to facilitate carrying out test procedures. Tables 1, 5, and 7 are Tables 47, 50, and 53, respectively, of Biometrika Tables for Statisticians, Vol. 2, by E. S. Pearson and H. O. Hartley; permission of the Biometrika Trustees is hereby acknowledged. Table 2 is made up from three tables prepared by A. W. Davis and published in Biometrika (1970a), Annals of the Institute of Statistical Mathematics (1970b) and Communications in Statistics, B. Simulation and Computation (1980). Tables 3 and 4 are Tables 6.3 and 6.4, respectively, of Concise Statistical Tables, edited by Ziro Yamauti (1977) and published by the Japanese Standards Association; this book is a concise version of Statistical Tables and Formulas with Computer Applications, JSA-1972. Table 6 is Table 3 of The Distribution of the Sphericity Test Criterion, ARL 72-0154, by B. N. Nagarscnker and K. C. S. PilIai, Aerospacc Research Laboratorics (1972). The author is indebted to the authors and publishers listed above for permission to reproduce these tables.
T. W. Stanford. California June
1984
ANDERSON
Preface to the First Edition
This book has been designed primarily as a text for a two-semester course in multivariate statistics. It is hoped that the book will also serve as an introduction to many topics in this area to statisticians who are not students and will be used as a reference by other statisticians. For several years the book in the form of dittoed notes has been used in a two-semester sequence of graduate courses at Columbia University; the first six chapters constituted the text for the first semester, emphasizing correlation theory. It is assumed that the reader is familiar with the usual theory of univariate statistics, particularly methods based on the univariate normal distribution. A knowledge of matrix algebra is also a prerequisite; however, an appendix on this topic has been included. It is hoped that the more basic and important topics are treated here, though to some extent the coverage is a matter of taste. Some llf the more recent and advanced developments are only briefly touched on in the late chapter. The method of maximum likelihood is used to a large extent. This leads to reasonable procedures; in some cases it can be proved that they are optimal. In many situations, however, the theory of desirable or optimum procedures is lacking. Over the years this manuscript has been developed, a number of students and colleagues have been of considerable assistance. Allan Birnbaum, Harold Hotelling, Jacob Horowitz, Howard Levene, Ingram OIkin, Gobind Seth, Charles Stein, and Henry Teicher are to be mentioned particularly. Acknowledgements are also due to other members of the Graduate Mathematical
xix
xx
PREFACE TO THE FIRST EDITION
Statistics Society at Columbia University for aid in the preparation of the manuscript in dittoed form. The preparation of this manuscript was supported in part by the Office of Naval Research.
T. W. Center for Advanced Study in the Behavioral Sciences Stanford, California December 1957
ANDERSON
CHAPTER 1
Introduction
1.1. MULTIVARIATE STATISTICAL ANALYSIS
Multivariate statistical analysis is concerned with data that consist of sets of measurements On a number of individuals or objects. The sample data may be heights and weights of some individuals drawn randomly from a population of schooi children in a given city, or the statistical treatment may be made on a collection of measurements, such as lengths and widths of petals and lengths and widths of sepals of iris plants taken from two species, or one may study the scores on batteries of mental tests administered to a number of students. The measurements made on a single individual can be assembled into a column vector. We think of the entire vector as an observation from a multivariate population or distribution. When the individual is drawn randomly, we consider the vector as a random vector with a distribution or probability law describing that population. The set of observations on all individuals in a sample constitutes a sample of vectors, and the vectors set side by side make up the matrix of observations. t The data to be analyzed then are thought of as displayed in a matrix or in several matrices. We shall see that it is helpful in visualizing the data and understanding the methods to think of each observation vector as constituting a point in a Euclidean space, each coordinate corresponding to a measurement or variable. Indeed, an early step in the statistical analysis is plotting the data; since tWhen data are listed on paper by individual, it is natural to print the measurements on one individual as a row of the table; then one individual corresponds to a row vector. Since we prefer to operate algebraically with column vectors, we have chosen to treat observations in terms of column vectors. (In practice, the basic data set may well be on cards, tapes, or disks.) An lntroductilm to Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0-471-36091-0 Copyright © 2003 John Wiley & Sons, Inc.
1
l.
INTRODUCTION
most statisticians are limited to two-dimensional plots, two coordinates of the observation are plotted in turn. Characteristics of a univariate distribution of essential interest are the mean as a measure of location and the standard deviation as a measure of variability; similarly the mean and standard deviation of a univariate sample are important summary measures. In multivariate analysis, the means and variances of the separate measurements-for distributions and for samples -have corresponding relevance. An essential aspect, however, of multivariate analysis is the dependence between the different variables. The dependence bet\veen two variables may involve the covariance between them, that is, the average products of their deviations from their respective means. The covariance standardized by the corresponding standard deviations is the correlation coefficient; it serves as a measure of degree of dependt:nce. A set of summary statistics is the mean vector (consisting of the univariate means) and the covariance matrix (consisting of the univariate variances and bivariate covariances). An alternative set of summary statistics with the same information is the mean vector, the set of· standard deviations, and the correlation matrix. Similar parameter quantities describe location, variability, and dependence in the population or for a probability distribution. The multivariate nornzal distribution is completely determined by its mean vector and covariance matrix, and the sample mean vector and covariance matrix constitute a sufficient set of statistics. The measurement and analysis of dependence between variables, between sets of variables, and between variables and sets of variables are fundamental to multivariate analysis. The multiple correlation coefficient is an extension of the notion of correlation to the relationship of one variable to a set of variables. The partial correlation coefficient is a measure of dependence between t\vo variables when the effects of other correlated variables have been removed. The various correlation coefficients computed from samples are used to estimate corresponding correlation coefficients of distributions. In this hook tests of hypotheses of indepemlence arc developed. The properties of the estimators and test proredures are studied for sampling from the multivariate normal distribution. A number of statistical problems arising in multivariate populations are straightforward analogs of problems arising in univariate populations; the suitable methods for handling these problems are similarly related. For example, in the univariate case we may wish to test the hypothesis that the mean of a variable is zero; in the multivariate case we may wish to test the hypothesiS that the vector of the means of several variables is the zero vector. The analog of the Student (-test for the first hypothesis is the generalized T 2 -test. The analysis of variance of a single variable is adapted to vector
1.2 THE MeLTIVARIATE NORMAL DISTRIBUTION
3
observations; in regression analysis, the dependent quantity may be a vector variable. A comparison of variances is generalized into a comparison of covariance matrices. The test procedures of univariate statistics are generalized to the multivariate case in such ways that the dependence between variables is taken into account. These methods may not depend on the coordinate system; that is, the procedures may be invariant with respect to linear transformations that leave the null hypothesis invariant. In some problems there may be families of tests that a~e invariant; then choices must be made. Optimal properties of the tests are considered. For some other purposes, however, it may be important to select a coordinate system so that the variates have desired statistical properties. One might say that they involve characterizations of inherent properties of normal distributions and of samples. These are closely related to the algebraic problems of canonical forms of matrices. An example is finding the normalized linear combination of variables with maximum or minimum variance (finding principal components); this amounts to finding a rotation of axes that carries the covariance matrix to diagonal form. Another example is Characterizing the dependence between two sets of variates (finding canonical correlations). These problems involve the characteristic roots and vectors of various matrices. The statistical properties of the corresponding sample quantities are treated. Some statistical problems arise in models in which means and covariances are restricted. Factor analysis may be based on a model with a (population) covariance matrix that is the sum of a positive definite diagonal matrix and a positive semidefinite matrix of low rank; linear str Jctural relationships may have a SImilar formulation. The simultaneous equations system of econometrics is another example of a special model.
1.2. THE MULTIVARIATE NORMAL DISTRIBUTION The statistical methods treated in this book can be developed and evaluated in the context of the multivariate normal distribution, though many of the procedures are useful and effective when the distribution sampled is not normal. A major reason for basing statistical analysis on the normal distribution is that this probabilistic model approximates well the distribution of continuous measurements in many sampled populations. In fact, most of the methods and theory have been developed to serve statistical analysis of data. Mathematicians such as Adrian (1808), Laplace (1811), Plana (1813), Gauss
4
INTRODUCTION
(1823), and Bravais (1846) studied the bivariate normal density. Francis Galton, th~ ~eneticist, introduced the ideas of correlation, regression, and homoscedasticity in the study 'of pairs of measurements, one made on a parent and OT]~ in an offspring. [See, e.g., Galton (1889).] He enunciated the theory of the multivariate normal distribution as a generalization of observed properties of s2.mples. Karl Pearson and others carried on the development of the theory and use of different kinds of correlation coefficients t for studying proble~ns in genetics, biology, and other fields. R. A. Fisher further developed methods for agriculture, botany, and anthropology, including the discriminant function for classification problems. In another direction, analysis of scores 0 ..1 mental tests led to a theory, including factor analysis, the sampling theory of which is based on the normal distribution. In these cases, as well as in agricultural experiments, in engineering problems, in certain economic problems, and in other fields, the multivariate normal distributions have been found to be sufficiently close approximations to the populations so that statistical analyses based on these models are justified. The univariate normal distribution arises frequently because the effect studied is the sum of many independent random effects. Similarly, the multivariate normal distribution often occurs because the multiple measurements are sums of small independent effects. Just as the central limit theorem leads to the univariate normal distriJution for single variables, so does the general central limit theorem for several variables lead to the .~: multivariate normal distribution. .. Statistical theory based on the normal distribution has the advantage that the multivariate methods based on it are extensively developed and can be studied in an organized and systematic way. This is due not only to the need for such methods because they are of practical us,:, but also to the fact that normal theory is amenable to exact mathematical treatment. The 'suitable methods of analysis are mainly based on standard operations of matrix algebra; the distributions of many statistics involved can be obtained exactly or at least characterized; and in many cases optimum properties of procedures can be deduced. The point of view in this book is to state problems of inference in terms of the multivariate normal distributions, develop efficient and often optimum methods in this context, and evaluate significance and confidence levels in these terms. This approach gives coherence and rigor to the exposition, but, by its very nature, cannot exhaust consideration of multivariate stc:tistical analysis. The procedures are appropriate to many nonnormal distributions,
tFor a detailed study of the development of the ideas of correlation, see Walker (1931).
1.2 THE MULTIVARIATE NORMAL DISTRIBUTION
5
but their adequacy may be open to question. Roughly speaking, inferences about means are robust because of the operation of the central limit theorem, but inferences about covariances are sensitive to normality, the variability of sample covariances depending on fourth-order moments. This inflexibility of normal methods with respect to moments of order greater than two can be reduced by including a larger class of elliptically contoured distributions. In the univariate case the normal distribution is determined by the mean and variance; higher-order moments and properties such as peakedness and long tails are functions of the mean and variance. Similarly, in the multivariate case the means and covariances or the means, variances, and correlations determine all of the properties of the distribution. That limitation is alleviated in one respect by consideration of a broad class of elliptically contoured distributions. That class maintains the dependence structure, but permits more general peakedness and long tails. This study leads to more robust methods. The development of computer technology has revolutionized multivariate statistics in several respects. As in univariate statistics, modern computers permit the evaluation of observed variability and significance of results by resampling methods, such as the bootstrap and cross-validation. Such methodology reduces the reliance on tables of significance points as well as eliminates some restrictions of the normal distribution. Nonparametric techniques are available when nothing is known about the underlying distributions. Space does not permit inclusion of these topics as welt ,as o.\her considerations of data analysis, such as treatment of outliers and)ransformations of variables to approximate normality and homoscedasticity'.. The availability of modern computer facilities makes possible the analysis of large data sets and that ability permits the application of multivariate methods to new areas, such as image analysis, and more effective a,nalysis of data, such as meteorological. Moreover, new problems of statistical analysis arise, such as sparseness of parameter or data matrices. Because hardware ~.',. and software development is so explosive and programs require specialized ~ knowledge, we are content to make a few remarks here and there about < computation. ~ackages of statistical programs are available for most of the methods.
CHAPTER 2
The Multivariate Normal Distribution
2.1. INTRODUCTION In this chapter we discuss the multivariate normal distribution and some of its properties. In Section 2.2 are considered the fundamental notions of multivariate distributions: the definition by means of multivariate density functions, marginal distributions, conditional distributions, expected values, and moments. In Section 2.3 the multivariate normal distribution is defined; the parameters are shown to be the means, variances, and covariances or the means, variances, and correlations of the components of the random vector. In Section 2.4 it is shown that linear combinations of normal variables are normally distributed and hence that marginal distributions are normal. In Section 2.5 we see that conditional distributions arc also normal with means that are linear functions of the conditioning variables; the coefficients are regression coefficients. The variances, covariances, and correlations-called partial correlations-are constants. The multiple correlation coefficient is the maximum correlation between a scalar random variable and linear combination of other random variables; it is a measure of association between one variable and a set of others. The fact that marginal and conditional distributions of normal distributions are normal makes the treatment of this family of distributions coherent. In Section 2.6 the characteristic function, moments, and cumulants are discussed. In Section 2.7 elliptically contoured distributions are defined; the properties of the normal distribution are extended to this larger class of distributions.
All Introductioll to Multivariate Statistical Allll(l'sis. "Ihird t;ditioll. By T. W. Anderson ISBN 0-471·36091·0 Copyright © 2003 John Wiley & Sons. Inc.
6
2.2
NOTIONS OF MULTIVARIATE DISTRIBUTIONS
7
2.2. NOTIONS OF MULTIVARIATE DISTRIBUTIONS 2.2.1. Joint Distributions
In this section we shall consider the notions of joint distributions of several variables, derived marginal distributions of subsets of variables, and derived conditional distributions. First consider the case of two (real) random variables t X and Y. Probabilities of events defined in terms of these variables can be obtained by operations involving the cumulative distribution function (abbrevialed as edt),
F(x, y) = Pr{X ~x, Y ~y},
(1)
defined for every pair of real numbers (x, y). We are interested in cases where F(x, y) is absolutely continuous;· this means that the following partial derivative exists almost everywhere:
a 2 F(x,y) a-~ay
(2)
=f(x,y),
and
(3)
F(x,y)=
f
y
x
f
_00
f(u,v)dudv. _00
The nonnegative function f(x, y) is called the density of X and Y. The pair of random variables (X, Y) defines a random point in a plane. The probability that (X, Y) falls in a rectangle is
(4)
Pr{x~X~x+~x,y~Y~y+~y}
'=F(x + ~x,y + ~y) - F(x + ~x,y) - F(x, y + ~y) + F(x,y) =f
y+aYfx+ax
Y
f(u,v)dudv
x
> 0, ~y > 0). The probability of the random point (X, Y) falling in any set E for which the following int~gral is defined (that is, any measurable set E) is
(~x
(5)
Pr{ (X, Y)
E
E} = f fj(x, y) dxdy.
tIn Chapter 2 we shall distinguish between random variables and running variables by use of capital and lowercase letters, respectively. In later chapters we may be unable to hold to this convention because of other complications of notation.
8
THE MULTIVARIATE NORMAL DISTRIBUTION
This follows from the definition of the integral [:-s the limit of sums of the sort (4)]. If f(x, y) is continuous in both variables, the probability element f(x,y)6.y6.x is approximately the probability that X falls between x and x + 6.x and Y falls between y and y + 6.y since
(6)
Pr{x~X~x+6.x,y~Y~y+6.y}
y+aYfx+ax =f f(u,v)dudv y x
for some x o, Yo (x ~xo ~x + 6.x, y ~Yo ~y + 6.y) by the mean value theorem of calculus. Since feu, {) is continuous, (6) is approximately f(x, y) 6.x 6.y. In fact,
(7)
lim
ay--+o ay-->o
~Ipr{x ~X ~x + 6.x, y ~ y ~y + 6.y} x y - f( x, y) 6.x 6.yl =
o.
Now we consider the case of p random variables Xl' X z , ... , Xp. The cdf is
(8) defined for every set of real numbers xJ, ... ,x p. The density function, if F(x l , ... , xp) is absolutely continuous, is
(9) (almost everywhere), and
(10)
F(xJ, ... ,x p) =
f:··· f'j(ul, ... ,up)du! ···du p.
The probability of falling in any (measurable) set R in the Euclidean space is
p-dirri~nsional
The probability element f(x!, ... , x p) 6.x! ... 6.xp is approximately the probability Pr{x! ~ Xl ~ Xl + 6.x l , ..• , xp ~ Xp ~ xp + 6.x p} if f(Xl' ... ' x p) is
9
2.2 NOTIONS OF MULTIVARIATE DISTRIBUTIONS
continuous. The joint moments are defined as t
t","· f_C/~" "'X~pf(Xl'"'' xp) dx 00
(12)
t 'if
rfXf'''' X:P =
00
l ...
dx p.
2.2.2. Marginal Distributions Given the cdf of two randcm variables X, Y as being F(x, y), the marginal cdf of X is (13)
Pr{X ~x} = Pr{X ~X, Y ~ co}
=F(x,oo). Let this be F(x). Clearly (14)
F(x)
=
r t' _00
_00
f(u,v) dvdu.
We call (15)
t' f(u,v)dv=f(u), _00
say, the marginal density of X. Then (14) is (16)
F(x) =
r
_00
feu) duo
In a similar fashion we define G(y), the marginal cdf of Y, and g(y), the marginal density of Y. Now we turn to the general case. Given F(x l , ... , x p ) as the cdf of Xl"'" X p ' we; wish to find the marginal cdf of some of Xl"'" X p ' say, of X1, ... ,X, (r
= F( Xl"'" x"oo, ... ,00). The marginal density of XI"'" X, is
tI
will be used to denote mathematical expectation.
10
THE MULTIVARIATE NORMAL DISTRIBUTION
The marginal distribution and density of any other subset of Xl' ... ' Xp are obtained in the obviously similar fashion. The joint moments of a subset of variates can be cOlllputed from the marginal distribution; for example,
(19)
ex;"
···X,h,=
ext'
···X>X,u+ 1 ···X~
=fX ···fx-xx~' ... x;'f(XI, ... ,xp)dxl ... dxp -x
x
f··· f
=
-x
oc
x~, ···x>
-x
2.2.3. Statistical Independence Two random variables X, Y with cdf F(x, y) are said to be independent if
F(x,y) =F(x)G(y),
(20)
where F(x) is the marginal cdf of X and G(y) is the marginal cdf of Y. This implies that the density of X, Y is (21 )
x ) = a 2 F ( x, y) = a 2 F ( x) G (y) f( ,y ax ay ax ay dF(x) dG(y)
=
----cJXd"Y
=f(x)g(y). Conversely, if Itx, y)
(22)
F(x,y) =
= f(x)g(y), then
f,J~J(u,V)dudv= f",fJ(u)g(V)dUdV
= fJ(u) du fxg(V) dv=F(x)G(y). Thus an equivalent definition of independence in the case of densities existing is that f(x, y) = f(x)g(y). To see the implications of statistical independence, given any XI <x 2, YI
= j Y2jX,J(u,v)dudv= YI
XI
jX2f(u)du jY2g(v)dv XI
YI
2.2
11
NOTIONS OF MULTIVARIATE DISTRIBUTIONS
The probability of X falling in a given inteIVal and Y falling in a given inteIVal is the product of the probability of X falling in the inteIVal and the probability of Y falling in the other inteIVal. If the cc'f of XI"'" Xp is F(x l , ... , xp), the set of random variables is said to be mutually independent if
(24) where Fj(xj) is the marginal cdf of Xi' i = 1, ... , p. The set XI"'" X, is said to be independent of the set X,+ I' ... , Xp if
(25)
F(XI""'X p) =F(xl, ... ,x"oo, ... ,oo) ·F(oo, ... ,oo,x,+"""xf:)'
One result of independence is that joint moments factor. For p-xample, if XI' ... , Xp are mutually indepc:ndent, then
(26)
foo ... foo _00.
Xh, ... Xhp f (x ) ... f (x ) dx ... dx
_00
I
p
I
I
P
pip
p
=
n {
j~1
2.2.4. Conditional Distributions If A and B are two events such that the probability of A and B occurring simultaneously is P(AB) and the probability of B occurring is P(B) > 0, then the conditional probability of A occurring given that B has occurred is P(AB)/P(B). Suppose the event A is X falling in the inteIVal [XI' xz] and the event B is Y falling in [YI'YZ]' Then the conditional probability that X falls in [XI' x z], given that Y falls in [YI' Yz], is
(27)
f
X2fY2 f(u,v)dvdu
XI
YI
Y2
jy, g( v) dv Now let YI = Y, Yz = Y + ~y. Then for a continuous density,
(28)
y+a y
jy
g(v) dv=g(y*)
~Y,
12
THE MULTIVARIATE NORMAL DISTRIBUTION
where y ::;y* ::;y + 6.y. Also
(29)
r+
6Y
f( u, v) dv = flu, y* (u)] 6.y,
y
where y ::;y*(u)::;y
+ 6.y. Therefore,
(30) It will be llOticed that for fixed y and 6.y (> 0), the integrand of (30) behaves as a univariate density function. Now for y such that g(y) > 0, we define Pr{x i ::; X::; \2iY=y}, the probability that X lies between XI and X z , given that Y is y, as the limit of (30) as 6.y -> O. Thus
(31)
Pr{xl::;X::;xziY=y}=
f
X2
f(uly)du,
XI
where f(uiy) = feu, y) /g(y). For given y, feu iy) is a density funct;on and is called the conditional density of X given y. We note that if X and Yare independent, f(xiy) = f(x). In the general case of XI> ... ' XI' with cdf F(x l , ... , xp), the conditional density of Xl' . .. , X" given X'+I =X,+l' ... ' Xp =x p' is .
-r~ '!:-'
(32)
f oo ···foo f(ul, ... ,u"x,+I, ... ,xp)£..ul ... du, _00
_00
For a more general discussion of conditional prObabilities, the reader is referred to Chung (1974), Kolmogorov (1950), Loeve (1977),(1978), and Neveu (1965). 2.2.5. Transformation of Variables Let the density of XI' ... ' Xp be f(x l , ... , x p). Consider the p real-valued,' functions ~
(33)
i= 1, ... ,p.
We assume that the transformation from the x-space to the y-space is one-to-one;t the inverse transformation is (34)
i = 1, ... ,p.
'More precisely. we assume this is true for the part of the x·space for which !(XI' ... 'X p ) is positive.
2.3
,
13
THE MULTIVARIATE NORMAL DISTRIBUTION
Let the random variables Y!, ... , y" be defined by (35)
i = 1, . .. ,p.
Y; = Yi( XI'···' Xp),
Then the density of YI , ... , Yp is
where J(y!, ... , Yp) is the Jacobian
(37)
J(Yl' ... 'Yp) =
aX I aYI
ax! ayz
aX I ayp
ax z
ax z ayz
axz ayp
axp ay!
axp ayz
axp ayp
~ mod
We assume the cJerivatives exist, and "mod" means modulus or absolute value of the expression following it. The probability that (Xl' ... ' Xp) falls in a region R is given by (11); the probability that (Y" ... , Yp) falls in a region S is
If S is the transform of R, that is, if each point of R transforms by (33) into a point of S and if each point of S transforms into R by (34), then (11) is equal to (3U) by the usual theory of transformation of multiple integrals. From this follows the assertion that (36) is the density of Yl , ••• , Yp.
2.3. THE MULTIVARIATE NORMAL DISTRIBUTION The univariate normal density function can be written
ke- t,,(x-.6)2 = ke- t(x-.6 ),,(x-.6), where a is positive ang k is chosen so that the integral of (1) over the entire x-axis is unity. The density function of a multivariate normal distribution of XI" .. , Xp has an analogous form. The scalar variable x is replaced by a vector
t ~.
(2)
14
THE MULTIVARIATE NORMAL DISTRIBUTION
the scalar constant {3 is replaced by a vector
( 3)
and the positive constant ex is replaced by a positive definite (symmetric) matrix
( 4)
A=
all
a l2
alp
a 21
an
a 2p
a pl
a p2
a pp
The square ex(x - (3)2 = (x - (3 )ex(x - (3) is replaced by the quadratic form p
(5)
=
(x-b)'A(x-h)
1: i.j~
a;Jex;-bJ(xj-bJ. 1
Thus the density function of a p-variate normal distribution is ( 6) where K (> 0) is chosen so that the integral over the entire p-dimensional Euclidean space of Xl"'" Xp is unity. Written in matrix notation, the similarity of the multivariate normal density (6) to the univariate density (1) is clear. Throughout this book we shall use matrix notation and operations. Th ~ reader is referred to the Appendix for a review of matrix theory and for definitions of our notation for matrix operations. We observe that [(Xl"'" Xp) is nonnegative. Since A is positive definite, (x-b)'A(x-b) <,:0,
( 7)
and therefore the density is bounded; that is, (::\)
Now let us determine K so that the integral of (6) over the p-dimensional space is one. We shall evaluate (9)
K*=f
:x: rx;
···r •
x
e--'lx-b)'A(X-bldxp···dx l ·
ry;
2.3
THE MULTIVARIATE NORMAL DISTRIBUTION
15
We use the fact (see Corollary A1.6 in the Appendix) that if A is positive definite, there exists a nonsingular matrix C such that
(10)
C'AC=[,
where I denotes the identity and C' the transpose of C. Let
(11)
x-b=Cy,
where YI
(12)
y=
Then
(13)
(x - b)'A(x - b)
=
y'C'ACy
=
y'y.
The Jacobian of the transformation is
(14)
J= modlCI,
where modi CI indicates the absolute value of the determinant of C. Thus (9) becomes (15) We have
(16) where exp(z) = e Z • We can write (15) as
p
=modlCI
n{&} i-I
= modICI(27T)ip
16
THE MULTIVARIATE NORMAL DISTRIBUTION
by virtue of
foo
-1-
( 18)
&
I' dt= l. e-,I
-00
Corresponding to (10) is the determinantal equation
(19)
IC'I·IAI·ICI = III.
Since
(20)
IC'I =ICI,
and since III = 1, we deduce from (19) that
(21)
mod ICI =
11M.
Thus
(22) The normal density function is
-M - , - e_l(x-b)'A(x-b) ' . (27T )'P
(23)
We shall now show the significance of b and A by finding tne first and second moments of XI"'" Xp' It will be convenient to consider these random variables as constituting a random vector
x-(U
(24)
We shall define generally a random matrix and the expected value of a random matrix; a random vector is considered as a special case of a random matrix with one column. Definition 2.3.1.
A random matrix Z is a matrix
(25) afrandom variables ZII"'"
g=l, ... ,m, Zmn'
h=I, ... ,n,
2.3
17
THE MULTIVARIATE NORMAL DISTRIBUTION
If the random variables Zu, ... , Zmn can take on only a finite number of values, the random matrix Z can be one of a finite number of matrices, say Z(l), ... , Z(q). If the probability of Z = Z(i) is Pi' then we should like to define tfZ as r,?_IZUh. Then tfZ = (tfZ gh ). If the random variables Zl1" .. , Zmn have a joint density, then by operating with Riemann sums we can define tfZ as the limit (if the limit exists) of approximating sums of the kind occurring in the diserete case; then again tfZ = (tfZgh ). Therefore, in general we shall use the following definition: Definition 2.3.2. (26)
i
The expected value of a random matrix Z is
g=l, ... ,m,
h=l, ... ,n.
In particular if Z is X defined by (24), the expected value
(27)
is the mean or mean vector of X. We shall usually denote this mean vector by ..... If Z is (X - .... XX - tJ.)', the expected value is (28)
C(X) = tf(X - tJ.)(X - .... )' = [tf(Xi - ILi)(Xj - ILj)],
the covariance or covariance matrix of X. The ith diagonal element of this matrix, tf(Xi - ILi)2, is the variance of Xi' and the i, jth off-diagonal element, tf(Xj - ILiXXj - ILj)' is the covariance of Xi and Xj' i j. We shall usually denote the covariance matrix by 1:. Note that
*'
(29)
&,(X) = tf(XX' - tJ.X' -XtJ.'
+ tJ.tJ.')
= tfXX' -
tJ.tJ.'.
The operation of taking the expected value of a random matrix (or vector) satisfies certain rules which we can summarize in the following lemma: Lemma 2.3.1. If Z is an m X n random matrix, D is an I X m real matrix, E is an n X q real matrix, and F is an I X q real matrix, then
(30)
tf(DZE+F) =D(tfZ)E+F.
18
THE MULTIVARIATE NORMAL DISTRIBUTION
Proof The element in the ith row and jth column of S(DZE
+ F) is
s( EdiJ,Zhgegj+fij) = Edih(~'Lhg)egj+fij'
(31)
h. g
h. g
which is the element in the ith row and jth column of D( SZ)E
Lemma 2.3.2.
tffY=DSX+J,
(33)
C(Y) =DC(X)D'.
Proof The first assertion follows directly from Lemma second from
C (Y)
=
•
If Y = DX + J, where X is a random vector, then
(32)
(34)
+F.
2~3.1,
and the
If ( Y - If Y)( Y - tff Y )'
= S[DX +J - (D If X + J)] [DX + J- (DSX + J)]' = If[D(X- tffX)][D(X- If X)], = tff[D(X- SX)(X- IfX)'D'], which yields the right-hand side of (33) by Lemma 2.3.1.
•
When the transformation corresponds to (11), that is, X = CY + b, then If X = C S Y + b. By the transformation theory given in Section 2.2, the density of Y is proportional to (16); that is, it is
(35) The expected value of tbe ith component of Y is (36)
= o.
19
2.3 mE MULTIV ARIATE NORMAL DISTRIBUTION
The last equality follows because t Yie- ty1 is an odd function of Yi' Thus IY= O. Therefore, the mean of X, denoted by tJ., is
tJ.=SX=b.
(37)
From (33) we see that C(X) = C( SYY')C'. The i, jth element of SYY' is (38)
SYjY = ,
f··· f 00
00
_0
_00
YiY·
n {I --e-ty~ }dYl ... dy & P
'h=l
P
because the density of Y 1s (35). If i = j, we have
(39) 1
= --
&
foo Yi2e-'Y' dYi I 2
_00
=1. The last equality follows because the next to last expression is the expected value of the square of a variable normally distributed with mean 0 and variance 1. If i j, (38) becomes
"*
. n {I --f [2; P
00
h=l
_00
e-ty~ dYh }
h*i,j
=0, since the first integration gives O. We can summarize (39) and (40) as
(41)
SYY' =1.
Thus
(42)
S(X- tJ.)(X- tJ.)' = CIC' = CC'.
From (10) we obtain A = (C')-IC- 1 by multiplication by (C')-l on the left and by C- 1 on the right. Taking inverses on both sides of the equality IAlternatively, the last equality follows because the next to last expression is the expected value of a normally distributed variable with mean O.
20
THE MULTIVARIATE NORMAL DISTRIBUTION
gives us ( 43)
CC ' =A- 1 •
Thus, the covariance matrix of X is (44) From (43) we see that I is positive definite. Let us summarize these results. Theorem 2.3.1. If the density of a p-dimensional random vector X is (23), then the expected value of X is b and the covariance matril is A- I. Conversely, given a vector .... and a positive definite matrix I, there is a multivariate normal
density (45) such that the expected value of the vector with this density is .... and the covariance matrix is I.
We sh,111 denote the density (45) as n(xl .... , :~) and the distribution law as N( .... , I).
The ith diagonal element of thc covariance matrix, ifii , is the variance of the ith component of X; we may sometimes denote this by u/. The co"elation cotfficient between Xi and Xj is defined as
(46) This measure of association is symmetric in Xi and Xj: Pij = Pji. Since
(47) is positive definite (Corollary A.l.3 of the Appendix), the determinant
(48)
UiUj Pij 2
I= Ui2Uj 2( 1 - Pij2)
Uj
is positive. Therefore, -1 < Pij < 1. (For singular distributions, see Sectio~ 2.4.) The multivariate normal density can be parametrized by the means iJ-/. i = 1, ... , p, the variances u/, i = 1, ... , p, and the correlations Pij' i i, j = 1, ... , p. "
<$
21
2.3 THE MI}LTIVARIATE NORMAL DISTRIBUTION
As a special case of the preceding theory, we consider the bivariate normal distribution. The mean vector is (49)
the covariance matrix may be written
~
(Xl - ILI)(X Z ILZ»)
(50)
(Xz-ILz)
.~.
where a} is the variance of Xl' a} the variance of Xz, and correlation between XI and Xz. The inverse of (50) is
(51)
p
the
~-l =_1_ 1 - pZ
The density function of Xl and X z is (52)
Theorem 2.3.2. The correlation coefficient p of any bivariate distribution is invariant with respect to transformations = b;X; + c;, b; > 0, i = 1,2. Every function of the parameters of a bivariate normal distribution that is invariant with respect to sueh transformations is a function of p.
xt
xt
Proof. The variance of is b?a'/, i = 1,2, and the covariance of Xi and is blbzalaz p by Lemma 2.3.2. Insertion of these values into the definition of the correlation between Xi and Xi shows that it is p. If f( ILl' ILz, ai' a z, p) is inval iant with respect to such transformations, it must • be f(O, 0, 1, 1, p) by choice of b; = 1/ a; and c; = - ILJ a;, i = 1,2.
Xi
22
THE MULTIV ARIATE NORMAL DISTRIBUTION
Th..: ..:orrelation codfici..:nt p is the natural measure of association between XI and X 2 • Any function of the parameters of the bivariate normal distribution that is indep..:ndent of the scale and location parameters is a function of p. The standardized variable· (or standard score) is Y; = (Xj - f.L,)/U"j. The mean squared difference between the two standardized variables is (53) The smaller (53) is (that is, the larger p is), the more similar Yl and Yz are. If p> 0, XI and X 2 tend to be positively related, and if p < 0, they tend to be negatively related. If p = 0, the density (52) is the product 0: the marginal densities of XI and X 2 ; hence XI and X z are independent. It will be noticed that the density function (45) is constant on ellipsoids (54)
for every positive value of c in a p-dimensional Euclidean space. The center of each ellipsoid is at the point J.l. The shape and orientation of the ellipsoid are determined by I, and the size (given I) is determined by c. Because (54) is a sphcr..: if l = IT 21, /I(xi J.l, IT 21) is known as a spherical normal density. Let us considcr in detail the bivariate case of the density (52). We transform coordinates by (Xi - P)/U"i = Yi' i = 1,2, so that the centers of the loci of constant density are at the origin. These loci are defined by 1 . 2 Z) _ -1--z(YI-2PYIYz+Yz -c. -p
(55)
The intercepts on the YI-axis and Y2-axis are ~qual. If p> 0, the major axis of the ellipse is along the 45° line with a length of 2 c( 1 + p) , and the minor axis has a length of 2 c (1 - p) . If p < 0, the major axis is along the 135° line with a length of 2/ce 1 - p) , and the minor axis has a length of 2/c( 1 + p) . The value of p determines the ratio of these lengths. In this bivariate case we can think of the density function as a surface above the plane. The contours of ..:qual dcnsity are contours of equal altitude on a topographical map; they indicate the shape of the hill (or probability surface). If p> 0, the hill will tend to run along a line with a positive slope; most of the hill will be in the first and third quadrants. When we transform back to Xi = U"iYi + ILi' we expand each contour by a factor of U"i in the direction of the ith axis and shift the center to (ILl' ILz).
I
I
2.4
liNEAR COMBINATIONS; MARGINAL DISTRIBUTIONS
23
The numerical values of the cdf of the univariate normal variable are obtained from tables found in most statistical texts. The numerical values of (56)
where YI = (Xl - J.L1)/U I and Yz = (x z - J.Lz)/uz, can be found in Pearson (1931). An extensive table has been given by the National Bureau of Standards (1959). A bibliography of such tables has been given by Gupta (1963). Pearson has also shown that 00
(57)
F(xl,x Z) =
E piTi(YI)Tj(YZ)' j=O
where the so-called tetrachoric functions T/Y) are tabulated in Pearson (1930) up to TI9(Y). Harris and Soms (1980) have studied generalizations of (57).
2.4. THE DISTRIBUTION OF LINEAR COMBINATIONS OF NORMALLY DISTRIBUTED VARIATES; INDEPENDENCE OF VARIATES; MARGINAL DISTRIBUTIONS One of the reasons that the study of normal multivariate distributions is so useful is that marginal distributions and conditional distributions derived from multivariate normal distributions are also normal distributions. Moreover, linear combinations of multivariate normal variates are again normally distributed. First we shall show that if we make a nonsingular linear transformation of a vector whose components have a joint distribution with a normal density, we obtain a vector whose components are jointly distributed with a normal density. Theorem 2.4.1.
Let X (with p components) be distributed according to
N(fL, I). Then
(1)
y= ex
is distributed according to N(CfL, eIe') for
e nonsingular.
Proof The density of Y is obtained from the density of X, n(xi fL, I), by replacing x by
(2)
24
THE MULTIVARIATE NORMAL DISTRIBUTION
and multiplying by the Jacobian of the transformation (2), III ICI·III·IC'I
IIli IClc'lt·
The quadratic form in the exponent of n(xl j.L, I) is
(4) The transformation (2) carries Q into
(5)
Q = (C-Iy - j.L)'I-I(C-ly - j.L)
= (C-Iy - C-ICj.L)'I-I(C-1y - C-1Cj.L) = [C-I(y-Cj.L)]'I-I[c l(y-Cj.L)]
= (y - Cj.L)'( C-1)'I-1C-1(y - Cj.L) = (y - Cj.L)'( CIC') -I(y - Cj.L) since (C-I), = (C') -I by virtue of transposition of CC- I = I. Thus the density of Y is (6) n(C-Iylj.L,I)modICI- 1 = (27r) -lPICIC'I- t exp[ - ~(y - Cj.L)'(CIC') -I(y - Cj.L) 1 =n(yICj.L,CIC').
•
Now let us consider two sets of random variables XI"'" Xq and
Xq+ I'" ., Xp forming the vectors
(7)
X(1)=
These variables form the random vector
(8)
X = (X(1)) X(2) =
(~11
.' Xp
Now let us assume that the p variates have a joint normal distribution with mean vectors
(9)
2S
2.4 LINEAR COMBINATIONS; MARGINAL DISTRIBUTIONS
and covariance matrices
(10)
C(X(I) - ,:a.(1»)(X(I) - ....(1»)' = In,
(11)
C(X(2) - ....(2»)(X(2) - ....(2»)' = 1 22 ,
(12)
C(X(I) - ....(I»)(X(2) - ....(2»), = 1 12 ,
We say that the random vector X has been partitioned in (8) into subvectors, that
= ( ....(1»)
(13)
....
....(2)
has been partitioned similarly into subvectors, and that
(14) has been partitioned similarly into submatrices. Here 121 = 1'12' (See Appendix, Section A~3.) We shall show that X(I) and X(2) are independently normally distributed if 112 = I~I = O. Then
In
1= ( o
(15) ;'t
~~t
:'Its inverse is
(16) Thus the quadratic form in the exponent of n(xl ...., I) is (17) Q "" (x - .... )' I
-I ( X -
~ [( X(I) - ....(1»)',
.... )
(X(2) -
=
[(X(I) - ....(1»)'1 111, (x(2) -
=
(X(l) -
(1»)'1 111(x(1) -
....
= QI +Q2,
I~I ) (:~:: =:~::)
(2»)'] (I!l
....
n(:~:: =::::)
(2»),I 2
....
....(1»)
+ (X(2) -
(2»)'I 2i(.r(2) - ....(2»)
....
26
THE MULTIVARIATE NORMAL DISTRIBUTION
say, where
(18)
Also we note that Il:l = Il:lll ·1l:22I. The density of X can be written
(19)
The marginal density of
X(I)
is given by the integral
Thus the marginal distribution of X(l) is N(1L(l), l: II); similarly the marginal distribution of X(2) is N(IL(2), l:22)' Thus the joint density of Xl"'" Xp is the product of the marginal density of XI"'" Xq and the marginal density of Xq+ I" .• , X p' and therefore the two sets of variates are independent. Since the numbering of variates can always be done so that X(I) consists of any subset of the variates, we have proved the sufficiency in the following theorem: Theorem 2.4.2. If XI"'" Xp have a joint nonnal distribution, a necessary and sufficient condition for one subset of the random variables and the subset consisting of the remaining variables to be independent is that each covariance of a variable from one set and a variable from the other set is O.
2.4
LINEAR COMBINATIONS; MARGINAL DISTRIBUTIONS
27
The nf cessity follows from the fact that if Xi is from one set and Xi from the other, then for any density (see Section 2.2.3) (21)
(J"ij = $ (Xi - IL;) ( Xj - ILj) =f'oo ···foo (Xi-IL;)(Xj-ILj)f(xl, ... ,xq) _00
_00
·f( Xq+l"'" xp) dx 1 •• , dx p =foo ···f"" (xi-ILi)f(xl,·.·,xq)dx l ···dx q _00 _00
=0.
'*
Since uij = UiUj Pij' and Uj , uj 0 (we tacitly assume that l: is nonsingular), the condition u ij = 0 is equivalent to Pij = O. Thus if one set of variates is uncorrelated with the remaining variates, the two sets are independent. It should be emphasized that the implication of independence by lack of correlation depends on the assumption of normality, but the converse is always true. . Let us consider the special case of the bivariate normal distribution. Then X(1) =Xl , X(2) =X2, j.L(1) = ILl' j.L(2) = ILz, l:ll = u ll = u?, l:zz = uzz = ul, and l:IZ = l:ZI = U 12 = UlUz PIZ' Thus if Xl and X z have a bivariate normal distribution, they are independent if and only if they are uncorrelated. If they are uncorrelated, the marginal distribution of Xi is normal with mean ILi and variance U/. The above discussion also proves the following corollary: Corollary 2.4.1. If X is distributed according to N(j.L, l:) and if a set of components of X is unco"elated with the other components, the marginal distribution of the set is multivariate nonnal with means, variances, and covariances obtained by taking the co"esponding components of j.L and l:, respectively.
Now let us show that the corollary holds even if the two sets are not independent. We partition X, j.L, and l: as before. We shall make a nonsingular linear transformation to subvectors
+ BX(Z),
(22)
y(1) = X(l)
(23)
y(Z) =X(Z),
choosing B so that the components of
yO)
are uncorrelated with the
28
THE MULTIVARIATE NORMAL DISTRIBUTION
components of y(2) = X(2). The matrix B must satisfy the equation
(24)
0 = tB'(y(1) - tB'y(I»(y{2) - tB'y(2) , = tB'(X(I)
+ EX(2) - tB'X(1) -BtB'X(2»(X(2) - tB'X(2» ,
= tB'[ (X(I) - tB' X(1» + B( X(2) - tB' x(2) 1(X(2) - tB' X(2», = "I l2 +B"I 22 • Thus B = - ~ 12l:221 and
(25) The vector
(Yy(2)(I») =y= (I0
(26)
-
~ ~-I) X
-"'-1[2"'-22
is a nonsingular transform of X, and therefore has a normal distribution with
tB'(Y(l») = tB'(I y(2) 0
(27)
=U =v,
say, and
(28)
C(Y)=tB'(Y-v)(Y-v)' = (tB'(y(l) - v(1»(yO) - vO»'
lff(y(2) - V(2»(y(1) - vO)'
tB'(y(l) - v(l»(y(2) _ V(2»') tB'(y(2) _ v(2»(y(2) _ V(2» ,
'I
:.1.4
LINEAR COMBINATIONS; MARGINAL DISTRIBUTIONS
29
since
(29)
B(y(l) - V(I»)(y(l) - V(l»)'
= B[(X(l) - j.L(l») - I12I2"2I(X(2) _ j.L(2»)] .[(X(l) _ j.L(I») - I12 I2"21 (X(2) _ j.L(2»)), = III - I12I2"2II21 - I12I2"2II21 + I12I2"21 I22I2"21 I21 =I ll -II2 I 2"lI 21 · Thus y(l) and y(2) are independent, and by Corollary 2.4.1 X(2) = y(2) has the marginal distribution N(j.L(2), I 22 ). Because the numbering of the components of X is arbitrary, we can state the following theorem: Theorem 2.4.3. If X is distributed according to N(j.L, I), the marginal distribution of any set of components of X is multivariate normal with means, variances, and co variances obtained by taking the corresponding components of j.L and I, respectively. Now consider any transformation
(30) : .:_~_; ~
Z=DX,
where Z has q components and D is a q X P real matrix. The expected value of Z is
(31) and the covariance matrix is
(32) The case q = p and D nonsingular has been treated above. If q Sop and Dis . of rank q, we can find a (p - q) X P matrix E such that
(33) is a nonsingular transformation. (See Appendix, Section A.3.) Then Z and W have a joint normal distribution, and Z has a marginal normal distribution by Theorem 2.4.3; Thus for D of rank q (and X having a nonsinguiar distribution, that is, a density) we have proved the following theorem:
30
THE MULTIVARIATE NORMAL DISTRIBUTION
Theorem 2.4.4. If X is distributed according to N(j.L, ~), then Z = DX is distributed according to N(Dj.L, D~D'), where D is a q Xp matrix of rank q 5.p. The remainder of this section is devoted to the singular or degenerate normal distribution and the extension of Theorem 2.4.4 to the case of any matrix D. A singular distribution is a distribution in p-space that is concentrated on a lower dimensional set; that is, the probability associated with any set not intersecting the given set is O. In the case of the singular normal distribution the mass is concentrated on a given linear set [that is, the intersection of a number of (p - I)-dimensional hyperplanes]. Let y be a set of cuonlinat.:s in the linear set (the number of coordinates equaling the dimensiunality of the linear set); then the parametric definition of the linear set can be given as x = Ay + A, where A is a p X q matrix and A is a p-vector. Suppose that Y is normally distributed in the q-dimensional linear set; then we say that (34)
X=AY+A
has a singular or degenerate normal distribution in p-space. If GY = v, then $X =Av + A = j.L, say. If G(Y- vXY- v)' = T, then (35)
$(X- j.L)(X-
j.L)'
= cffA(Y-v)(Y-v)'A' =ATA'
=~,
say. It should be noticed that if p > q, then ~ is singular and therefore has no inverse, and thus we cannot write the normal density for X. In fact, X cannot have a density at all, because the fact that the probability of any set not intersecting the q-set is 0 would imply that the density is 0 almost everywhere. Now. conversely, let us see that if X has mean j.L and covariance matrix ~ of rank r, it can be written as (34) (except for 0 probabilities), where X has an arbitrary distribution, and Y of r (5. p) components has a suitable distributiun. If 1 is of rank r, there is a p X P nonsingular matrix B such that (36)
B~B'=(~ ~),
where the identity is of order r. (See Theorem A.4.1 of the Appendix.) The transformation (37)
BX=V=
V(l) ) (
V (2)
2.4
31
LINEAR COMBINATIONS; MARGINAL DISTRIBUTIONS
defines a random vector V with covariance matrix (36) and a mean vector (38)
SV=BJ1.=v=
V(I») ( V(2) ,
say. Since the variances uf the elements of probability 1. Now partition
(39)
V(2)
are zero,
V(2)
= V(2) with
B- 1 = (C D),
where C consists of r columns. Then (37) is equivalent to
(40) Thus with probability 1
(41) which is of the form of (34) with C as A, V(1)as Y, and Dv(2) as A. Now we give a formal definition of a normal distribution that includes the singular distribution. Definition 2.4.1. A random vector X of p components with S X = J1. and S(X - J1.XX - J1.)' = I is said to be normally distributed [or is said to be distributed according to N(J1., I») if there is a transfonnation (34), where the number of rows of A is p and the number of columns is the rank of I, say r, and Y (of r components) has a nonsingular nonnal distribution, that is, has a density
(42) It is clear that if I has rank p, then A can be taken to be I and A to be 0; then X = Y and Definition 2.4.1 agrees with Section 2.3. To avoid redundancy in Definition 2.4.1 we could take T = I and v = o.
Theorem 2.4.5. If X is distributed according to N(J1., I), then Z = DX is distributed according to N(DJ1., DI.D'). This theorem includes the cases where X may have a nonsingular or a singular distribution and D may be nonsingular or of rank less than q. Since X can be represented by (34), where Y has a nonsingular distribution
32
THE MULTI VARIATE NORMAL DISTRIBUTION
N( v, T), we can write
( 43)
Z=DAY+DA,
where DA is 1 X r. If the rank of DA is r, the theorem is proved. If the rank is less than r, say s, then the covariance matrix of Z,
(44)
DATA'D' =E,
say, is of rank s. By Theorem A.4.1 of the Appendix, there is a nonsingular matrix (45) such that (46) FEF' =
F EF' 1 1 ( F2EF;
= (FjDA)T(FjDA)' (F2 DA) T( F j DA)'
Thus F j DA is of rank s (by the COnverse of Theorem A.Ll of the Appendix); and F2DA = 0 because each diagonal element of (F2DA)T(F2DA)' is il quadratic form in a row of F2DA with positive definite matrix T. Thus the covariance matrix of FZ is (46), and
(47)
FZ =
(~: )DAY+FDA = (Fl~AY) +fDA = (~l) +FDA,
say. Clearly U j has a nonsingular normal distribution. Let F- j = (G 1 G 2 ). Then
(48) which is of the form (34).
•
The developments in this section can be illuminated by considering the geometric interpretation put forward in the previous section. The density of X is constant on the ellipsoids (54) of Section 2.3. Since the transformation (2) is a linear transformation (Le., a change of coordinate axes), the density of
~
2.5
33
CONDITIONAL DISTRIBUTIONS; MULTIPLE CORRELATION
Y is constant On ellipsoius (49) The marginal distribution of X(l) is the projection of the mass of the distribution of X onto the q-dimensional space of the first q coordinate axes. The surfaces of constant density are again ellipsoids. The projection of mass on any line is normal.
2.5. CONDITIONAL DISTRIBUTIONS AND MULTIPLE CORRELATION COEFFICIENT 2.5.1. Conditional Distributions In this section we find that conditional distributions derived from joint normal distribution are normal. The conditional distributions are of a particularly simple nature because the means depend only linearly on the variates held fixed, and the variances and covariances do not depend at all on the values of the fixed variates. The theory of partial and multiple correlation discufsed in this section was originally developed by Karl Pearson (1896) for three variables and ex(ended by Yule (1897" 1897b). Let X be distributed according to N(j.L"~:) (with :1: nonsingular). Let us partition
( 1)
X
= (X(l») X(2)
as before into q- and (p - q)-component subvectors, respectively. We shall use the algebra developed in Section 2.4 here. The joint density of y(l) = X(I) -:1:I~ 1221 X(2) and y(2) = X(2) is n(y(1)1
j.L(I) - :1: 12 l :221j.L(2), :1: 11
-
:1: 12:1:221:1:21 )n(y(2)1 j.L(2), :1: 22 ).
The density of X(I) and X(2) then can be obtained from this expression by substituting X(I) - 112:1: 221 X(2) for y(l) and X(2) for y(2) (the Jacobian of this transformation being 1); the resulting density of XO) and X(2) is t2)
f( X(I),X(2») =
11
(21T)'q~
exp{-.!.[(x(I)_j.L(I»)-11- 1 (x(21_p.(2»)], 2 12 22 .1 1112[( X(I) - j.L(I») - 1 12 :1:2"l( X(2)
•
I
1
(21T)'(P-q)~
-
p.(2»)]}
exp[-.!.(x(2)-j.L(2»)'1-I(x(2)-j.L(2»)], 2
22
34
THE MULTIVARIATE NORMAL DISTRIBUTION
where (3) given that at the point which is n(x(2) \ j.L(2), l:22)' the second factor of (2). The quotient is
This density must be n(x\
j.L,
l:). The conditional density of
XI~) = XI~1 is the quotient of (2) and the marginal density of Xl!).
X(I)
X(2)
(4) f(X(ll\XI!l) =
,1
(27rrq~
exp{-.l[(x(l)_j.L(I»)-l: 'l:-I(X(2)_,,(2»)]' 2 12 22 ..... 'l:lllz[(x(l) - j.L(I)) -l: 12 l:Z21 (X(2) - j.L(2»)1}.
It is understood that xl!) consists of p - q numbers. The density f(x(l)\x(2») is a q-variate normal density with mean
say, and covariance matrix cS'{[X(I) - v(x(Z»)] [X(I) - v(x(2»)]'ix(2)} = l:1l.2 = l:1l -l: 12 l: Z21l:21'
(6)
It should be noted that the mean of X(I) given x(2) is simply a linear function of X(2), and the covariance matrix of X(I) given X(2) does not depend on X(2) at all. Definition 2.5.1. The matrix ~ ficients of X(ll on X(2).
= l: 12l:Z21 is the matrix of regression coef-
The element in the ith row and (k - q )th column of ~ = l: 12 l:Z21 is often denoted by (7)
i=l, .. "q,
{3ik.q+ I ..... k-I.k+ I ..... p'
k=q+l, ... ,p.
The vector j.LII) + ~(x(2) - 1-1-(2») is called the regression function. Let u i ",+1. .. .,' be the i,jth element of l:11'2' We call these partial cuuarial/ces; if,i'" + I. .. "" is a partial variance. Definition 2.5.2
(8)
Pij·q + I .... p
yu
U i j-q+I, .... /1
1
"-q+ .···,P
.Iu.. V }}'q+ I .···.P
i,j=l, ... ,q, '
is the partial correlation between Xi and Xj holding Xq + 1, ... , Xp fixed.
2.5
CONDITIONAL DISTRIBUTIONS; MULTIPLE CORRELATION
35
The numbering of the components of X is arbitrary and q is arbitrary. Hence, the above serves to define the conditional distribution of any q components of X given any other p - q components. In the case of partial covariances and correlations the conditioning variables are indicated by the. subscripts after the dot, and in the case of regression coefficients the dependent variable is indicated by the first SUbscript, the relevant conditioning variable by the second subscript, and the other conditioning variables by the subscripts after the dot. Further, the notation accommodates the conditional distribution of any q variables conditional on any other r - q variables (q 5,r 5.p). Theor~m 2.5.1. Let the components of X be divided into two groups composing the sub vectors X(1) and X(2). Suppose the mean J.l is similarly divided into J.l(l) and J.l(2), and suppose the covariance matrix l: of X is divided into l:11' l:12' l:22, the covariance matrices of X(I), of X(I)and X(2), and of X(2), respectively. Then if the distribution of X is nonnal, the conditional distribution of x(I) given X(2) = X(2) is nonnal with mean J.l(l) + l:12l:Z2l (X(2) - J.l(2») and covariance matrix l: 11 -l: 12 'i. Z2l "i. 21 ·
As an example of the above considerations let us consider the bivariate normal distribution and find the conditional distribution of Xl given X 2 =x 2 • In this case J.l(l) = J.tl' ....(2) = J.t2' l:11 = a}, l:12 = a l a2 p, and l:22 = ai- Thus the 1 X 1 matrix of regression coefficients is l:J2l:zl = a l pi a2' and the 1 X 1 matrix of partial covariances is
(9)
l:U'2
= l:u -l:J2l:Zll:2l = a l2 - a l2al p2 I al = a 12 (1 - p2).
The density of Xl given x 2 is n[xll J.tl + (a l pi ( 2)(x 2 - J.t2)' a 12(1 - p2)]. The mean of this conditional distribution increases with X 2 when p is positive and decreases with increasing x 2 when p is negative. It may be noted that when a l = a2' for example, the mean of the conditional distribution of Xl does not increase relative to J.tl as much as X 2 increases relative to J.t2' [Galton (1889) observed that the average heights of sons whose fathers' heights were above average tended to be less than the fathers' heights; he called this effect "regression towards mediocrity."] The larger I pi is, the smaller the variance of the conditional distribution, that is, the more information x 2 gives about x I' This is another reason for considering p a measure of association between XI and X 2 • A geometrical interpretation of the theory is enlightening. The density f(x t , x 2 ) can be thought of as a surface z = f(x l , x 2 ) over the Xl' x 2-plane. If we intersect this surface with the plane x 2 = c, we obtain a curve z = t(x l , c) over the line X 2 = c in the Xl' x 2-plane. The ordinate of this curve is
36
THE MULTIVARIATE NORMAL DISTRIBUTION
proportional to the conditional density of XI given X 2 = c; that is, it is proportional to the ordinate of the curve of a univariate normal distribution. In the more general case it is convenient to consider the ellipsoids of constant density in the p-dimensional space.' Then the surfaces of constant density of f(xl .... ,xqicq+I, ... ,cp) are the intersections of the surfaces of constant density of f(xl, ... ,x) and the hyperplanes Xq+1 =cq+l>''''x p = c p ; these are again ellipsoids. Further clarification of these ideas may be had by consideration of an actual population which is idealized by a normal distribution. Consider, for example, a population of father-son pairs. If the population is reasonably homogeneous, the heights of fathers and the heights of corresponding sons have approximately a normal distribution (over a certain range). A,. conditional distribution may be obtained by considering sons of all faLlers whose height is, say, 5 feet, 9 inches (to the accuracy of measurement); the heights of these sons will have an approximate univariate normal distribution. The mean of this normal distribution will differ from the mean of the heights of SOns whose fathers' heights are 5 feet, 4 inches, say, but the variances will be about the same. We could also consider triplets of observations, the height of a father, height of the oldest son, and height of the next oldest son. The collection of heights of two sons given that the fathers' heights are 5 feet, 9 inches is a conditional distribution of two variables; the correlation between the heights of oldest and next oldest sons is a partial correlation coefficient. Holding the fathers' heights constant eliminates the effect of heredity from fathers; however, one would expect that the partial correlation coefficient would be positive, since the effect of mothers' heredity and environmental factors would tend to cause brothers' heights to vary similarly. As we have remarked above, any conditional distribution obtained from a normal distribution is normal with the mean a linear fum,tion of the variables held fixed and the covariance matrix constant In the case of nonnormal distributions the conditional distribution of one set of variates on another does not usually have these properties, However, one can construct nonnormal distributions such that some conditional distributions have these properties. This can be done by taking as the density of X the product n[x(l)1 j.L(1) + ~(X(2) - j.L(2», ~ 1I'21f(x(2», where f(X(2» is an arbitrary density.,
2.5.1. The Multiple Correlation Coefficient
We dgain c;r)flsider properties of ~(2),
X
partitioned into
X(I)
and
X(2),
We shalt study some
2.5
37
CONDITIONAL DISTRIBUTIONS; MULTIPLE CORRELATION
Definition 2.5.3. The vector X(I'2) = X(I) - j.L(I) tOl of residuals of x(1) from its regression on X(2).
-
~(X(2) -
j.L(2»
is the vec-
Theorem 2.5.2. The components of X(I'2) are unco"eiated with the components of X(2).
Theorem 2.5.3.
For every vector a
Proof By Theorem 2.5.2
(11)
r( Xi -'- a 'X(2») = $[X; - l L i - a'(X(2)
-
j.L(2))f
$XP'2) + (~(;) - a )'(X(2) -
=
$[ XP-2) -
=
r[ XP-2)] + (~(i) - a)' $( X(2) -
j.L(2») (
j.L(2»)f
X(2) -
j.L(2»)' (~(i) -
a)
= r( XP'2») + (~(i) - a )'Id~(i) - a). Since I22 is positive definite, the quadratic form in • and attains it~ minimum of 0 at a = ~(i)'
~(i)
-
a is nonnegative
Since $ X(!'2) = 0, r(x?,2» = $(Xp·2»2. Thus ILl + ~(i)(X(2) - j.L(2» is the best linear predictor of XI in the sense that of all functions of X(2) of the form a' X (2) + c, the mean squared error of the above is a minimum. Theorem 2.5.4.
(12)
For every vector a Corr(x.,'tJ'(I) Il'. X(2») ..... Corr(X. a' X(2)) ~ " •
Proof Since the correlation between two variables is unchanged when either or both is multiplied by a positive constant, we can assume that
38
THE MULTI VARIATE NORMAL DISTRIBUTION
(13)
crj, - 2 G( X j - J.LJ~(i)( X(2) - j.L(2») + r(~(,)X(2») ::>;
cra - 2 G(Xj - J.Lj)a'(X(2) - j.L(2») + r( a'X(2».
This leads to (14)
G( Xi - J.LJ~(i)( X(2) - j.L(2») > G( Xi - J.Lj) a '( X(2) _ j.L(2»)
v' cr,j r( ~(j)X(2»)
...; cra r( a' X (2)
-
.
•
Definition 2.5.4. The maximum co"elation between Xi and the linear combination a' X(2) is called the multiple correlation coefficient between Xi and X(2).
It follows that this is (15)
R j ·q + l ,
",p
vi 0'(i):1: 22
1
O'(i)
,fU; A useful formula is (16)
l-R?'Q+1, .. ,p
where Theorem A,3,2 of the Appendix has been applied to (17)
:1:.= ,
cr..
"
( O'(i)
Since (18) it follows that (19)
cr..ll·q+I ..... p =
(1 - J?2
"q+I" .. ,p
) a:II' ..
This shows incidentally that any partial variance of a component of X cannot be greater than the variance. In fact, the larger Rj • Q + 1,,,.,P is, the greater the
2.5
CONDmONAL DISTRIBUTIONS; MULTIPLE CORRELATION
39
reduction in variance on going to the conditional distribution. This fact is another reason for considering tiIe multiple correlation coefficient a measure of association between Xi and X(2). That ~(i)X(2) is the best linear predictor of Xi and has the maximum correlation between Xi and linear functions of X(2) depends only on the covariance structure, without regard to normality. Even if X does not have a normal distribution, the regression of X(I) on X(2) can be defined by joL(J) + l: 12l:221 (X(2) - joL(. I); the residuals can be defined by Definition 2.5.3; and partial covariances and correlations can be defined as the covariances and correlations of residuals yielding (3) and (8). Then these quantities do not necessarily have interpretations in terms of conditional distributions. In the case of normality P-i + ~(i)(X(2) - joL(2») is the conditional expectation of Xi given X(2) = X(2). Without regard to normality, Xi - GX i IX(2) is uncorrelated with any function of X(2), GX i IX(2) minimizes G[Xi - h(X(2»)]2 with respect to functioas h(X(2») of X(2), and GXi IX(2) maximizes the correlation between Xi and functions of X(2). (See Problems 2.48 to 2.51.) 2.5.3. Some Formulas for Partial Correlations
We now consider relations between several conditional distributions o~tained by holding several different sets of variates fixed. These relations are useful because they enable us to compute one set of conditional parameters from another st:t. A very special ca',e is
(20)
this follows from (8) when P = 3 and q = 2. We shall now find a generalization of this result. The derivation is tedious, but is given here for completeness. Let
(21)
X(I) X(2)
X= (
1 ,
X(3)
where X(I) is of PI components, X(2) of P2 components, and X(3) of P3 components. Suppose we have the conditional distribution of X(l) and X(2) given X(3) = x(3); how do we find the conditional distribution of X(I) given X(2) = X(2) and X(3) = x(3)? We use the fact that the conditional density of X(1)
40
given
THE MULTIVARIATE NORMAL D1STR[9UTION X(2)
= X(2) and
X(3)
= X(3) is
. f( x (1)1 x (2) ,x(3»
(22)
=
f(
(I)
(2)
x , x ,x
(3»
f( X(2), x(3» _ f( X(I), X(2), x(3» If( x(3» -
f( x (2) , x(3» If( x(3»
_ f(x(1),x(2)lx(3» -
f( x(2)lx(3»
In tr.e case of normality the conditional covariance matrix of givel' X(3) =. X(3) is
X(I)
and
X(2)
(23)
say, where
(24)
The conditional covariance of X(I) given X(2) = X(2) and X(3) = x(3) is calculated from the conditional covariances of X(I) and X(2) given X(3) = X(3) as
This result permits the calculation of aji'p, + I, .,., P' i, j = 1, ... , PI' fro;.,.,.' aii'P, +p" ... ,p' i, j = 1, ... , PI + P2' In particular, for PI = q, P2 = 1, and P3 = P - q - 1, we obtain
(26)
ai. q+ l.q+2, ... , P aj, q + I'q +2, ... • P aii'q+I, ... ,p
=
a i i'q+2 .... ,P-
aq+l.q+I'q+2, ... ,p
i,j= 1, ... ,q. Since
(27)
a.·,,·q+I, ... ,p =a:lI·q+2, .. 2 ... ,p (1- pl,q+l·q+2, ... ,p ),
41
THE CW RACTERISTIC FUNCTION; MOMENTS
2.6
we obtain
,
(28)
Plj-q+l, ... ,p =
Pij·q+2 ..... p - Pi.q+l·q+2 ..... pPj.q+l·q+2 ..... p .2 .2 • P'.q+l·q+2 .... ,p P),q+l·q+2 ..... p
VI _
VI-
This is a useful recursion formula to compute from {pi/.pl, { Pij'p-l,
{Pijl
in succession
pl, ... , PI2·3, ... , p'
2.6. THE CHARACTERISTIC FUNCTION; MOMENTS 2.6.1. The Characteristic Function The characteristic function of a multivariate normal distribution has a form similar to the density function. From the characteristic function, moments and cumulants can be found easily. Definition 2.6.1. The characteristic function of a random vector X is
4>(t)
(1)
= $el/'x
defined for every real vector t.
, To make this definition meaningful we need to define the expected value of a complex-valued function of a random vector.
=
Definition 2.6.2. Let the complex-valued function g(x) be written as g(x) glCe) + igix), where gl(X) andg 2(x) are real-valued. Then the expected value
Ofg(X) is
(2) In particurar, since ei8 = cos (J + i sin
(3)
(J,
$ei/'X = $cost'X+i$sint'X.
To evaluate the characteristic function of a vector X, it is often convenient to use the following lemma: Lemma 2.6.1. Let X' = (X(l)' X(2)'). If X(l) and X(2) are independent and g(x) = g(l)(X(l»)g(2)(X(2»), then
'"
(4)
42
THE MULTIVARIATE NORMAL DISTRIBUTION
Proof If g(x) is real-valued and X has a density,
= J~:c'" J~xgtl)(X(l)gt2)(x(2)f(l)(x(1»f(2l(x(2) dx 1 ... dxp = (" .,. -x
JX gtl)(x(ll)f(ll(x(I) -x
dx l ... dx q
If g(x) is complex-valued,
g(x)
(6)
=
(g\I)(X(l)) +igi l)(x(lI)][g\2)(x(2)) +igfl(x(2l )]
= gpl( X(I)) g\"l( X(2)) - gil) ( x(ll) gi2l( X(2» + i (g~l)( x(l)) g\2)( X(2)) + gpl( x(ll) gi2'( X(2l)] . Then
lffg(X) = lff(g\I)(X(l))g\2)(X(2)) -g~Il(X(l))gi2l(X(2l)]
(7)
+ i lff(gil'( X(l)) g\2)( X(2» + g\ll( X(ll)g~2l( X(2))] =
lffg\l)(X(ll) lffgf)( X(2)) -
$g~I)(X(ll)
is'gi2)(X(2))
+ i( lffgill(XII)) lffgF)(X(2) + lffg\l)(X(l» lffgi2)(X(2)]
= [ $gP)( X(l) + i lffgi\) ( X(I»)][ $g\2l ( X(2» + i lffgf)(X(2»] = lffg(l)(X(I)) lffg(2)(X(2l).
•
By applying Lemma 2.6.1 successively to g(X) = eit'X, we derive Lemma 2.6.2.
If the components of X are mutually independent, p
(8)
lffeit'x =
fllffeitl'~i. j~l
We now find the characteristic function of a random vector with a normal distribution.
2.6 THE CHARACTERISTIC FUNCTION; MOMENTS
Theorem 2.6.1.
43
The characteristic function of X distributed according to
N(IL, !,)is
(9) for every real vector t. Proof From Corollary A1.6 of the Appendix we know there is a nonsingular matrix C such that
(10) Thus
(11) Let
(12)
X-IL =CY.
Then Y is distributed according to N(O, J). Now the characteristic function of Y is p
(13)
I/I(u) =
Ge iu ' Y =
n
GeiuJYj.
j~l
Since lj is distributed according to N(O, 1), p
(14)
I/I(U) =
ne-tuJ=e-tu'u. /-1
Thus
(15) = eit'u Geit'CY = eit'''"e- tit'CXt'C)'
for t'C = u'; the third equality is verified by writing both sides of it as integrals. But this is
(16)
by (11). This proves the theorem.
•
44
THE MULTIVARIAT, NORMAL DISTRIBUTION
The characteristic function of the normal distribution is very useful. For example, we can use this method of proof to demonstrate the results of Section 2.4. If Z = DX, then the characteristic function of Z is (17)
= ei,(D'I)'IL- ~(D'I)'I(D'I)
= eil'(DIL)- ~I'(DID')I , which is the characteristic function of N(DtJ., DI.D') (by Theorem 2.6.1). It is interesting to use the characteristic function to show that it is only the multivariate normal distribution that has the property that every linear combination of variates is normally distributed. Consider a vector Y of p components with density f(y) and characteristic function
and suppose the mean of Y is tJ. and the covariance matrix is I.. Suppose u'Y is normally distributcd for every u. Then the characteristic function of such linear combination is
(19) Now set t = 1. Since the right-hand side is then the characteristic function of I.), the result is proved (by Theorem 2.6.1 above and 2.6.3 below).
N(tJ.,
Theorem 2.6.2. If every linear combination of the components of a vector Y is normally distributed, then Y is normally distributed.
It might be pointed out in passing that it is essential that every linear combination be normally distributed for Theorem 2.6.2 to hold. For instance, if Y = (Yl , Y 2 )' and Yl and Y 2 are not independent, then Yl and Y2 can each have a marginal normal distribution. An example is most easily given geometrically. Let Xl' X 2 have a joint normal distribution with means O. Move the same mass in Figure 2.1 from rectangle A to C and from B to D. It will be seen that the resulting distribution of Y is such that the marginal distributions of Yl and Y2 are the same as Xl and X 2 , respectively, which are normal, and yet the joint distribution of Yl and Y2 is not normal. This example can be used also to demonstrate that two variables, Yl and Y2 , can be uncorrelated and the marginal distribution of each may be normal,
i.
45
2.6 THE CHARACfERISTIC FUNCfION; MOMENTS
Figure 2.1
but the pair need not have a joint normal distribution and need not be independent. This is done by choosing the rectangles so that for the resultant distribution the expected value of YI Y 2 is zero. It is clear geometrically that this can be done. For future reference we state two useful theorems concerning characteristiC functions. Theorem 2.6.3. If the random vector X has the density f(x) and the characteristic function cfJ(t), then
(20)
f(x) =
1 --p
(21T)
00
f _00
00.,
... fe-'I XcfJ(t) dt l
..•
dtp-
_00
This shows that the characteristic function determines the density function uniquely. If X does not have a density, the characteristic function uniquely defines the probability of any continuity interval. In the univariate case a 'continuity interval is an interval such that the cdf does not have a discontinuity at an endpoint of the interval. , Theorem 2.6.4. Let (.Fj(x)} be a sequence of cdfs, and let (cfJlt)} be the of corresponding characteristic functions, A necessary and sufficient condition for ~(x) to converge to a cdf F(x) is that, for every t, cfJlt) converges 10 a limit cfJU) that is continuous at t = 0, When this condition is satisfied, the limit cfJ(t) is identical with the characteristic function of the limiting distribution F(x).
~·equence
For the proofs of these two theorems, the reader is referred to Cramer H1946), Sections 10.6 and 10.7.
~f
46
THE MULTIVARIATE NORMAL DISTRIBUTION
2.6.2. The Moments and Cumulants The moments of XI" .. ' Xp with a joint normal distribution can be obtained from the characteristic function (9). The mean is (21)
=
f {-
I>hjtj
+ ilLh }cP(t)\
J
1=0
= ILh· The second moment is (22)
=
~ {( - ~ (7hk tk + ilLh)( - ~ (7kj tk + ilLj) -
(7hj
}cP(t)llao
= (7hj + ILh ILj· Thus (23) (24)
Variance( Xj) = $( X, - 1L,)2 =
(7j"
Covariance(Xj , Xj) = $(Xj -lLj)(Xj - ILJ = (7ij.
Any third moment about the mean is (25) The fourth moment about the mean is
Every moment of odd order is O. Definition 2.6.3. If all the moments of a distribution exist, then the cumuIan ts are the coefficients K in
(27)
47
2.7 ELLIPTICALLY CONTOURED DISTRIBUTIONS
In the case of the multivariate normal distribution
= ILl"'" KO'" 01 The cumulants for
K IO .·. 0
=lLp,K20 ... 0=0'1l, ••• ,KO ... 02=O'pp,KIl0 ... 0=0'12' ••••
which LSi> 2 are O.
2.7. ELLIPTICALLY CONTOURED DISTRIBUTIONS 2.7.1. Spherically and Elliptically Contoured Distributions It was noted at the end of Section 2.3 that the density of the multivariate normal distribution with mean .... and covariance matrix I. is constant on concentric ellipsoids
(x- .... )'I.-I(X- .... ) =k.
( 1)
A general class of distributions .vith this property is the class of elliptically contoured distributions with density
IAI-~g[(x- v)' A -I(X- v)],
(2)
where A is a positive definite matrix,
foo ... foo
(3)
-co
gO ~ 0, and
g(y'Y) dYI ... dyp
= 1.
-co
If C is a nonsingular matrix such that C'A-IC=I, the transformation x - v = Cy carries the density (2) to the density g(y y). The contours of constant density of g(y y) are spheres centered at the origin. The class of such densities is known as the spherically contoured distributions. Elliptically contoured distributions do not necessarily have densities, but in this exposition only distributions with densities will be treated for statistical inference. A spherically contoured density can be expressed in polar coordinates by the transformation I
I
(4)
YI =rsin
°
1,
Y2 =rcos 8 1 sin
02'
Y3 = r cos
O2 sin 03'
Yp-I Yp
° cos 1
° cos =rcos ° cos
= r cos
cos
0p-2
sinOp _ 2 '
1
02 ...
1
O2 "'COS 0p_2 cos 0p_I'
48
THE MULTIVARIATE NORMAL. DISTRIBUTION
where -~7T
°
(5) Note that R,0 1, ••• ,0p _ 1 are independently distributed. Since
(6) (Problem 7.2), Ihe margillal Jellsily of R i~
(7) where (8) 2'Ttp
C( P ~ = f( ~p )-
The margiual density of O; is n~(p - i)jcos P-; -18/{rq)r[!(p - i -1)]}, i = 1, ... , P - 2, and of 0p-l is 1/(27T). In the normal case of N(O, I) the density of Y is
g(y'y) = (27T) -tp exp( -h'y), and the density of R=(yly)4 is r P- 1 exp(- ~r2)/[2tp--If(~)]. The density of r2 = v is vtp-1e- tU/[2tpr(~)]. This is the x2-density with p degrees of freedom. The constant C(p) is the surface area of a sphere of unit radius in p dimensions. The random vector U with coordinates sin 0» cos 0 1 sin 2 " " , cos 0 1 cos O2 "'cos0p_ 1 , where 0 1, ... ,0p _ 1 are independently distributed each with the uniform distribution over (-7T/2, 7T/2) except for 0 p _ 1 having the uniform distribution over ( - 7T, 7T), is said to be uniformly distributed on the unit sphere. (This is the simplest example of a spherically contoured, distribution not having a density.) A stochastic representation of Y with th&
°
'."~
2.7
49
ELLIPTICALLY CONTOURED DISTRIBUTIONS
density g(y'y) is d
(9)
Y=RU,
where R has the density (7). Since each of the densities of
(10)
°
1,,,,,
0 p _ 1 are even,
tCU= O.
Because R anJ U are independent,
tCy=o
(11) if tCR
< 00.
Further,
(12) if ~R2 < 00. By symmetry tCU I2 = ... = tCU/ = lip because Ef_IU/ = 1. Again by symmetry tCU1U2 = tCUP3 = '" = tCUp_1 Up. In particular tCU1U2 ... ~ sin 0 1 cos 0 1 sin O2 , the integrand of which is an odd function of 0 1 and of (Jz. Hence, tCll;~ = 0, i"* j. To summarize,
tCuu' = (llp)lp
(13) and
(14) (if tC R2
< 00).
The distinguishing characteristic of the class of spherically contoured d!stributions is that OY g, Y for every orthogonal matrix O. Theorem 2.7.1. If Y has the density g(y'y), then Z = OY, where 0'0 =1, has the density g(z'z). ( Proof. The transformation z = Oy has Jacobian 1.
•
We shall extend the definition of Y being spherically contoured to any distribution with the property OY g, Y. Corollary 2.7.1. If Y is spherically contoured with stochastic representation Y g, RU with R2 = Y'Y, then U is spherically contoured. Proof. If Z = OY and hen:e Z g, Y, and Z has the stochastic representa• tion Z = SV, where S2 = Z'Z, then S = Rand V= OU g, U.
50
THE MULTIVARIATE NORMAL DISTRIBUTION
The density of X= v + CY is (2). From (11) and (14) we derive the following theorem: Tbeorem 2.7.2.
(15)
If X has the density (2) and GR 2 < 00,
GX= .... =v,
C(X) = G(X - .... )(X - .... )' = I = (lip) GR2A.
In fact if GR m < 00, a moment of X of order h (s;m) is G(XI - i-LI)h, ... (Xp-JLp)hP=GZ~'···Z;pGRhIG(x;)th, where Z has the distributior..
N(O, I) and h = hi + ... +h p • Tbeorem 2.7.3. If X has the density (2), GR 2 < 00, and f[cC(X)] = f[ C(X)] for all c > 0, then f[ C(X)] = f(I).
In particular pJX) =
CTijl
VCTii CTjj = /..;jl VA;; Ajj , where I
=
(CTij )
and A =
(Ai/
2.7.2. llistributions of Linear Combinations; Marginal Distributions
First we consider a spherically contoured distribution with density g(y' y). Let y' = (Y'I' Y2), where YI and Y2 have q and p - q components, respectively. The marginal density of Y2 is ( 16)
Express YI in polar coordinates (4) with r replaced by r l and p replaced by q. Then the marginal density of Y2 is
( 17) This expression shows that the marginal distribution of Y2 has a density which is sphcrically contoured. . Now consider a vector X' = (X(\)', X(2)') with density (2). If GR 2 < 00, the covariance matrix of X is (15) partitioned as (14) of Section 2.4. Let Z(\) = X(\) - II2I22IX(2) = X(1) - A 12 A 21x(2), Z(2) = X(2), T(1) = v O) - I 12 I 221 V(2) = V(I) - A 12 A;} v(2), T(2) = v(2). Then the density of Z' = (Z(l)', Z(2) ') is (18)
1A 11.2\- tl A 221-
tg [ (z(l) -
T(l»)' A 1l.2(Z(I) - T(I»)
+ (Z(2) -
V(2»), A'22( Z(2) - V(2»)].
2.7
51
ELLIPTICALLY CONTOURED DISTRIBUTIONS
Note that Z(1) and Z(2) arc uncorrclatcd cvcn though possibly dependent. Let C I and C2 be q X q and (p - q) X (p - q) matrices satisfying CIAll2c~ = Iq and C2 A;1 c; = I p ._ q . Define y(1) and y(2) by Z(I) - T(1) = c l l ! ) and Z(2) - V(2) = C y(2). Then y(l) and y(2) have the density g(y(I)'y(l) + y(2)'y(2». 2 The marginal density of y(2) is (17), and the marginal density of X(2) = Z(2) is
The moments of Y2 can be calculated from the moments of Y. The generalization of Theorem 2.4.1 to elliptically contoured distributions is the following: Let X with p components have the density (2). Then Y = ex has the density ICAC'I- tg[(x - Cv)'(CAC,)-I(X - Cv)] for C nonsingular. The generalization of Theorem 2.4.4 is the following: If X has the density (2), then Z = DX has the density (20)
IDAD'I- tg2 [ (z - Dv )'( DA D') -I (z - Dv)
1,
where D is a q xp matrix of rank q 5.p and g2 is given by (17). We can also characterize marginal distributions in terms of the representation (9). Consider y(1)) d ( V(1) ) Y= ( y(2) =RV=R V(2) ,
(21)
where y(1) and V(1) have q components and y(2) and V(2) have p components. Then R~ = y(2)'y(2) has the distribution of R 2 V(2)'V(2), and V (2)'V(2)
(22)
=
V(2)'V(2) V'V
q
y(2)'y(2)
1: - - -
y'y·
In the case Y - NCO, Ip), (22) has the beta distribution, say B(p - q, q), with density
(23) ,
r(p/2) f( q/2)r[(p - q)/2]
Hence, in general,
(24)
zt(p-q)-I(1-
Z)~q-I
,
Oszsl.
52
THE MULTIVARIATE NORMAL DISTRIBUTION
where R~!b R 2 b, b ~ B( p - q, q), V has the uniform distribution of v'v = 1 in P2 dimensions, and R2, b, and V are independent. All marginal distributions are elliptically contoured ..
2.7.3. Conditional Distributions and Multiple Correlation Coefficient The density of the conditional distribution of YI given Y2 when Y = (YI' Y2)' has the spherical density g(y' y) is g(y'IYI +Y2Y2) _ g(YIYI +ri) g2(YZY2) g2(ri)
(25)
where the marginal density giY2Y2) is given by (17) and ri = Y2Y2' In terms of YI, (25) is a spherically contoured distribution (depending on Now consider X = (X;, X;)' with density (2). The conditional density Of X(I) given X(2) = X(2) is :k
rD.
;:t'
(26)
IA 11'21 - tg{ [(xCI) -v(l»' - (x(2) - V(2», B']A J/ 2 [x(l) - v(1) - B(x(2) - V(2»]
+ (x(2) _v(2»' A;:i(x(2) -V(2»} -;-g2[(X(2) _v(2», A;:i (X(2) - v(2»]
= IA II .21- tg{[x(l) - V(I) -B(X(2) - V(2»), Aj1.2[X(I) -v(l) -B(x(2) - V(2»]
+rn
-;-g2(ri),
where ri = (X(2) - v(2», A ;:i(X(2) - v(2» and B = A 12 Azr The density (26) is elliptically contoured in x(1) - V(I) - B(X(2) - V(2» as a function of X(I). The conditional mean of X(I) given X(2) = X(2) is
(27)
rD
if tC(Ri1YzY2 = < 00 in (25), where Ri = Y{YI. Also the conditional covariance matrix is (tCri!q)A I12 . It follows that Definition 2.5.2 of the partial correlation coefficient holds when (O"ij-q+I •...• p)=I. U . 2 =I. U + I. 12I.ill21 and I. is the parameter matrix given above. . Theorems 2.5.2, 2.5.3, and 2.5.4 are true for any elliptically contoured distribution for which tCR 2 < 00.
2.7.4. The Characteristic Function; Moments The characteristic function of a random vector Y with a spherically contoured distribution tCe it ' Y has the property of invariance over orthogonal
2.7
ELLIPTICALLY CONTOURED DISTRIBUTIONS
53
transformations, that is, (28)
where Z = Of also has the density g(y' y). The equality (28) for all orthogonal 0 implies tffeil'Z is a function of t't. We write
tffeil'Y = cP(t't).
(29) Then for X = IL + CY (30)
= e't"'cP(t'CC't) =eil· ...cP(t'At) when A = ce'. Conversely, any characteristic function of the form e il ''''cP(t' At) corresponding to a density corresponds to a random vector X with the density (2). The moments of X with an elliptically contoured distribution can be found from the characteristic function eil· ...cP(t'It) or from the representalion X = .... + RCU, where C' A -I C = 1. Note that
,':;;.
(31)
tffR2 = C(p) fcorP+Ig(r2) dr = -2pcP'(0), o
(32)
tffR4 = C(p) fcorP+3g(r2) dr= 4p(p + 2) cP" (0). o
J ;
Consider the higher-order moments of f = RU. The odd-order moments 61 Rare 0, and hence the odd-order moments of fare O. We have (33)
In fact, all moments of X - .... of odd order are O. , Consider tffUPPkU/, BecauseU'U = 1, P
(34)
1=
L i,j=l
tffWU/=ptffUI4+p(p-1)tffUI2uf.
54
THE MULTIVARIATE NORMAL DISTRIBUTION
Integration of tff sin~ 0\ gives tffU\4 = 3/[p(p + 2)]; then (34) implies ,fU\=U,: = l/[p(p + 2)]. Hence tffy;4 = 3tffR 4/[p(p + 2)] and tffy\2Yl = ,fR 4 /[p(p+2)]. Unless i=j=k=l or i=j*k=l or i=k*j=l or i = I j = k. we have tffUiUPkUI = O. To summarize tffU;UPkUI = (8i /'kl + aikajl + DjfDjk)/[p(p + 2)]. The fourth-order moments of X are
*
(35)
tff(Xi -lLi)(Xj -lLj)(Xk -lLd(XI-ILI) tffR4 p( P + 2) (\jAkf + tffR4
Aik
Ajl + AjfAjk )
p
= ---,
- +2 (O'ij·(Tk/ + O'ikO'j'l + (J'jf(J'jk)'
(tffR2r P
The fourth cumulant of the ith component of X standardized by its standard deviation is
(36)
= 3K, say. This is known as the kurtosis. (Note that K is ~tff{(Xi - ILY/ [tff(X i - ILYFl - 1.) The standardized fourth cumulant is 3K for every component of X. The fourt" cumulant of Xi' Xj' X k, and XI is (37)
K"kl = tff( Xi - IL,)( Xj - ILJ( X k - ILk)( XI - ILl) - «(J'ij(J'kl + (J'ik (J'jl + (J'j/(J'jd
For the normal distribution
K
= O. The fourth-order moments can be written
(38)
More detail about elliptically contoured distributions can be found in Fang and Zhang (1990).
2.7
ELLIPTICALLY CONTOURED DISTRIBUTIONS
55
The class of elliptically contoured distributions generalizes the normal distribution, introducing more flexibility; the kurtosis is not required to be O. The typical "bell-shaped surface" of IAI-tg[(x-v)'A-1(x-v)] can be more or less peaked than in the case of the normal distribution. In the next subsection some eXflmpies are given. 2.7.5. Examples (l) The multivariate t-distribution, Suppose Z - N(O, lp), ms 2 1: x~, and Z
and
S2
(39)
are independent. Define Y= (l/s)Z. Then the density of Y is
r(m;p) ( y,y)_m;p r( ~ )m m '
--:-=,:,-..::..........:....- 1 + P / 27T P / 2
and (40) If X = ..... + CY, the density
0"
X is
(41)
(2) Contaminated normal. The contaminated normal distribution is a mixture of two normal distributions with proportional covariance matrices and the same mean vector. The density can be written ( 42)
where c > 0 and 0 S; e S; 1. Usually e is rather small and c rather large. (3) Mixtures of /lormal distributions. Let w(v) be a cumulative distribution function over 0 S; v S; 00. Then a mixture of normal densities is defined by (43)
56
THE MULTIVARIATE NORMAL D!STRIBUTION
which is an elliptically contoured density. The random vector X with this density has a representation X = wZ, where Z - N(tJ., I) and w - ;.y(w) are independent. Fang, Kotz, and Ng (1990) have discussed (43) and have given other examples of elliptically c(.ntoured distributions.
PROBLEMS 2.1. (Sec. 2.2)
Let f(x,y) = 1, O:sx:s 1, O:sy:s 1, = 0, otherwise.
Find: (a) F(x, y).
(b) F(x). (c) f(x). (d) f(xiy). [Note: f(xoiyo) = 0 if f(x o, Yo) = 0.1 (e) rf;'X"Y"'.
(f) Prove X and Yare independent.
2.2. (Sec. 2.2)
Lei f(x, y)
= =
2,O:sy:sx:s 1, 0, olhelwise.
Find: (a) F(x, y).
(f) f(xiy).
(b) F(x).
f(yix). gX"y"'. (j) Are X and Y independent?
(c) f(x). (d) G(y).
(g)
(h)
(e) g(y).
Let f(x, y) = C for x 2 + y2 :s k 2 and 0 elsewhere. Prove C = 1/(1re), JfX= (ty=O, gX 2 = gy2=k 2/4, and gxy=O. Are X and Y
2.3. (Sec. 2.2)
mdeprrdent? 2.4. (Sec. 2.2) Let F(x l , x 2 ) be the joint cJf of Xl' X 2 , and let Fj(x) be the marginal cdf of Xi' i = 1,2. Prove that if f~(x) is continuous, i = 1,2, then F(x I, t2) is continuous. 2.5. esc-c. 2.2) Show that if the set XI"'" X, is independent of the set X,+I"'" Xp, then
57
IOBLEMS
;1.6. (Sec.
2.3) Sketch the ellipsl!s f(x, y) normal density with
=
0.06, where f(x, y) is the bivariate
o.
(a) J.l.x = 1, J.l.y = 2, u}
= 1, u/ = 1, Pxy = (b) J.l.x = 0, J.l.y = 0, u} = 1, u/ = 1, Pxy = O.
(c) J.l.x = 0, J.l.y = 0, u} = 1, u/ = 1, Pxy = 0.2. (d) J.l.x = 0, J.l.y = 0, u} = 1, u/ = 1, Pxy = 0.8. (e) J.l.x = 0, J.l.y = 0, u} = 4, u/ = 1, Pxy = 0.8.
2.7. (Sec. 2.3) Find b and A so that the following densities can be written in the form of (23). Also find J.l.x, J.l.y, ux' cr." and Pxy(a) 21'TTexp{ -
(b)
H<x _1)2 + (y -
2)2]).
2 1 ( X /4-1.6.l.Y/2+ y2 ) 2.4'TT exp 0.72 .
1 (c) 2'TTexp[ -j(x 2 + y2 + 4x - 6y + 13)], (d)
2~exp[ - ~(2X2 + y2 + 2xy -
22x - 14y + 65)].
2.8. (Sec. 2.3) For each matrix A in Problem 2,7 find .C so that C'AC =
~
/,
2.9. (Sec, 2.3) Let b = O.
(a) Write the density (23). (b) Find 1:. 2.10. (Sec. 2.3) Prove that the principal axes of (55) of Section 2.3 are along the 45 0 and 1350 lines with lengths 2';c(1 + p) and 2';c(1 - p), respectively, by transforming according to Yt = (ZI + Z2)/ Ii, Y2 = (ZI - Z2)/ Ii. 2.11. (Sec. 2.3) Suppose the scalar random variables XI"'" Xn are· independent and have a density which is a function only of x? + ... +x;. Prove that the Xi are normally distributed with mean 0 and common variance. Indicate the mildest conditions on the density for your proof.
58
THE MULTIVARIATE NORMAL DISTRIBUTION
2.12. (Sec. 2.3) Show that if Pr{X ~
o.
y
~
OJ
=
a fOJ the distribution
then p = cos( I - 2 a hr. [Hint: Let X = U, Y = pU + cos 2 7T( ~ - a) geometrically.] 2.13. (Sec. 2.3)
Prove that if Pij = p, i
4' j,
J1 -
p2 V and verify p =
i, j = 1, ... , p, then P ~ -l/(p -1).
2.1-1. (Sec. 2.3) COllcentratioll ellipsoid. Let the density of the p-component Y be f(y)=nw+ 1)/[(p+2)7T]~P for y'y5,p+2 and 0 elsewhere. Then $Y=O and "IT' = I (Problem 7.4). From this result prove that if the density of X is g(x) = v'iAfnw + 1l/[(p + 2)7T ]~P for (x - .... )'A(x - ....) 5,p + 2 and 0 elsewhere, then ,fX= .... and ,f(X- .... )(X- .... )' =A- I • 2.15. (Sec. 2.4) Show that when X is normally distributed the components are mutually independent if and only if the covariance matrix is diagonal. 2.16. (Sec. 2.4) Find necessary and sufficient conditions on A so that AY + X has a continuous cdf. 2.17. (Sec. 2.4) Which densities in Problem 2.7 define distributions in which X and Yare independent? 2.18. (Sec. 2.4) (a) Write the marginal density of X for each case in Problem 2.6. (b) Indicate the marginal distribution of X for each case in Problem 2.7 by notation N(a. b). (cl Write the marginal densitY of XI and X 2 in Problem 2.9. 2.19. (Sec. 2.4) What is the distribution of Z the densities in Problem 2.6?
=
th~
X - Y when X and Y have each of
2.20. (Sec. 2.4) What is the distribution of XI + 2X2 - 3X3 when XI' X 2 , X3 have the distribution defined in Problem 2.9? 2.21. (Sec. 2.4) Let X = (XI' X 2 )'. where XI = X and X 2 = aX + b and X has the distribution N(O,1). Find the cdf of X. 2.22. (Sec. 2.4) IV( /L,
(T
Let XI"'" X N be independently distributed, each according to
2).
(a) What is the distribution of X = (Xi"'" X N )'? Find the vector of means and the covariance matrix. (h) Using Theorem 2.4.4, find the marginal distribution of X = LX';N.
PROBLEMS
59
2.23. (Sec. 2.4) Let XI' ... ' X N be independently distributed with Xi having distribution N( f3 + 1Zi, (T 2), where :t;i is a given number, i = 1, ... , N, and ~izi = o. (a) Find the distribution of(XI, ... ,XN )'. (b) Find the distribution of X and g = ~xizJ~zl for ~zl > o. 2.24. (Sec. 2.4) Let (XI'YI)',(X2'Y2)',(X3,Y3)' be independently distributed, (Xi' Y)' according to
i = 1,2,3.
(a) Find the distribution of the six variables. (b) Find the distribution of (X, Y)'. 2.25. (Sec. 2.4) Let X have a (singular) normal distribution with mean 0 and covariance matrix
(a) Prove I is of rank l. (b) Find a so X = a'Y and Y has a nonsingular normal distribution, ane! give the density of Y. 2.26. (Sec. 2.4) Let
(a) Find a vector u '1= 0 so that Iu = O. [Hint: Take cofactors of any column.) (b) Show that an] matrix of the form G = (H u), where H is 3 X 2, has the property
(e) Using (a) and (b), find B to satisfy (36). (d) Find B- 1 and partition according to (39). (e) Verify that CC' = I. 2.27. (Sec. 2.4) Prove that if the joint (marginal) distribution of XI and X 2 is singular (that is, degenerate); then the joint distribution of XI' X 2, and X) is singular.
60
THE MULTIVARIATE NORMAL DISTRIBUTION
2.28. (Sec. 2.5) In each part of Prohlem 2.6, find the conditional distribution of X given Y=y, find the conditional distribution of Y given X=x, and plot eaeh rcgr"s~'ion linc on the appr()priatc graph in Problem 2.6. 2.29. (Sec. 25)
Let J..'- = 0 and
0.80 I. -0.56
1. ~ =
lUm ( -0.40
-0.40) -0.56 . 1.
(a) Find the conditional distribution of Xl and X 3 , given X 2 = X2' (b) What is the partial correlation between XI and X3 given X 2? 2.30. (Sec. 2.5)
X3
In Problem 2.9, find the conditional distribution of Xl and X 2 given
=X3'
2.31. (Sec. 2.5)
Verify (20) directly from Theorem L.5.1.
2.32. (Sec. 2.5) (a) Show that finding Ot to maXimize the absolute value of the correlation hetween Xi and Ot' X(2) is equivalent to maximizing (u;i)Ot)l subject to Ot'l:220t constant. (b) Find Ot by maximizing (u(i)OtJ2 - A(Ot '::;22 Ot - C), where c is a constant and A is a Lagrange multiplier. 4J
~
2.33. (Sec. 2.5) lnuariallceofthe mulliple cOITelalion coefficient. Prove that Ri . q + l , "'~;P is an invariant characteristic of the multivariate normal distribution of Xi and X(2) under the transformation xi = bix; + c; for b; 0 and X(2)* = HX(2) + k for H nonsingular and th~ every function of J..'-;, CTii' U(;» I.lP), and 122 that is invariant is a function of R j • q + I, ,,' p'
*"
2.34. (Sec. 2.5)
Prove that
l-Rl.,!+I.
..• p
1 =_1_1 I Pkjl
Pki
Pi}
I,
k,j =q
+ 1, ... ,p .
Pkj
2.35. (Sec. 2.5) Find the multiple correlation coefficient between Xl and (X 2 , X 3 ) in Problem 2.29. 2.36. (Sec. 2.5)
Prove explicitly that if I is positive definite,
f
61
P~OBLEMS
2.37. (Sec. 2.5) PrOve Hadamard's inequality p
III::; nCTii' i-I
[Hint: Using Problem 2.36, prove III::;CTII II 22 I, where I (p - 1), and apply induction.]
zz
is (p-l)x
2.38. (Sec. 2.5) Prove equality holds in Problem 2.37 if and only if I is diagonal. 2.39. (Sec.2.5) Prove {3IZ.3 = CTI2'3/CTZZ'3 = PI3.ZCTI.Z/CT3.Z and {313.2 = CTI).2/U33.Z = PI3.2 CTI'Z/ CT3·2' where CT/k = CTjI·k· 2.40. (Sec. 2.5) Let (Xl' Xz) have the density n (xl 0, I) = I(x l , x 2 ). Let the density of Xl given Xl =x l be f(x2Ixl)' Let the joint density of Xl' X 2 • X3 be I(XI' xz)/(x 3 Ix l)' Find the covariance matrix of Xl. X 2 • X3 and the partial correlation between X 2 and X3 for given XI' Prove 1 - R~'23 = (1 - Pf3X1 - Pf2.3)' [Hint: Use the fact that the variance of Xi'in the conditional distribution given X2 and X3 is (1- Rf.23)CTlI']
2.41. (Sec. 2.5)
2.42. (Sec. 25) If P = 2, c~n there be a difference between the simple correlation
between Xl and Explain. 2.43. (Sec. 2.5)
x2
and the multiple correlation between Xl and
X(Z) =
Xl?
Prove CTik·q-I ..... k-I. k+I ..... P
{3ik.q+I ..... k-l.k+I •. .• P
CTkk.q+ I .... • k-l.k+I ..... p CTi·q + I ..... k-l.k+l, ... ,p
= Pik'q-I .... k-l.k+I ..... p CTk.q+I, .... k-l.k+I, .... p.
i
=
1., ..• q,
k
=q+
1, ... , p. where CT/q+I ..... k-l.k+I ..... p.., j=i.k. [Hint: Prove this for the special case k=q+l by using Problem 2.56 with PI = q. P2 = 1. P3 = P - q - 1.] C1jj.q+I ..... k-l.k+I ..... p.
~.44. .,
(Sec. 2.5) Give a necessary and sufficient condition for of CT1.q+I ..... CTjp ' .
:.45. (Sec. 2.5)
Show
[Hint: Use (19) and (27) successively.]
R
j • q + l ..... P
=
0 in terms
62
THE MULTIVARIATE NORMAL DISTRIBUTlCN
2.46. (Sec. 2.5)
Show Pi].q+l, ",p = f3ij.q+l, ... ,pf3j l. q +I, .... p·
2.47. (Sec. 2.5)
Prove
_u i2
Pi2·3
"p
=
Vu llu22
.
[Hint: Apply Theorem A.3.2 of the Appendix to the cofactors used to calculate u~ . 2.48. (Sec. 2.5) Show that for any joint distribution for which the expectations exist and any function h( X(2») tha t
[Hiw: In the above take the expectation first with respect to Xi conditional 011
X( 2).]
2.49. (Sec. 2.5) Show that for any function h(X(2») and any joint distribution of Xi and Xl2l for which the relevant expectations exist, J'[Xi - h(X{2))f = J'[Xi g(X(2»)]2 + J'[g(X(2») - h(X(2»)]2, where g(X(2») = J'Xj IX(2) is the conditional expectation of Xi given X(2) = x(2). Hence g(X(2») minimizes the mean squared error of prediction. [Hint: Use Problem 2.48.] 2.50. (Sec. 2.5) Show that for any function h(X(2» and any joint distribution of Xi and X(2) for which the relevant expectations exist, the correlation between Xi and h(X(2») is not greater than the correlation between Xi and g(X(2»), where g(X(2») = J' X i lx(2). 2.51. . (Sec. 2.5)
Show that for any vector functicn h(x(2»)
C [X(l) - h(X(2»)] [X(l) - :1(X(2»)], - J'[X(1) - C X(l)IX(2)][X(I) - J' X(I)IX(2»), is positive semidefinite. Note this generalizes Theorem 2.5.3 and Problem 2.49. 2.52. (Sec. 2.5) Verify that I12Ii21 similarly to I. 2.53. (Sec. 2.5)
=
-"'Ill "'12' where'" = I-I is partitioned
Show
where 13 = I12Iit [Hint: Use Theorem A.3.3 of the Appendix and the fact that I -1 is symmetric.]
63
PROBLEMS
2.54. (Sec. 2.5) Use Problem 2.53 to show that
x'I -IX =
(X(I) -
I12Ii2IX(2»)'I'/2( x(1) - I12Ii2IX(2») + x(2)'Iib(2).
2.55. (Sec. 2.5) Show
c8'( x(I)lx t2), x(3»)
=
fL(l) + I 13I;:l (X(3)
-
fL(3»)
+ (I12 - II3 I 331I 32 )( I22 - I23 I 331I 32 . [ X(2) - fL(2) - I23 I ;} (X(3) - fL(3»)] .
rl
2.56. (Sec. 2.5) Prove by matrix algebra that
I22 III-(I 12 I 13) ( I32
I23)-I(I21) _ -I I33 I31 -I11- I 13 I 33 I 31
- (I12 - II3I331 I32)(I22 - I23I331 I 32 ) -1(I21 - I23I31 I 31 ). 2.57. (Sec. 2.5) Inuariance of the partial correlation coefficient. Prove that P12.3, ... , P. is invariant under the transformations xi = ajx j + b;X(3) + cj ' a j > 0, i = 1,2, x(3)* = ex(3) + d, where x(3) = (x 3 , ••. , x p )', and that any function of fL and I that is invariant under these transformations is a function of P12.3, ... , p' 2.58. (Sec. 2.5) Suppose X(I) and have the density
of q and p - q components, respectively,
X(2)
where
Q=
(x O) -
+ ( X(2)
fL(I») ,A II ( x(l) - fL(I») + (x(l) - fL(I») ,A 12 ( x(2) -
fL(2») ,A 21 ( x(l) -
fLO»)
+ (X(2)
-
-
fL(2»)
fL(2») ,An( X(2) - fL(2»).
Show that Q can be written as QI + Q2' where
fL(I») + Ai/Ad X(2) -
Q1 =
[(x(l) -
Q2 =
(X(2) - fL(2»)'(A22 -A 21 A'IIA I2 )(x(2) -
fL(I») + A,/AnCX(2) - fL(2»)].
fL(2»)),A!I[(x(I) -
Show that the marginal density of
X(2)
fL(2»).
is
IA22 - A21Ai1IA12lt
~~--~~~~e
_lQ 2
2
(21T) t(p-q) Show that the conditional density of
X(I)
given
IAlIit _lQ --,-e 2
X(2)
= x(2) is
I
(21T )'q
(without using the Appendix). This problem is meant to furnish an alternative proof of Theorems 2.4.3 and 2.5.1.
64
THE MULTIY ARIATE NORMAL DISTRIBUTIl
2.59. (Sec. 2.6)
Prove Lemma 2.6.2 in detail.
'"
2.60. (Sec. 2.6) Let Y be distributed according to N(O, I). Differentiating the characteristic function, verify (25) and (26). 2.61. (Sec. 2.6) Verify (25) and (26) by using the transformation X - IJ. = CY, where 1= CC', and integrating the density of Y. 2.62. (Sec. 2.6) Let the density of (X, Y) be 2n(xIO, l)n(yIO, 1),
O!>y!>x
< 00, O!>
O!> -y!> -x
o
-x!>y
< 00,
O!>x!>
-y
otherwise.
Show that X, Y, X + Y, X - Y each have a marginal normal distribat;on. 2.63. (Sec. 2.6) Suppose X is distributed according to N(o, I). Let 1= (0"1"'" Prove
O"p)'
, 0"10"1
G(XX' ®XX')
=
I
®
I + vee I (vee I)' +
[
:,
O"IO"p
=
(I + K)(I
®
I) + vee I (vee I)',
where EI.E'I
K=
: [
and
E;
EIE~
is a column vector with 1 in the ith position and O's elsewhere.
2.64. Complex normal distribution. Let (X', Y')' have a normal distribution with mean vector (,ix, ,iy)' and covariance matrix
1=
(~
- III) f
'
where f is positive definite and III = - III' (skew symmetric). Then. Z = X + iY is said to have a complex normal distribution with mean 6 = I~x ilJ.y and covariance matrix G(Z - 6)(Z - 6)* = P = Q + iR, where Z* =X' - iY'. Note that P is Hermitian and positive definite.
+
(a) Show Q = 2r and R = 2111. (b) St,ow IPI 2 = 1211. [Hint: If+illli
=
If-illll.]
PROBLEMS
6S
(c) Show
Note that the inverse of a Hermitian matrix is Hermitian. (d) Show that the density of X and Y can be written
'. ," i ,,"'
Complex no/mal (continued). If Z has the complex normal distribution of Problem 2.64, show that W = AZ, where A is a nonsingular complex matrix, has
,
the complex normal distribution with mean AS and covariance matrix C(W) = APA*.
5, •
2.66. Show that the characteristic function of Z defined in Problem 2.64 is
where l?ll(x + iy) = x. 2.6'/. (Sec. 2.2) Show that f''-.e- x'/2dx/& is approximately (l_e- 2a '/,,)1/2. [Hint: The probability that (X, Y) falls in a square is approximately the probability that (X, Y) falls in an approximating circle [P6lya (1949)].] 2.68. (Sec. 2.7) For the multivariate I-distribution with density (41) show that GX= V- and C(X) = [m/(m - 2)]A.
CHAPTER 3
Estimation of the Mean Vector and the Covariance j\Jlatrix
3.1. INTRODUCTION The multivariate normal distribution is specified completely by the mean vector f.L and the covariance matrix l:. The first statistical problem is how to estimate these parameters on the basis of a sample of observations. In Section 3.2 it is shown that the maximum likelihood estimator of .... is the sample mean; the maximum likelihood estimator of l: is proportional to the matrix of sample variances and covariances. A sample variance is a sum of squares of deviations of observations from the sample mean divided by one less than the number of observations in the sample; a sample covariance is similarly defined in terms of cross products. The sample covariance matrix is an unbiased estimator of I. The distribution of the sample mean vector is given in Section 3.3, and it is shown how one can test the hypothesis that .... is a given vector when l: is known. The case of l: unknown will he treated in Chapter 5. Some theoretical properties of the sample mean are given in Section 3.4, and the Baves estimator of the population mean is derived for a normal a priori distribution. In Section 3.5 the James--Stein estimator is introduced; impwwments llver the sample mean for the mean squared error loss function arc discussed. In Section 3.6 estimators of the mean vector and covariance matrix of elliptically contoured distributions and the distributions of the estimators art. treated.
An inrroJuction to Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0-471-36091-.0 Copyright © 2003 John Wiley & Sons, Inc.
66
3.2
ESTIMATORS OF MEAN VECTOR AND COVARIANCE MATRIX
67
3.2. THE MAXIMUM LIKELIHOOD ESTIMATORS OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
Given a sample of (vector) observations from a p-variate (nondegenerate) normal distribution, we ask for estimators of the mean vector IJ. and the covariance matrix I of the jistribution. We shall deduce the maximum likelihood estimators. It turns out that the method of maximum likelihood is very useful in various estimation and hypothesis testing problems concerning the multivariate normal distribution. The maximum likelihood estimators or modifications of them often have some optimum properties. In the particular case studied here, the estimators are asymptoti-;ally efficient [Cramer (1946), Sec. 33.3]. Suppose our sample of N observations on X distributed according to N( .... ,!) is x 1' ... 'X N , where N>p. The likelihood function is N
(1)
L=On(xal .... ,I) «=1
In the likelihood function the vectors Xl"'" X N are fixed at the sample values and L is a function of IJ. and I. To emphasize that these quantities are variable~ (and not parameters) we shall denote them by ....* and I*. Then the logarithm of the likelihood function is
(2)
logL= -!pNlog2'lT-!NlogII*1 N
-! L
(xa - ....*),I*-l(X a
-
*).
....
«=1
Since log L is an increasing function of L, its maximum is at the same point in the space of IJ.*, I* as the maximum of L. The maximum l.ikelihwd estimators of .... and I are the vector IJ.* and the positive definite matrix '1:* that maximize log L. (It rcmains to bt: st:cn that thc suprcmum oll')g L is attained for a positivc definitc matrix I *.) Let the sample mean vector be 1
N
N LX la «=1
(3) 1
N
N
L
Ot=1
. xpa
68
ESTIMATION OF THE MEAN VECTOR AND TIiE COVARIANCE MATRIX
where Xu =(Xla""'xpa )' and i j = E~_lxja/N, and let the matrix of sums of squares and cross· products of deviations about the mean be N
(4)
A=
E (x,,-i)(x,,-i)'
It will be convenient to use the following lemma: ,\'\
Lemma 3.2.1. defined
hr
Let XI"'" XN be N (p-component) vectors, and let i be (3). Then for any vector b
IN
;E
(5)
N
(x .. -b)(xa- b)'
E (xa-i)(xa-i)' +N(i-b)(i-b)'.
=
a-1
a-I
Proof
(6) N
N
E (xa'-b)(xa- b)'= E
[(Xa-i) + (i-b)][(xa-i) +(i-b»)'
N
=
E
[(x,,-i)(x,,-i)'+(x,,-i)(i-b)'
a-I
+(i-b)(xa-i)' + (x-xb)(i-b)'] =
E(Xu
-i)(xa -i)' + [
a-I
E
(Xa -i)](i -b)'
a-I N
+(i-b)
.
E (xu-i)' +N(i':"b)(i-h)'.
""t.':·'.,
The second and third terms on the right-hand ,ide are 0 becauSe' E(xa -i) Ex" - NX = 0 by (3). • , " '~ When we let b = p.* , we have
.
(7) N
E (x,,- p.*)(x" a-I
N
p.*)'
=
E (xa -i)(x" -i), +N(i- p.*)(~C p.*)' a-I
=A + N(i- p.*)(i- p.*)'.
69
3.2 ESTIMATORS OF MEAN VECTOR AND COVARlANCEMATRlX
Using this result and the properties of the trace of a matrix (tr CD = = tr DC), we ha Ie
'Ecijd ji
~<.;"'. ::
(8) N
E (x
, a
N
E (x a -p.*)'1:*-I(xa -p.*)
-p.*)'1:*,-I(x a -p.*)=tr
a-I
a-I N
,t
=tr
f
=
E 1:*-I(Xa -
p.*)(x a - 110*)'
tr 1:*-IA + tr 1:*-1 N(i - p.*)(i - 110*)'
= tr 1:*-IA + N(i - 110*)'1:*-1 (X - 110*). Thus we can write (2) as log L = - tPN log(2'IT) - tN 10gl1:*1
(9)
- ttr 1:*-IA - tN(i - 110* )'1:*-1 (x - 110*)' Since 1:* is positive definite, 1:*-1 is positive definite, and N(ip.*)'1:*-I(X - 110*) ~ 0 and is 0 if and only if 110* =i. To maximize the second and third terms of (9) we use the following lemma (which is also used in later chapters): Lemma 3.2.2.
If D is positive definite of order p, the maximum of f(G) =NllogIGI-trG-ID
(10)
with respect to positive definite matrices G exists, occurs at G = (lIN)D, and has the value (11)
f[(IIN)D] =pNlogN-NlogIDI-pN.
, Proof, Let D =EE' artd E'G-IE =H. Ther, G =EH-IE', and IGI = lEI 'IH- 11 'IE'I = IH-II ·IEE'I = Inll IHI, and tr G-ID = tr G-lEE' = ~E'~:-\''';'trH. Then the function to be maximized (with respect to positive defiriiteH) is , " ,~ '.~. ' . ;(12):~~;:, i$ f=-NlogIDI +NlogIHI-trH .
'!:~,7Jr~'~here
.
.
,;
~
, .-; .
'( I3~:::'" :, ..
:\~~;' "
I'
:; '>Ii
T is lower triangular (Corollary A.I.7). Then the
f= -NlogiDI +NlogITl 2 -tr1T' p
= -NloglDI +
E (Nlogt~-t~)- Et~ i-I
i>j
70
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
occurs at tj~ = N, t ij = 0, i (l/N)D. •
*" j;
that is, at H = NI. Then G = (1/N)EE'
=
Theorem 3.2.1. If Xl' ... , XN constitute a sample from N(JL, :£) with p < N, the ma;'(imum likelihood estimators of JL and :£ are v.. = i = (1/ N)r.~~ 1 xa and i = (l/N)r.~jxa -iXx a -x)', respectively. Other methods of deriving the maximum likelihood estimators have been discussed by Anderson and Olkin (1985). See Problems 3.4, 3.8, and 3.12. Computation of the estimate i is made easier by the specialization of Lemma 3.2.1 (b = 0) N
(14)
N
E (xa-x)(xa- x )'= E xax~-lVxi'.
An element of r.~·=lXaX~ is computed as r.~~lXjaXja' and an element of ~:ii' is computed as NXji j or (r.~~lXjJ(r.~~lXja)jN. It should be noted that if N > p. the probability is 1 of drawing a sample so that (14) is positive definite; see Problem 3.17. The covariance matrix can be written in terms of the variances or standard deviations and correlation coefficients. These are uniquely defined by the variances and covariances. We assert that the maximum likelihood estimators of functions of the parameters are those functions of the maximum likelihood estimators of the parameters.
Lemma 3.2.3. Let fUJ) be a real-valued function defined on il set S, and let rjJ be a single-valued function, with a single-valued inverse, on S to a set S*; that
is, to each (J E S there co"esponds a unique (J* 8* E S* there co"esponds a unique (J E S. Let
E
S*, and, conversely, to each
(15)
Then if f(8) attains a maximum at (J= (Jo, g«(J*) attains a maximum at 11* = IIJ' = rjJ(8 u). If the maximum of f(lI) at 8u is uniquO!, so is the maximum of g(8" ) at 8S· Proof By hypothesis f(8 0 ) '?f(8) for all 8 E S. Then for any (J*
E
S*
Thus g«(J*) attains a maximum at (J;r. If the maximum of f«(J) at (Jo is unique, there is strict inequality above for 8 80 , and the maximum of g( (J*) is unique. •
*"
3.2 ESTIMATORS OF MEAN VECfOR AND COVARIANCE MATRIX
71
We have the following corollary: Corollary 3.2.1. If on the basis of a given sample 81 " " , 8m are maximum likelihood estimators of the parameters 01, . " , Om of a distribution, then cP 1C0 1, .( •• , Om), . .. , cPmC0 1, ••• , Om) are maximum likelihood estir.l1tor~ of cPIC 01" " , Om)'.'" cPmCOI"'" Om) if the transformation from 01" , . , Om to cPI"'" cPm is one-to-one.t If the estimators of 0" ... , Om are unique, then the estimators of cPI"'" cPm are unique. Corollal1 3.2.2.
If
XI"'"
XN
constitutes a sample from NCJL, :£), where
x
Pij Cpjj = 1), then the maximum likelihood estimator of JL is fJ. = = (l/N)LaX a ; the maximum likelihood estimator of a/ is = (l/N)LaCXia= (l/N)CLaxta - Nit), where x ia is the ith component of Xa and Xi is the ith component of x; and the maximum likelihood estimator of Pij is a ij
=
ai aj
6/
xY
(17)
Proof The set of parameters f..Li = f..Li' a j 2 = ifij , and Pij = a j / vajj~j is a one-to-one transform of the set of parameters f..Li and aij. Therefore, by Coronary 3.2.1 the estimator of f..Lj is ilj' of a? is a-jj , and of Pij is
(18)
•
Pearson (1896) gave a justification for this estimator of Pij' and (17) is sometimes called the Pearson co"elation coefficient. It is also called the simple co"eiation coefficient. It is usually denoted by rir tThe assumptIOn that the transformation is one-to-one is made so that the sct >" ..• , >m uniquely dcfine~ the likelihood. An ahcrnative in casc 0* = '/>( 0) docs not havc a unique inverse is to define s(O*) = (0: >(0) = O*} and g(O*) = sup f(o)1 OE S(o'), which is considered the "induced likelihood" when f(O) is the likelihood function. Then 0* = >(0) maximizes g(O*), for g(O*)=supf(O)IOES(O')~supf(O)IOES=f(O)=g(O*) for all 0* ES*. [See, e.g., Zehna (1966).]
72
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
•
IIj
dllj
Figure 3.1
A convenient geometrical interpretation of this sample (Xl' X2"'" is in terms of the rows of X. Let
XN )
=~ .
(19)
u:
that is, is the ith row of X. The vector Uj can be considered as a vector in an N-dimensional space with the ath coordinate of one endpoini being x ja and the other endpoint at the origin. Thus the sample is represented by p vectors in N-dimensional Euclidean space. By defmition of the Euclidean metric, the squared length of Uj (that is, the squared distance of one endpoint from the other) is u:Uj = r.~_lX;a' Now let us show that the cosine of the an le between Uj and Uj is U,Uj/ yuiujujuj = EZ_1XjaXja/ EZ-lx;aEZ_1X}a' Choose the scalar d so the vector du j is orthogonal to Uj - dUj; that is, 0 = duj(uj - du j ) = d(ujuj duju j }. Therefore, d = uJu;lujuj" We decompose Uj into Uj - duj and du j [Uj = (Uj - dUj) + du j 1 as indicated in Figure 3.1. The absolute value of the cosine of the angle between u j and Uj is the length of du j divided by the length of
Uj;
that is, it is Vduj(duj)/uiuj = ydujujd/uiuj; the cosine is
U'jU j / p;iljujUj. This proves the desired result.
To give a geometric interpretation of ajj and a jj / y~jjajj' we introdu~' the equiangular line, which is the line going through the origin and the point (1,1, ... ,1). See Figure 3.2. The projection of Uj on the vector E = (1, 1, ... ~ 1)' is (E 'U;lE 'E)E = (E a x j ,,/E a l)E =XjE = (xj ' xj , ••• , i)'. Then we dec.ompose Uj into XjE, the projection on the equiangular line, and u,-XjE, ~e projection of U j on the plane perpendicular to the equiangular line. The squared length of Uj - XjE is (Uj - X,E}'(U j - XjE) = Ea(x 'a _X()2; this is NUjj = ajj' Translate u j - XjE and Uj - X/E, so thaf,each vector has anenppoint at the origin; the ath coordinate of the first vector is Xj" - Xj' .and of
!.2 ESTIMATORS OF MEAN VEcrOR AND COVARIANCE MATRIX
73
I
Figure 3.1
~e
s(:cond is Xja -Xj' The cOsine of the angle between these two vectors is .
(20)
. (Uj -XjE)'(Uj -XJE)
N
,E'
(Xia-Xi)(Xj,,-Xj)
a-l
N
N
a-l
a-l
E (Xia _X;)2 E (Xja _X )2 j
,
As an example of the calculations consider the data in Table 3.1 and graphed in Figure 3.3, taken from Student (1908). The measurement Xu = 1.9 on the first patient is the increase in the number of hours of sleep due to the use of the sedative A, X21 = 0.7 is the increase in the number of hours due to
Table 3.1. Increase in Sleep
.r
.. Patient 1 2
Drug A.
Drug B
Xl
Xl
1.9 0.8
,
3
1.1
4
"
5 6
0.1 -0.1
7
8 9 10
4.4
5.s 1.6 4.6 3.4
0.7 -1.6 -0.2 -1.2 -0.1
3.4 3.7 0.8
0.0 2.0
74
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
5 4
•
3
•
•
2
••
-2
• Figure 3.3. Increase in sleep.
sedative B, and so on. Assuming that each pair (i.e., each row in the table) is an observation from N(IJ.,"I), we find that A
(21 )
___
(2.33) 0.75 '
IJ.
-x -
I
= (3.61 2.56
S = (4.01
2.85
and
PI2 = r l2 = 0.7952. (S
2.56 ) 2.88 ' 2.85 )
3.20 '
will be defined later.)
3.3. THE DISTRIBUTION OF THE SAMPLE MEAN VECTOR;
INFERENCE CONCERNING THE MEAN WHEN THE COVARIANCE MATRIX IS KNOWN 3.3.1. Distribution Theory
In the univariate case the mean of a sample is distributed normally and independently of the sample variance. Similarly, the sample mean X defined in Section 3.2 is distributed normally and independently of I.
3.3 THE DISTRIBUTION OF THE SAMPLE MEAN VECTOR
75
To prove this result we shall make a transformation of the set of observation vectors. Because this kind of transformation is used several times in this book, we first prove a more general theorem. Theorem 3.3.1. Suppose XI"'" X N are independent, where Xa is distributed according to N(IL a , I). Let C = (c a,,) be an N X N orthogonal matrix. Then Ya=r:~_ICa"X" is distributed according to N(v a, I), where va= r:~=ICa"IL", a = 1, ... , N, and YI, ... , YN are independent. Prool The set of vectors YI , ... , YN have a joint normal distribution, because the entire set of components is a set of linear combinations of the components of XI"'" X N , which have a joint normal distribution. The expected value of Ya is N
(1 )
tG'Ya =
N
tG' E ca"X" = E ca" tG'X" /3-1
,,-I
N
~ Ca"IL,,=Va· /3-1
The covariance matrix between Ya and Yy is
(2)
C(Ya , Y;) = tG'(Ya - va)(Yy - v y )'
=
tG'L~1 ca,,(X,,- IL")][e~ cys(Xe- ILer] N
E
ca"cyetG'(X,,-IL,,)(Xe-IL.)'
/3, s-I N
E
/3,s-1
c a" cye o"e 2'
where 0ay is the Kronee ker delta (= 1 if a = 'Y and = 0 if a '1= 'Y). This shows that Ya is independent of YY ' a 'Y, and Ya has the covariance • matrix 2'.
*
76
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
We also use the following general lemma: Lemma3.3.1. IfC=(c",,) is orthogonal, then where y" = r:~_1 c""x", a = 1, ... , N.
r:~_IX"X~=r:~_IY"Y~'
Proof N
(3)
L
,,-I
y"y: =
L LCa"x" LCa'YX~ a
= =
"
'Y
L (Lc""Ca'Y )x"x~ {3,'Y " L
{30'Y
{j"'Yx,,x~
• Let Xl>'." X N be independent, each distributed according to N(JL, I). There exists an N X N orthogonal matrix B = (b ",,) with the last row (4)
(l/m, ... ,l/m).
(See Lemma A.4.2.) This transformation is a rotation in the N-dimensional space described in Section 3.2 with the equiangular line going into the Nth coordinate axis. Let A = NI, defined in Section 3.2, and let (5)
N
Z,,=
L
b""X".
{3~1
Then
(6) By Lemma 3.3.1 we have N
(7)
A=
LX"x:-NXi' a-I N
=
L a-I
Z"Z~ - ZNZ'rv
77
3.3 THE DISTRIBUTION OF THE SAMPLE MEAN VECTOR
Since ZN is independent of Z\"",ZN_\, the mean vector of A. Since N
(8)
tCz N =
L
1
N
f3~1
bNf3 tCXf3
=
L "' .... =/N .... ,
f3~1 yN
X = (1/ /N)ZN
ZN is distributed according to N(/N .... , I) and according to N[ .... ,(1/NYI]. We note N
(9)
tCz a =
L
X is independent
is distributed
N
baf3 tCXf3 =
f3~1
L
baf3 ....
f3~1
N
=
L
baf3 b Nf3 /N Il-
13=1
=0,
a*N.
Theorem 3.3.2. The mean of a sample of size N from N(Il-, I) is distributed according to N[ .... ,(1/N)I] and independently of t, the maximum likelihood estimator of I. Nt is distributed as r.~:IIZaZ~, where Za is distributed according to N(O, I), ex = 1, ... , N - 1, and ZI"'" ZN_\ are independent. Definition 3.3.1. only if tCet = e.
An estimator t of a parameter vector
e is
unbiased if and
Since tCX= (1/N)tCr.~~lXa = .... , the sample mean is an unbiased estimator of the population mean. However, ,IN-I
tCI = N tC
(10)
L
N-l
ZaZ~ = ! r I o
a~1
Thus
t
is a biased estimator of I. We shall therefore define N
( 11)
s=
N~lA= N~l L
(xa-i)(xa- i )'
a=1
as the sample covariance matrix. It is an unbiased estimator of I and the diagonal elements are the usual (unbiased) sample variances of the components of X. 3.3.2. Tests and Confidence Regions for the Mean Vector When the Covariance Matrix Is Known A statistical problem of considerable importance is that of testing the hypothesis that the mean vector of a normal distribution is a given vector.
78
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
and a related problem is that of giving a confidence region for tile unknown vector of means. We now go on to study these problems under the assumption that the covariance matrix "I is known. In Chapter 5 we consider these problems when the covariance matrix is unknown. In the univariate case one bases a test or a confidence interval on the fact that the difference between the sample mean and the population mean is normally distributed with mean zero and known variance; then tables of the normal distribution can be used to set up significance points or to compute confidence intervals. In the multivariate case one uses the fact that the difference between the sample mean vector and the population mean vector is normally distributed with mean vector zero and known covariance matrix. One could set up limits for each component on the basis of the distribution, but this procedure has the disadvantages that the choice of limits is somewhat arbitrary and in the case of tests leads to tests that may be very poor against some alternatives, and, moreover, such limits are difficult to compute because tables are available only for the bivariate case. The procedures given below, however, are easily computed and furthermore can be given general intuitive and theoretical justifications. The proct;dures and evaluation of their properties are based on the following theorem: Theorem 3.3.3. If the m-component vector Y is distributed according to N(v,T) (nonsingular), then Y'r-Iy is distributed according to the noncentral X"-distribution with m degrees of freedom and noncentrality parameter v 'T- I v. If v = 0, the distribution is the central X 2-distribution. Proof Let C be a nonsingular matrix such that CTC' = I, and define Z = CY. Then Z is normally distributed with mean tC Z = C tC Y = C v = A, say, and covariance matrix tC(Z - AXZ - A)' = tCC(Y - v XY - v)'C' = CTC' = I. Then Y'T- 1 Y= Z'(C')-Ir-IC-IZ = Z'(CTC')-IZ = Z'Z, which is the sum of squares of the components of Z. Similarly v'r-Iv = A'A. Titus y'T-Iy is distributed as Li~ I zl, where ZI' ... ' Zm are independently normally distributed with means A\, ... , Am' respectively, and variances 1. By definition this distributir n is the noncentral X 2-distribution with noncentrality parameter Lr~ AT. See Section 3.3.3. If Al = ... = Am = 0, the distribution is central. (See Problem 7.5.) •
\
Since IN(X - /J.) is distributed according to N(O, "I), it follows from the theorem that (12)
3.3 THE DISTRIBUTION OF THE SAMfLE MEAN VECfOR
79
has a (central) X2-distribution with p degrees of freedom. This is the fundamental fact we use in setting up tests and confidence regions concerning ..... Let xi( ex) be the number such that
(13)
Pr{xi> xi(ex)} =
ex.
Thus (14) To test the hypothesis that .... = .... 0' where .... 0 is a specified vector, we use as our critical region
(15) If we obtain a sample such that (15) is satisfied, we reject the null hypothe~is. It can be seen intuitively that the probability is greater than ex of rejecting the hypothesis if .... is very different from .... 0' since in the space of i (15) defines an ellipsoid with center at JLo' and when JL is far from .... 0 the density of i will be concentrated at a point near the edge or outside of the ellipsoid. The quantity N(X- .... oYI-1(X- .... 0) is distributed as a noncentral X 2 with p degrees of freedom and noncentrality parameter N( .... - .... O)'~-I( .... - .... 0) when X is the mean of a sample of N from N( .... , I) [given by Bose (1936a), (1936b)]. Pearson (1900) first proved Theorem 3.3.3 for v = O. Now consider the following statement made on the basis of a sample with mean i: "The mean of the distribution satisfies
(16) as an inequality on ....*." We see from (14) that the probability that a sample will be drawn such that the above statement is true is 1 - ex because the event in (14) is equivalent to the statement being false. Thus, the set of ....* satisfying (16) is a confidence region for .... with confidence 1 - ex. In the p-dimensional space of X, (15) is the surface and exterior of an ellipsoid with center .... 0' the shape of the ellipsoid depending· on I -I and the size on (1/N) X;( ex) for given I -I. In the p-dimensional space of ....* (16) is the surface and interior of an ellipsoid with its center at i. If I -I = I, then (14) says that the rob ability is ex that the distance between x and .... is greater than X; ( ex) / N .
80
.
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
Theorem 3.3.4. If x is the mean of a sample of N drawn from N(fJ., I) and I is known, then (15) gives a critical region of size ex for testing the hypothesis fJ. = fJ.o' and (16) gives a confidence region for fJ. ,)f confidence 1 - ex. Here xi( ex) is chosen to satisfy (13). The same technique can be used for the corresponding two-sample problems. Suppose we have a sample (xr)}, ex = 1, ... , N 1, from the distribution N( 1L(1), 1:), and a sample {x~;)}, ll' = 1, ... , N z, from a second normal population N( IL(Z), ~) with the same covariance matrix. Then the two sample means I
(17)
xtl)
N,
= -N "t..." X(I) a' I a~l
are distributed independently according to N[ 1L(1), (1/N1 )I] and N[ 1L(2), (1/ Nz):I], respectively. The difference of the two sample means, y=x(!)-x(Z), is distributed according to Nl'v,[(1/N!)
v
=
+ (1/Nz)]I}, where
fJ.(!) - fJ.(2). Thus
(18) is a confidence region for the difference v of th-:: two mean vectors, and a critical region for testing the hypothesis fJ.(I) = fJ.(Z) is given by (19) Mahalanobis (1930) suggested (fJ.(1) - fJ.(2))"I -I (fJ.(l) - fJ.(2») as a measure of the distance squared between two populations. Let C be a matrix such that I = CC' and let v(il = C-! fJ.(i), i = 1,2. Then the distance squared is (V(I) v(2»)'( v(1) - v(Z»), which is the Euclidean distance squared.
3.3.3. The Noncentral X2-Distribution; the Power Function The power function of the test (15) of the null hypothesis that fJ. = fJ.o can be evaluated from the noncentral x 2-distribution. The central x2-distribution is the distribution of the sum of squares of independent (scalar) normal variables with means 0 and variances 1; the noncentral xZ-distribution is the generalization of this when the means may be different from O. Let Y (of p components) be distributed according to N(A, I). Let Q be an orthogonal
3.3
81
THE DISTRIBUTION OF THE SAMPLE MEAN VECfOR
matrix with elements of the first row being
(20)
i
= 1, ... ,p.
Then Z = QY is distributed according to N( 'T, J), where
(21)
and T=~. Let V=;"Y=Z'Z=Lf~lz1- Then W=Lf~2Z? has a X 2distribution with p - 1 degrees of freedom (Problem 7.5), and Z\ and W have as joint density
(22)
where C- I = 2-)Phn~(p - 1)]. The joint density of V = W + Z~ and ZI is obtained by substituting w = v - zf (the Jacobian being 1):
(23)
The joint density of V and U = ZI/ IV is (dz\
=
IV du)
(24)
The admissible range of z\ given v is - IV to IV, and the admissible range of u is - 1 to 1. When we integrate (24) with respect to u term by term, the terms for a odd integrate to 0, since such a term is an odd function of u. In
82
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
the other integrations we substitute u = Ii (du = ~ds /
Ii) to obtain
=B[Hp-l),/3+~l f[Hp-l)lf(/3+~) r(~p + /3)
by the usual properties of the beta and gamma functions. Thus the density of V is (26)
We can use the duplication formula for the gamma function [(2/3 + 1) = (2/3)! (Problem 7.37), (27)
f(2/3 + 1) = r( /3 +
nrc /3 + 1)22/l/(;,
to rewrite (26) as (28)
This is the density of the noncentral X2-distribution with p degrees of freedom and noncentrality parameter T2. Theorem 3.3.5. If Y of p components is distributed according to N(>", 1), then V = Y' Y has the density (28), where T 2 = A' A. To obtain the power function of the test (15), we note that IN (X - .... 0) has the distribution N[ IN (.... - .... 0)' :£]. From Theorem 3.3.3 we obtain the following corollary: Corollary 3.3.1.
If X is the mean of a random sample of N drawn from
NC .... , :£), then N(X - ""0)':£-1 (X - .... 0) has a noncentral X 2-distribution with p degrees of freedom and noncent:ality parameter N( .... - .... 0)':£ -1( .... - .... 0).
83
3.4 THEORETICAL PROPERTIES OF ESTIMATORS OF THE MEAN VECTOR
3.4. THEORETICAL PROPERTIES OF ESTIMATORS OF THE MEAN VECTOR 3.4.1. Properties of l\laximum Likelihood Estimators It was shown in Section 3.3.1 that x and S are unbiased estimators of fl. and "I, respectively. In this subsection we shall show that x and S are sufficient statistics and are complete. Sufficiency
A statistic T is sufficient for a family of distributions of X or for a parameter (J if the conditional distribution of X given T = t does not depend on 0 [e.g., Cramer (1946), Section 32.41. In this sense the statistic T gives as much information about 0 as the entire sample X. (Of course, this idea depends strictly on the assumed family of distributions.) Factorization Theorem. A statistic t(y) is sufficient for 0 density fey 10) can be factored as
(1)
if and only if the
f(yIO) =g[t(y),O]h(y),
where g[t(y), 01 and h(y) are nonnegative and h(y) does not depend on O.
x
Theorem 3.4.1. If XI' ... , XN are observations from N(fl., "I), then and S are sufficient for fl. and "I. If f1. is given, r.~_I(Xa - fl.)(x a - fl.)' is sufficient for "I. If "I is given, x is sufficient for fl.. Proof The density of XI' ... ' X N is N
(2)
n n(xal fl.,"I)
a-I
= (2'lTf·
tNP I
"I1-
tN
exp [ -1 tr "I-I
= (2'lT) - tNPI "II- tN exp{ -
a~l (Xa -
fl.) ( Xa - fl.)']
HN( x - fl. )'"I -1 (X -
fl.) + (N - 1) tr "I -IS]}.
The right-hand side of (2) is in the form of (1) for x, S, fl., I, and the middle is in the form of (1) for r.~_I(Xa - fl.XXa - fl.)', I; in each case h(x l ,·.·, x N ) = 1. The right-hand side is in the form of (1) for x, fl. with h(x l ,···, x N ) = exp{ - 1(N -l)tr I-Is). • Note that if "I is given, i is sufficient for fl., but if fl. is given, S is not sufficient for I.
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
3_
Completeness To prove an optimality property. of the T 2-test (Section 5.5), we need the result that (x, S) is a complete sufficient set of statistics for (IJ., I).
v
84
Definition 3.4.1. A family of distributions of y indexed by 9 is complete if for every real-valued function g(y),
(3) identically in 9 implies g(y) = 0 except for a set of y of probability 0 fur every 9. If the family of distributions of a sufficient set of statistics is complete, the set is called a complete sufficient set. Theorem 3.4.2. The sufficient set of statistics x, S is complete for IJ., I when the sample is drawn from N(IJ., I).
Proof We can define the sample in terms of x and ZI' .•. ' Zn as in Section 3.3 with n = N - 1. We assume for any function g(x, A) = g(x, nS) that
f ... f KI I I - 4N g ( x, a~1 Za z~ )
( 4)
.exp{
-~[N(X- IJ.)'I-I(x- IJ.) + a~1 Z~I-IZa]}
n
·dX
n dza=O,
a=1
where K = m(2rr)- jpN, dX = nf=1 di;, and dZ a = nf=1 dz;a. If we let I-I = I - 20, where- 0 = 0' and 1-20 is positive definite, and let IJ. = (I - 20)-l t , then (4) is
(5)
0=
J-.-JKII-20IiNg(x'atZaZ~) .e;;p{ -
~[tr(I -
20)
C~I zaz~ + ~xX,) -2Nt'x + Nt '(I - 20) -I t]} dX
D
dZ a
f··· f g( x, B -NiX') ·exp[tr 0B +t'(Ni)]n[xIO, (l/N)I] n n(z"IO, 1) dX n dz a , a-I
= II - 201 iN exp{ - ~Nt'(I - 20) -I t}
n
u=1
n
.4 THEORETICAL PROPERTIES OF ESTIMATORS OF THE MEAN VECTOR
:6)
85
0==tC'g(x,B-N""ri')exp[tr0B+t'(N.X)1
= j ... jg(x, B - Nii') exp[tr 0B + t'(N.X)]h(x, H) dXdB, where hex, B) is the ,ioint density of x and B and dB = 0'5 i dbij' The right-hand side of (6) is the Laplace transform of g(x, B - Nii')h(x, Bl. Since this is 0, g(x, A) = 0 except for a set of measure O. •
Efficiency If a q-component random vector Y has mean vector tC'Y = v and covariance matrix tC'(Y - v Xy - v)' = qr, then
(7)
(Y-V),qr-I(y-V) =q+2
i~
called the concentration ellipsoid of Y. [See Cramer (1946), p. 300.J The defined by a uniform distribution over the interior of this ellipsoid has the same mean vector and covariance matrix as Y. (See Problem 2.14,) Let 9 be a vector of q parameters in a distribution, and let t be a vector of unbiased estimators (that is, tC't = 9) based on N observations from that distribution with covariance matrix qr. Then the ellipsoid d~nsity
(8)
N( t - 9)' tC' (
a ~o: f) ( iJ ~o: f), (t -
9) = q + 2
lies entirely within the ellipsoid of concentration of t; a log f / a9 denotes the column vector of derivatives of the density of the distribution (or probability function) with respect to the components of 9. The discussion by Cramer (1946, p. 495) is in terms of scalar observations, but it is clear that it holds true for vector observations. If (8) is the ellipsoid of concentration of t, then t is said to be efficient. In general, the ratio of the volume of (8) to that of the ellipsoid of concentration defines the efficiency of t. In the case of the multivariate normal distribution, if 9 = ~, then x is efficient. If 9 includes both I.t and l:, then x and S have efficiency [(N - O/N]PIJ.+ 1)/2. Under suitable regularity conditions, which are satisfied by the multivariate normal distribution,
(9)
tC'(aIOgf)(alo gf a9 a9
),= _tC'aa9 Iogf a9' . 2
This is the information matrix for one observation. The Cramer-Rao lower
86
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
bound is that for any unbiased t:stimator t the matrix ( 10)
Nc:C'(t-O)(t-O)
, r
,,2 IOg ]-1
-l-c:C' aoao'f
is positive semidefinite. (Other lower bounds can also be given,) Consistency
Definition 3.4.2. A sequence of vectors til = (II,,' ... , t mil)', n = 1,2, ... , is a consistent estimator of a = (0 1 , ••• , On,)' if plim., _ oclill = Oi' i = 1, ... , m.
By the law of large numbers each component of the sample mean x is a consistent estimator of that component of the vector of expected values I.l. if the observation vectors are hdependently and identically distributed with mean I.l., and hence x is a consistent estimator of I.l.. Normality is not involved. An element of the sample covariance matrix is (11) s,; = N
~I
t
(x;" - J.L;)(x j" - J.Lj) -
a=1
N~ 1 (Xi -
J.Li)(X j - J.Lj)
by Lemma 3.2.1 with b = I.l.. The probability limit of the second term is O. The probability limit of the first term is lJij if XI' x 2 " " are independently and identically distributed with mean I.l. and covariance matrix !.. Then S is a consistent estimator of I. Asymptotic Nonnality
First we prove a multivariate central limit theorem. Theorem 3.4.3. Let the m-component vectors YI , Y2 , ••• be independently and identically distributed with means c:C'Y" = l' and covariance matrices c:C'0-:. - v XY" - v)' = T. Then the limiting distribution of 0/ rnn:':~I(Ya - v) as 11 --> 00 is N(O, T). Proof Let
(12)
(M t , u) = c:C' exp [ iut'
1
c1 " E (Ya - v) , yn
,,~I
where II is a scalar and t an m-component vector. For fixed t, cP,,(t, u) can be considered as the characteristic function of (1/rn)L.:~I(t'Ya - c:C't'Ya ). By
I
3.4 THEORETICAL PROPERTIES OF ESTIMATORS OF THE MEAN VECfOR
87
the univariate central limit theorem [Cramer (1946), p. 215], the limiting distribution is N(O, t'Tt). Therefore (Theorem 2.6.4), (13) for every u and t. (For t = 0 a special and obvious argument is used.) Let u = 1 to obtain
(14) for every t. Since e- tt'Tt is continu0us at t = 0, the convergence is uniform in some neighborhood of t = O. The theorem follows. • Now we wish to show that the sample covariance matrix is asymptotically normally distributed as the sample size increases. Theorem 3.4.4. Let A(n)=L.~=I(Xa-XNXXa-XN)/, where X 1,X2 , ... are independently distributed according to N(fJ-, I) and n = N - 1. Then the limiting distribution of B(n) = 0/ vn)[A(n) - nI] is normal with mean 0 and covariances (15)
Proof As shown earlier, A(n) is distributed as A(n) = L.:_IZaZ~, where Zl' Z2' . .. are distributed independently according to N(O, I). We arrange the elements of ZaZ~ in a "ector such as
(16)
the moments of Y a can be deduced from the moments of Za as given in Section 2.6. We have ,cZiaZja=U'ij' ,cZiaZjaZkaZia=U'ijU'k/+U'ikU'j/+ U'i/U'jk, ,c(ZiaZja - U'ij)(ZkaZ/a - U'k/) = U'ikU'jt+ U'i/U'jk' Thus the vectors Y a defined by (16) satisfy the conditions of Theorem 3.4.3 with the elements of v being the elements of I arranged in vector form similar to (6)
88
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
and the elements of T being .given above. If the elements of A(n) are arranged in vector form similar to (16), say the vector Wen), then Wen) - nv = L.~-l(Ya - v). By Theorem 3.4.3, 0/ vn)[W(n) - nv 1has a limiting normal II distribution with mean 0 and the covariance matrix of Ya . The elements of B(n) will have a limiting normal distribution with mean .0 if XI' x 2 ' ••. are independently and identically distributed with finite fourthorder moments, but the covariance structure of B(n) will depend on the fourth-order moments. 3.4.2. Decision Theory It may he enlightening to consider estimation in terms of decision theory. We review SOme of the concepts. An observation X is made on a random variable X (which may be a vector) whose distribution P8 depends on a parameter 0 which is an element of a set 0. The statistician is to make a deci~ion d in a set D. A decision procedure is a function 8(x) whose domain is the set of values of X and whose range is D. The loss in making decision d when the distribution is P8 is a nonnegative function L(O, d). The evaluation of a procedure 8(x) is on the basis of the risk function
(17) For example, if d and 0 are univariate, the loss may be squared error, L(O, d) = (0 - d)2, and the risk is the mean squared error $8[8(X) A decision procedure 8(x) is as good as a procedure 8*(x) if
(18)
R(O,8):::;R(O,8*),
of. '<10;
8(x) is better than 8*(x) if (18) holds with a strict inequality for at least one value of O. A procedure 8*(x) is inadmissible if there exists another procedure 8(x) that is better than 8*(x). A procedure is admissible if it is not inadmissible (i.e., if there is no procedure better than it) in terms of the given loss function. A class of procedures is complete if for any procedure not in the class there is a better procedure in the class. The class is minimal complete if it does not contain a proper complete subclass. If a minimal complete class exists, it is identical to the class of admissible procedures. When such a class is available, there is no (mathematical) need to use a procedure outside the minimal complete class. Sometimes it is convenient to refer to an essentially complete class, which is a class of procedures 'iuch that for every procedure outside the class there is one in the class that is just as good.
1.4 THEORETICAL PROPERTIES OF ESTIMATORS OF THE MEAN VECTOR
89
For a given procedure the risk function is a function of the parameter. If the parameter can be assigned an a priori distribution, say, with density pte), th om the average loss from use of a decision procedure 8(x) is
(19) Given the a priori density p, the decision procedure 8(x) that minimizes r( p, 8) is the Bayes procedure, and the resulting minimum of r( p, 8) is the Bayes risk. Under general conditions Baycs procedures are admissible and admissible procedures are Bayes or limits of Bayes procedures. If the tknsity of X given e is f(xl e), the joint density of X and e is I(xl e)p(e) and the average risk of a procedure 8(x) is
(20)
r(p,8)= ffL[e,8(x)U(xle)p(e)dxde o x = f {f L[e,8(x)]g(elx)de}/(x)dx;
x
0
here
(21)
f(x) = ff(xle)p(e)dO, o
g
(el )=/(xle)p(e) x f(x)
are the marginal density of X and the a posteriori density of 11 given x. The procedure that minimizes r( p, 8) is one that for each x minimizes the expression in braces on the right-hand side of (10). that is. the expectation of L[ 0, 8(x)] with respect to the a posteriori distrib:ltion. If e and d are vectors (a and d) and LeO, d) = (a - d)'Q(O - d), where Q is positive definite. then
(22)
J,"olxL[O, d(x)]
=
cf01xlo -
cf(Olx») 'Q[O - J'( Olx»)
+ [
= J,"(Olx), the mean of the a posteriori distribu-
Theorem 3.4.5. [fx 1 , ••• , xN are independently distributed, each Xa according to N(p.., I), and if p.. has an a priori distribution N(v, «1», then the a posteriori distlibution of p.. given x I' ... , X N is normal with mean
(23) and covariance matrix
(24)
90
ESTl~ATlON OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
Proof Since i is sufficient for "", we need only consider i, which has the distribution of "" + v, where v has the distribution N[O, (1/ N)!,] and is independent of "". Then the joint distribution of "" and i is
(25)
The mean of the conditional distribution of "" given i is (by Theorem 2.5.1) ( 26 )
.
11
which reduces to (23).
( 1)-1 ( i -
+ «I> «I> + N!'
11 ) ,
•
Corollary 3.4.1. If XI' ••• ' Xl>' are independently distributed, each xa according to N( "", !,), "" has an a priori distribution N( 11, «1», and the loss function is (d - ",,)'Q(d - ",,), then the Bayes estimator of"" is (23). The Bayes estimator of "" is a kind of weighted average of i and 11, the prior mean of "". If (l/N)!, is small compared to «I> (e.g., if N is large), 11 is given little weight. Put another way, if «I> is large, that is, the prior is relatively uninformative, a large weight is put on i. In fact, as «I> tends to 00 in the sense that «1>-1 ..... 0, the estimator approaches i. A decision procedure oo(x) is minimax if (27)
supR(O,oo) = infsupR(O,o). 9
8
9
Theorem 3.4.6. If X I' ... , X N are independently distributed each according to N("",!,) and the loss function is (d - ",,)'Q(d - ""), then i is a minimax estimator.
Proof This follows from a theorem in statistical decision theory that if a procedure Do is extended Bayes [i.e., if for arbitrary e, r( p, Do) :5 r( p, op) + e for suitable p, where op is the corresponding Bayes procedure] and if R(O, Do) is constant, then Do is minimax. [See, e.g., Ferguson (1967), Theorem 3 of Section 2.11.] We find (28)
R("", i) = $(i - ",,)'Q(i - "") = G tr Q( i - "")( i - ",,)' 1
=NtrQ!'.
3.5
91
IMPROVED ESTIMATION OF THE MEAN
Let (23) be d(i). Its average risk is (29)
th",rth"...{trQ[d(i) -I.l.][d(i)
-l.l.l'li}
For more discussion of decision theory see Ferguson (1967). DeGroot (1970), or Berger (1980b).
3.5. IMPROVED ESTIMATION OF THE MEAN 3.5.1. Introduction
,
The sample mean i seems the natural estimator of the population mean I.l. based on a sample from N(I.l., I). It is the maximum likelihood estimator, a sufficient statistic when I is known, and the minimum variance unbiased estimator. Moreover, it is equivariant in the sense that if an arbitrary vector v is added to each observation vector and to I.l., the error of estimation (x + v) - (I.l. + v) = i - I.l. is independent of v; in other words, the error does not depend on the choice of origin. However, Stein (1956b) showed the startling fact that this conventional estimator is not admissible with respect to the loss function that is the sum of mean squared errors of the components Vlhen I = I and p ~ 3. James and Stein (1961) produced an estimator which has a smaller sum of mean squared errors; this estimator will be studied in Section 3.5.2. Subsequent studies have shown that the phenomenon is widespread and the implications imperative. 3.5.2. The James-Stein Estimator The loss function p
(1)
L(p."m) = (m -I.l.),(m -I.l.) =
I: (m; -
,11-;)2 =lIm -1.l.1I 2
;=1
is the sum of mean squared errors of the components of the estimator. We shall show [James and Stein (1961)] that the sample mean is inadmissible by
92
ESTIMATION OF THE MEAN VEcrOR AND THE COVARIANCE MATRIX
displaying an alternative estimator that has a smaller expected loss for every mean vector "". We assume that the normal distribution sampled has covariance matrix proportional to I with the constant of proportionality known. It will be convenient to take this constant to be such that Y= (1/N)L:~lXa =X has the distribution N("", I). Then the expected loss or risk of the estimator Y is simply tS'IIY - ",,11 2 = tr I = p. The estimator proposed by James and Stein is (essentially)
m(Y)=(l-
(2)
P-2 )(y-V)+v,
lIy -vII 2
where v is an arbitrary fixed vector and p;;::. 3. This estimator shrinks the observed y toward the specified v. The amount of shrinkage is negligible if y is very different from v and is considerable if y is close to v. In this sense v is a favored point. Theorem 3.5.1. With respect to the loss function (1), the risk of the estimator (2) is less than the risk of the estimator Y for p ;;::. 3. We shall show that the risk of Y minus the risk of (2) is positive by applying the following lemma due to Stein (1974). Lemma 3.5.1.
If f(x) is a function such that
(3)
f(b)-f(a)= tf'(x)dx a
for all a and b (a < b) and if
f
(4)
OO
1 '( )' dx
_00
{2;
then
(5)
f
OO
-00
f(x)(x - 0) _1_e-~(x-8)2 dx=
{2;
foo
f'(X) _1_e-~(x-8)2 dx.
_00
{2;
3.5
93
IMPROVED ESTIMATION OF THE MEAN
Proof of Lemma. We write lhe left-hand side of (5) as
(6)
f
OO
1
I
,
[f(x) -f(O)](x- O)_e-'(x-O) dx
&
8
+J9
[f(x) -f(O)](x- o)_l_e-t(x-O)' dx
&
_00
1 I . ' dydx = f OOfX f(y)(x- O)-e-'(x-O)
&
98
-
9
J
fef
-co x
-
e
J
_00
JY _cc
( y) ( x -
0) - I
&
e - ,I( x - 0 )' dy dx
f'(y)(x - 0) - 1e - 'I( x-o )' dxdy,
&
which yields the right-hand side of (5). Fubini's theorem justifies the inter• change of order of integration. (See Problem 3.22.) The lemma can also be derived by integration by parts in special cases. Proof of Theorem 3.5.1. The difference in risks is
94
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
Now we use Lemma 3.5.1 with (8) f(yj)
=
P
Yi-
Vi
I: (Yj -
,f'(y;)
= -p~---
I: (Yj -
IIj )2
}= I
IIj )2
}=1
[For p:;:: 3 the condition (4) is satisfied.] Then (7) is (9)
)-,c{2(
.tl.R( ,...
-
fl
P-
= (p - 2)
2)t[ i=1
2 ,cfl
1
IIY-vIl 2
1
IIY- vii
2
> O.
-
2(Yi- IIi)21 4
1I1'-v1l
-
(P_2)2} IIY-vIl 2
•
This theorem states that }. is inadmissible for estimating ,... when p:<: 3, since the estimator (2) has a smaller risk for every ,... (regardless of the choice of l'),
The risk is the sum of the mean squared errors ,c[m;CY) - ,u;F. Since YI •... , Yp are independent and only the distribution of Y; depends on ,ui' it is puzzling that the improved estimator uses all the Yj's to estimate ,ui; it seems that irrelevant information is being used. Stein explained the phenomenon by arguing that the sample distance squared of Y from v, that is, IIY - v1l 2 , overestimates the squared distance of ,... from l' and hence that the estimator Y could be improved by bringing it nearer v (whatever v is). Berger (1980a), following Brown, illustrated by Figure 3.4. The four points Xl' x 2 , X 3 , X 4 represent a spherical distribution centered at ,.... Consider the effects of shrinkage. The average distance of m(x l ) and m(x 3 ) from,... is a little greater than that of XI and x 3 , but m(x 2 ) and m(x 4 ) are a little closer to ,... than x 2 and X~ are if the shrinkage is a certain amount. If p = 3, there are two more points (not on the line 11,,...) that are shrunk closer to ,....
m(Z3)
Figure 3.4. Effect of shrinkage.
3.5
95
IMPROVED ESTIMATION OF THE MEAN
The risk of the estimator (2) is
( 10)
iif
2
1
2
tf."p.llm( f) - ""II = p - (p - 2) c:C'p. Ilf _ vll 2 '
where IIf - vll 2 has a noncentral X2-distribution with p degrees of freedom 2 and noncentrality parameter II"" - v1l . The farther"" is from v, the less the improvement due to the James-Stein estimator, but there is always some improvement. The density of IIf - vll 2 = V, say, is (28) of Section 3.3.3, where 'T2 = II"" - v1I2. Then
(11)
for p ~ 3. Note that for"" = v, that is, r2 = 0, (11) is 1/(p - 2) and the mean squared error (10) is 2. For large 1-' the reduction in risk is considerable. Table 3.2 gives values of the risk for p = 10 and (J' 2 = 1. For example, if r2 = II"" - vII2 is 5, the mean squared error of the James-Stein estimator is 8.86, compared to 10 for the natural estimator; this is the case if JLi - Vi = 1/ Ii = 0.707, i = 1, ... ,10, for instance.
Table 3.2t. Average Mean Squared Error of the James-Stein Estimator for p = 10 and (J"2 = 1
,.2 = II .... - vl1 2
Gp.llm(Y) _ .... 11 2
0.0 0.5
2.00 4.78 6.21 7.S\ 8.24 8.62 ,8.86 9.03
1.0
2.0 3.0 4.0 5.0 6.0 t
From Efron and Morris (1977).
96
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
An obvious question in using an estimator of this class is how to choose the vector v toward which the observed mean vector is shrunk; any v yields an estimator better than the natural one. Ho~ever, as seen from Table 3.2, the improvement is small if II"" - vII is very large. Thus, to be effective some knowledge of the position of "" is necessary. A disadvantage of the procedure is that it is not objective; the choice of v is up to the investigator. A feature of the estimator we have been studying that seems disadvantageous is that for small values of IIY - vII, the mUltiplier of Y - v is negative; that is, the estimator m(Y) is in the direction from v opposite to that of Y. This disadvantage can be overcome and the estimator improved by replacing the factor by when the factor is negative.
°
Definition 3.5.1.
For any function g(u), let
g(u):<:o,
(12)
g(u) < 0.
=0, Lemma 3.5.2.
When X is distributed according to N("", 1),
Proof The right-hand side of (13) minus the left-hand side is
(14) plus 2 times
(15)
cS'11",,'X[g+ (IIXII) - g(IIXII)]
= 1I""llf~oo ... f~J'l [g+ (llyll) - g(lIyll)] ._1_ 1 ex p { -
(21T)2P
t[ t y? - 2ydl",,11 i~l
+ 1I""1I2]} dy,
where y' = x' P, (11",,11,0, ... ,0) = ",,' P, and PP' = T. [The first column of P is 0/11""11)",,.] Then (15) is 11",,11 times
(16)
e-tIlILII2f~oo'" f~ooIaooYl[g+(IIYII) -g(lIyll)][e IlILIIY , -e- IlILIIY ,] ' - -1- , e-''l:P.• ,y,2 dy dy '" dy :<:
(21T)2P
(by replacing Yl by
1
-Yl for Yl < 0).
2
p
•
°
3.5
97
IMPROVED ESTIMATION OF THE MEAN
Theorem 3.5.2.
The estimator
(17) has smaller risk thay. m(y) defined by (2) and is minimax. Proof In Lemma 3.5.2, let g(u) = 1 - (p - 2)/u 2 and X = Y - v, and replace J.l. by J.l. - v. The second assertion in the theorem follows from Theorem 3.4.6. •
The theorem shows that m(Y) is not admissible. However, it is known that m+(Y) is also not admissible, but it is believed that not much further improvement is possible. This approach is easily extended to the case where one observes x I' ... , Xv from N(J.l.,'1) with loss function L(J.l., m) = (m - J.l.)''1 -l(m - J.l.). Let "I = CC' for some nonsingular C, Xa = Cx~, a = 1, ... , N, J.l. = CJ.l.*, and L* (m* , J.l.*) = II m* - J.l.* 112. Then xi, ... , xt are observations from N( J.l.* • J), and the problem is reduced to the earlier one. Then
(18)
p-2 ( 1- N(x-v)''1 l(X-V)
)+ _
(x-v)+v
is a minimax estimator of J.l.. 3.5.3. Estimation for a General Known Covariance Matrix and un Arbitrary Quadratic Loss Function
Let the parent distribution be N(J.l.,'1), where "I is known, and let the loss function be ( 19)
L(J.l., m)
= (m - J.l.),Q(m - J.l.),
where Q is an arbitrary positive definite matrix which reflects the relative importance of errors in different directions. (If the loss function were singular, the dimensionality of x could be reduced so as to make the loss matrix nonsingular.) Then the sample mean x has the distribution N(J.l.,(1/N)'1) and risk (expected loss) (20)
SCi - J.l.) 'Q( x - J.l.) = tS' tr Q( x - J.l.)( x - J.l.)' = iJtr Q'1,
whieh is constant, not depending on J.l..
98
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
Several estimators that improve on i have been proposed. First we take lip an estimator proposed independently bv Berger (975) and Hudson (1974). TheOl'em 3.5.J. Lct r( z). () ~ z < oc, he l/ lIondccreasing differentiable Junc(ion such (ha( 0::5: r(z) ::5: 2(p - 2). Then for p ~ 3
has smaller risk (han i and is minimax. Proof There exists a matrix C such that C'QC=[ and O/N)"I= CI1C' where A is diagonal with diagonal elements 0 1 ~ O2 ~ ... ~ Op > 0 (Theorem A.2.2 of the Appendix). Let i = Cy + v and I.l. = CI.l.* + v. Then y has the distribution N(I1-*, 11), and the transformed loss function is
L* (m*, 11-*) = (m* - 11-*)'( m* - 11-*) = Ilm* - 11-* 112.
(22)
The estimator (21) llf 11- is transformeJ tll the estimator of 11-*
=
C- 1(11- - v),
(23) We now proceed as in the proof of Theorem 3.5.1. The difference in risks bet\veen y and m* is (24)
Since r(z) is differentiable, we use Lemma 3.5.1 with (x - (J) = (Yi - JLj)Oi and (25)
(26)
.
!(Yi) f'() Y,
=
l'(y'A- 2y) 'A 2 Yi'
Y
Y
2
= r(y'A- y)
y' A
2Y
2r'(y'A-2y)
+
Y 'A
2Y
Y? 0/ -
2r(y'I1-2y ) Y? (y' 11- 2Y ) 2 0/'
3.5
99
IMPROVED ESTIMATION OF THE MEAN
Then
(27) 6.R( *)=G.{2( I.l.
,.
P
Corollary 3.5.1.
2 _2)r(Y'6,-2y) +4r'(Y'6,-2y)- r (Y'6,-2y)} >0 Y'6,-2y Y'6,- 2 y -
For P ~ 3
(28) {/ -
min[p-2,N2(x-v)'I-1Q-t:~:-I(x-v)1 N(x _ v)'I-1Q-1I-1(x _ v) Q
has smaller risk than
-I I -I}-
(x - v) + v
x and is minimax.
Proof the function r(z) = min(p - 2, z) is differentiable except at z = P - 2. The function r(z) can be approximated arbitrarily closely by a differentiable function. (For example, the corner at z = p - 2 can he smoothed by a circular arc of arbitrary small radius.) We shall not give the details of the proof. • In canonical form y is shrunk by a scalar times a diagonal matrix. The larger the variance of a component is, the less the effect of the shrinkage. Berger (1975) has proved these results for a more general density, that is, for a mixture of nOLnals. Berger (1976) has also proved in the case of normality that if
(29)
r( z) =
1'"
z U W - C + 1 e- !uz du --"0_ _ _ _ __
fa'" u w-c e- !uz du
for 3 - iP :5 c < 1 + iP, where a is the smallest characteristic root of IQ, then the estimator m given by (21) is minimax, is admissible if c < 2, and is proper Bayes if c < 1. Another approach to minimax estimators has been introduced by Bhattacharya (1966). Let C be such that C- 1 (1/N)I(C- 1 ), =/ and C'QC=Q*, which is diagonal with diagonal elements qf ~ qr ~ ... ~ > O. Then y =
q;
100
c- 1i
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
has the distribution N(IJ.*, n, and the loss function is p
L*(m*,IJ.*) = L,qt(m;-1l-i)2
(30)
;=1 p
p
E E aj(ml -
=
1l-i)2
;=1 j=;
j
p
= L,aj L,(mj-Il-f)2 j=1
;=1
p
= L, a,lIm*{}) - IJ.* (f)1I 2 , j=1
where a j = qt - qj+ l ' j = 1, ... , p - 1, ap = q;, m*{}) = (mj, ... , mj)', and IJ.*(J) = (Il-i, ... , Il-j)', j = 1, ... , p. This decomposition of the loss function suggests combining minimax estimators of the vectors 1J.*(j), j = 1, ... , p. Let y{}) = (Y 1, ••• , Y/ . Theorem 3.5.4. If h{})(y(J» = [h\j)(y(J», ... , h~j)(y(j»l' is a minimax estimator Of 1J.*(j) under the loss function IIm*(J) - IJ.*U)1I 2 , j = 1, ... , p, then
~tajh\j)(y(j)),
(31)
,
i=l, ... ,p,
j=1
is a minimax estimator of Il-i,···, Il-;. Proof First consider the randomized estimator defined by j = i, ... ,p,
(32) for the ith comporient. Then the risk of this estimator is p
(33)
P
p
L, q'! ,cp.' [G;( Y) - 1l-;]2 = L, q'! L, ;=1
;=1
j=;
P
j
a.
-:f ,cp.' [h\il( y
= L, a j L, ,cp.' [ h\j)( y(j» - Il-; j=1
;=1
Il-I] 2
r
P
= L, a j ,cp..IWj)( yU» - 1J.*(i)1I 2 j=1
P S L, ajj j-1
=
P
= L, qj j=1
,cp..L*(Y,IJ.*)*,
and hence the estimator defined by (32) is minimax.
l.6
101
ELLIPTICALLY CONTOURED DISTRIBUTIONS
Since the expected value of Gi(Y) with respect to (32) is (31) and the loss function is convex, the risk of the estimator (31) is less than that of the • randomized estimator (by Jensen's inequality).
3.6. ELLIPTICALLY CONTOURED DISTRIBUTIONS 3.6.1. Observations Elliptically Contoured Let x I' ... , XN be N ( = n + 1) independent observations on a random vector ~g[(xa - 1:)' A -I (x" - V)]. The density of the sample is
X with density I AI-
N
IAI-~N f[g[(x-v)'A-I(x-v»).
(1)
a=1
The slmple mean i and covariance matrix S = (1/n)[L.~~I(Xa - ,...Xx" - ,...)' -- N(i - ,...)(i -,...),] are unbiased estimators of the mean ,... = v and the covariance matrix!' = [J:'R" /p]A, where R" = (x - v)' A -I(X - v). Theorem 3.6.1. N from
The covariances of the mean and covariance of (l sample of ~g[(x - v)' A -I(X - v)] with J:'R 4 < 00 are
I AI-
"( X-,... ) ( X-,... )' =1il"" 1~
(2)
Cn
i,j=1.. .. ,p,
1
+ Ii ( (J",k Ojl + (J"il Ujk ), Lemma 3.6.1.
i, j, k,l = 1.... , p.
The second-order moments of the elements of S are
(5) i,j,k,l, = l, .... p. Proof of Lemma 3.6.1. We have N
(6)
C
I:
(xia - ILi)(X ja - ILj)(X k{3
-
ILk)(SlfJ - ILl)
a, {3=!
= N C( Xia -
ILJ( Xj" - ILJ( Xka - ILd( Xla - ILl)
+ N(N -1) J:'(Xia =N(1
4-
K)( (J"ij(J"kl
ILJeXja - ILj) Jf(XkfJ - ILd(x lfJ - ILl)
+ (J"ik(J"jl + (J"iIOjk) + N(N - 1) (J"ij(J"kl'
102
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
1
N
(8)
cf
L
(X ia -/-Li)(X ju -/-Lj)7il
a= I
N
L
(X k{3-/-Lk)(Xly-/-Ly)
{3.y=1
It will be convenient to use more matrix algebra. Define vee B, B ® C (the Kronecker product), and Km " (the commutator matrix) by
(10)
( 11)
Kill" vee
n = vee B'.
See. c.g., Magnus and Neudecker (1979) or Section A.S of the Appendix. We can rewri te (4) as ( 12)
6? (vee S) = cC' (vee S - vee I) ( vee S - vee I)'
= n Kn~N ([p' + Kpp)( I ® l;) + ;'vec l(vec I)'. Theorem 3.6.2
(13)
m[n vee(xS -vee - J.l.) ] I ~
!!, N [ ( ) , (
! (+ K
1) ( [p' + K p D ) (I
~ I) +
K
vee I (vee I)' ) ] .
103
3.6 ELLIPTICALLY CONTOURED DISTRIBUTIONS
This theorem follows from the central limit theorem for independent identically distributed random vectors (with finite fourth moments). The theorem forms the basis for large-sample inference. 3.6.2. Estimation of tb e Kurtosis Parameter To apply the large-sample distribution theory derived for normal distributions to problems of inference for elliptically contoured distributions it is necessary to know or estimate the kurtosis parameter K. Note that
Since i.£. J.l. and S.£. 'I,
(15)
1
N
2 P
N E [(Xa- i )'S-l(Xa- i )] ->p(p+2)(1+K). a=l
A consistent estimator of
K
is
(16)
Mardia (1970) proposed using M to form a consistent estimator of
K.
3.6.3. Maximum Likelihood Estimation We have considered using S as an estimator of 'I = ($ R2 /p) A. When the parent distribution is normal, S is the sufficient statistic invariant with respect to translations and hence is the efficient unbiased estimator. Now we study other estimators. We consider first the maximum likelihood estimators of J.l. and A when the form of the density gO is known. The logarithm of the likelihood function is
(17)
N
logL= -zloglAI +
N
E logg[(x a ':"J.l.)'A- 1 (xa -J.l.)]. ,,~1
104
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
The derivatives of log L with respect to the components of J.l are
(18) Setting the vector of derivatives equal to 0 leads to the equation
(19)
N g'[(Xa-ft.),A.-1(Xu-ft.)] g[(xa-ft.)'A.-1(xu-ft.)]
a~l
A
N g'[(X,,-ft.)'A.-1(xa-ft.)] g[(Xu-ft.)'A-'(xa-ft.)T·
Xa=Jl.a~1
Setting equal to 0 the derivatives of log L with respect to the elements of A -I gives ( 20)
N
A)'A.-I(
,[(
A)]
A.=-~" g Xu-J.l Xu-Jl. (x _A)(X _A)'. N a~l L." [( _A)'AA_I( Xa _A)] u J.l a J.l g Xu J.l J.l
The estimator
A. is a kind of weighted average of the rank 1 matrices
(xu - ft.Xxa - ft.)'. In the normal case the weights are 1/N. In most cases (19) and (20) cannot be solved explicitly, but the solution may be approximated by iterative methods. The covariance matrix of the limiting normal distribution of /N(vec A.vec A) is
where
pep + 2)
(22) O"lg
= 4C g'(R:J. 2]2' R g(R2)
r
l
(23)
20"1g(1 - O"lg) 0"2g= 2+p(1-0"Ig)'
See Tyler (1982). 3.6.4. Elliptically Contoured Matrix Distributions Let
(24)
3.6
105
ELLIPTICALLY CONTOURED DISTRIBUTIONS
be an Nxp random matrix with density g(Y'Y)=g(L~~1 YaY:). Note that the density g(Y'Y) is invariant with respect to orthogonal transformations Y*= ON Y. Such densities are known as left spherical matrix densities. An example is the density of N observations from N(O,Ip )'
(25) In this example Y is also right spherical: YOp i!. Y. When Y is both left spherical and right spherical, it is known as ~pherical. Further, if Y has the density (25), vec Y is spherical; in general if Y has a density, the density is of tr..e form
(26)
geL Y'Y)
=
g
(a~l ;~ Y?a )
=
g(tr YY')
= g[(vec Y)'vec Y] = g [(vec Y')' vec Y']. We call this model vector-sphen·cal. Define
(27) "here C' A -I C = Ip and E'N = (1, ... , 1). Since (27) is equivalent to Y = (X- ENJ.l,)(C')-1 and (C')-IC- I = A-I, the matrix X has the density
(28)
IAI-N/2g[tr(X-ENJ.l')A-I(X-ENJ.l')']
= IAI-N/2g[a~1 (x" - J.l)' A -I(X" - J.l+ From (26) we deduce that vec Y has the representation vec Y i!. R vec U,
(29) where w = R2 has the density
~Np
(30)
r(~P/2) w lNp -
1
g(w),
vec U has the uniform distribution on L~~ 1Lf~ 1u~a = 1, and Rand vec U are independent. The covariance matrix of vee Y is
(31)
106
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
Since vec FGH = (H' ® F)vec G for any conformable matrices F, G, and H, we can write (27) as (32) Thus (33) (34) (35) (36)
GR 2 C(vecX) = (C®IN)C(vecY)(C'cHN ) = Np A®IN ,
G(row of X) = J.l', GR 2 C(rowof X') = Np A.
The rows of X are uncorrelated (though not necessarily independent). From (32) we obtain (37)
vec X g, R ( C ® IN) vec U + J.l ® EN'
X g, RUC' + ENJ.l'.
(38)
Since X - ENJ.l' = (X - ENX') + EN(i' - J.l)' and E'N(X - ENX') = 0, we can write the density of X as
where x = (1/ N)X'E N' This shows that a sufficient set of statistics for J.l and A is x and nS = (X - ENX')'(X - ENX'), as for the normal distribution. The maximum likelihood estimators can be derived from the following theorem, which will be used later for other models. Theorem 3.6.3.
Suppose the m-component vector Z has the density
I~I- ~h[(z - v )'~-l(Z - v)], where w~mh(w) has a finite positive maximum at wlr and ~ is a positive definite matrix. Let be a set in the space of (v, ~) such that if (v,~) E then (v, c~) E for all c > O. Suppose that on the basis of an observation z when h( w) = const e - iw (i.e., Z has a nonnnl
n
n
n
distlibution) the maximum likelihood estimator (v, (i» E n exists and is unique with iii positive definite with probability 1. Then the maximum likelihood estimator of ( v, ~) for arbitrary h(·) is (40)
v=v,
3.6
107
ELLIPTICALLY CONTOURED DISTRIBUTIONS
and t~ maximum of the likelihood is I~I- 4h(w,) [Anderson, Fang, and Hsu (1986)]. Proof Let 'IT = I({>I
-1/ m ({>
and
(41) Then (v,
({> ) E n and I'IT I = 1. The likelihood is
(42) Under normality h(d) = (21T)- i m e- td, and the maximum of (42) is attained at v = V, 'IT = 1ii = I(lil-I/ m(li, and d = m. For arbitrary hO the maximum of (42) is attained at v = v, Ii = Ii, and d = Who Then the maximum likelihood estimator of ({> is
( 43) Then (40) follows from (43) by use of (41).
•
Theorem 3.6.4. Let X (N xp) have the density (28), where wiNPg(w) has a finite positive maximum at wg • Then the maximum likelihood estimators of J.l and A are ( 44)
A=
ft =i,
Np A wg '
Corollary 3.6.1. Let X (N X p) have the density (28). Then the maximum likelihood estimators of v, (All'"'' App), and Pij' i,j=I, ... ,p, are .~, (p /wgXa lJ , ••• , a pp ), and aiJ JOjjOjj, i, j = 1, ... , p.
Proof Corollary 3.6.1 follows from Theorem 3.6.3 and Corollary 3.2.1. Theorem 3.6.5. that
(45)
Let j(X) be
0
•
vector-valued function of X (NXp) such
108
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
for all v and . j(cX) =j(X)
(46)
for all c. Then the distribution of j(X) where X h:zs an arbitrary density (28) is the same as its distribution where X has the normal density (28). Proof Substitution of the representation (27) into j(X) gives ( 47)
j(X) =j(YC'
+ ENf1.') =j(YC')
by (45). Let j(X) = h(vec X). Then by (46), h(cX) = h(X) and
(48)
j(YC') =h[(C®IN)vecYj =h[R(C®IN)vecUj = h[(C®IN ) vec uj.
•
Any statistic satisfying (45) and (46) has the same distribution for all gO. Hence, if its distribution is known for the normal case, the distribution is valid for all elliptically contoured distributions. Any function of the sufficient set of statistics that is translation-invariant, that is, that satisfies (45), is a function of S. Thus inference concerning I can be based on S.
Corollary 3.6.2. Let j(X) be a vector-valued function of X (N X p) such that (46) holds for all c. Then the distribution of j(X) where X has arbitrary density (28) with f1. = 0 is the same as its distribution where X has normal density (28) with f1. = o. Fang and Zhang (1990) give this corollary as Theorem 2.5.8.
PROBLEMS 3.1. (Sec. 3.2) Find ji, Frets (1921). 3.2. (Sec. 3.2)
i,
and (Pi}) for the data given in Table 3.3, taken from
Verify the numerical results of (21).
3.3. (Sec. 3.2) Compute ji, i, S, and P for the following pairs of observations: (34,55), (12, 29), (33, 75), (44, 89), (89, 62), (59, 69), (50, 41), (88, 67). Plot the observations. 3.4. (Sec. 3.2) Use the facts that I C* I = n A;, tr C* = ~Ai' and C* = I if Al = ... = Ap = 1, where AI' ... ' Ap are the characteristic roots of C*, to prove Lemma 3.2.2. [Hint: Use f as given in (12).)
109
PROBLEMS Table 3.3 t . Head Lengths and Breadths of Brothers Head Length, First Son,
Head Breadth, First Son,
Head Length, Second Son,
Head Breadth, Second Son,
XI
X2
X3
X4
191 195 181 183 176
155 149 148 153 144
179 201 185 188 171
145 152 149 149 142
208 189 197 188 192
157 150 159 152 150
192 190 189 197 187
152 149 152 159 151
179 183 174 190 188
158 147 150 159 151
186 174 185 195 187
148 147 152 157 158
163 195 186 181 175
13" 153 145 140
161 183 173 182 165
130 158 148 146 137
192 174 176 197 190
154 143 139 167 163
185 178 176 200 187
152 147 143 158 150
ISS
tThese data, used in examples in the first edition of this book, came from Rao (1952), p. 2,45. Izenman (1980) has indicated some entries were apparemly incorrectly copied from Frets (1921) and corrected them (p. 579).
3.5. (Sec. 3.2) Let Xl be the body weight (in kilograms) of a cat and weight (in grams). [Data from Fisher (1947b).]
X2
(a) In a sample of 47 female cats the relevant data are
110.9)
~xa = ( 432.5 ' Find jl,
t,
S, and
p.
1029.62 ) 4064.71 .
the heart
110
ESTIMATION OF THE MEAN VECTOR AND THE COVARIANCE MATRIX
Table 3.4. Four Measurements on Three Species of Iris (in centimeters) Iris setosa
Iris versicolor
Sepal length
Sepal width
Petal length
Petal width
Sepal length
Sepal width
5.1 4.9 4.7 4.6 5.0
3.5 3.0 3.2 3.1 3.6
1.4 1.4 1.3 1.5 1.4
0.2 0.2 0.2 0.2 0.2
7.0 6.4 6.9 5.5 6.5
3.2 3.2 3.1 2.3 2.8
4.7 4.5 4.9 4.0 4.6
5.4 4.6 5.0 4.4 4.9
3.9 3.4 3.4 2.9 3.1
1.7
0.4 0.3 0.2 0.2 0.1
5.7 6.3 4.9 6.6 5.2
2.8 3.3 . 2.4 2.9 2.7
5.4 4.8
3.7 3.4
5.0 5.9 6.0 6.1 5.6
1.4 1.5 1.4 1.5
Iris virginica
Petal Petal length width
Petal Petal length width
Sepal length
Sepal width
1.4 1.5 1.5 1.3 1.5
6.3 5.8 7.1 6.3 6.5
3.3 2.7 3.0 2.9 3.0
6.0 5.1 5.9 5.6 5.8
2.5 1.9 2.1 1.8 2.2
4.5 4.7 3.3 4.6 3.9
1.3 1.6 1.0 1.3 1.4
7.6 4.9 7.3 6.7 7.2
3.0 2.5 2.9 2.5 3.6
6.6 4.5 6.3 5.8 6.1
2.1 1.7 1.8 1.8 2.5
2.0 3.0 2.2 2.9 2.9
3.5 4.2 4.0 4.7 3.6
1.0 1.5 1.0 1.4 1.3
6.5 6.4 6.8 5.7 5.8
3.2 2.7 3.0 2.5 2.8
5.1 5.3 5.5 5.0 5.1
2.0 1.9 2.1 2.0 2.4
3.2 3.0 3.8 2.6 2.2
5.3 5.5 6.7 6.9 5.0
2.3 1.8 2.2 2.3 1.5
4.8
3.0
1.5 1.6 1.4
4.3 5.8
3.0 4.0
1.2
0.2 0.2 0.1 0.1 0.2
5.7 5.4 5.1 5.7 5.1
4.4 3.9 3.5 3.8 J.8
1.5 1.3 1.4 1.7 1.5
0.4 0.4 0.3 0.3 0.3
6.7 5.6 5.8 6.2 5.6
3.1 3.0 2.7 2.2 2.5
4.4 4.5 4.1 4.5 3.9
1.4 1.5 1.0 1.5 1.1
6.4 6.5 7.7 7.7 6.0
5..\ 5.1 4.6 5.1 4.8
3.4 3.7 3.6 3.3 3.4
1.7
1.5 1.0 1.7 1.9
0.2 0.4 0.2 0.5 0.2
5.9 6.1 6.3 6.1 6.4
3.2 2.8 2.5 2.8 2.9
4.8 4.0 4.9 4.7 4.3
1.8 1.3 1.5 1.2 1.3
6.9 5.6 7.7 6.3 6.7
3.2 2.8 2.8 2.7 3.3
5.7 4.9 6.7 4.9 5.7
2.3 2.0 2.0 1.8 2.1
5.0 5.0 5.2 5.2
3.0 3.4 3.5 3.4
4.7
3.2
1.6 1.6 1.5 1.4 1.6
0.2 0.4 0.2 0.2 0.2
6.6 6.8 6.7 6.0 5.7
3.0 2.8 3.0 2.9 2.6
4.4 4.8 5.0 4.5 3.5
1.4 1.4 1.7 1.5 1.0
7.2 6.2 6.1 6.4 7.2
3.2 2.8 3.0 2.8 3.0
6.0 4.8 4.9 5.6 5.8
1.8 1.8 1.8 2.1 1.6
4.8 5,4 5.2 5.5 4.9
3.1 3.4 4.1 4.2 3.1
1.6 1.5 1.5 1.4 1.5
0.2 0.4 0.1 0.2 0.2
5.5 5.5 5.8 6.0 5.4
2.4 2.4 2.7 2.7 3.0
3.8 3.7 3.9 5.1 4.5
1.1
1.0 1.2 1.6 1.5
7.4 7.9 6.4 6.3 6.1
2.8 3.8 2.8 2.8 2.6
6.1 6.4 5.6 5.1 5.6
1.9 2.0 2.2 1.5 1.4
5.0 5.5 4.9 4.4 5.1
3.2 3.5 3.6 3.0 3,4
1.2 I.3 1.4 1.3 1.5
0.2 0.2 0.1 0.2 0.2
6.0 6.7 6.3 5.6 5.5
3.4 3.1 2.3 3.0 2.5
4.5 4.7 4.4 4.1 4.0
1.6 1.5 1.3 1.3 1.3
7.7 6.3 6.4 6.0 6.9
3.0 3.4 3.1 3.0 3.1
6.1 5.6 5.5 4.8 5.4
2.3 2.4 1.8 1.8 2.1
1.1
III
PROBLEMS
Table 3.4. (Continued) Iris setosa Sepal Sepal length width 5.0 4.5 4.4 5.0 5.1
3.5 2.3 3.2 3.5 3.8
4.8 5.1 4.6 5.3 5.0
3.0 3.8 3.2 3.7 3.3
Iris versicolor
Petal length
Petal width
1.3
1.9
0.3 0.3 0.2 0.6 0.4
5.5 6.1 5.8 5.0 5.6
1.4 1.6 1.4 1.5 1.4
0.3 0.2 0.2 0.2 0.2
5.7 5.7 6.2 5.1 5.7
1.3 1.3 1.6
Iris uirginica
Petal length
Petal width
2.6 3.0 2.6 2.3 2.7
4.4 4.6 4.0 3.3 4.2
1.2 1.4 1.2 1.0 1.3
6.7 6.9 5.8 6.8 6.7
3.0 2.9 2.9 2.5 2.8
4.2 4.2 4.3 3.0 4.1
1.2 1.3 1.3 1.1 1.3
6.7 6.3 6.5 6.2 5.9
Sepal Sepal length width
Sepal Sepal length width
Petal length
Petal width
3.1 3.1 2.7 3.2 3.3
5.6 5.1 5.1 5.9
2.4 2.3 1.9 2.3
5.7
2.5
3.0 2.5 3.0 3.4 3.0
5.2 5.J 5.2 5.4 5.1
2.1 1.9 2.0 2.3 1.8
(b) In a sample of 97 male cats the relevant data are 281.3)
"<'
Find
,_ (
~xaxa -
LXa = ( 1098.3 '
836.75 3275.55
3275.55 ) 13056.17 .
fL, l;, S, and p.
3.6. Find fL, 1:, and ( P;j) for Iris setosa from Table 3.4, taken from Edgar Anderson's famous iris data [Fisher (1936)]. 3.7. (Sec. 3.2) In variance of the sample correlation coefficient. Prove that r 12 is an invariant characteristic of the sufficient statistics i and S of a bivariate sample under location and scale transformations (Xfa = bix;a + Ci' b i > 0, i = 1,2, a = 1, ... , N) and that every function of i and S that is invariant is a function of r 12 • [Hint: See Theorem 2.3.2.] 3.B. (Sec. 3.2)
Prove Lemma 3.2.2 by induction. [Hint: Let HI _ (H;_I
H;-
h'
=
h ll ,
i=2, ... ,p,
(1)
and use Problem 2.36.] 3.9. (Sec. 7.2) Show that
(Note: When p observations.)
=
1, the left-hand side is the average squared differences of the
112
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
3.10. (Sec. 3.2) Estimation of l: when jl. is known. Show that if XI"'" XN constitute a sample from N(jl., l:) and jl. is known, then (l/N)r.~_I(Xa - jl.XX a - jl.)' is the maximum likelihood estimator of l:. Estimation of parameters of a complex normal distribution. Let be N obseIVations from the complex normal distributions with mean 6 and covariance matrix P. (See Problem 2.64.)
3.11. (Sec. 3.2) ZI"'" ZN
(a) Show that the maximum likelihood estimators of 0 and Pare 1
A
N
N
o =z= N L a=1
za'
P= ~ L
(z.-z)(za-z)*,
a=l
(b) Show that Z has the complex normal distribution with mean 6 and covariance matrix (l/N)P. (c) Show that and P are independently distributed and that NP has the distribution of r.:_ 1Wa Wa*' where WI"'" w" are independently distributed, each according to the complex normal distribution with mean 0 and covariance matrix P, and n = N - 1.
z
3.12. (Sec. 3.2) Prove Lemma 3.2.2 by using Lemma 3.2.3 and showing N log I CI tr CD has a maximum at C = ND - I by setting the derivatives of this function with respect to the elements of C = l: -I equal to O. Show that the function of C tends to -00 as C tends to a singular matrix or as one or more elements of C tend to 00 and/or - 00 (nondiagonal elements); for this latter, the equivalent of (13) can be used. 3.13. (Sec. 3.3) Let Xa be distributed according to N( 'Yca, l:), a = 1, ... , N, where r.c~ > O. Show that the distribution of g = (l/r.c~)r.caXa is N[ 'Y,(l/r.c~)l:]. Show that E = r.a(Xa - gcaXXa - gc a )' is independently distributed as r.~.::l ZaZ~, where ZI'"'' ZN are independent, each with distribution N(O, l:). [Hint: Let Za = r.bo/lX/l' where bN/l = C/l/,;r;'I and B is orthogonal.]
3.14. (Sec. 3.3) Prove that the power of the test in (J 9) is a function only of p and [N I N 2 /(N I + N 2)](jl.(11 - jl.(21),l: -1(jl.(11 - jl.(21), given VI. 3.15. G'ec.. 3.3)
Efficiency of the mean. Prove that i is efficient for estimating jl..
3.16. {Sec. 3.3) Prove that i and S have efficiency [(N -l)/N]p{p+I)/2 for estimating jl. and l:.
3.17. (Sec. 3.2) Prove that Pr{IAI = O} = 0 for A defined by (4) when N > p. [Hint: Argue that if Z; = (ZI"'" Zp), then Iz~1 "" 0 implies A = Z;Z;' + r.;::p\ 1 ZaZ~ is positive definite. Prove Pr{l ztl = Zjjl zt-II + r.t::11Zij cof(Zij) = O} = 0 by induction, j = 2, ... , p.]
113
PROBLEMS
3.18. (Sec. 3.4) Prove l_~(~+l:)-l =l:(41+.l:)-1,
~_~(~+l:)-l~=(~-l +l:-l)-l.
3.19. (Sec. 3.4) Prove (l/N)L~_l(Xa - .... )(X a - .... )' is an unbiased estimator of I when .... is known. 3.20. (Sec. 3.4) Show that
3.21. (Sec. 3.5) Demonstrate Lemma 3.5.1 using integration by parts. 3.22. (Sec. 3.5) Show that
f OOfOOI f'(y)(x II
y
f 6 f'" _00
_:xl
I
'I
1 ,,(,-Il)' . IJ) --edxdy
&
=
fX 1f'(y)I--e-,(r1 , ' dy, & Il )'
t)
'I
f'(y)(IJ-x)--e-,(I-H) I,. d\'dy=
&
f"
1f'(y)I--e-;l,-/l1 I , . , dy,
-x
&
3.23. Let Z(k) = (Zij(k», where i = 1, ... , p. j = 1. ... , q and k = 1. 2..... be a sequence of random matrices. Let one norm of a matrix A be N1(A) = max i . j mod(a), and another he N 2(A) = L',j a~ = tr AA'. Some alternative ways of defining stochastic convergence of Z(k) to B (p x q) are (a) N1(Z(k) - B) converges stochastically to O. (b) N 2 (Z(k) - B) converges stochastically to 0, and (c) Zij(k) -=- bij converges stochastically to 0, i = 1, ... , p, j = 1,.,., q. Prove that these three definitions are equivalent. Note that the definition of X(k) converging stochastically to 11 is that for every arhitrary positive I'i and e, we can find K large enough so that for k> K Pr{IX(k)-al <8}>I-s.
3.24. (Sec. 3.2) Covariance matrices with linear structure [Anderson (1969)]. Let q
( i)
l: =
L g
{1
(T,G,.
114
ESTIMATION OF THE MEAN VECfOR AND THE COVARIANCE MATRIX
where GO"'" Gq are given symmetric matrices such that there exists at least one (q + 1)-tuplet uo, u I ,. .. , uq such that (j) is positive definite. Show that the likelihood equations based on N observations are
g=O,l, ... ,q.
(ii) Show that an iterative (scoring) method can be based on (1'1"1')
~ tr i--IGi--1G A(i)_.!.. tr ~-IG~-IA .... i-l g .... i-I hUh - N ~i-I g~i-l ,
I.....
11-0
g=O,l, ... ,q,
CHAPTER 4
The Distributions and Uses of Sample Correlation Coefficients
4.1. INTRODUCTION In Chapter 2; in which the multivariate normal distribution was introduced, it was shown that a measure of dependence between two normal variates is the correlation coefficient Pij = (Fiji (Fa Ujj' In a conditional distribution of Xl>"" Xq given X q+ 1 =x q+ 1 , ••• , Xp =xp' the partial correlation Pij'q+l •...• P measures the dependence between Xi and Xj' The third kind of correlation discussed was the mUltiple correlation which measures the relationship between one variate and a set of others. In this chapter we treat the sample equivalents of these quantities; they are point estimates' of the population qtiimtities. The distributions of the sample correlalions are found. Tests of hypotheses and confidence intervais are developed. In the cases of joint normal distributions these correlation coefficients are the natural measures of dependence. In the population they are the only parameters except for location (means) and scale (standard deviations) pa· rameters. In the sample the correlation coefficients are derived as the reasonable estimates of th ~ population correlations. Since the sample means and standard deviations are location and scale estimates, the saniple correlations (that is, the standardized sample second moments) give all possible information about the popUlation correlations. The sample correlations are the functions of the sufficient statistics that are invariant with respect to location and scale transformations; the popUlation correlations are the functions of the parameters that are invariant with respect to these transformations.
V
An Introduction to Multivariate Stat.stical Analysis, Third Edition. By T. W. Andersor. ISBN 0-471-36091·0. Copyright © 2003 John Wiley & Sons, Inc.
115
116
SAMPLE CORRE LATION COEFFICIENTS
In regression theory or least squares, one variable is considered random or dependent, and the others fixed or independent. In correlation theory we consider several variables as random and treat them symmetrically. If we start with a joint normal distribution and hold all variables fixed except one, we obtain the least squares model because the expected value of the random variable in the conditional distribution is a linear function of the variables held fIXed. The sample regression coefficients obtained in least squares are functions of the sample variances and correlations. In testing independence we shall see that we arrive at the same tests in either caf.( (i.e., in the joint normal distribution or in the conditional distribution of least squares). The probability theory under the null hypothesis is the same. The distribution of the test criterion when the null hypothesis is not true differs in the two cases. If all variables may be considered random, one uses correlation theory as given here; if only one variable is random, one mes least squares theory (which is considered in some generality in Chapter 8). In Section 4.2 we derive the distribution of the sample correhtion coefficient, first when the corresponding population correlation coefficient is 0 (the two normal variables being independent) and then for any value of the population coefficient. The Fisher z-transform yields a useful approximate normal distribution. Exact and approximate confidence intervals are developed. In Section 4.3 we carry out the same program for partial correlations, that is, correlations in conditional normal distributions. In Section 4.4 the distributions and other properties of the sample multiple correlation coefficient are studied. In Section 4.5 the asymptotic distributions of these correlations are derived for elliptically contoured distributions. A stochastic representation for a class of such distributions is found.
4.2. CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE 4.2.1. The Distribution When the Population Correlation Coefficient Is Zero; Tests of the Hypothesis of Lack of Correlation In Section 3.2 it was shown that if one has a sa-uple (of p-component vectors) Xl"'" xN from a normal distribution, the maximum likelihood estimator of the correlation between Xi and Xj (two components of the random vector X) is
( 1)
4.2
117
CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE
where Xj" is the ith component of x" and _
1
N
xi=N L
(2)
Xj".
a=\
In this section we shall find the distribution of 'ij when the population correlation between Xj and Xj is zero, and we shall see how to use the sample correlation coefficient to test the hypothesis that the population coefficient is zero. For convenience we shall treat '12; the same theory holds for each 'iJ' Since '12 depends only on the first two coordinates of each xu' to find the distribution of '12 we need only consider the joint distribution of (X Il ,X 2t ), (X I2 , xn),"" (X 1N ' X2N )· We can reformulate the problems to be considered here, therefore, in terms of a bivariate normal distribution. Let xi, ... , x~ be observation vectors from
N[(J-LI),(
(3)
J-L2
a} (1'2 (1'1
ala~p)l. P
(1'2
We shall consider ( 4)
where N
(5)
a jj
L
=
l.j = 1,2.
(xj,,-Xj){Xj,,-X j ),
a=1
x:.
and Xi is defil1ed by (2), Xio being the ith component of From Section 3.3 we see that all' a 12 , and an are distributed like n
(6)
aij
=
L
ZiaZju'
i,j = 1,2,
a=l
where n = N - 1, (zla' zZa) is distributed according to
(7) and the pairs
(Zll' ZZI),"
., (ZIN' Z2N)
are independently distributed.
118
SAMPLE CORRELATION COEFFICIENTS
Figure 4.1
Define the n-component vector Vi = (Zil"'" zin)" i = 1,2. These two vectors can be represented in an n-dimensional space; see Figure 4.1. The correlation coefficient is the cosine of the angle, say (J, between VI and v 2 • (See Section 3.2.) To find the distribution of cos (J we shall first find the distribution of cot (J. As shown in Section 3.2, if we let b = v~vl/v'lrl!> then I'" -- hr' l is orthogonal to l'l and
( 8)
cot
(J=
bllvlli IIv2 -bvlll'
If l'l is fixed, we can rotate coordinate axes so that the first coordinate axis lies along VI' Then bl'l has only the first coordinate different from zero, and l'" - hl'l has this first coordinate equal to zero. We shall show that cot (J is proportional to a t-variable when p = O. W..: us..: thl: following lemma.
Lemma 4.2.1. IfY!> ... , Yn are independently distributed, if Yo = (y~I)', y~2)') has the density f(yo)' and if the conditional density of y~2) given YY) = y~l) is f(y~"'l.v,;ll), ex = 1, ... , n, then in the conditional distribution of yrZ), ... , yn(2) given rill) = Yil), ... , y~l) = y~l), the random vectors y l(2), ..• , YP) are independent and the density of Y,;2) is f(y~2)ly~I», ex = 1, ... , n.
Proof The marginal density of Y?), . .. , YY) is Il~~ I fl(y~I», where Ny~l» is the marginal density of Y';I), and the conditional density of y l(2), ••• , yn(2) given l'11 11 = y\I), ... , yn(l) = y~l) is (9)
n:~J(yo)
n~~ I [IV,ll)
n O (x=
f(yo)
I
(I») f I (Y tr
=
On f( cr=!
(2)1
(1»)
Yo Yo
.
•
4.2
CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE
119
Write V; = (2;1' ... , 2;,)', i = 1,2, to denote random vectors. The conditional distribution of 2 2a given 2 1a = zia is N( {3zla' a 2), where {3 = puzlal and a 2 = a 22 (1 - p2). (See Se~tion 2.5.) The density of V2 given VI = VI is N( {3"'I' a 2I) since the 2 2" are independent. Let b = V2V;/V~VI (= a 21 /a U )' so that bv'I(V2 - bvl ) = 0, and let U = (V2 - bv l )'(V2 - bv l ) = V2V 2 - b 2v'IVI (=a22-aiziall). Then cotO=bVau/U. The rotation of coordinate axes involves choosing an n X n orthogonal matrix C with first row (1jc)v~, where C2=V~VI·
We now apply Theorem 3.3.1 .vith X" = 2 2a . Let Ya = L/3c a/32 2/3' a = 1, ... , n. Then YI , ••• , Y,. are inde.pendently normally distributed with variance a 2 and means
(10) n
(11)
CYa =
L
n
Cay {3Zly = {3c
y~1
L
CayC ly = 0,
y~1
We have b=L:'r~122"zl"jL:~~IZ~"=CL:~~122,,cl,.lc2=Yljc and, from Lemma 3.3.1, "
(12)
U=
L
2ia - b
II
L
2
a~1
a=1
"
L
zia =
2 Ya - Y?
a=1
which is independent of b. Then U j a 2 has a X 2-distribution with degrees of freedom.
n- 1
Lemma 4.2.2. If (2 1 ", 2 2a ), a = 1, ... , n, are independent, each pair with density (7), then the conditional distributions of b = L:~122a21ajL:~12ia and Uja2=L:~1(22a-b2Ia)2ja2 given 2 Ia =zla, a=l, ... ,n, are N({3,a 2jc 2) (c2=L:~IZla) and X2 with n-1 degrees of freedom, respectively; and band U are independent.
If p = 0, then {3 = 0, and b is distributed conditionally according to N(O, a 2jc Z ), and (13)
cbja
JUla
cb 2
n-1
120
SAMPLE CORRELATION COEFFICIENTS
has a conditional t-distribution with n - 1 degrees of freedom. (See Problem 4.27.) However, this random yariable is (14)
~
..;a;;a 1au Va 22 -ai2la u 12
=~
a 12
V1 -
=~_r
/,r;;;;a:;;
[aid(a U a 22 )]
Ii - r2
.
Thus ..;n=T r I ~ has a conditional t-distribution with n - 1. degrees of freedom. The density of t is
(15) and the density of W = rI ~ is
(16)
r(!n)
r[!
(1+w 2 )-tn
Since w = r{1- r2)- t, we have dw Idr = (1- r2)- i. Therefore the density of r is (replacing n by N - 1) (17) It should be noted that (17) is the conditional density of r for VI fixed. However, since (17) does not depend on VI' it is also the marginal density of r. Theorem 4.2.1. Let X I' ... , X N be independent, each with distribution N( fl., I). If Pij = 0, the density of rij defined by (1) is (17).
From (17) we see that the density is symmetric about the origin. For N> 4, it has a mode at r = 0 and its order of contact with the r-axis at ± 1 is !(N - 5) for N odd and !N - 3 for N even. Since the density is even, the odd moments are zero; in particular, the mean is zero. The even moments are found by integration (letting x = r2 and using the definition of the beta function). That Gr 2m = r[!(N -1)]f(m + !)/{y';r[!(N -1) + m]} and in particular that the variance is 1j(N - 1) may be verified by the reader. The most important use of Theorem 4.2.1 is to find significance points for testing the hypothesis that a pair of variables are not correlated. Consider the
4.2
CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE
121
hypothesis (18) for some particular pair (i, j). It would seem reasonable to reject this hypothesis if the corresponding sample correlation coefficient were very different from zero. Now how do we decide what we mean by "very different"? Let us suppose we arc interested in testing H against the alternative hypotheses Pij > O. Then we reject H if the sample correlation coefficient rif is greater than some number '0' The probability of rejecting H when H is true is
(19) where k N(r) is (17), the density of a correlation coefficient based on N observations. We choose ro so (19) is the desired significance level. If we test H against alternatives Pi} < 0, we reject H when 'i} < -roo Now suppose we are interested in alternatives Pi}"* 0; that is, Pi! may be either positive or negative. Then we reject the hypothesis H if rif > r 1 or 'i} < - ' I ' The probability of rejection when H is true is (20) The number r l is chosen so that (20) is the desired significance level. The significance points r l are given in many books, including Table VI of Fisher and Yates (1942); the index n in Table VI is equal to OLir N - 2. Since ,;N - 2 r / ~ has the t-distribution with N - 2 degrees of freedom, t-tables can also be used. Against alternatives Pi}"* 0, reject H if (21)
where t N _ 2 (a) is the two-tailed significance point of the I-statistic with N - 2 degrees of freedom for significance level a. Against alternatives Pij > O. reject H if (22)
l22
SAMPLE CORRELATION COEFFICIENTS
h-
From (13) and (14) we see that ..; N - 2 r / r2 is the proper statistic for testing the hypothesis that the regression of V 2 on VI is zero. In terms of the original observation (Xiel. we have
where b = [,;:=I(X:,,, -x:,)(x I " -x l )/1:;;=I(X I " _x l )2 is the least squares regression coefficient of XC" on XI,,' It is seen that the test of PI2 = 0 is equivalent to the test that the regression of X 2 on XI is zero (i.e., that PI:'
uj UI
=
0).
To illustrate this procedure we consider the example given in Section 3.2. Let us test the null hypothesis that the effects of the t\VO drugs arc llncorrelated against the alternative that they are positively correlated. We shall use the 5lfC level of significance. For N = 10, the 5% significance point (ro) is 0.5494. Our observed correlation coefficient of 0.7952 is significant; we reject the hypothesis that the effects of the two drugs are independent.
~.2.2.
The Distribution When the Population Correlation Coefficient Is Nonzero; Tests of Hypotheses and Confidence Intervals
To find the distribution of the sample correlation coefficient when the population coefficient is different from zero, we shall first derive the joint density of all' a 12 , and a 22 . In Section 4.2.1 we saw that, conditional on VI held fixed, the random variables b = aU/all and U/ u 2 = (a 22 - ai2/all)/ u 2 arc distrihuted independently according to N( {3, u 2 /e 2 ) and the X2-distribution with II _. 1 degrees of freedom, respectively. Denoting the density of the X 2-distribution by g,,_I(U), we write the conditional density of band U as n(bl{3, u 2/a ll )gn_I(II/u 2)/u 2. The joint density of VI' b, and U is lI(l'IIO. u IC[)n(bl{3, u2/all)g,,_1(1l/u2)/u2. The marginal density of V{VI/uI2='all/uI2 is gll(u); that is, the density (\f all is
(24)
where dW is the proper volume element. The integration is over the sphere ~"ll'l = all; thus, dW is an element of area on this sphere. (See Problem 7.1 for the use of angular coordinates in
4.2
123
CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE
defining dW.) Thus the joint density oi b, U, and all is
(25) _ gn( allla})n( bl i3,
Now let b = aId all' U = a22
(T
2la u )gn-I (u I (]" 2)
(]"12(]" 2
-
-
aid all. The Jacobian is
o (26) 1
Thus the density of all' a 12 , and a 22 for all ~ 0, a 22 ~ 0, and a ll a 22 is
(27)
where
-
ai2 ~ 0
124
SAMPLE CORRELATION COEFFICIENTS
The density can be written
(29) for A positive definite, and. 0 otherwise. This is a special case of the Wishart density derived in Chapter 7. We want to find the density of
· (30) where ail = all / a}, aiz = azz / a}, and aiz = a12 /(a l a z). The tra:lsformation is equivalent to setting 171 = az = 1. Then the density of all> azz , and r = a 12 /,ja ll aZZ (da 12 = drJalla zz ) is (31) where
all - 2prva;;
(32)
va:;; + azz
1 -p z
Q=
To find the density of r, we must integrate (31) with respect to all and azz over the range 0 to 00. There are various ways of carrying out the integration, which result in different expressions for the density. The method we shall indicate here is straightforward. We expand part of the exponential:
(33)
exp [
va:;; ] =
prva;; ( 1-p Z)
va:;;)"
;. (prva;; '-' a!(1-pZ) a
a~O
Then the density (31) is
(34)
.{exp[-
all ]a(n+al/Z-l}{exp[_ azz ]a(n+al/z-J}. 2(1- pZ) 11 2(1- pZ) 22
4.2
125
CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE
Since (35)
laoa~(n+a)-lexp[o
a
2
2(1- p )
]da=r[~(n+a)1[2(1-P2)r\,,+a),
the integral of (34) (term-by-term integration is permissible) is
(36)
(1- p2)~n2nV:;;:f(~n)f[Hn - 1)1
. -£ <>=0
,(pr)a af2[Hn+a)12n+a(1-p")n+a a .(1 - p 2) 2 ~n
2 ~(n - 3)
=(1-p) (1-r) I 1 [I V'1Tf("2n)f "2(n -1)
x
a
,,(2pr) 1 <>=0 '-' a.,
f2[1.( ?
-
n+a
)1 .
The duplication formula for the gamma function is (37)
f(2z) =
22Z-lf(z)(z+~)
v:;;:-
It can be used to modify the constant in (36). Theorem 4.2.2. The correlation coefficient in a sample of Nfrom a bivariate normal distribution with correlation p is distributed with density
(38) -lsrs1,
where n = N - 1.
The distribution of r was first found by Fisher (1915). He also gave as another form of the density,
(39) See Problem 4.24.
126
SAMPLE CORRELATION COEFFICIENTS
Hotelling (1953) has made an exhaustive study of the distribution of ,.. He has recommended the following form: ( 40)
11 - 1
rc n)
·(1- pr)
2 ) i"( 1
(1 _
,f2; f(n+~) -n
+i
_
2) i(n - 3 )
p
r
(I I.
I.
1 + pr ) F "2'2,11 +"2' - 2 - ,
where ( 41)
... _
F(a,b,c,x) -
x
f(a+j) f(b+j) f(c) xi f(b) f(c+j) j!
j~ f(a)
is a hypergeometric function. (See Problem 4.25.) The series in (40) converges more rapidly than the one in (38). Hotelling discusses methods of integrating the density and also calculates moments of r. The cumulative distribution of r, (42)
Pr{r s r*} = F(r*IN, p),
has been tabulated by David (1938) fort P = 0(.1).9, '1/ = 3(1)25, SO, 100, 200, 400, and r* = -1(.05)1. (David's n is our N.) It is clear from the density (38) that F(/"* IN, p) = 1 - F( - r* IN, - p) because the density for r, p is equal to the density for - r, - p. These tables can be used for a number of statistical procedures. First, we consider the problem of using a sample to test the hypothesis (43)
H: p= Po.
If the alternatives are p > Po, we reject the hypothesis if the sample correlation coefficient is greater than ro, where ro is chosen so 1 - F(roIN, Po) = a, the significance level. If the alternatives are p < Po, we reject the hypothesis if the sample correlation coefficient is less than rb, where ro is chosen so F(r~IN, Po) = a. If the alternatives arc p =1= Pu, thc rcgion of rejection is r> r l and r < r;, where r l and r; are chosen so [1- F(rIIN, Po)] + F(rIIN, Po) = a. David suggests that r l and r; be chosen so [l-F(r\IN,po)]=F(r;IN,po) = ~a. She has shown (1937) that for N;:: 10, Ipl s 0.8 this critical region is nearly the region of an unbiased test of H, that is, a test whose power function has its minimum at Po' It should be pointed out that any test based on r is invariant under transformations of location and scale, that is, xia = biXia + C j ' b i > 0, i = 1,2, 'p = ()(.1l.9 means p = 0,0.1,0.2, ... ,0.9.
4.2
CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE
127
Table 4.1. A Power Function p
Probability
- 1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0
0.0000 0.0000 0.0004 0.0032 0.0147 0.0500 0.1376 0.3215 0.6235 0.9279 1.0000
a = 1, ... , N; and r is essentially the only invariant of the sufficient statistics (Problem 3.7). The above procedure for testing H: p = Po against alternatives p > Po is uniformly most powerful among all invariant tests. (See Problems 4.16, 4.17, and 4.18.) As an example suppose one wishes to test the hypothesis that p = 0.5 against alternatives p"* 0.5 at the 5% level of significance using the correlation observed in a sample of 15. In David's tables we find (by interpolation) that F(0.027 I 15, 0.5) = 0.025 and F(0.805115, 0.5) = 0.975. Hence we reject the hypothesis if our sample r is less than 0.027 or greater than 0.805. Secondly, we can use David's tables to compute the power function of a test of correlation. If the region of rejection of H is r> rl and r < r;, the power of the test is a function of the true correlation p, namely [1- F(rIIN, p) + [F(r;IN, p)J; this is the probability of rejecting the null hypothesis when the population correlation is p. As an example consider finding the power function of the test for p = 0 considered in the preceding section. The rejection region (one-sided) is r ~ 0.5494 at the 5% significance level. The probabilities of rejection are given in Table 4.1. The graph of the power function is illustrated in Figure
4.2.
Thirdly, David's computations lead to confidence regions for p. For given N, r; (defining a ~ignificance point) is a function of p, say fl( p), and r l is another function of p, say fz( p), such that
(44)
Pr{ft( p) < r
Clearly, fl( p) and fz( p) are monotonically increasing functions of p if r1 and r; are chosen so 1- F(rIIN p) =!a = F(r;IN, p). If p = fil(r~ if, the
128
SAMPLE CORRELATION COeFFICIENTS
Prob.
p
Figure 4.2. A power function.
inverse of r = [..( p), i = 1,2, then the inequality fI( p) < r is equivalent tot
p
(45) This equation says that the probability is 1 - a that we draw a sample such that the interval (fi I (r )Ji' (r)) covers the parameter p. Thus this interval is a confidence interval for p with confidence coefficient 1 - a. For a given N and a the curves r = fl( p) and r = fz( p) appear as in Figure 4.3. In testing the hypothesis p = Po, the intersection of the line p = Po and the two curves gives the significance points and In setting up a confidence region for p on the basis of a sample correlation r*, wc find thc limits I:;' (r*) and
r,
r;.
Figure 4.3 tThe point (f,( p), p) on the first curve is to the left of (r, pl, and the point (r, flier»~ is above (r,
pl.
4.2
CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE
129
fJ I (r*) by the intersection of the line r = r* with the two curves. David gives these curves for a = 0.1, 0.05, 0.02, and 0.01 for various values of N. Onesided confidence regions can be obtained by using only one inequality above. The tables of F(rIN, p) can also be used instead of the curves for finding the confidence interval. Given the sample value r*, f~ I (r*) is the value of p such that !a= Pr{r:s;r*lp} =F(r*IN,p), and similarly /2 1(r*) is the value of P such that !a = Pr{r;::. r* I p} = 1 - F(r* IN, p). The interval between these two values of p, ([2 1 (r* ),f~ I (r* )), is the confidence interval. As an example, consider the confidence interval with confidence coefficient 0.95 based on the correlation of 0.7952 observed in a sample of 10. Usi"1g Graph II of David, we j ind the two limits are 0.34 and 0.94. Hence we state that 0.34 < P < 0.94 with confidence 95%. Definition 4.2.1. Let L(x, 6) be the likelihood function of the observation vector x and the parameter vector 6 E fl. Let a null hypothesis be defined b.y a proper subset w of fl. The likelihood ratio criterion is
A( x) =
(46)
_sU...oP...o0-".E-".w-:;-L-;-(x_,--::6:-7-) SUPeE Il
L (x,6) .
The likelihood ratio test is the procedure of rejecting the null hypothesis wizen A(x) is less than a predetermined constant.
Intuitively, one re.iects the null hypothesis if the density of the observations under the most favorable choice of parameters in the null hypothesis is much less than the density under the most favorable unrestricted choice of the parameters. Likelihood ratio tets have some desirable features; see Lehmann (195?), for example. Wald (1943) has proved some favorable asymptotic properties. For most hypotheses concerning the multivariate normal distribution, likelihood ratio tests are appropriate and often are optimal. Let us consider the likelihood ratio test of the hypothesis that p = Po based on a sample XI' ••• ' x N from the bivariate normal distribution. The set fl consists of ILl' ILz, Up 172 , and P such that 171 > 0, 172 > 0, - 1 < P < l. The set w is the subset for which p = Po. The likelihood maximized in fl is (by Lemmas 3.2.2 and 3.2.3) NNe-N
( 47)
maxL 11
= ---N---,~l,\-.-.-,-.-".
(21T) (I -,-)' al\/-a~i-
130
SAMPLE CORRELATION COEFFICIENTS
LJ nder the null hypothesis the likelihood function is (48)
where u 2 = U 1U 2 and occurs at T=
T=
u l /U 2 . The maximum of (48) with respect to
.;a:; / ..;a;;. The concentrated likelihood is
T
(49)
the maximum of (49) occurs at
(50)
The likelihood ratio criterion is, therefore,
(51 ) The likelihood ratio test is (1- ptX1 - r 2 X1 - por)-2 < e, where e is chosen so the probability of the inequality when samples are drawn from normal populations with correlation Po is the prescribed significanc~ level. The critical region can be written equivalently as (52) or
(53) r<
poe - (1 - pt)v"f=C pJe + 1 - p~
---"---+-"';""':''"--;;---
'*
Thus the likelihood ratio test of H: p = Po against alternatives p Po has a rejection region of the form r> r l and r < r;; but r l and r; are not chosen so that the probability of each inequality is Ci/2 when H is true, but are taken to be of the form given in (53), where e is chosen so that the probability of the two inequalities is Ci.
4.2 CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE
131
4.2.3. The Asymptotic Distribution of a Sample Correlation Coefficient and Fisher's Z In this section we shall show that as the sample size increases, a sample correlation coefficient tends to be normally distributed. The distribution of a particular function. of a sample correlation, Fisher's z [Fisher (1921)], which has a variance approximately independent of the population correlation, tends to normajty faster. We are particularly interested in the sample correlation coefficient
(54) for some i and j, i
'* j. This can also be written
(55) where CgJi(n) =Agh(n)/ VUggUhh' The set CuCn), Cjj(n), and Ci/n) is distributed like the distinct elements of the matrix
where the
(zta' Z/;) are
independent, each with distribution
where U ij
p=--VUjjU'jj .
Let
(57)
(58)
132
SAMPLE CORRELATION COEFFICIENTS
Then by Theorem 3.4.4 the vector vn[U(n) - b] has a limiting normal distribution with mean and covariance matrix
°
(59)
2p ) 2p . 1 + p2
Now we need the general theorem: Theorem 4.2.3. Let Wen)} be a sequence of m-component random vectors and b a fixed vector such that m[ U(n) - b] has the limiting distribution N(O, T) as n --> 00. Let feu) be a vector-valued function of u such that each component fj(u) has a nonzero differential at u =b, and let iJfj(u)/iJUjlu~b be the i,jth component of
Proof See Serfling (1980), Section 3.3, or Rao (1973), Section 6a.2. A function g(u) is said to have a differential at b or to be totally differentiable at b if the partial derivatives ag(u) / au, exist at u = b and for every e> 0 there exists a neighborhood Ne(b) such that
(60)
Ig(U)-g(b)-j~ iJ~~~)(Ui-b;)lsellu-bll
for all u E NeC b ).
•
It is cIeJr that U(n) defined by (57) with band T defined by (58) and (59), respectively, satisfies the conditions of the theorem. The function
(61)
satisfier, the conditions; the elements of
(62)
4.2
133
CORRELATION COEFFICIENT OF ABlY ARIATE SAMPLE
and f(b)
=
p. The vanance of the limiting distribution of {,l [r(n) - p] is
(63)
I
- iP -
I
.~p
Thus we obtain the following: Theorem 4.2.4.
If r(n) is the sample correlation coefficient of a sample of N
(= n + 1) from a rlOrmal distribution with correlation p, then {,l[r(n) - p l! (1- p2) [or m[r(n) - p]/(l- p2)] has the limiting distribution N(O, 1).
It is clear from Theorem 4.2.3 that if f(x) is differentiable at x = p. then {flU(r) - f( p)] is asymptotically normally distributed with mean zero and
variance
A useful function to consider is one whose asymptotic variance is constant (here unity) independent of the parameter p. This function satisfies the equation
(64) Thus f( p) is taken as Wog(1 + p) -log(l - p)] = t logO + p)/O - p)]. The so-called Fisher's z is z=
(65) where r = tanh z = (e Z
(66)
-
tlog 11 +_ rr = tanh -
e- Z ) I(e' + e-=). Let
1
r,
134
SAMPLE CORRELATION COEFFICIENTS
Theorem 4.2.5. Let z be defined by (65), where r is the correlation coefficient of a sample of N (= n + 1) from a bivariate normal distribution with con·elarion p: let?: be defined by (66). Then In (z - {) has a limitb'g normal distribution with mean 0 and variance 1. It can be shown that to a closer approximation
(67) ( 68)
Cz -
?: + 2Pn .
, --1 - - c ( z-?:-P )2 . C(z-o n- 2 2n
The latter follows from (69)
, 1 8 - p2 C(z-?:t=-+-,-+ ... n
4n-
25 one take z as normally distributed with mean and variance given by (67) and (68). Konishi (1978a, 1978b, 1979) has also studied z. [Ruben (1966) has suggested an alternative approach, which is more complicated, but possibly more accurate.] We shall now indicate how Theorem 4.2.5 can be used. a. Suppose we wish to test the hypothesis p = Po on the basis of a sample of N against the alternatives p'* Po. We compute,. and then z by (65). Let ( 70)
1 1 + Po ?:o = '\og-l--· • - Po
Then a region of rejection at the 5% significance I ;ve\ is
(71)
VN - 3 Iz - ?:ol > 1.96.
A better region is (72)
I
VN - 3 z - ?:o - N~Po _ 1 > 1.96.
I
b. Suppose we have a sample of Nl from one population and a sample of
N2 from a second population. How do we test the hypothesis that the two
135
4.2 CORRELATION COEFFICIENT OF A BIVARIATE SAMPLE.
correlation coefficients <..re equal, Pl = P2? From Theorem 4.2.5 we know that if the null hypothesis is true then Zl - Z2 [where Zl and Zz are defined by (65) for the two sample correlation coefficients] is asymptotically normally distributed with mean 0 and variance l/(Nl - 3) + lj(Nz - 3). As a critical region of size 5%, we use
(73)
c. Under the conditions of paragraph b, assume that Pl = pz = p. How do we use the results of both samples to give a joint estimate of p? Since Zl and Zz have variances Ij(N, - 3) and Ij(Nz - 3), respectively, we can estimate ~ by (Nl - 3)Zl
(74)
Nl
+ (Nz - 3)zz
+Nz - 6
and convert this to an estimate of P by the inverse of (65). d. Let r be the sample correlation from N observations. How do we obtain a confidence interval for p? We know that approximately
(75)
Pr{ -1.96 ~ VN - 3 (z -
n ~ 1.96} = 0.95.
From this we deduce that [ -1.961 VN - 3 + z, 1.96 j VN - 3 + z] is a confiFrom this we obtain an interval for p using the fact dence interval for p = tanh ~ = (e' - e-{ )j(e{ + e-q, which is a monotonic transformation. Thus a 95% confidence interval is
r
(76)
tanh(z -1.96jvN -
3) ~p~ tanh(z + 1.961VN- 3).
The bootstrap method has been developed to assess the variability of a sample quantity. See Efron (1982). We shall illustrate the method on the sample correlation coefficient, but it can be applied to other quantities studied in this book. Suppose Xl"'" X N is a sample from some bivariate population not necessarily normal. The approach of the bootstrap is to consider these N vectors as a finite population of size N; a random vector X has the (discrete) probability
(77)
I
Pr{X=x,,} = N'
a=l, ... ,N.
136
SAMPLE CORRELATION COEFFICIENTS
A randJm sample of size N drawn from this finite population ha3 a probability distribution, and the correlation coeffici( nt calculated from such a sample has a (discrete) probability distribution, say PN(r). The bootstrap proposes to use this distribution in place of the unobtainable distribution of the correlation coefficient of random samples from the parent population. However, it is prohibitively eXpensive to compute; instead PN(r) is estimated by the empirical distribution of r calculated from a large number of random samples from (77). Diaconis and Efron (1983) have given an example of N = 15; they find the empirical distribution closely resembles the actual distribution of r (essentially obtainable in this special case). An advantage of this approach is that it is not necessary to assume knowledge of the parent population; a disadvantage is the massive computation.
4.3. PARTIAL CORRELATION COEFFICIENTS; CONDmONAL DISTRIBUTIONS 4.3.1. Estimation of Partial Correlation Coefficients Partial correlation coefficients in normal distributions are correlation coefficients in conditional distributions. It was shown in Section 2.5 that if X is distributed according to N(I-I-, I), where
(1)
X(I») (
X= X(Z) ,
1-1-(1) )
1-1-=
(
I-I-(Z) ,
then the conditional distribution of X(l) given X(Z) = x(Z) is N[I-I-(I) + P(x(Z) I-I-(Z», Iu.zl, where
(2) (3) The partial correlations uf X(1l given x(Z) are the correlations calculated in the usual way from I ll . z. In this section we are interested in statistical problems concerning these correlation coefficients. First we consider the problem of estimation on the basis of a sample of N from N(I-I-, I). What are the maximum likelihood estimators of the partial correlations of X(l) (of q components), Pij-q+l •...• p? We know that the
42
137
PARTIAL CORRELATION COEFFICIENTS
maximum likelihood estimator of I is (l/N)A, where N
(4)
A=
L
(xa-x)(xa- x )'
a~l
and x=(1/N)I::~~lXa=(X(l)' X(2 ),),. The correspondence between I ll . 2 , p, and 122 is one-to-one by virtue of (2) and (3) and
and
I
(5) (6) We can now apply Corollary 3.2.1 to the effect that maximum likelihood estlmators of functions of pa rameters are those functions of the maximum likelihood estimators of those parameters.
Theorem 4.3.1. Let XI'""XN be a sample from N( .... ,I), where .... and I are partitioned as in (1). Define A by (4) and (x ll )' Xl:)') = (1/N)L.~~ l(X~I)' x~) '). Then the maximum likelihood estimators of ....(\) .....("1. p, I ll . 2 , and 122 are fJ,(1) =x(l\ fJ,(2) =X(2),
(7) and
in = (l/N)A 22 ,
respectively.
In turn, Corollary 3.2.1 can be used to obtain the maximum likelihood estimators of ....(Il, ....(2\ p, In' (1'ii.'I+I. .... I'. i=l, ... ,q. and Pij.q+I .... p. i, j = 1, ... , q. It follows that the maximum likelihood estimators of the partial correlation coefficients are Uij-q+l .... p Pij'q+l •..• p
where
Uij.q+l ..... p
is the i,jth element of ill.:!'
i,j=l. .... q.
138
SAMPLE CORRELATION COEFFICIENTS
Theorem 4.3.2.
Let XI"'" x N
be a sample of
N from N(tJ., l:). The
maximum likelihood estimators of .Pij'q + I, .. " p' the partial correlations of the first q components conditional on the last p - q components, are given by ~ aij.q+I,,,.,p Pij'q+I ..... p = ",fa.. a.. ' lI·q+I, .. "p jJ·q+I, ... ,p
(9)
i,j=l, ... ,q,
where
( 10)
The estimator i\.q + I, ... , p' denoted by r ij .q+ \. .... p' is called the sample pm1ial correlation coefficient between Xi and Xj holding X q+ I , ... , Xp fixed. It is also called the sample partial correlation coefficient between Xi and Xj having taken account of Xq + I' ... , Xp. Note that the calculations can be done in terms of (r i / The matrix A ll . 2 can also be represented as ( 11)
A II ~ =
N
L
[X~I)
-
P{ X~2) -
ill) -
i
(2
))] [
X~I)
-
i(l) -
P{ x~)
-
i
(2
))] ,
a~1
The vector X~I) - ill) - p(X~2) - i (2 )) is the residual of x~l) from its regression on X~2) and 1. The partial correlations are simple correlations between these residuals. The definition can be used also when the distributions involved are not normal. Two geometric interpretations of the above theory can be given. In p-dimcnsional ~pacc XI"'" x N represent N points. The sample regression function (12)
is a (p - q )-dimensional hyperplanc which is the intersection of q (p - 1)dimensional hyperplanes, P
L
Xi=X i +
(13)
~ij(Xj-XJ,
i = 1, ... ,q,
j~q+1
. .
~
.
~
~
~
-I
where Xi' Xj are runnmg vanables. Here f3ij IS an element of P = l:12l:22 = A\2 A 2 The ith row of is (~i, q+ I " ' " ~iP)' Each right-hand side of (13) is the least squares regression function of Xi on Xq + l ' ... , x p; that is, if we project the points XI"'" x N on the coordinate hyperplane of Xj,Xq+I, ... ,x p'
-r
P
4.3
139
PARTIAL CORRELATION COEFFICIENTS
2
N Figure 4.4
then (13) is th{ regression plane. The point with coordinates p
(14)
Xi=Xi+
L j~q
~ij(xia-Xj)'
i = 1, . .. ,q,
+1
j=q+l, ... ,p, is on the hyperplane (13). The difference in the ith coordinate of Xu fl.nd the point (14) is Yia = Xia - [Xi + LJ-q+ I ~;/xiU - x j )] for i = 1, ... , q anJ 0 for the other coordinates. Let i. = (YI,,' ... ' Y'I,J These points can be repr~ sented as N points in a q-dimensional space. Then A Il .z = L~~IYaY~. We can also interpret the sample as p points in N-space (Figure 4.4). Let u j = (xii' ... ' Xi,'i)' be the jth point, and let I:: = (1, ... ,1)' be another point. The point with coordinates Xi' ... ' Xi is XiI::. The projection of u i on the hyperplane spanned by U q + I, ... , up, I:: is p
(15)
uj=xil::+
L
~ij(ui-Xil::);
j~q+1
this is the point on the hyperplane that is at a minimum distance from U i • Let uj be the vector from ui to u i , that is, Ui - Ui , or, equivalently, this vector translated so that one endpoint is at the origin. The set of vectors u1'; ... , are the projections of u 1' •.• , u q on the hyperplane orthogonal to
u:
140
SAMPLE CORRELATION COEFFICIENTS
Then ui'u; =ajj.q+I •...• P, the length squared of u; (Le., the square ofthe distance of u from a). Then u; 'uj / ..juj'u; uj'uj = 'if'q+l ..... p is the cosine of the angle between ui and uj. As an example of the use of partial correlations we consider some data [Hooker (1907)] on yield of hay (Xl) in hundredweights per acre, spring rainfall (X2 ) in inches, and accumulated temperature above 42°F in the spring (X3 ) for an English area over 20 years. The estimates of JLi' U i (= and Pjj are
Uq+I, ••• ,Up,E.
..;u:;),
v.. =.r =
n
(16)
::
(~"
P3l
P12
=
(442) 8~'1O ,
po) (
1
P23
P32
1
=;
(59~·91 W02) ,
100 0.80 -0.40
0.80 1.00 -0.56
-O~)
-0.56 . 1.00
From the correlations we observe that yield and rainfall are positively related, yield and temperature are negatively related, and rainfall and temperature are negatively related. What interpretation is to be given to the apparent negative relation between yield and temperature? Does high temperature tend to cause low yield, or is high temperature associated with low rainfall and hence with low yield? To answer this question we consider the correlation between yield and temperature when rainfall is held fixed; that is, we use the data given above to estimate the partial correlation between Xl and X3 with X 2 held fixed. It ist (17)
Thus, J thl! effect of rainfall is removed, yield and temperature :lfe positively correlated. The conclusion is that both hi[;h raninfall and high temperature increase hay yield, but in most years high rainfall occurs with low temperature and vice versa. tWe com,Jute with
i
as if it were I.
4.3
141
PARTIAL CORRELATION COEFFICIENTS
4.3.2. The Distribution of the Sample Partial Correlation Coefficient
In order to test a hypothesis about a population partial correlation coefficient we want the distribution of the sample partial correlation coefficient. The partial correlations are computed from AJl.2 =A II -AI2AZ2IA21 (as indicated in Theorem 4.3.1) in the same way that correlations are computed from A. To obtain the distribution of a simple correlation we showed that A was distributed as L~~lzaz~, where ZI"",ZN_I are distributed independently according to N(O, l:) and independent of X (Theorem 3.3.2). Here we want to show that AII-2 is distributed as L~~II-( p-q)VaV~, where VI' ... , VN - I -( p_q) an~. distributed independently according to N(O, l: 11.2) and independently of p. The distribution of a partial correlation coefficient will follow from the characterization of the distribution of A Jl2' We state the theorem in a general form; it will bc used in Chapter 8, where we treat regression'in detail. The following corollary applies it to AII'~' expressed in terms of residuals. Theorem 4.3.3. Suppose YI , .•. , Y,,, are independent with Ya distributed according to N(fwa , ~), where wa is an r-component vector. Let H = L::'~ l waw~. assumed nonsingular, G = L;;~I Yaw~ H- I , and m
(18)
C=
L
m
(Ya - GWa)(Ya - Gwa )' =
a=1
L
YaY~ - GHG'.
a=1
Then C is distributed as L;;~~VaV~, where VI,,,,,Vm _ r are independently distributed according to N(O, ~) a~d independently of G. Proof. The rows of Y = (YI , ... , Ym) are random vectors in an m-dimensional space, and the rows of W = (wl>"" w m ) are fixed vectors in that space. The idea of the proof is to rotate coordinate axes so that the last r axes are in the space spanned by the rows of W. Let E2 = FW, where F is a square matrix such that FHF' = 1. Then m
(19)
E2E2 =FWW'F' =F
L
waw~F'
a=1
=FHF' =1. Thus the m-component rows of E2 are orthogonal and of unit length. It is possible to find an (m - r) X m matrix E I such that (20)
E=
(!:)
142
SAMPLE COR RELATION COEFFICIENTS
is orthogonal. (Sec Appendix, Lemma A.4.2.) Now let U = YE' (i.e., Ua = L.~~l enllfll )· By Theorem 3.3.1 the columns of U=(U1, ... ,Um) are independcntly and normally distributed, each with covariance matrix <1>. The means are given by GU = GYE' = rWE'
PI)
=
rr 1E 2 (E;
E 2)
= (0 rp-l) by orthogonality of E. To complete the proof we need to show that C transforms to L.~::;U"U~. We have m
E
(22)
a~
m
fJ~ = IT' = UEE'U' = UU' =
I·
E
UaU~,
a=1
Note that
G = YW'H- 1 = UEE 2(F- 1)'F'F
( 2:1)
=U(!:)ESF
=U(~)F=U(2)F,
(24) a=m-r+l
Thus C is . m-r
m
(25)
Eu.,u,:-
El:,l'::-GHG'= "
I
This proves the theorem.
"
U"U,:= n=m-r+ I
I
E UnU,:. ,,~l
•
It follows from the above considerations that when we obtain the following:
r = 0, the
G U = 0, and
Corollary 4.3.1. If r = 0, the matrix GHG' defined in Theorem 4.3.3 is distributed as L.:~m-r+l UaU~, where Um-r+1,,,,,Um are independently distributed, each according to N(O, <1».
4.3
PARTIAL CORRELATION
143
COEFFICIEN'I~S
We now find the distribution of A II .2 in the same form. It was shown in Theorem 3.3.1 that A is distributed as L~:fzaz~, where ZI"",ZN_I are independent, each with distributi01 N(O, :I). Let Za be partitioned into two subvectors of q and p - q components, respectively:
(26) Then
Za
=(
~i::) .
Aij = L~~ I Z~i)Z~j),. By Lemma 4.2.1, conditionally on
zF) =
Z\2)"",Z~~1 =Z~~I' the random vectors ZP)"",Z~~I are independently distributed, with Z~I) distributed according to N(pz~), :I ll .2), where p = :I12:I;l and :I ll .2 = :III -:I 12 :I Z/:I 21 . Now we apply Theorem 4.3.3 with Z~I) = Y", z~) = IVa' N -1 = m, p - q = r, 13 = r, :I 11.2 =~, All = L~:f YaY~, A12AZ21 = G, A22 =H. We find that the conditional distribution of All (A12Azi)A22(Az21A'12)=All.2 given Z~2)=Z~), a=1, ... ,N-1, is that of L~:t-(P-q)VaV~, where VI, ... ,VN_I_(p_q) are independent, each with distribution N(O,:I 11.2)' Since this distribution does not depend on (z~2)}, we obtain the following theorem:
Theorem 4.3.4. The matrix A 11 . 2 =AII -A12Az21A21 is distributed as L~:f -( p-q)VaV~, where VI'"'' V N_ 1_( p-q) are independently distributed, each according to N(O, :I11.2), and independently of AI2 and A 22 . Corollary 4.3.2.
If
:I 12 =
° (P = 0),
then A 11.2
is distributed as
L~:I-( p-q)V"V~ and A12AZ21A21 is distributed as L~'::~_( p_qp"V~, where VI"'" V N- 1 are independently distributed, each according to N(O, :I11.2)'
Now it follows that the distribution of rij'q+l ..... p based on N observations is the same as that of a simple correlation coefficient based on N - (p - q) observations with a corresponding population correlation value of Pij'q+l •...• p. Theorem 4.3.5. If the cdf of rij based on a sample of N from a normal distribution with correlation Pij is denoted by F(r!N, Pij)' then the cdf of the sample partial correlation rij .q + I • .... 1' hased Oil a sample of N from a normal distribution with partial con'elation coefficient Pij-q+ I • ...• 1' is F[rlN (p -q), Pij.q+l ..... p]· This distribution was dcrivcd hy Fisher (1924). 4.3.3. Tests of Hypotheses and Confidence Regions for Partial Correlation Coefficients Since the distribution of a sample partial correlation rij . q + l •...• p based on a sample of N from a distribution with population correlation Pij'q+l •...• P
144
SAMPLE CORRELATION COEFFICIENTS
equal to a certain value, P, say, is the same as the distribution of a simple correlation, based on a sample of size N - (p - q) from a distribution with the corresponding population correlation of p, all statistical inference procedures for the simple population correlation can be used for the partial correlation. The procedure for the partial correlation is exactly the same except that N is replaced by N - (p - q). To illustrate this rule we give two examples. Example 1. Suppose that on the basis of a sample of size N we wish to obtain a cClllfidence interval for Pij-q+l, ... ,p' The sample partial correlation is 'ij'q+I, ... ,p' The procedure is to use David's charts for N-(p-q). In the example at (he end of Section 4.3.1, we might want to find a confidence interval for P12'3 with confidence coefficient 0.95. The sample partial correlation is '12.3 = 0.759. We use the chart (or table) for N - (p - q) = 20 - 1 = 19. The int.!rval is 0.50 < P12'3 < 0.88. Example 2. Suppose that on the basis of a sample of size N we use Fisher's z for an approximate significance test of Pij-q + I, ... P = Po against two-sided alternatives. We let
.. I , .... p II 1 +,IJ·q+ z = 2' og 1 - ,.. ' IJ·q+I, ... ,p
(27)
1 + Po bo = 2'log 1 - Po' 1
IN -
Then (p - q) - 3 (z - bo) is compared with the significance points of the standardized normal distribution. In the example at the end of Section 4.3.1, we might wish to test the hypothesis P13.2 = 0 at the 0.05 level. Then bo = 0 and ";20 - 1 -:- 3 (0.0973) = 0.3892. This value is clearly nonsignificant (\ 0.3892\ < 1.96), and hence the data do not indicate rejection of the null hypothesis. To answer the question whether two variables Xl and x 2 are related when both may be related to a vector X(2) = (X3"'" x p )' two approaches may be used. One is to consider the regression of X 1 on x 2 and X(2) and test whether the regression of XI on X 2 is O. Another is t) test whether P12.3, ... , P = O. Problems 4.43-4.47 show that these approaches lead to exactly the same test. 4.4. THE MULTIPLE CORRELATION COEFFICIENT 4.4.1. Estimation of the Multiple Correlation Coefficient
The population multiple correlation between one variate and a set of variates was defined in Section 2.5. For the sake of convenience in this section we shall treat the case of the multiple correlation between XI and the vector
145
4.4 THE MULTIPLE CORRELATION COEFFICIENT
= (X2 •••.• X p )'; we shall not need subscripts on R. The variables can always be numbered so that the desired multiple correlation is this one (any irrelevant variables being omitted). Then the multiple correlation in the population is
X(2)
U;I)'1
(1)
where
22
I
U(I)
0'11
13.
u(l)'
and
are defined by
'1 22
(2) (3) Given a sample
x N (N > p), we estimate "I by S = [N /( N -
XI."',
J)1i
or
(4) and we estimate
13
by
j3 = i221 U(1) =A 22I a(I)' We define the
sample multiple
correlation coefficient by
(5) That this is the maximum likelihood estimator of R is justified by Corollary 3.2.1, since we can define R, U(I)' '1 22 as a one-to-one tra.nsformation of :I. Another expression for R [s(:e (16) of Section 2.5] follows from
(6)
\A\ a ll \A 22 \
.
The quantities Rand j3 have properties in the sample that are similar to those Rand 13 have in the population. We have analogs of Theorems 2.5.2, 2.5.3, and 2.5.4. Let XI a = i 1 + j3 '(x~) - i(2»), and xf a = X1 a - i 1 a be the residual. Theorem 4.4.1. The residuals xf a are uncorrelated in the sample with the components of x~), a = 1, ... , N. For every vel tor a
(7)
146
SAMPLE CORRELATION COEFFICIENTS
The sample correlation between x I" and a' X~.21, ex = I, ... , N, is maximized for a = ~, llIzd that maximum corre/ation is R. Proof Since the sample mean of the residuals is 0, the vector of sample covariances between xi a and X~2) is proportional to
~ 8)
v
t
[I. x lo -
.i\) -
~'(X~2) -
."(l2))]
(X~2) -
i
(2
))'
= a;l) - ~'A22 = o.
0:-=1
The right-hand side of (7) can be written as the left-hand side plus
N
= (~- a)'
L.
(X~2) - i (2 »)( X~2) _i(2))'(~ - a),
a=1
which is 0 if and only if a = ~. To prove the third assertion we consider the vector a for which L;~·=z[a'(x~2) - i (2 ))]2 = L~=1[~'(X~2) -i(2))j2, since the correlation is unchanged when the linear function is mulitplied by a positive constant. From (7) we obtain N
(10)
L.
a ll -2
(xlo-il)~'(x~;)-i(1))+
0=1
-2
[~'(x~2)-i(2))]
2
a=1
N
:'>a ll
N
L.
L. o~l
(xI,,-il)a'(x~2)-il2))+
N
L.
2
[a'(x~2)-il2))],
a=1
from which we deduce
( II)
i
, I I
which is (5).
•
Thus i I + ~ '(X~2) - i (2 )) is the best linear predictor of X la in the sample, and ~'x;;) is the linear function of X~2) that has maximum sample correlation
147
4.4 THE MULTIPLE CORRELATION COEFFICIENT
with Xl". The minimum sum of squares of deviations [the left-hand side of (7)] is N
(12)
L
[(X la -XI) -
~'(X~2) -i(2))f =a ll
-
J3'A22J3
a=1
as defined in Section 4.3 with q = 1. The maximum likelihood estimator of (T11.2 is 0- 11 . 2 = a ll . 2 /N. It follows that
(13) Thus 1 - R2 measures the proportional reduction in the variance by using residuals. We can say that R2 is the fraction of the ·variance explained by X(2). Thelarger R2 is, the more the variance is decreased by use of the explanatory variables in X(2). In p-dimensional space XI' ••• ' x N represent N points. The sample regression function Xl =i l + J3'(X(2) -i(2)) is the (p -1)-dimensional hyperplane that minimizes the squared deviations of the points from the hyperplane, the deviations being calculated in the X I-direction. The hyperplane goes through the point i. In N-dimensional space the rows of (Xl' ..• ' X N ) represent p points. The N-component vector with ath component Xja -Xi is the projection of the vector with a th component Xi a on the plane orthogonal to the equiangular line. We have p such vectors; a'(x~2) -i(2)) is the ath component of a vector in the hyperplane spanned by the last p - 1 vectors. Since the right-hand side of (7) is the squared distance between the first vector and the linear combination vf the last p - 1 vectors, J3 '(X~2) - i(2)) is a component of the vector which minimizes this squared distance. The interpretation of (8) is that the vector with ath component (Xl a -Xl) - J3'(X~2) -i(2)) is orthogonal to each of the last p - 1 vectors. Thus the vector with ath component J3'(x~) i(2)) is the projection of the first vector on the hyperplane. See Figure 4.5. The length squared of the projection vector is N
(14)
.
'\' [,,( (2) -(2))]2 -13 _ "A ' - , A-I L.. 13 xa - x 2213 - all) 22 all)' a=l
and the length squared of the first vector is L~';'I(Xla -X I )2 = all. Thus R is the cosine of the angle between the first vector and its projection.
148
SAMPLE CORRELATION COEFFICIENTS
N Figure 4.5
In Section 3.2 we saw that the simple correlation coefficient is the cosine of the angle between the two vectors involved (in the plane orthogonal to the equiangular line). The property of R that it is the maximum correlation between X 1a and linear combinations of the components of X!,2) .:orresponds to the geometric property that R is the cosine of the smallest angle between the vector with components X la -XI and a vector in the hyperplane spanned by the other p - 1 vectors. The geometric interpretations are in terms of the vectors in the (N - 1)dimensional hyperplane orthogonal to the equiangular line. It was shown in Section 3.3 that the vector (Xii - Xj, ... , XiN - x) in this hyperplane can be designated as (Zjl, ... ,Zj,N_I)' where the Zja are the coordinates referred to an (N - I)-dimensional coordinate system in the hyperplane. It was shown that the new coordinates are obtained from the old by the transformation Zja='[.~=lba/3xj/3' a=I, ... ,N, where B=(ba /3) is an orthogonal matrix with last row 0/ IN, ... ,1/ IFh Then N
(15)
a jj =
L a=J
N-l
(Xja-i;)(Xja-Xj)
=
L
ZjaZja'
a=J
It will be convenient to refer to the multiple correlatior. defined in terms of as the multiple correlation without subtracting the means. The popUlation multiple correlation R is essentially the only function of the parameters JL and l: that is invariant under changes of location, changes of scale of XI' and nonsingular linear transformations of X(2), that is, transformations xt = cX I + d, X(2)* = CX(2) + d. Similarly, the sample multiple correlation coefficient R is essentially the only function of x and t, the
Zja
4.4
149
THE MULTIPLE CORRELATION COEFFICIENT
sufficient set of statistics for ..... and ~, that is invariant under these transformations. Just as the simple correlation r is a measure of association between two scalar variables in a sample, the multiple correlation R is a measure of association between a scalar variable and a vector variable in a sample. 4.4.2. Distribution of the Sample Multiple Correlation Coefficient When the Population Multipll~ Correlation Coefficient Is Zero From (5) we have (16) then
(17)
all -a;I)A
I
22 a(l)
all
= a ll ' 2 all'
and (18) For q = 1, Corollary 4.3.2 states that when (:l = 0, that is, when R = O. a l12 is distributed as I:~:fVa2 and a(I)A 22I a(l) is distributed as I:~':~_P+I Va2• where VI"'" VN _ I are independent, each with distribution NCO'O"II)' Then I a ll ' 2 /0"1l.2 and a(I)A 22 a(I)/0"1l'2 are distributed independently as x2-variables with N - P and p - 1 degrees of freedom, respectively. Thus
(19)
has the F-distribution with p - 1 and N - P degrees of freedom. The density of F is
(20)
r[l(N-I)] (-I'!(P-I) 2 P ,11'-1)-1 r[Hp-l)]r[HN--p)] N-p) fI
(
1+ p
-I
N-pf)
_1"_1)
150
SAMPLE CORRELATION COEFFICIENTS
Thus the density of
p-1
N-F P p-I,N-p
R=
(21)
p-1
1+ N_pFp-I,N-P is (22)
2
r[!.(N-1)] '(N ) 2_ RP-2(1-R 2 f - P - I r[t(p - l)]r[HN - p)] ,
05,R5,1.
Theorem 4.4.2. Let R be the sample multiple correlation coefficient [defined by (5)] between XI and X(2), = (X 2" .. , Xp) based on a sample of N from N(Il, :I.). If Ii .= 0 [that is, if (0'12"'" O'IP)' = 0 = 13], then [R 2/O - R 2)]. [(N - p )/(p - 1)] is distributed as F with p - 1 and N - P degrees of freedom.
It should be noticed that p - 1 is the number of components of X(2) and that N _. P = N - (p - 1) - 1. If the multiple correlation is between a component Xi and q other components, the numbers are q and N - q - 1. It might be observed that R2 /(1- R2) is the quantity that arises in regression (or least squares) theory for testing the hypothesis that the regression of XI on X 2" .. , Xp is zero. If R"" 0, the distribution of R is much more difficult to derive. This distribution will be obtained in Section 4.4.3. Now let us consider the statistical problem of testing the hypothesis H: Ii = 0 on the basis of a sample of N from N("., l:). [Ii is the population multiple correlation hetween XI and (X 2 , •.. , X p)'] Since R ~ 0, the alternatives considered are Ii> O. Let us derive the likelihood ratio test of this hypothesis. The likelihood function is (23)
L (".*" . I * ) -_
1 L N (x a - " .* ) ,l: *-1 ( x" -".* ) ] . , 1N ,exp [ - -2 (27T)2P \:I.*\ ,N a=1
The observations are given; L is a function of the indeterminates ".*, l:*. Let (tJ be the region in the parameter space specified by the null hypothesis. The likelihood ratio criterion is
n
(24)
4.4 THE MULTIPLE CORRELATION COEFFICIENT
151
Here a is the space of JL*, 1* positive definite, and w is the region in this space where R'= ,; u(I)I;:;}u(1) /,[ii";; = 0, that is, where u(I)1 2iu(1) = O. Because 1221 is positive definite, this condition is equivalent to U(I) = O. The maximum of L(JL*, 1 *) over n occurs at JL* = fa. = i and 1 * = i = (1/N)A =(1/N)I:~_I(Xa-i)(xa-i)' and is
(25)
In w the likelihood function is
The first factor is maximized at JLi = ill =x I and uti = uli = (1/N)au, and the second factor is maximized at JL(2)* = fa.(2) = i(2) and 1;2 = i22 = (l/N)A 22 • The value of the maximized function is
(27) Thus the likelihood ratio criterion is [see (6)]
(28) The likelihood ratio test consists of the critical region A < Au, where Ao is chosen so the probability of this inequality when R = 0 is the significance level a. An equivalent test is
(29) Since [R2 /(1- R2)][(N - p)/(p -1)] is a monotonic function of R, an equivalent test involves this ratio being· larger than a constant. When R = 0, this ratio has an Fp_1• N_p·distribution. Hence, the critical region is
(30)
R2 N-p l-R2' p-l >Fp_l.N_p(a),
where Fp _ l • N_/a) is the (upper) significance point corresponding to the a significance level.
152
SAMPLE CO RRELATION COEFFICIENTS
Theorem 4.4.3. Given a sample x I' ... , X N from N( fl., l:), the likelihood ratio test at significance level ex for the hypothesis R = 0, where R is the population multiple correlation c~efficient between XI and (X2 , ..• , X p), is given by (30), where R is the sample multiple correlation coefficient defined by (5). As an example consider the data given at the end of Section 4.3.1. The samille multiple correlation coefficient is found from
r "--;-1-l-rI--''31
(31) 1- R2
=
1
32
.... 2J
r32
1
I
1.00 0.80 - 0040
0.80 -0040 I 1.00 - 0.56 - 0.56 1.00 = 0.357. 1.00 - 0. 56 1 1.00 1 -0.56
Thus R is 0.802. If we wish to test the hypothesis at the 0.01 level that hay yield is independent of spring rainfall and temperature, we compare the observed [R2 /(1- R 2)][(20 - 3)/(3 - 1)] = 15.3 with F2 17(0.01) = 6.11 and find the result significant; that is, we reject the null hyp~thesis. The test of independence between XI and (X2 , ••• , Xp) =X(2), is equivalent to the test that if the regression of XI on X(2) (that is, the conditional . X 2 -x X p -- : " ).IS ILl + ... Il /( (2) (2» expected vaIue 0 f X I gIVen x fl., t he 2 ,···, vector of regression coefficients is O. Here 13 = A2"2Ia(l) is the usual least squares estimate of 13 with expected value 13 and covariance matrix 0'1l.2A2"l (when the X~2) are fixed), and all.z/(N - p) is the usual estimate of 0'11.2' Thus [see (18)] (32) is the usual F-statistic for testing the hypothesis that the regression of XI on is O. In this book we are primarily interested in the multiple correlation coefficient as a measure of association between one variable and a vector of variables when both are random. We shall not treat problems of univariate regression. In Chapter 8 we study regression when the dependent variable is a vector.
X 2 , ••• , xp
Adjusted Multiple Correlation Coefficient The expression (17) is the ratio of a U ' 2 ' the sum of squared deviations from the fitted regression, to all. the sum of squared deviations around the mean. To obtain unbiased estimators of 0'11 when 13 = 0 we would divide these quantities by their numbers of degrees of freedom, N - P and N - 1,
4.4
THE MULTIPLE CORRELATION COEFFICIENT
153
respectively. Accordingly we can define an adjusted multiple con'elation coefficient R* by
(33) which is equivalent to
(34) This quantity is smaller than R2 (unless p = 1 or R2 = 1). A possible merit to it is that it takes account of p; the idea is that the larger p is relative to N, the greater the tendency of R" to be large by chance. 4.4.3. Distribution of the Sample Multiple Correlation Coefficient When the Population Multiple Correlation Coefficient Is Not Zero
In this subsection we shall find the distribution of R when the null hypothesis Ii = 0 is not true. We shall find that the distribution depends only on the population multiple correlation coefficient R. First let us consider the conditional distribution of R 2 /O - R2) = a(I)A2'ia(l)/all'2 given Z~2) = z;l, a = 1, ... , n. Under these conditions ZII"'" Zln are-independently distributed, Zla according to N(!3'z~2), 0'11-2)' where 13 = :I.2'21 U(l) and 0'11.2 = 0'11 - U(I):I.~2IU(I)' The conditions are those of Theorem 4.3.3 with Ya=Zla' r=!3', wa=z~2), r=p-1, $=0'11_2, m = n. Then a ll -2 = all - a~l)A~21a(l) corresponds to L::'_I 1';, 1';: - GHG'. and a n ' 2 /0'1l-2 has a X2-distribution with n - (p - 1) degrees of freedom. a(I)A2'21a(l) = (A2'21a(I»)' A22(A2'21a(1) corresponds to GHG' and is distributed as LaUa2, a = n - (p - J) + 1, ... , n, where Var(U,,) = 0'11-2 and
(35) where FHF' =1 [H=F-1(F,)-I]. Then a(l)A2'21a(l/0'1l_Z is distributed as La(Ua/ where Var(Uj = 1 and
;;::;y,
(36)
ru::;)
154
SAMPLE CORRELATION COEFFICIENTS
p - 1 degrees of freedom and noncentrality parameter WA22I3/ulJ-2' (See Theorem 5.4.1.) We are led to the following theorem: Theorem 4.4.4. Let R be the sample multiple correlation coefficient between and X(Z)' = U:2 , •. _, Xp) based on N observations (X lJ , X~2»), ... , (XlIV' x~»). The conditional distribution of [R 2/(1 - R2)][N - p)/(p - 1)] given X~2) fixed is lIoncentral F with p - 1 and N - p degrees of freedom and noncentrality parameter WAzZI3/UlJ-2' X(I)
The conditional density (from Theorem 5.4.1) of F p)/(p - 1)] is
= [R 2/(1- R2)][(N-
(p - l)exp[ - tWA2213/ull-2] (37)
(N-p)r[-}(N-p)]
oc
I WA z2 13)U[(P-l)f]1(P-Il+U-1 -Nrh-(N-l)+a] ( -2 U Il -2 p
and the conditional density of W = R Z is (df = [(N - p)/(p - 1)](1 w)-2 dw)
(38)
cxp[ - ~WA2213/ U1l2] (1 _ w) ~(IV-p)-l r[-}(N - p)]
To obtain the unconditional density we need to multiply (38) by the density of Z(2), ... , Z~2) to obtain the joint density of W and Z~2), ... , Z~2) and then integrate with respect to the latter set to ohtain the marginal density of W. We have (39)
l3'A2213 UIl-2
=
WL~~lz~2)z~2)'13
Ull-2
1 !
4.4
155
THE MULTIPLE CORRELATION COEFFICIENT
Since the distribution of Z~) is N(O, l:22)' the distribution of WZ~) / VCTlI.2 is normal with mean zero and variance
(40)
c( WZ~) ).2
CWZ~2)Z~2)/(l
Vl:1I·2
CTll •2
=
Wl:dl Wl:22(l/CTll W'Idl = 1 - W'I 22 (l/CT II
CT II -
jp I-jP· Thus (WA 22 (l/CTll.2)/[}F/(l-lF)] has a X2-distribution with n degrees of freedom. Let R2 /(1- R2) = cpo Then WA 22 (l/ CTI I .2 = CPX;. We compute
cpa. r(~nl+a)fOO 1 utn+a-le-t"du (l+cp)f n+a r("2n) 0 2,n+ar(~n+a) I
cpa r(!Il+a) (1 + cp)tn+a r(~n) Applying this result to (38), we obtain as the density of R2
(1- R 2 )t
r[!(n-p+l)]r(!n)
00
IL~O
(IF((R 2 )t
Fisher (I928) found this distribution. It can also be written
(43)
where F is the hypergeometric function defined in (41) of Section 4.2.
156
SAMPLE CORRELATION COEFFICIENTS
Another form of the density can be obtained when n - p + 1 is even. We have (44)
The density is therefore (1_iF)t n(R 2)t(P-3l(1_R 2)t(n- p-I)
(45)
r[!(n-p+ 1)]
Theorem 4.4.5. The density of the square of the multiple correiation coefficient, R2, between XI and X 2, ... , Xp based on a sample of N = n + 1 is given by (42) or (43) [or (45) in the case of n - p + 1 even], where iF is the corresponding population multiple correlation coefficient. The moments of Rare
.t
(1- R2) t(n-p+1 )-1 (R2) t(p+h-I )+1-'-1 d( R2)
o
(1_R"2)tn =
f(~n)
00
Eo
(R"2(r2(~n + JL)r[Hp +h -1) + JL] JLlr[!(p-1) +JL]r[!(n+h) +JL]
The sample multiple correlation tends to overestimate the population multiple correlation. The sample multiple correlation is the maximum sample correlation between XI and linear combinations of X(2) and hence is greater
157
4.4 THE MULTIPLE CORRELATION COEFFICIENT
than the sample correlation between x 1 and 13' x(2); however, the latter is the simple sample correlation corresponding to the simple population correlation between XI and Il'x(2), which is R, the population multiple correlation. Suppose Rl is the multiple correlation in the first of two samples and ~I is the estimate of 13; then the simple correlation between x 1 and ~'l X(2) in the second sample will tend to be less than R I and in particular will be less than R 2 , the multiple ccrrelation in the second sample. This has been called "the shrinkage of the multiple correlation." Kramer (1963) and Lee (1972) havc given tables of the upper significance points of R. Gajjar (1967), Gurland (1968), Gurland and Milton (970), Khatri (1966), and Lee (191Th) have suggested approximations to the distributions of R2 /(1- R2) and obtained large-sample results.
4.4.4. Some Optimal Properties of the Multiple Correlation Test Theorem 4.4.6. Given the observations Xl' ... , X N from N(IJ., l:), of all tests of R = at a given significance level based on i and A = [.~_l(Xa - iXxa - i)' that are invariant with respect to transformations
°
(47)
any critical rejection region given by R greater than a constant is unifonll/y most powerful. Proof The multiple correlation coefficient R is invariant under the transformation, and any function of the sufficient statistics that is invariant is a flmction of R. (See Problem 4.34.) Therefore, any invariant test must be based on R. The Neyman-Pearson fundamental lemma applied to testing the null hypothesis R = 0 against a specific alternative R = Ro > tells us the most powerful test at a given level of significance is based on the ratio of the density of R for R = Ro, which is (42) times 2R [because (42) is the density of R2], to the density for R = 0, which is (22). The ratio is a positive constant times
°
(48) Since (48) is an increasing function of R for R :;:: 0, the set of R for which (48) is greater than a constant is an interval of R greater than a constant.
.
158
SAMPLE CORRELAnON COEFFICIENTS
Theorem 4.4.7. On the basis of observations Xl"'" X N from N("., l:), of all tests of R = 0 at a given significance level with power depending only on R, the test with critical region given by R greater than a constant is unifOlmly most powerful. Theorem 4.4.7 follows from Theorem 4.4.6 in the same way that Theorem 5.6.4 follows from Theorem 5.6.1.
4.5. ELLIPTICALLY CONTOURED DISTRIBUTIONS ~.5.1.
Observations Elliptically Contoured
Suppose Xl •... ' with density
XN
are N independent observations on a random p-vector X
(1) The sample covariance matrix S is an unbiased estimator of the covariance matrix ~=[/'R2IplA, where R 2 =(X-v)'A- I (X-v) and IiR 2 <00. An estimator of flij = (fiji"; O'iiO'jj = A;/ ..; Aii\j is r ij = Si/ VS;S;j-' i, j = 1. .... p. The small-sample distribution of rij is in general difficult to obtain, but the asymptotic distribution can be obtained from the limiting normal distribution of IN(S -:1) given in (13) of Section 3.6. First we prove a general theorem on asymptotic distributions of functions of the sample covariance matrix S using Theorems 4.2.3 and 3.6.5. Define s = vec S,
(2)
0'
= vec l:.
Theorem 4.5.1. Let J(s) be a vector-valued function such that each component of J(s) has a nonzero differential at s = 0'. Suppose S is the covariance of a sample from (1) such that Ii R4 < 00. Then
(3)
IN[J(s) - f( 0')] = a~~~) IN (s - 0') + op(l)
~ N{O, a~~~) [2(1 + K)(l: ®l:) + KO'O"]( a~~~))} Corollary 4.5.1. (4)
If J(cs)=f(s)
for all c > 0 and all positive definite S and the conditions of Theorem 4.5.1 hold, then
4.5
159
ELLIPTICALLY CONTOURED DISTRIBUTIONS
Proof. From (4) we deduce
0= af(cs) = af(cs) a(CS) = af(cs)S ac as' ac as"
(6) That is,
a~~~)
(7)
(J
= O.
•
The conclusion of Corollary 4.5.1 can be framed as
The limiting normal distribution in (8) holds in particular when the sample is drawn from the normal distribution. The corollary holds true if K is replaced by a consistent estimator R. For example, a consistent estimator of 1 + R given by (16) of Section 3.6 is
1 + R=
(9)
f
[(xa _x)'S-i(xa _x)]2/[Np(p + 2)].
a=1
A sample correlation such as f(s) = rjj = Sjj/ ";SjjSjj or a set of such correlations is a function 01 S that is invariant under scale transformations; that is, it satisfies (4). Corollary 4.5.2.
Under the conditions of Theorem 4.5.1,
(10)
As in the :::ase of the observations normally distributed,
(11)
IN ( 1 1 + rjj 1 1 + pjj) d V"i'+"R 210g1-rjj -2: log l_pjj --+N(O,I).
Of course, any improvement of (11) over (10) depends on the distribution samples. Partial correlations such as rjj .q + 1, ... ,p' i,j=I, ... ,q, are also invariant functions of S. Corollary 4.5.3. ( 12)
Under the conditions of Theorem 4.5.1,
V1 NA(r.. + K
lj.q+i, .... p
-p.. lj.q+i, ... ,p
)~N(O,I).
160
SAMPLE CORRELATION COEFFICIENTS
Now let us consider the asymptotic distribution of R2, the square of the multiple correlation, when iF, the square of the population multiple correlation, is O. We use the notation of Section 4.4. iF = 0 is equvialent to (1(1) = O. Since the sample and population mUltiple correlation coefficients between Xl and X(2) = (X 2 , ••• , XpY are invariant with respect to linear transformations (47) of Section 4.4, for purposes of studying the distribution of R2 we can assume IJ. = 0 and I. = Ip. In that case S11 41, S(I) 4 0, and S22 4Ip _ 1 ' Furthermore, for k, i 1 and j = I = 1, Lemma 3.6.1 gives
"*
(13)
tS'S(1)S(I) =
Theorem 4.5.2.
(* + N)Ip_I'
Under the conditions of Theorem 4.5.1
(14)
Corollary 4.5.4.
Under the conditions of Theorem 4.5.1
(15)
4.5.2. Elliptically Contoured Matrix Distributions
Now let us turn to the model (16)
based on the vector spherical model g(tr Y'Y). The unbiased estimators of v and !'=(tS'R 2 /p)A are i=(I/N)X'E N and S=O/n)A,where A= (X ~ Enx'HX - ENX'). Since
(17)
(X - ENV')'(X - ENV') =A +N(i - v)(x - v)',
A and x are a complete set of sufficient statistics. The maximum likelihood estimators of v and A are 11 = and A = (p /wg)A. The maximum likelihood estimator of Pij = A;J A;/ljj =
x
CT;/VCT;;~j is
Pij = a;/Va;;a jj =S;JVSiiSjj
V
(Theorem 3.6.4). The sample correlation rij is a function f(X) that satisfies the conditions (45) and (46) of Theorem 3.6.5 and hence has the same distribution for an arbitrary density g[trOl as for the normal density g[trOl = const e- ttr('). Similarly, a partial correlation r;j.q+l •... ,p and a multiple correlation R2 satisfy the conditions, and the conclusion holds.
4.5
Hil
ELLIPTICALLY CONTOURED DISTRIBUTIONS
Theorem 4.5.3. When X has the vector elliptical density (6), the distributions of rii , rij .q + I' and R2 are the distributions derived for normally discribllted observations. It follows from Theorem 4.5.3 that the asymptotic distributions of rij . r;j.q+l, ... ,f!' and R2 are the same as for sampling from normal distributions. The class of left spherical matrices Y with densities is the class of g(Y'Y). Let X= YC' + ENV', where C' A -IC = I, that is, A = CC'. Then X has the density
(18) We now find a stochastic representation of the matrix Y. Let V = (VI"'" vp ), where Vi is an N-component vector. 1, ... , p. Define recursively WI = VI'
Lemma 4.5.1. i
=
(19) Let
uj = w;/Ilwili.
(20)
i = 2 ..... p.
Then
Ilu;11 =
1, i = 1, ... , p, and u;u j = 0, i
"* j.
FlI11her,
V=VT',
where V = (u" ... , up); tii = Ilwill, i = 1. .... p; t ij 1, ... , i-I, i = 1, ... , p; and t ij = 0, i <j.
=
l·;w/llwjll = l';/I!,
j
=
The proof of the lemma is given in the first part of Section 7.2 and as the Gram-Schmidt orthogonalization in the Appendix (Section A.S.l). This lemma generalizes the construction in Section 3.2; see Figure 3.1. See also Figure 7.1. Note that'T is lower triangular, V'V = Ip, and V'V= TT'. The last equation, t ij ~ 0, i = 1, ... , p, and lij = 0, i <j, can be solved uniquely for T. Thus T is a function of V'V (and the restrictions). Let Y (Nxp) have the density g(Y'Y). and let 0,\ he an orthogonal Nx N matrix. Then Y* = ON Y has the density g(Y* 'Y*). Hence Y* = ON Y 4. Y. Let Y* = V* T*', where t~ > 0, i = 1, ... , p, and t'!j = 0, i <j. From Y* 'Y* = Y'Y it follows that T* T*' = TT' and hence T* = T, Y* = V* T, and U* = ON V 4. V. Let the space of V (N x p) such that V'V = I p be denoted O(Nxp).
Definition 4.5.1. If V (N x p) satisfies V'V = I p and 0.\. V 4 V for all orthogonal ON' then U is uniformly distrihuted on O( N x p).
162
SAMPLE CORRELATION COEFFICIENTS
The space of V satisfying V'V = Ip is known as a Steifel manifold. The probability measure of Definition 4.5.1 is known as the Haar invariant distribulioll. The property 0NV g, V for all orthogonal ON defines the (normalized) measure uniquely [Halmos (1956)]. Theorem 4.5.4. If Y (NXp) has the density g(Y'Y), then V defined by Y = UT', V'V = I p ' I" > 0, i = 1, ... , p, and t'i = 0, i <j,. is uniformly distributed on O(N xp). The proof of Corollary 7.2.1 shows that for arbitrary Tis
gO
the density of
p
(21 )
n {C[ HN + 1 - i)] t,~-')g(tr IT'),
'~l
where CO is ddined in (8) of Section 2.7. The stochastic representation of Y (N X p) with density geY/Y) is (22)
Y=VT',
where V eNXp) is uniformly distributed on O(Nxp) and T is lower triangular with positive diagonal elements and has density (21). Theorem 4.5.5. that (23)
Let J(X) be a vector-valued function of X (N X p) such
J(X + EN V ')
= J(X)
for all v and (24)
J(XG/) = J(X)
for all G (p X p). Then the distn'bution of J(X) where X has an arbitrary density (18) is the same as the distribution of J(X) where X has the normal density (18). Proof From (23) we find that J(X) = J(YC'), and from (24) we find J(YC') =J(UT'C') =J(U), which is the same for arbitrary and normal densi.. ties (18). Corollary 4.5.5. Let JeX) be a vector-valued function of X (N X p) with the density (18), where v = O. Suppose (24) hold~ for all G (p X p). Then the distribution ofJ(X) for an arbitrary density (18) is the same as the distribution of j(X) when X has the normal density (18).
163
PROBLEMS
The condition (24) of Corollary 4.5.5 is that f(X) is invariant with respect to linear transformations X --> XG. The density (18) can be written as
ICI-1g{C- 1[~+N(i - v)(x - v),](C') -I},
(25)
which shows that A and i are a complete set of sufficient statistics for A=CC' and v.
PROBLEMS 4.1. (Sec. 4.2.1) Sketch
for (a) N = 3, (b) N = 4, (c) N
=
5, and (d) N = 10.
4.2. (Sec. 4.2.1)
Using the data of Problem 3.1, test the hypothesis that Xl and X 2 are independent against all alternatives of dependence at significance level 0.01.
4.3. (Sec. 4.2.1)
Suppose a sample correlation of 0.65 is observed in a sample of 10. Test the hypothesis of independence against the alternatives of positive correlation at significance level 0.05.
4.4. (Sec. 4.2.2) Suppose a sample correlation of 0.65 is observed in a sample of 20. Test the hypothesis that the population correlation is 0.4 against the alternatives that the population correlation is greater than 0.4 at significance level 0.05. 4.5. (Sec. 4.2.0 Find the significance points for testing p = 0 at the 0.01 level with N = 15 observations against alternatives (a) p *- 0, (b) p> 0, and (c) p < O. 4.6. (Sec. 4.2.2) Find significance points for testing p = 0.6 at the 0.01 level with N = 20 observations against alternatives (a) p *- 0.6, (b) p> 0.6, and (c) p < 0.6. 4.7. (Sec. 4.2.2) Tablulate the power function at p = -1(0.2)1 for the tests in Problf!m 4.5. Sketch the graph of each power function. 4.8. (Sec. 4.2.2) Tablulate the power function at p = -1(0.2)1 for the tests in Problem 4.6. Sketch the graph of each power function. 4.9. (Sec. 4.2.2)
Using the data of Problem 3.1, find a (two-sided) confidence interval for P12 with confidence coefficient 0.99.
4.10. (Sec. 4.2:2) Suppose N = 10, , = 0.795. Find a one-sided confidence interval for p [of the form ('0,1)] with confidence coefficient 0.95.
164
SAMPLE CORRELATION COEFFICIENTS
4.11. (Sec. 4.2.3) Use Fisher's Z to test the hypothesis P = 0.7 against alternatives O.i at the 0.05 level with' r = 0.5 and N = 50.
{' *"
4.12. (Sec. 4.2.3) Use Fisher's z to test the hypothesis PI = P2 against the alternatives PI P2 at the 0.01 level with r l = 0.5, NI = 40, r2 = 0.6, N z = 40.
*"
4.13. (Sec.4.2.3) Use Fisher's z to estimate P based on sample correlations of -0.7 (N = 30) and of - 0.6 (N = 40). 4.14. (Sec. 4.2.3) Use Fisher's z to obtain a confidence interval for p with confidence 0.95 based on a sample correlation of 0.65 and a sample size of 25. 4.15. (Sec. 4.2.2). Prove that when N = 2 and P = 0, Pr{r = l} = Pr{r = -l} =
!.
4.16. (Sec. 4.2) Let kN(r, p) be the density of the sample corrclation coefficient r for a given value of P and N. Prove that r has a monotone likelihood ratio; that is, show that if PI > P2' then kN(r, PI)/kN(r, P2) is monotonically increasing in r. [Hint: Using (40), prove that if
F[U;n+U(1+pr)]=
L
ca (1+pr)a=g(r,p)
a=O
has a monotone ratio, then kN(r, p) does. Show
if (B 2/BpBr)Iogg(r, p) > 0, then g(r, p) has a monotone ratio. Show the numerator of the above expression is positive by showing that for each IX the sum on f3 is positive; use the fact that c a + 1 < !c a .] 4.17. (Sec.4.2) Show that of all tests of Po against a specific PI (> Po) based on r, the procedures for which r> c implies rejection are the best. [Hint: This follows from Problem 4.16.] 4.18. (Sec. 4.2) Show that of all tests of P = Po against p> Po based on r, a procedure for which r> c implies rejection is uniformly most powerful. 4.19. (Sec. 4.2) Prove r has a monotone likelihood ratio for r > 0, P > 0 by proving her) = kN(r, PI)/kN(r, P2) is monotonically increasing for PI > P2' Here her) is a constant times O:~:;:~Oca prra)/(r::~Oca pfr a ). In the numerator of h'(r), show that the coefficient of r {3 is positive. ' 4.20. (Sec. 4.2) Prove that if l: is diagonal, then the sets rij and aii are independently distributed. [Hint: Use the facts that rij is invariant under scale transformations and that the density of the observations depends only on the a ii .]
165
PROBLEMS
4.21. (Sec. 4.2.1) Prove that if p = 0
$r
2
m
=
r[HN-1)]r(m+t) j;r[t(N-l) +m]
---';=-'--=-:-----'-"--'-_-':-'-
4.22. (Sec. 4.2.2) Prove Up) and f2( p) are monotonically increasing functions of p. 4.23. (Sec. 4.2.2) Prove that the density of the sample correlation r [given by (38)] is
[Hint: Expand (1 - prx)-n in a power series, integrate, and use the duplication
formula for the gamma ft:,nction.] 4.24. (Sec. 4.2) Prove that (39) is the density of r. [Hint: From Problem 2.12 show 00
1o f
00
e -'{Y'-2Xyz+z'ld " , y dz
l
=
-ie )
cos -x ~
.
Then argue
(yz ) 11 o 00
00
0
n-I
dn - I cos e_l(y'-2X}'z+z')d ' y dz = n I
dx
-
I(
-x ) "
VI-x'
Finally show that the integral of(31) with respect to a II (= y 2 ) and a 22 (= z') is (39).] 4.25. (Sec. 4.2)
Prove that (40) is the density d r. [Hint: In (31) let ~ v < 00) and r ( -1 ~ r
a 22 = ue u ; show that the density of v (0
Use the expansion
;r(j+~)j t.-
j-O
Show that the integral is (40).]
r( '21)'1 Y , J.
all
=
~ 1)
ue- L' and
is
166
SAMPLE CORRELATION COEFFICIENTS
4.26. (Sec. 4.2)
Prove for integer h
f!r~"""'1 =
(l_p~)J:n
E(2p)2.8+
1
r2[Hn+l)+J3]r(h+/3+~)
;;;:rOIl) .8-11 (2J3+1)!
Sr~"=
(l_p2)ln
E
(2p)"
r(tn+h+J3+1)
r20n+J3)r(h+J3+i)
;;;:rOn) .8-0 (2J3)!
rOn+h+J3)
4.27. (Sec. 4.2) The I-distribution. Prove that if X and Yare independently distributed, X having the distnbution N(O,1) and Y having the X2-distribution with m degrees of freedom, then W = XI JY1m has the density r[Hm
+ I)] (I +
.;m ;;;:r( ~m)
:':')-1
1 ",+1)
m
[Hint: In the joint density of X and Y, let x = tw1m- J: and integrate out w.]
4.28. (Sec. 4.2)
Prove
[Him: Use Problem 4.26 and the duplication formula for the gamma function.]
4.29. (Sec. 4.2) Show that In ( 'ij - Pij)' (i, j) = (1,2), (1, 3), (2, 3), have a joint limiting distribution with variances (1 - Pi~)2 ann covaliances of rij and rik' j '" k being i(2pjk - PijPjk X1 - Pi~ - p,i - PP + pli,o 4.30. (Sec. 4.3.2) Find a confidence interval for rUe = 0.097 and N = 20.
P13.2
with confidence 0.95 based on
4.31. (Sec. 4.3.2) Use Fisher's = to test the hypothesis P12'34 = 0 against alternatives Plc. l • '" 0 at significance level 0.01 with r 12.34 = 0.14 and N = 40. ·U2. (Sl·C. 4.3) Show that the inequality rf~.3 s I is the same as the inequality Irijl ~ 0, where Irijl denotes the determinant of the 3 X 3 correlation matrix. 4.33. (See. 4.3) II/variance of Ihe sample partial correiatioll coefficient. Prove that r lc .3 ..... p is invariant under the transformations x;a = aix ia + b;x~) + c i' a i > 0, t' = 1, 2, x~')' = Cx~') + b" a = 1, ... , N, where x~') = (X3a,"" xpa )', and that any function of i and l: that is invariant under these transformations is a function of r 12.3.... p. 4.34. (Sec. 4.4) Invariance of the sample multiple correlation coefficient. Prove that R is a fUllction of the sufficient statistics i and S that is invariant under changes of location and scale of x I a and nonsingular linear transformations of x~2) (that is. xi" = ex I" + d, x~~)* = CX~2) + d, a = 1, ... , N) and that every function of i and S that is invariant is a function of R.
167
PROBLEMS
Prove that conditional on ZI" = ZI,,' a = 1, ... , n, R Z/0 - RZ) is distributed like T 2 /(N* - 1), where T Z = N* i' S-I i based on N* = n observations on a vector X with p* = p - 1 components, with mean vector (c / 0"11)0"(1) (nc z = EZT,,) and covariance matrix l:ZZ'1 = l:zz -1l/0"1l)0"(1)0"(1)' [Hint: The conditional distribution of Z~Z) given ZI" =ZI" is N[O/O"ll)O"(I)ZI,,' l:22.d. There is an n X n orthogonal matrix B which carries (z II" .. , Z In) into (c, ... , c) and (Zi!"'" Zili) into CY;I"'" l'ill' i = 2, ... , p. Let the new X~ be (YZ" , ••• , Yp,,}.l
4.35. (Sec. 4.4)
4.36. (Sec. 4.4)
Prove that the noncentrality parameter in the distribution in Problem 4.35 is (all/O"II)lF/(l-IF). Find the distribution of R Z/0 - RZ) by multiplying the density of Problem 4.35 by the dcnsity of all and intcgrating with respect to all'
4.37. (Sec. 4.4)
4.38. (Sec. 4.4)
Show that thl: density of rZ derived from (38) of Section 4.2 is identical with (42) in Section 4.4 for p = 2. [Hint: Use the duplication formula for the gamma function.l
4.39. (Sec. 4.4) Prove that (30) is the uniformly most powerful test of on r. [Hint: Use the Neyman-Pearson fundamentallemma.l 4.40. (Sec. 4.4)
Prove that (47) is the unique unbiased estimator of
R2
R=
0 based
based on R2.
4.41. The estimates of .... and l: in Problem 3.1 are
i = ( 185.72
S
=
151.12
95.2933 .52:~6.8?. ( 69.6617 46.1117
183.84
149.24)',
52.8683: 69.6617 46.1117] ?~.??~~ ; ..5~ :3.1.1? .. ~5:?~3.3 .. 51.3117' 100.8067 56.5400 35.0533: 56.5400 45.0233
(a) Find the estimates of the parameters of the conditional distribution of (X3,X 4 ) given (xl,xz); that is, find SZISIII and S22'1 =S2Z -SZISI;ISIZ' (b) Find the partial correlation r~4'12' (e) Use Fisher's Z to find a confidence interval for P34'IZ with confidence 0.95. (d) Find the sample multiple correlation coefficients between x:, and (XI' xz) and between X4 and (XI' X2~' (e) Test the hypotheses that X3 is independent of (XI' x 2 ) and x 4 is inJependent of (XI' X2) at significance levels 0.05. 4.42. Let the components of X correspond to scores on tests in arithmetic speed (XI)' arithmetic power (X 2 ), memory for words (X3 ), memory for meaningful
symbols (X.1 ), and memory for meaningless symbols (X;). The observed correla-
168
SAMPLE CORRELATION COEFFICIENTS
tions in a sample of 140 are [Kelley (1928») 1.0000 0.4248 0.0420 0.0215 0.0573
0.4248 1.0000 0.1487 0.2489 0.2843
0.0420 0.1487 1.0000 0.6693 0.4662
0.0215 0.2489 0.6693 1.0000 0.6915
0.0573 0.2843 0.4662 0.6915 1.0000
(a) Find the partial correlation between X 4 and X s, holding X3 fixed. (b) Find the partial correlation between Xl and X 2 , holding X 3, X 4 , and Xs fixed. (c) Find the multiple correlation between Xl and the set X 3 , X. I , and Xs. (d) Test the hypothesis at the 1% significance level that arithmetic speed is independent of the three memory scores. 4.43. (Sec. 4.3)
Prove that if Pij-q+I, ... ,p=O, then ..;N-2-(p-q)rij.q+I ..... pl ';1 - r;}.q+ I , ... ,p is distributed according to the t-distribution withN - 2 - (p - q) degrees of freedom. Let X' = (Xl' X 2 , X(2)') have the distribution MII-,:n The conditional distribution of XI given X 2 = x 2 and X(2) = X(2) is
4.44. (Sec. 4.3)
where
The estimators of 1'2 and 'Yare defined by
Show C2 = a 12 .3, . substitute.) 4.45. (Sec. 4.3)
. , pla 22 . 3, ... ,p.
[Hint: Solve for c in terms of c2 and the a's, and
In the notation of Problem 4.44, prove
=
all ·3..... p
-
c?a22.3, .... p·
169
PROBLEMS
Hint: Use
4.46. (Sec. 4.3)
Prove that 1/a 22 .3..
.. 1'
is the element in the upper left-hand corner
of
4.47. (Sec. 4.3) PI2.3 ....• p
Using the results in Problems 4.43-4.46, prove that the test for 1'2 = O.
= 0 is equivalent to the usual I-test for
4.48. Missing observations. Let X = (Y' Z')', where Y has p components and Z has q components, be distributed according to N(Il-, l:), where
Let M observations be made on X, and N - M additional observations be made on Y. Find the maximum likelihood estimates of Il- and l:. [Anderson (1957).] [Hint: Express the likelihood function in terms of the marginal density of Yand the conditional density of Z given Y.] 4.49. Suppose X is distributed according to N(O, l:), where P
p2)
1
P
p
1
.
Show that on the basis of one observation, x' = (Xl' x 2 • X 3 ). we can obtain a confidence interval for p (with confidence coefficient 1 - a) by using as endpoints of the interval the solutions in I of
where xi(a) is the significance point of the x2-distribution with three degrees of freedom at significance level a.
CHAPTER 5
The Generalized T 2 -Statistic
5.1. INTRODUCTION One of the most important groups of problems in univariate statistics relates to the m.::an of a given distribution when the variance of the distribution is unknown. On the basis of a sampk one nlay wish to decide whether the mt:an is .::qual to a number specified in advance, or one may wish to give an interval within which the mean lies. The statistic usually used in univariate statistics is the difference between the mean of the sample i and the hypothl:tieal population m.::an j.L divided by the sample standard deviation s. if the distribution sampled is N( j.L, (T ~), then
has the well-known t-distribution with N - 1 degrees of freedo n, where N is the number of observations in the sample. On the basis of this fact, one can set up a test of the hypothesis j.L = J1.o, where J-Lo is specified, or one can set up a confidenre interval for the unknown parameter J-L. The multivariate analog of the square of t given in (1) is
( 2) where x is the m.::an vector of a sample of N, and S is the sample covariance matrix. It will be shown how this statistic can be used for testing hypotheses ahout the mean vector /J. of the popUlation and for obtaining confidence regions for the unknown /J.. The distribution of T2 will be obtained when /J. in (2) is the mean of the distribution sampled and when /J. is different from
All [mroc/lletioll to Multivariate Statistical Analysis. Third Edition. ISBN 0-471-36091-0
170
Copyright © 2003 John Wiley & Sons, Inc.
By T. W. Anderson
5.2
DERIVATION OF THE T 2 -STATISTIC AND ITS DISTRIBUTION
171
the population mean. Hotelling (1931) proposed the T 2 -statistic for two samples and derived the distribution when fL is the population mean. In Section 5.3 various uses of the T 2 -statistic are presented, including simultaneous confidence intervals for all linear combinations of the mean vector. A James-Stein estimator is given when l: is unknown. The power function 0: the T2-test is treated in Section 5.4, and the multivariate Behrens-Fisher problem in Section 5.5. In Section 5.6, optimum properties of the T 2-test are considered, with regard to both invariance and admissibility. Stein's criterion for admissibility in the general exponential fam]y is proved and applied. The last section is devoted to inference about the mean in elliptically contoured distributions.
5.2. DERIVATION OF THE GENERALIZED T 2 -STATISTIC AND ITS DISTRIBUTION 5.2.1. Derivation of the T 2-Statistic As a Function of the Likelihood Ratio Criterion Although the T 2-statistic has many uses, we shall begin our discussion by showing that the likelihood ratio test of the hypothesis H: fL = fLo on the basis of a sample from N(fL, l:) is based on the T 2 -statistic given in (2) of Section 5.1. Suppose we have N observations X l ' " ' ' X N (N > p). The likelihood. function is
The observations are given; L is a function of the indeterminates fL, l:. (We shall not distinguish in notation between the indeterminates and the parameters.) The likelihood ratio criterion is
(2)
that is, the numerator is the maximum of the likelihood function for fL, l: in the parameter space restricted by the null hypothesis (fL = fLo, l: positive definite), and the denominator is the maximum over the entire parameter space (l: positive definite). When the parameters are unrestricted, the maximum occurs when fL' l: are defined by the maximum likelihood estimators
172
THE GENERALIZED T 2 -STATISTIC
(Section 3.2) of JL and l:,
(3)
fl.n =x,
(4)
in=~ I: (xa-x)(xa-x)'.
N
a~l
When JL = JLo, the likelihood function is maximized at ~
(5)
1
l:,.= /Ii
N
I:
(x" - JLo)(x" - JLo)'
a-I
by Lemma 3.2.2. Furthermore, by Lemma 3.2.2
(6) (7) Thus the likelihood ratio criterion is
(8)
A=
l~nl~N = IL(Xa-X)(xa-x)'lt~ 1l:.,I;:N
IL(Xa-JLo)(Xa-JLo)'I;:N IAlfN
where N
(9)
A=
I:
(Xa -X)(Xa -x)'
=
(N -l)S.
a=1
Application of Corollary A.3.1 of the Appendix shows (10)
A2/N=
IAI IA + [IN ( x - JLo) 1[IN ( x - JLo) 1'I 1
1 + N(X - JLo)'A-I(X - JLo) 1
where
(11)
T2 = N(X - JLo) 'S-l (i - JLo) = (N - l)N( x - JLo)' A -l( x - JLo).
5.2
2 DERIVATION OF THE T -STATISTIC AND ITS DISTRIBUTION
. 173
The likelihuod ratio test is defined by the critical region (region of rejection)
(12) where Ao is chosen so that the probability of (12) when the nul1 hypothesis is true is equal to the significance level. If we take the ~Nth root of both sides of (12) and invert, subtract I, and mUltiply by N - I, we obtain (13) where
(14)
""0
Theorem 5.2.1. The likelihood ratio test of the hypothesis,... = for the distribution N(,..., l:) is given by (13), where T2 is defined by (11), i is the mean of a sample of N from N(,..., l:), S is the covan'ance matrix of the sample, and To" is chosen so that the probability of (13) under the null hypothesis is equal to the chosen significance level. The Student t-test has the property that when testing J.L = 0 it is invariant with respect to scale transformations. If the scalar random variahle X is distributed according to N( J.L, a 2), then X* = cX is distributed according to N(c J.L, c 2a 2), which is in the same class of distributions, and the hypothesis If X = 0 is equivalent to If X* = If cX = O. If the observations Xu are transformed similarly (x: = cxa)' then, for c > G, t* computed from x! is the same as t computed from Xa' Thus, whatever the unit of measurement the statistical result is the same. The generalized T2-test has a similar property. If the vector random variable X is distributed according to N(,..., l:), then X* = CX (for \ C\ '" 0) is distributed according to N(C,..., Cl: C'), which is in the same class of distributions. The hypothesis If X = 0 is equivalent to the hypothesis If X" = IfCX = O. If the observations Xa are transformed in the same way, x! = Cx a , then T*" c'Jmputed on the basis of x! is the same as T2 computed on the basis of xu' This fol1ows from the facts that i* = ex and A = CAC' and the following lemma: Lemma ,5.2.1. vector k,
(15)
For any p x p nonsingular matrices C alld H alld allY
k' H-1k = (Ck)'( CHC') -1 (Ck).
THE GENERAlIZED T 2 -STATISTIC
174 Proof The right-hand side of (IS) is (16)
(Ck)'( CHC') -I (Ck) = k'C'( C') -I H-1C- 1Ck =k'H-lk.
•
We shall show in Section 5.6 that of all tests invariant with respect to such transformations, (13) is the uniformly most powerful. We can give a geometric interpretation of the ~Nth root of the likelihood ratio criterion, (17)
A2/N
=
I [~~I(X" -i)(x" -i)'1 1[~~I(X,,-fLo)(X,,-fLo)'1 '
in terms of parallelotopes. (See Section 7.5.) In the p-dimensional representation the numerator of A21 N is the sum of squares of volumes of all parallelotopes with principal edges p vectors, each with one endpoint at i and the other at an x". The denominator is the sum of squares of volumes of all parallelotopes with principal edges p vectors, each with one endpoint at. fLo and the other at xo' If the sum of squared volumes involving vectors emanating from i, the "center" of the x,,, is much less than that involving vectors emanating from fLo, then we reject the hypothesis that fLo is the mean of the distribution. There is also an interpretation in the N-dimensional representation. Let Yi=(Xil"",X iN )' be the ith vector. Then N
1N.r.;= E
(18)
",~l
1 ",Xi"
vN
is the distance from the origin of the projection of Yi on the equiangular line (with direction cosines 1/ IN, ... , 1/ IN). The coordinates of the projection are (Xi"'" X). Then (Xii - Xi"'" XiN - X) is the projection of Yi on the plane through the origin perpendicular to the equiangular line. The numerator of AliII' is the square of the p-dimensional volume of the parallelotope with principal edges, the vectors (Xii - Xi' ... , Xi N - X). A point (X il !LOi"'" XiN - !Lo) is obtained from Yi by translation parallel to the equiangular line (by a distance IN !Lo). The denominator of A2/ N is the square of the volume of the parallelotope with principal edges these vectors. Then A2/ N is the ratio of these squared volumes. 5.2.2. The Distribution of T 2
In this subsection we will find the distribution of T2 under general conditions, including the case when the null hypothesis is not true. Let T2 = Y'S-J Y where Y is distributed according to N(v, I) and nS is distributed independently as [7,: I Z" with Z I" .. , Z" independent, each with distribution
Z:,
5.2
DERIVATION OF 'fHE
r 2-STATlSTIC AND
175
ITS DISTRIBUTION
N(O, l:). The T2 defined in Section 5.2.1 is a special case of this with Y= m(x - .... 0) and v = m( .... - .... 0) and n = N -1. Let D be a nonsingular matrix such that Dl:D' = I, and define
(19)
Y* =DY,
S* =DSD',
v* =Dv.
Then T2 = y* 'S* -I Y* (by Lemma 5.2.1), where Y* is distributed according to N( v* , I) and nS* is distributed independently as 1::~ IZ: Z: ' = 1::_ 1 DZa(DZ,.)' with the Z: = DZa independent, each with distribution N(O, I). We note v 'l: -I v = v* '(I)-I v* = v* 'v* by Lemma 5.2.1. Let the Lrst row of a p X P orthogonal matrix Q be defined by
y*
qlj=~'
(20)
i = 1, ... ,p;
this is permissible because 1:f_1 qlj = 1. The other p - 1 rows can be defined by some arbitrary rule (Lemma A,4.2 of the Appendix). Since Q depends on Y*, it is a random matrix, Now let U=QY* ,
(21)
B = QnS*Q'.
From the way Q was defined,
VI = 1:quY;* = ';y* 'y* , (22)
~ = 1:qjjY;* = ,;y* 'Y* 1:qjjqlj = 0.
j"" 1.
Then
(23)
~2
= U'B-IU= (V1,0, ... ,0)
=
Vl2 b ll ,
b ll b 21
b 12 b 22
b lp b 2p
bpi
b P2
b PP
VI
° °
where (b ji ) = B- 1 • By Theorem A,33 of the Appendix, 1jb ll = b ll b(I)B;2 Ib (l) = b ll . 2•. .• p' where
-
(24) and T2 In = V/ jb ll . 2.. '" P = Y* 'Y* jb ll . 2..... p. The conditional distribution of B given (~ is that of 1::_1VaV;, where conditionally the Va = QZ: are
176
THI: GENERALIZED T"-STATlSTIC
independent, each with distribution N(O, I). By Theorem 4.3.3 b ll . 2•...• p is conditionally distributed as L::(/-I)W,}, where conditionally the W" are independent, each with the distribution N(O,1); that is, b ll . 2..... p is conditionally distributed as X 2 with n - (p - 1) degrees of freedom. Since the conditional distribution of b ll . 2..... P does not depend on Q, it is unconditionally distributed as X z. The quantity Y* 'Y* has a noncentral XZ-distribution with p degrees of freedom and noncentrality parameter v*'v* =v'l;-lv. Then T 2 /n is distributed as the ratio of a noncentral X 2 and an independent X
2
•
Theorem 5.2.2. Let T2 = Y' S-I Y, where N(v,l;) and I1S is independently distributed independent, each with' distribution N(O, l;). distributed as a noncentral F with p and 11 noncentrality parameter v'l; -I v. If v = 0, the
Y is distributed according to as L';"'IZ"Z;, with Zp".,Z" Then (T 2/n)[n - p + 1) /p] is P + 1 degrees of freedom and distribution is central F.
We shall call this the TZ-distribution with n degrees of freedom. Corollary 5.2.1. Let XI"", x N be a sample from N(I'-, l;), and let T2 = N(i-l'-o)'S-I(i-l'-o)' The distribution of [T z/(N-1)][(N-p)/p] is noncentral F with p and N - p degrees of freedom and noncentrality parameter N(I'- - I'-o)'l; -1(1'- - 1'-0)' If I'- = 1'-0' then the F-distribution is central.
The above derivation of the TZ-distribution is due to Bowker (1960). The noncentral F-density and tables of the distribution are discussed in Section 5.4. For large samples the distrihution of T2 given hy Corollary 5.2.1 is approximately valid even if the parent distribution is not normal; in this sense the T2-test is a robust procedure. Theorem 5.2.3. Let {X), a = 1,2,.", be a sequence of independently identically distributed random vectors with mean vector I'- and covariance matrix l;; letXN=(1/N)L~~IX", SN=[l/(N-I)]L~~I(X,,-XN)(X,,-XN)" and TJ=N(XN-l'-o)'S,vI(XN-I'-O)' Then the limiting distribution of TJ as N --+ ()() is the X Z-distribution with p degrees of freedom if I'- = I'- o. Proof By the central limit theorem (Theorem 4.2.3) ,he limiting distribution of ..[N(5iN - 1'-) is N(O, l;). The sample covariance matrix converges stochastically to l;. Then the limiting distribution of T~ is the distribution of Y'l;-Iy, where Y has the distribution N(O, l;). The theorem follows from Theorem 3.3.3. •
USES OF THE T 2-STATISTIC
5.3
177
When the null hypothesis is true, T21n is distributed as X} I X';-P + I ' and A2/N given by (10) has the distribution of X,~-p+I/(x,~-p~1 + x}). The density of V = X; I( Xa2 + X;), when Xa" and X; are independent. is
(25)
r [ h a + b )1 ~a r(~a)r(~o)v.
_ I
:\b - 1
_
(1
v)
_
. 1
1.
•
-f3(v'Za,"b),
this is the density of the beta distriblltion with parameters ~a and ~b (Problem 5.27). Thus the distribution of A2/N =(1 + T1ln)-1 is the beta distribution with parameters ~p and ~(n - p + ]). 5.3. USES OF THE T 2-STATISTIC 5.3.1. Testing the Hypothesis That the Mean Vector Is a Given Vector The likelihood ratio test of the hypothesis fL = fLo on the basis of a sample of N from N(fL,:I) is equivalent to
(1) as given in Section 5.2.1. If the significance level is a, then the 100 a '7c point of the F-distribution is taken, that is,
(2) say. The choice of significance level may depend on the power of the test. We shall discuss this in Section 5.4. The statistic T2 is computed from i and A. The vector A - Iti - Ill) = b is the solution of Ab = i-fLo' Then T2 I(N - 1) = N(i - fLo)'b. Note that'T 2 I(N - 1) is the nonzero root of
(3) Lemma 5.3.1. If v is a vector of p components and if B is a nonsinguiar p X P malli,;, then v' B- 1 V is the nonzero root of
(4)
Ivv'-ABI =0.
Proof The nonzero root, say A1, of (4) is associated with a characteristic vector Il satisfying
(5)
THE GENERALIZED T 2 -STATlSTIC
178
Figure 5.1. A confidence ellipse.
Since A\ '" 0, v'll '" O. Multiplying on the left by v' B-1, we obtain
•
(6) In the case above v =
IN (i -
JLo) and B
= A.
5.3.2. A Confidence Region for the Mean Vector If JL is the mean of N(JL, l;), the probability is 1 - a of drawing a sample of N with mean i and covariance matrix S such that (7)
Thus, if we compute (7) for a particular sample, we have confidence 1 - a that (7) is a true statement concerning JL. The inequality ( 8)
is the interior and boundary of an ellipsoid in the p-dimensional space of m with center at i and with size and shape depending on S-l and a. See Figure 5.1. We state that JL lies within this ellipsoid with confidence 1 - a. Over random samples (8) is a random ellipsoid. 5.3.3. Simultaneous Confidence Intervals for All Linear Combinations of the Mean Vector From the confidence region (8) for JL we can obtain confidence intervals for linear functions "Y 'JL that hold simultaneously with a given confidence coefficient. Lemma 5.3.2 (Generalized Cauchy-Schwarz Inequality). definite matrix S, (9)
For a positive
5.3
USES OF THE T 2-STATISTIC
179
Proof. Let b = "I 'y/'Y 'S'Y. Then (10)
O:$; (y - bS'Y),S-l(y - bS'Y)
= y'S-ly - b'Y'SS-ly - y'S-lS'Yb + b 2 'Y 'SS-lS'Y =y'S-ly _
('Y'Y)~
'Y'S'Y '
which yields (9).
•
When y = i - IL, then (9) implies that
(11)
1'Y'(i- IL)I :$; V'Y'S'Y(i- IL)'S-l(i- IL) 2
:$; h's'Y JTp • N _ I ( ex)/N
holds for all "I with probability 1 - ex. Thus we can assert with confidence 1 - ex that the unknown parameter vector satisfies simultaneously for all "I the inequalities (12) The confidence region (8) can be explored by setting "I in (12) equal to simple vectors such as (1,0, ... ,0)' to obtain m l , (1, - 1,0, ... ,0) to yield m 1 - m 2 , and so on. It should be noted that if only one linear function "I'lL were of interest, J7~2.N_I(ex) =,jnpFp • n _ p + l (ex)/(n-p+l) would be replaced by t n ( ex). 5.3.4. Two-Sample Problems Another situation in which the T 2-statistic is used is one in which the null hypothesis is that the mean of one normal population is equal to the mean of the other where the covariance matrices are assumed equal but unknown. Suppose y~il, ... , y~) is a sample from N(IL(il, I), i = 1,2. We wish to test the null hypothesis IL(J) = IL(2). The vector ,(il is distributed according to N[IL(i), (1/N)IJ. Consequently VNI N 2 /(N I + N 2 ) (,(I) - ,(2» is distributed according to N(O, I) under the null hypothesis. If we let
THE GENERALIZED T 2-STATISTIC
180
then (N\ + N2 - 2)S is distributed as L~;.~N2-2 Z"Z~, where Z" is distributed according to MO, I). Thus
(14) is distributed as T2 with N\ + N2 - 2 degrees of freedom. The critical region is
(15) with significance level a. A confidence region for ....(1) vectors m satisfying
(16)
(y(I) -
y(2) -
.:<;;
-
....(2)
with confidence level 1 - a is the set of
m)'S-I (YO) - y(2) - m)
N J +N2 2 ~Tp.Nl+N2-2(a) I
2
Simultaneous confidence intervals are
(17) An example may be taken from Fisher (1936). Let XI = sepal length, x 2 = sepal width, x3 = petal length, x 4 = petal width. Fifty observations are taken from the population Iris versicolor (1) and 50 from the population Iris setosa (2). See Table 3.4. The data may be summarized (in centimeters) as
( 18)
( 19)
-(I)
=
(~:~~~)
-(2)
=
(;:~~~l
x
x
4.260' 1.326
1.462' 0.246
.
5.3
USES OF THE T 2 -STATISTIC
(20)
98S=
181
19.1434 9.0356 9.7634 3.2394
9.0356 11.8658 4.6232 2.4746
9.7634 4.6232 12.2978 3.8794
3.2394 2.4746 3.8794 2.4604
The value of T2/98 is 26.334, and (T 2/98) X .'If = 625.5. This value is highly significant CJmpared to the F-value for 4 and 95 degrees of freedom of 3.52 at the 0.01 significance level. Simultaneous confidence intervals for the differences of cO!TIponent means J.L)1) - J.L)21, i = 1,2,3,4, are 0.930 ± 0.337, - 0.658 ± 0.265, - 2.798 ± 0.270. and 1.080 ± 0.121. In each case 0 does not lie in the interval. [Since (9~(.01) < T4 98(.01), a univariate test on any component would lead to rejection of the n~ll hypothesis.) The last two components show the most significant differences from O. 5.3.5. A Problem of Several Samples After considering the above example, Fisher considers a third sample drawn from a popUlation assumed to have the same covariance matri.':. He treats the same measurements on 50 Iris virginica (Table 3.4). There is a theoretical reason for believing the gene structures of these three species to be such that the mean vectors of the three popUlations are related as (21)
where fL(3) is the mean vector of the third population. This is a special case of the following general problem. Let (X~.i'). a = 1, ... ,N;, i= 1, ... ,q, be samples from N(fL(i),I), i= J, ..•• q, respectively. Let us test the hypothesis q
(22)
H:
L (3; ....(i) = .... , i~l
where {31"'" {3q are given scalars and .... is a given vector. Thc criterion is (23)
where (24)
2 THE GENERALIZED T -STATISTIC
182
( 25)
(26)
r.
This T2 has the Y"-distribution with f_1 M - q degrees of freedom. Fisher actually assumes in his example that the covariance matrices of the three populations may be different. Hence he uses the technique described in Section 5.5. 5.3.6. A Problem of Symmetry Consider testing the hypothesis H: J.LI = J.L2 = ... = J.Lp on the basis of a sample x I' ••• , XIV from N(p.., I), where p..' = ( J.LI' •.• , J.Lp )' Let C be any (p -1) xp matrix of rank p - 1 such that CE =0,
(27)
where
l"
= (1, ... ,
n Then
(28)
0'.=
1, ... ,N,
has mean Cp.. and covariance matrix Cle. The hypothesis H is Cp.. = O. The statistic to be used is
(29) where 1
(30)
y=
N
N
E Ya=CX, a-I
s=
N~ 1 a-I t (y,,-Y)(Y,,-y)'
= N~
N
1C
L
(x,,-i)(x,,-i),C'.
a-I
This statistic has the T 2-distribution with N - 1 degrees of freedom for a (p - I)-dimensional distribution. This T 2-statistic is invariant under any linear transformation in the p - 1 dimensions orthogonal to E. Hence the statistic is independent of the choice of C. An example of this sort has been given by Rao (l948b). Let N be the amount of cork in a boring from the north into a cork tree; let E, S, and W be defined similarly. The set of amounts in four borings on one tree is
2 USES OF THE T -STATISTIC
5.3
183
considered as an observation from a 4-variate normal distribution. The question is whether the cork trees have the same amount of cork on each side. We make a transformation
Yl=N-E-W+S,
(32)
Y2 = S - W, . Y3=N-S.
The number of observations is 28. The vector of means is
y=
(33)
(
8.86 ) 4.50 ; 0.86
the covariance matrix for y is S=
(34)
128.72 61.41 ( - 21.02
61.41 56.93 -28.30
-21.02) -28.30 . 63.53
The value of T2 /(N -1) is 0.768. The statistic 0.768 X 25/3 = 6.402 is to be compared with the F-significance point with 3 and 25 degrees of freedom. It is significant at the 1% level. 5.3.7. Improved Estimation of the Mean
In Section 3.5 we considered estimation of the mean when the covariance matrix was l':nown and showed that the Stein-type estimation based on this knowledge yielded lower quadratic risks than did the sample mean. In particular, if the loss is (m - ,...)'l;-l(m - ,..,), then (35)
p-2 ( 1- N(x-v)'l;-l(X_V)
)+ _
(x-v)+v
is a minimax estimator of ,.., for any v and has a smaller risk than i when p ~ 3. When l; is unknown, we consider replacing it by an estimator, namely, a multiple of A = nS.
p
Theorem 5.3.1. 3 given by
When the loss is (m - ,..,)'l;-l(m - ,..,), the estimator for
~
(36)
(1- N(i-V)'~-l(i-vJ(i-V) +v
has smaller risk than i and is minimax for 0 < a < 2(p - 2) /(n - p + 3), and the risk is minimized for a = (p - 2)/(n - p + 3).
THE GENERALIZED T 2-STATISTIC
184
Proof As in the case when I is known (Section 3.5.2), we can make a transformation that carries (l/N)I to I. Then the problem is to estimate IL based on Y with the distribution N(IL,J) and A = L:=IZaZ~, where ZI'.·.' Zn are independently distributed, each according to N(O, I), and the loss function is (m - 1L)'(m - IL). (We have dropped a factor of N.) The difference in risks is
(37) LlR(IL)=,g'..{/lY-IL/l2-11(1- (Y_V),:-I(Y_V))(Y-V)+V-lLn
The proof of Theorem 5.2.2 shows that (Y - v)' A -) (Y - v) is distributed as /lY-v/i2/Xn2_p+l' where the Xn2_p+l is independent of Y. Then the difference in risks is
2 2 2 =,g' {2a(p - 2)Xn _p+l _ a ( Xn _p+l /}
..
IIY-vIl 2
IIY-v/l 2
= {2(p - 2)(n -p + l)a -[2(n-p+I)+(n-p+I)2ja 2},g' _ 1 _2.
"IIY-vIl
The factor in braces is n - p + 1 times 2(p - 2)a - (n - p + 3)a 2, which is positive for 0 < a < 2( p - 2) /(n - p + 3) and is maximized for a = (p-2)/(n-p+3). • The improvement over the risk of Y is (n - p + 1)( p - 2)2 /(n - p + 3)· ,g'.. /lY_v/l- 2 , as compared to the improvement (p - 2)2,g'.. IIY- v/l- 2 of m(y) of Section 3.5 when I is known.
5.4
185
DISTRIBUTION UNDER ALTERNATIVES; POWER FUNCTION
Corollary 5.3.1. (39)
Thl' l'.I'/illlll/or./in· fI
~ .\
(1- N(x_v)'~-I(X_V)r(i-V)+V
has smaller risk than (36) and is minimax for 0 < a < 2(p - 2)/(n - p + 3).
Proof This corollary follows from Theorem 5.3.1 and Lemma 3.5.2.
•
The risk of (39) is Hot necessarily minimi~ed at a = (p - 2)/(n - p + 3), but that value seems like a good choice. This is the estimator (] R) of Section 3.5 with I replaced by [1/(/1 - P + 3)]A. When the loss function is (m - fl.)'Q(m - fl.), where Q is an arbitrary positive definite matrix, it is harder to present a uniformly improved estimator that is attractive. Th( estimators of Section 3.5 can be used with I replaced by an estimate.
5.4. THE DISTRIBUTIOI\ OF T2 UNDER ALTERNATIVE HYPOTHESES; THE POWER FUNCTION In Section 5.2.2 we showed that (T~ /1l)(N - p)/p has a noncentral F-distribution. In this section we shall discuss the noncentral F-distribution. its tabulation, and applications to procedures based on T2. The noncentral F-distribution is defined .IS the distrihution of the ratio of a noncentral X 2 and an independent X 2 divided hy the ratio of corresponding degrees of freedom. Let V have the noncentral X2-distrihution with p degrees of freedom and noncentrality parameter T2 (as given in Theorem 3.3.5), and let W be independently distributed as X2 with m degrees of freedom. We shall find the density of F = (V/p)/lW/nz), which is the noncentral F with noncentrality parameter T2. The joint density of V and U' is (28) of ·Section 3.3 multiplied by the density of W, which is 2- tmr-l(-~m)wtm-le- tw. The joint density of F and W (dv = pwdf/m) is
( 1)
.~ !3~J ~2 to
(
f3!f( ~~ + f3) ( ~:) ~I'+Il-·1 w~(J'~m iTll-I .
The marginal density, obtained by integrating (1) with respect to 00, is
(2)
pe--~T'
~ (T~/2)Il(pj//1I)!J'+Il-·lr[~(p+m)+f3l
mf( ~m) f3~O
f3!r( ~p + f3)( 1 + pf/m) !(p-m l-r;
)1·
from (]
THE GENERALIZED T 2 -STATlSTIC
186
Theorem 5.4.1. If V has a noncentrai X2-distribution with p degrees of freedom alld noncentrality parameter r2, and W has an independent X 2-distribution with m degrees of freedom, then F = (V/p)/(W /m) has the density (2). The density (2) is the density of the noncentral F-distribution. If T~ = N(i - J.lO)'S-I(X - J.lo) is based on a sample of N from N(J.l, ~), then (T 2 /nXN - p)/p has the noncentral F-distribution with p and N - P degrees of freedom and noncentrality parameter N( J.l - J.lo)':t -I (J.l - J.lo) = r2. From (2) we find that the density of T2 is
e-~T' (3)
:x:
(N-l)r[HN-p)]
Il~O
(r 2 /2)Il[t 2 /(N-l)]W+Il-If(t N +{3) !31r(tp+{3)!1+t2/(N-l)]1N+/l
where (4)
••
_:xl
IFICa,b,x) -
Il~O
rca + ,3)r(b)x/l r(a)r(b + {3){31·
The density (3) is the density of the noncentral T 2-distribution. Tables have been given by Tang (1938) of the probability of accepting the null hypothesis (that is, the probability of Type II error) for various values of r2 and for significance levels 0.05 and 0.01. His number of degrees of freedom fl is our P [1(1)8], his f2 is our n - p + 1 [2,4(1)30,60,00], and his noncentrality parameter cP is related to our r2 by (5)
cP= __r _
IP+T
[l( ~ )3U)8]. His accompanying tables of significance points are for T 2 /(T 2 + N-l).
As an example, suppose p = 4, n - p + 1 = 20, and consider testing the null hypothesis J.l = 0 at the 1% level of significance. We would like to know the probability, say, that we accept the null hypothesis when cP = 2.5 (r2 =, 31.25). It is 0.227. If we think the disadvantage of accepting the null hypothesis when N, J.l, and ~ are such that r2 = 31.25 is less than the disadvantage of rejecting the null hypothesis when it is true, then we may find it
5.5
TWO-SAMPLE PROBLEM WITH UNEQUAL COVARIANCE MATRICES
187
reasonable to conduct the test as assumed. However, if the disadvantage of one type of error is about equal to that of the other, it would seem reasonable to bring down the probability of a Type II error. Thus, if we use a significance level of 5%, the probability of Type II error (for cp = 2.5) is only 0.043. Lehmer (1944) has computed tables of cp for given significance level and given probability of Type II error. Here tables can be used to see what value of T 2 is needed to make the probability of acceptance of the null hypothesis sufficiently low when IL O. For instance, if we want to be able to reject the hypothesis IL = 0 on the basis of a sample for a given IL and I, we may be able to choose N so that NIL/I -IlL = T2 is sufficiently large. Of course, the difficulty wi~ these considerations is that we usually do not know exactly the values of IL and I (and hence of T2) for which we want the probability of rejection at a certain value. The distribution of T2 when the null hypothesis is not true was derived by different methods by Hsu (1938) and Bose and Roy (1938).
*'
5.5. THE lWO-SAMPLE PROBLEM WITH UNEQUAL COVARIANCE MATRICES
If the covariance matrices are not the same, the T 2-test for equality of mean vectors has a probability of rejection under the null hypothesis that depends on these matrices. If the difference between the matrices is small or if the sample sizes are large, there is no practical effect. However, if the covariance matrices are quite different and/or the sample sizes are relatively small, the nominal significance level may be distorted. Hence we develop a procedure with assigned significance level. Let (x~)}, a = 1, ... , N;, be samples from N(IL(i), I), i = 1,2. We wish to test the hypothesis H: IL(I) = 1L(2). The mean i(l) of the first sample is normally distributed with expected value Gi(l)
(1)
= IL(I)
and covariance matrix
(2) Similarly, the mean expected value
(3)
i(2)
of the second sample is normally distributed with
2
188
THE GENERALIZED T -STATlSTlC
and covariance matrix
(4) Thus i(l) -i(2) has mean IL(I) -1L(2) and covariance matrix (l/NI)I 1 + (1/N2 ):.t 2 • We cannot use the technique of Section 5.2, however, because N,
L
(5)
N2
(x~)
- i(l»)( x~l) - i(l»)' + L
a~j
(X~2) - i(2»)( X~2) - i(2»),
a~j
does not have the Wishart distribution with covariance matrix a multiple of (l/Nj)II + O/N2 )I 2 • If NI = N2 = N, say, we can use the T 2-test in an obvious way. Let Ya = x~) - X~2) (assuming the numbering of the observations in the two samples is independent of the observations themselves). Then Ya is normally distributed with mean IL(I) - 1L(2) and covariance matrix II + I 2 , and YI, ... ,YN are independent. Let y=(1/N)L:~~ly,,=i(l)-i(2), ll.nd define S by N
(6)
(N-l)S=
L
(y,,-y)(y,,-y),
a=l N
=
L
(x~)
- X~2) -
i(l)
+ i(2») ( X~I) -
X~2) - i(1)
+ ,i(2»)' .
a~l
Then
(7) is suitabl,;! for testing the hypothesis 1L(l) - 1L(2) = 0, and has the T 2-distribution with N - 1 degrees of freedom. It should be observed that if we had known II = I 2 , we would have used a T 2-statistic with 2N - 2 degrees of freedom; thus we have lost N - 1 degrees of fr~edom in constructing a test which is independent of the two covariance matrices. If NI = N2 = 50 as in the example in Section 5.3.4, then T/ 49 (.On = 15.93 as compared to T/ 9S (.0l)
= 14.52. Now let-us turn our attention to the case of N #- N 2 • For cor,venience, let NI < N 2 • Then we define J
( 8)
Y
= X(I) -
""
{fi
~ X(2)
N2
a
1 N, 1 N2 + - - - " X(2) - - " X(2) IN J N 2 f3~ t... I
{3
N 2 -y~ t... J
-y'
a= 1, ... ,N
J
•
5.5
TWO-SAMPLE PROBLEM WITH UNEQUAL COVARIANCE MATRICES
189
The covariance matrix of Yo and Y{3 is
(10) Thus a suitable statistic for testing J.l(I) tion with N] - 1 degrees of freedom, is
J.l(2)
= 0, which has the T 2-distribu-
(11) where (12) and N,
(13)
(N]-1)S=
L
N,
(y,,-y)(y,,-y)'
=
L
(u"-u)(u,,-u),,
a=1
a=l
where U = (1/NI)L,~~IUa and u a =x~l) - ';N]/N2X~2), a = 1, ... , N]. This procedure was suggested byScheffe (1943) in the univariate case. Scheffe showed that in the univariate case this technique gives the shortest confidence intervals obtained by using the t-distribution. The advantage of the method is that i(l) - i(2) is used, and this statistic is most relevant to J.l(I) - J.l(2). The sacrifice of observations in estimating a covariance matrix is not so important. Bennett (1951) gave the extension of the procedure to the multivariate case. This appr~ach can be used for more general cases. Let (x~)}, a = 1, ... , !Ii;, i = 1, ... , q, be samples from N(J.l(i),1:), i = 1, ... , q, respectively. Consider testing the hypothesis q
(14)
H:
L
/3;J.l(i)
= J.l,
;=1
where {31"'" {3q are given scalars and unequal, take N] to be the smallest. Let
J.l
is a given vector. If the !Ii; are
THE GENERALIZED T 2 -STATlSTlC
190
Let ji and S be defined by N
-(i)=~ ~ N £..,
(17)
X
(i)
X fJ '
1/3=1 N,
(18)
(N(-I)S=
L
(Yu-y)(y,,-ji)'·
a=l
Then ( 19) is suitable for testing H, and when the hypothesis is true, this statistic has the T"-distribution for dimension p with NI - 1 degrees of freedom. If we let u" = L:!~I (3iJNI/NiX~), a = 1, ... , N 1, then S can be defined as N,
(20)
(NI -1)S=
L
(ua-u)(ua-u)'.
a=l
Another problem that is amenable to this kind of treatment is testing the hypothesis that two subvectors have equal means. Let x = (X(I)" x(2)'), be distributed normally with mean J.l = (J.l(I)" J.l(2),), and covarirnce matrix
(21) We assume that x(l) and X(2) are each of q components. Then y = X(I) - X(2) is distributed normally with mean J.l(l) - J.l(2) and covariance matrix :£y = :£11 - :£21 - :£ 12 + :£22' To test the hypothesis J.l(I) = J.l(2) we use a T 2-statistic Nj'S;.-ly, where the mean vector and covariance matrix of the sample are partitioned similarly to J.l and :£.
5.6. SOME OPTIMAL PROPERTIES OF THE T2-TEST 5.6.1. Optimal Invariant Tests In this section we shall indicate that the T 2 -test is the best in certain classes of tests and sketch briefly the proofs of these results. The hypothesis J.l = 0 is to be tested on the basis of the N observations XI' ... , X N from N(J.l, 1:). First we consider the class of tests based on the
2 5.6 SOME OPTIMAL PROPERTIES OF THE T -TEST
191
statistics A = L(X" - i)(x" - i)' and i which are invariant with respect to the transformations A* = CAe' and i* = ex, where C is nonsingular. The transformation x~ = Cx" leaves the problem invariant; that is, in terms of x~ we test the hypothesis tC'x~ = 0 given that xi, ... , x~ are N observations from a multivariate normal population. It seems reasonable that we require a solution that is also invariant with respect to these transformations; that is, we look for a critical region that is not changed by a nonsingular linear transformation. (The defin.tion of the region is the same in different coordinate systems.) Theorem 5.6.1. Given the observations XI"",XN from N(p.,I,), of all tests of p. = 0 based on i and A = L(x" - i)(x" - i)' that are invariant with respect to transformations i* = ex, A* = CAC' (C nonsingular), the T 2-test is uniformly most powerful. Proof First, as we have seen in Section 5.2.1, any test based on T2 is invariant. Second, this function is essentially the only invariant, for if f(i, A) is invariant, then f(x, A) = f(x*, J), where only the first coordinate of x* is different from zero and it is ...;x/A -I X. (There is a matrix C such that ex = x* and CAC' = I.) Thus f(x, A) depends only on x/A -1 X. Thus an invariant test must be based on X/A-IX. Third, we can apply the NeymanPearson fundamental lemma to the distribution of T2 [(3) of Section 5.4] to find the uniformly most powerful test based on T2 against a simple alternative r 2 = Nfl: I, -\ p.. The most powerful test of r2 = 0 is based .on the ratio of (3) of Section 5.4 to (3) with r2 = O. The critical region is ( 1)
(t2/ n )4P-I(1 + t2/nft(n+l)r[~(n + 1)]
rcip)
/
rcip)
r[hn+l)]
e-~T'
£ (r2/2)"r[hn+l)+a] (
,,~O
a!rc~p+a)
2
t /n
)"
l+t 2 /n
The right-hand side of (1) is a strictly increasing function of (t2/ n )/(l + t 2In), hence of t 2 • Thus the inequality is equivalent to t 2 > k for k suitably chosen. Since this does not depend on the alternative r 2, the test is uniformly most • powerful invariant.
2 THE GENERALIZED T -STATISTIC
192
Definition 5.6.1. A critical function IjJcr, A) is a function with values between 0 and 1 (inclusive) such that CIjJ(i, A) = e, the significance level, when IL = O. A randomized test consists of rejecting the hypothesis with probability IjJ(x, B) when i = x and A = B. A nonrandomized test is defined when IjJ(i, A) takes on only the values 0 and 1. Using the form of the
Neyman-Pearson lemma appropriate for critical functions, we obtain the following corollary: Corollary 5.6.1. On the basis of observations XI' ••. ' X N from N(IL, 1:), of all randomized tests based on i and A that are invariant with respect to transformations i* = ex, A* = CAC' (C nonsingular), the T 2 -test is uniformly most poweiful. Theorem 5.6.2. On the basis of observations X I' ... , X N from N( IL, 1:), of all tests of IL = 0 that are invariant with respect to transformations x~ = CXa (C nonsingular), the T 2-test is a uniformly most poweiful test; that is, the T 2-test is at least as poweiful as any other invariant test. Proof Let IjJ(x l , ••• , x N ) be the critical function of an invariant test. Then
Since i, A are sufficient statistics for 1L,1:, the expectation tS'[ IjJ(x l , ••• , A] depends only on i, A. It is invariant and has the same power as IjJ(x l , ••• , x N ). Thus each test in this larger class can be replaced by one in the smaller class (depending only on i and A) that has identical power. Corollary 5.6.1 completes the proof. •
XN )Ii,
Theorem 5.6.3. Given observations x l' ... ' X N from N(IL, 1:), of all tests of IL = 0 based on i and A = L( x a - i)( X a - i)' with power depending only on NIL'1: -IlL, the T 2-test is uniformly most poweiful. Proof We wish to reduce this theorem to Theorem 5.6.1 by identifying the class of tests with power depending on NIL'1:- 1 1.A. with the class of invariant tests. We need the following definition:
Definition 5.6.2.
A test IjJ(x l , ••. , x N
)
for all XI' ... ' X N except for a set of XI' ... ' exception set may depend on c.
is said to be almost invariant if
XN
of Lebesgue measure zero; this
5.6
2
193
SOME OPTIMAL PROPERTIES OF THE T -TEST
It is clear that Theorems 5.6.1 and 5.6.2 hold if we extend the definition of invariant test to mean that (3) holds except for a fixed set of .:1' .... Xv of measure 0 (the set not depending on C). It has been shown by Hunt and Stein [Lehmann (1959)] that in our problem almost invariancc implies invari· ance (in the broad sense). Now we wish to argue that if IjJCx, A) has power depending only on NIL'I -I J.l, it is almost invariant. Since the power of IjJ(i, A) depends only on NJ.l'I -IlL, the power is
(4) ==
tffp..I. IjJ(ex,
CAC').
The second and third terms of (4) are merely different ways of writing the same integral. Thus
(5) identically in J.l, I. Since i, A are a complete sufficient set of statistics for IL, I (Theorem 3.4.2), f(i, A) = IjJ(i, A) - IjJ(ex, CAe') = 0 almost every· where. Theorem 5.6.3 follows. • As Theorem 5.6.2 folloy, s from Theorem 5.6.1, so does the following theorem from Theorem 5.6.:::
Theorem 5.6.4. On the basis of observations XI"'" x N from N(J.l, ~), of all tests of IL = 0 with power depending only on NIL'I -IlL, the T 2-test is a uniformly most poweiful test.
Theorem 5.6.4 was first proved by Simaika (1941). The results and proofs given in this section follow Lehmann (I959). Hsu (1945) has proved an optimal property of the T 2-test that involves averaging the power over IL and I. 5.6.2. Admissible Tests We now turn to the question of whether the T 2-test is a good test compared to all possible tests; the comparison in the previous section was to the restricted class of invariant tests. The main result is that the T~-test is admissible in the class of all tests; that is, there is no other procedure that is better. Definition 5.6.3. A test T* of the null hypothesis Ho : wE Do against the alternative wED I (disjoint from Do) is admissible if there exists no other test T
THE GENERALIZID T 2 -STATISTIC
194
sllch that
(6)
Pr{Reject Hoi T,
~ Pr{Reject Hoi T*,
w},
(7)
Pr{Reject Hoi T, (v};:: Pr{Reject Hoi T*,
w),
w}
with strict inequality for at least one w.
The admissibility of the T 2-test follows from a theorem of Stein (1956a) that applies to any exponential family of distributions. An e:rponelltiai family of distributions (11, Y'.8, m, n, P) consists of a finitedimensional Euclidean space 'ii, a measure In on the u-algebra @ of all ordinary Borel sets of -;11, a subset n of the adjoint space (~II' (the linear space of all real-valued linear functions on '11) such that
t/J( w) =
(8)
f
eW'y
dm(y) < 00,
wEft,
0/
and P. the function on 0 to the set of probability measures on
P",(A) = t/J/w)
!.r'.I
dm (y),
@
given by
AE.c1l.
The family of normal distributions N(IL, l:) constitutes an exponential family, for the density can be written
(9) We map from d'to 11; the vector y = (y(1)" y(2),), is composed of y(1; = x and y(2)=(xL2xlX2, ... ,2xlXP,xi" .. ,x~)'. The vector 00=(00(1)',00(2),), is composed of oo(I)=I-11L and 00(2)= _t(Ull,UI2, ... ,Ulp,u22, ... ,upp)" where (u;j) = I -I; the transformation of parameters is one to one. The measure m(A) of a set A E: y,{j is the ordinary Lebesgue measure of the se', of x that maps into the set A. (Note that the prohability measure in all is not defined by a density.) Theorem 5.6.5 (Stein). Let (",iI, :1(1, m, 0, P) be an exponential family and a nonempty proper subset of O. (i) Let A be a subset of 0/ that is closed and convex. (ii) Suppose that for every vector 00 E '11' and real c for which {yl oo'y > c} alld A are di,ljoilll, there exists 00 1 E 0 such that for arbitrarily large A the vector 00 1 + Aoo E 0 - 0 0 , Theil the test with acceptance region A is admissible jbr testin!? the hypothesis that 00 E 0 0 against the alternative 00 E 0 - 0 0 ,
no
5.6 SOME OPTIMAL PROPERTIES OF THE T 2 -TEST
195
) A
Figure 5.2
The cond~tions of the theorem are illustrated in Figure 5.2, which is drawn simultaneously in the space 0/ and the set fl. Proof The critical function of the test with acceptance region A is
(10)
(11)
J
00
E fl- flo,
with strict inequality for some 00; we shall show that this assumption leads to a contradiction. Let B = {yl
{yl o} =AIIB, where A is the complement of A. The m-measure of the set (12) is positive; otherwise c} has positive m-measure. (Since A is closed, A is open and it can be covered with a denumerable collection of open spheres, for example, with rational radii and centers with rational coordinates. Because there is a hyperplane separating A and each sphere, there exists a denumerable coHection of open half-spaces H j disjoint from A that covers A. Then at least one half-space has an intersection with A liB with positive m-measure.) By hypothesis there exists 001 Efland an arbitrarily large A such that (13)
THE GENERALIZED T 2-STATISTIC
196 Then
(14)
f[ CPA(Y) - cp(y)] dP",Jy) =
I/I(~J
f[CPA(Y) - cp(y)]e""'Y dm(y)
=
~~::~
f[ cpAY) - cp(y)]eA""Y dP",,(y)
=
~~::~ e Ac f[ CPA(y) -
cp(y)]eA(""Y-C)dP",,(y)
= 1/1(00 1) eAC{f [CPA(y) - cp(y)]eA(""Y-C)dP",,(y) I/I(wJ w'y>c + f,
[CPA(Y) - cp(y)] eA("":'-C) dP",,(y)}.
w ysc
For w'y > c we have CPA(Y) = 1 and CPA(Y) - cp(y)? 0, and (yl cpiy) - cp(y) > O} has positive measure; therefore, the first integral in the braces approaches 00 as A -> 00. The second integral is bounded because the integrand is bounded by 1, and hence the last expression is positive for sufficiently large A. This contradicts (11). • This proof was given by Stein (1956a). It is a generalization of a theorem of Birnhaum (1955). Corollary 5.6.2. If the conditions of Theorem 5.6.5 hold except that A is not necessarily closed, but the boundary of A has m-measure 0, then the conclusion of Theorem 5.6.5 holds. Proof The closure of A is convex (Problem 5.18), and the test with acceptance region equal to the closure of- A differs from A by a set of probability 0 for all 00 En. Furthermore,
(15)
An{ylw'y>c}=0
~
Ac{ylw'y:sc}
~
closure A c {y I00 I Y :S c} .
Then Theorem 5.6.5 holds with A replaced by the closure of A.
•
Theorem 5.6.6. Based on observations xp ... , x N from Hotelling's T 2-test is admissible for testing the hypothesis IL = O.
N(IL, 1;),
5.6
197
SOME OPTIMAL PROPERTIES OF THE ["-TEST
Proof To apply Theorem 5.6.5 we put the distribution of the observations into the form of an exponential family. By Theorems 3.3.1 and 3.3.2 we can transform x I " " , X N to Zc<=L~~lc"llxll' where (c"ll) is orthogonal and Zs = {Fii. Then the density of ZI'" ., ZN (with respect to Lebesgue measure) is
(16)
e-!:jJ.~'i.-IjJ.,
(27T)'P 1:lI,N
eXP[{FiIL':l-lz,v+tr(-t:l-I)
t Zoz~1.
(X~I
The vector y = (y(l), y(2)')' is composed of ylll = z", (= (b ll , 2b I2 , ..• , 2b lp , b 22 , ..• , bpp )', where
viii)
and y':\ =
N
(17)
L
B=
ZUZ:l
,,~l
The vector 00=(00(1)',00(2)')' is composed of 00(11 = {Fi:l-IIL and w(2)= _~«(T",fTI2, ••• ,fTlp,fT22 ••..• (TPP)'. The measure m(A) is the Lehesgue measure of the set of z I' ... , Z N that maps into the set A. Lemma 5.6.1. (18)
Let B =A + tvXi'. Then Ni
x
'A- I --
Ni'B-li
x- I-M'B-li'
Proof of Lemma. If we let B = A + {Fi i{Fi i' in (10) of Section 5.2. we obtain by Corollary A.3.I
(19)
1 = 1 + T 2 /(N -1)
).."/,'1
= IB - {Fii{Fii'l
IBI
= I-M'B-li.
•
Thus the a~ceptance region of a T -test is 2
(20)
A={zN,Blz:VB-'z,\~k, B positive definite)
for a suitable k. The function z:VB-I ZN is convex in (z. B) for B positive definite (Problem 5.17). Therefore, the set Z'N B -I ZN ~ k is convex. This shows that the set A is convex. Furthermore, the closure of A is convex (Problem 5.18). and the probability of the boundary of A is O. Now consider the other condition of Theorem 5.6.5. Suppose A is disjoint with the half-space
(21)
c<w'y=v'ZN-~trAB,
THE GENERALIZED T 2 -STATISTlC
198
where :\ is a symmetric matrix and B is positive semidefinite. We shall take A) = 1. We want to show that (0) + Aw E fl _. flo; that is, that VI + Av *" 0 (which is trivial) and /\.1 + AI\. is positive definite for A > O. This is the case when ;\ is positive semidefinite. Now we shall show that a half-space (21) disjoint with A and I\. not positive semidefinit =implies a contradiction. If I\. is not positive semidefinite, it can be written (by Corollary A.4.1 of thf' Appendix)
o -1
(22)
o where D is nonsingular. If I\. is not positive semidefinite, -1 is not vacuous, because its order is the number of negative characteristic roots of A. Let :.\ = (1/1');:0 and
(23)
B~(D')'[:
~lD'
0 1'1
0
Then
(24)
1 1 w'y = -v'zo + ztr I'
[-1 0 0
0 1'1
0
n
which is greater than c for sufficiently large y. On the other hand
(25)
which is less than k for sufficiently large y. This contradicts the fact that (20) and (21) are disjoint. Thus the conditions of Theorem 5.6.5 are satisfied and the theorem is proved. • This proof is due to Stein. An alternative proof of admissibility is to show that the T 2 -test is a proper Bayes procedure. Suppose an arbitrary random vector X has density I(xl (0) for 00 E fl. Consider testing the null hypothesis Ho: 00 E flo against the alternative H) : 00 E fl - flo. Let ITo be a prior finite measure on flo, and IT) a prior finite measure on fl 1• Then the Bayes procedure (with 0-1 loss
5.7
199
ELLIPTICALLY CONTOURED DISTRIBUTIONS
function) is to reject Ho if
(26)
Jf( xl w )IIo( dw)
>c
for some c (O:s; c :s; 00). If equality in (26) occurs with probability 0 for all w E flo, then the Bayes procedure is unique and hence admissible. Since the measures are finite, they can be normed to be probability measures. For the T 2-test of Ho: J.l = 0 a pair of measures is suggested in Problem 5.15. (This pair is not unique.) Th(' reader can verify that with these measures (26) reduces to the complement of (20). Among invariant tests it was shown that the T2-test is uniformly most powerful; that is, it is most powerful against every value of J.l'l: -I J.l among invariant tests of the specified significance level. We can ask whether the T 2-test is "best" against a specified value of J.l'l:-IJ.l among all tests. Here "best" can be taken to mean admissible minimax; and "minimax" means maximizing with respect to procedures the minimum with respect to parameter values of the power. This property was shown in the simplest case of p = 2 and N = 3 by Giri, Kiefer, and Stein (1963). The property for .~enelal p and N was announced by Salaevski'i (1968). He has furnished a proof for the case of p = 2 [Salaevski'i (1971)], but has not given a proof for p > 2. Giri and Kiefer (1964) have proved the T 2-test is locally minimax (as J.l'l: -I J.l --+ 0) and asymptotically (logarithmically) minimax as IL';'£ -I J.l--> 00.
5.7. ELLIPfiCALLY CONTOLRED DISTRIBUTIONS 5.7.1. Observations Elliptically Contoured When
( 1)
xl>""
XN
constitute a sample of N from
IAI--lg[(x - v)' A -I(X- v)],
the sample mean i and covariance S are unbiased estimators of the distribution mean J.l = v a nd covariance matrix I = ( $ R2 Ip) A, where R2 = (X - v), A -I(X - v) has finite expectation. The T 2-statistic, T2 = N(i J.l)' S-I (i - J.l), can be used for tests and confidence regions for J.l when I (or A) is unknown, but the small-sample distribution of T2 in general is difficult to obtain. However, the limiting distribution of T2 when N --+ 00 is obtained from the facts that IN (i - J.l) !!. N(O, l:) and S.!!.., l: (Theorem 3.6.2).
THE GENERALIZED T 2 -STATlSTlC
200
Theorem 5.7.1.
Let
XI"'"
XN
be a sample from (1). Assume {fR 2 < 00.
Then T2!!. X;. Proof Theorem 3.6.2 implies that N(i - 1L)'l;-I(i - IL)!!. Xp2 and N(i - 1L)'l;-I(i - IL) - T2 g, O. •
Theorem 5.7.1 implies that the procedures in Section 5.3 can be done on an asymptotic basis for elliptically contoured distributions. For example, to test the null hypothesis IL = lLo, reject the null hypothesis if
(2) where X;( a) is the a-significance point of the X2 -distribution with p degrees of freedom. the limiting probability of (2) when the null hypothesis is true and N --- 00 is a. Similarly the confidence region N(i - m)' S-I (i - m) ::;; X;( ex) has li.niting confidence 1 - a.
5.7.2. Elliptically Contoured Matrix Distributions Let X (N X p) have the density (3)
\C\-Ng [ C-I(X- ENV')'(X - ENV')(C') -I]
based on the left spherical density g(Y'Y). Here Y has the representation yf!: UR', where U (NXp) has the uniform distribution on O(Nxp), R is lower triangular, and U and R are independent. Then Xf!: ENV' + UR'C'. The T 2-criterion to test the hypothesis v = 0 is Ni'S-li, which is invariant with respect to transformations X ---XG. By Corollary 4.5.5 we obtain the following theorem.
Theorem 5.7.2. Suppose X has the density (3) with v = 0 and T2 = Ni'S-li. Then [T 2/(N-1)][(N-p)/p] has the distribution of Fp,N_p = (xi /p)/[ X~_p/(N - p)]. Thus the tests of hypotheses and construction of confidence regions at stated significance and confidence levels are valid for left spherical distributions. The T 2 -criterion for H: v = 0 is
(4) since X f!: UR'C',
(5)
201
'ROBLEMS
mel
:6)
S=
N~ 1 (X'X-tv:ri')
=
N~ 1 [CRU'URC' -CRuu'(C'R)']
=CRSu(CR)'.
5.7.3. Linear Combinations Ui'Jter, Glimm, and Kropf (1996a, 1996h. 1996c) have observed that a statistician can use X' X = CRR'C' when v = 0 to determine a p x if matrix LJ and base a T-test on the transform Z = XD. Specifically, define
(7) (8)
Sz = N
~ 1 (Z'Z -
NV.') = D'SD,
(9) Since QNZ £. QNUR'C' £. UR'C' = Z, the matrix Z is based on the leftspherical YD and hence has the representation Z = JIR* " where V (N x q) has the uniform distribution on O(N xp), independent of R*' (upper triangular) having the distribution derived from R* R*' = Z' Z. The distribution of T2 I(N -1) is Fq,N_qql(N - q). The matrix D can also involve prior information as well as knowledge of X' X. If p is large, q can be small; the power of the test based on TJ may be more pow
PROBLEMS 5.1. (Sec. 5.2) Let x" be distributed according to N(". + \3(z" - z), l;), ex = 1, ... , N, where z = (1/N)Ez". Let b =[l/Hz" -z)2]Ex (za -Z),(N - 2)S = E[x,,-i-b(za-z)][xa-i-b(za-Z)l'. and r 1 =Hz a -Z) 2 b'S-lb. Show that T2 has the T 2 -distribution with N - 2 degrees of freedom. [Hinr: See Q
Problem 3.13.] 5.2. (Sec. 5.2.2) Show that r2 /(N - 1) can be written as R 2 /O - R2) with the correspondences given in Table 5.1.
THE GENERALIZED T 2-STATISTIC
202 Table 5.1
Section 5.2 XOa Xa
=
1/{N
Section 4.4 Zia z~)
a O) = LZlaZ~) B = LXaX~
A22 = Lzi2)Z~2)'
1 = LX5a
all =
T2
N-l
LZfa
R2 1-R2
P -1 n
P N
5.3. (Sec. 5.2 2) Let
R2
I - R2 =
LUaX~(LXaX~) -ILUaXa
LII,~ - LII"x~(LrnX~)
I LUaXa '
where U \> ••. , II N are N numbers and x\> ... , X N are independent, each with the distribution N(O, l:). Prove that the distribution of R 2 /O - R2) is independent of U I , ... , UN' [Hint: There is an orthogonal N X N matrix C that carries ( I l l " ' " LIN) into a vector proportional to (1/ {N, ... , 1/ {N).] 5.4. (Sec. 5.2.2) Use Problems 5.2 and 5.3 to show that [T 2 /(N - 1)][(N - p)/p] has the Fp. N_p-distribution (under the null hypothesis). [Note: This is the analysis that corresponds to Hotelling's geometric proof (1931).) 5.5. (Sec. 5.2.2) Let T 2 =Ni'S-li, where i and S are the mean vector and covariance matrix of a sample of N from N(fL, l:). Show that T2 is distributed the same when fL is repla<..ed by A = (T, 0, ... ,0)', where T2 = fL' l: -I fL, and l: is replaced by 1. 5.6. (Sec. 5.2.2) Let U = [T 2 /(N - Dl![l + T 2 /(N - 0]. Show that 'YV'(W,)-IV'Y', where 'Y = (1/ {N, ... , 1/ {N) and
U
=
PROBLEMS
,5.7. (Sec. 5.2.2)
21)3 Let
i *- 1,
v'-
(n
Prove that U = s + (1 - s)w, where
Vi V;') (Vi) -1
:
:
"/1* "I'*,
V"*
y*'.
Hint: EV,= V*, where 1
0
11211; -
V I/'1 I
o o
E=
, vpvl ---, VI V 1
5.S.
0
(Sec. 5.2.2) Prove that w has the distribution of the square of a multiple correlation between one vector and p - I vcctors in (N - O-space without subtracting means; that is, it has density
[Hint: The transformation o~ Problem 5.7 is a projection of (N -I)-space orthogonal to
VI']
v
2
, •.. , v
p
,
Y on the
5.9. (Sec.5.2.2) Verify that r=s/(1-s) multiplied by (N -0/1 has the noncentral F-distribution with 1 and N - 1 degrees of freedom and noncentrality parameter NT 2.
THF GENERALIZED T 2 -STATiSTiC
204
5.10. (Sec. 5.2.2) From Problems 5.5-5.9, verify Corollary 5.2.1. 5.11. (Sec. 5.3) Use the data in Section 3.2 to test the hypothesis that neither drug has a soporific effect at significance level 0.01. 5.12. (Sec. 5.3) Using the data in Section 3.2, give a confidence region for fl. with confidence coefficient 0.95. 5.13. (Sec. 5.3) Prove the statement in Section 5.3.6 that the T 2-statistic is independent of the choice of C. 5.14. (Sec. 5.5) Use the data of Problem 4.41 to test the hypothesis that the mean head length and breadth of first sons are equal to those of ~econd sons at significance level 0.01. 5.15. (Sec. 5.6.2) T 2-test as a Bayes procedure [Kiefer and Schwartz (1965)]. Let XI' ... , X N be independently distributed, each according to N( fl., I). Let TI 0 be defined by [fl.,::£] = [0,([ + TJTJ')-I] with TJ having a density proportional to II+TJTJ'I-~N, and let TIl be defined by [fl.,::£)=[([+TJTJ,)-ITJ,([+TJTJ,)-I] with TJ having a density proportional to
(a) Show that the lleasures are finite for N> p by showing TJ'([ + TJTJ')-ITJ!O 1 and verifying that the integral of II + TJTJ'I- tN = (1 + TJTJ')- tN is finite. (b) Show that the inequality (26) is equivalent to Ni'(L~_IXaX~)-IX"2!k. Hence the T 2 -test is Bayes and thus admissible. 5.16. (Sec. 5.6.2) Let get) = f[lyl + (1 - t)Y2]' where fey) is a real.yalued functiun of the vector y. Prove that if get) is convex, then fey) is convex. 5.17. (Sec. 5.6.2) Show that z'B-1z is a convex function of (z,B), where B is a positive definite matrix. [Hint: Use Problem 5.16.] 5.18. (Sec. 5.6.2)
Prove that if the set A is convex, then the closure of A is convex.
5.19. (Sec. 5.3) Let x and S be based on N observations from N(fl., ::£), and let x be an additional observation from N(fl., ::£). Show that x - x is distributed according to N[0,(1+1IN)::£].
Verify that [N I(N + l)](x - X)'S-I(X - x) has the T 2 -distribution with N - 1 degrees of freedom. Show how this statistic can be used to give a prediction region for x based on x and S (i.e., a region such that one has a given confidence that the next observation will fall into it).
205
PROBLEMS
5.20. (Sec. 5.3) Let x~) be observations from N(jJ.(i), ::£,), a the likelihood ratio criterion for testing the hy')othesis 5.21. (Sec. 5.4) Prove that verifying
is larger for
jJ.,::£-IjJ.
Discuss the power of the test 1-'-2 = O.
1-'-1
=
1, ... , N i • i
1.2. Find
=
jJ.(l} = jJ.(2).
jJ.' = (1-'-1,1-'-2)
than for
jJ. = 1-'-1
= 0 compared to the power of the test
1-'-1
by
= 0,
5.22. (Sec. 5.3) (a) Using the data of Section 5.3.4, test the hypothesis (b) Test the hypothesis I-'-V) = I-'-(?).I-'-~) = I-'-S2). 5.23. (Sec. 5.4)
Prove
2 1-'-\1) = 1-'-(1 ).
Let
jJ.'::£ -1jJ.
~ jJ.l l),::£ 111jJ.(I).
Give a condition for strict inequality to hold.
[Hint: This is the vector analog of Problem 5.21.]
5.24. Let Xli)' = (Yl i )', Z(i)'), i = 1,2, where y(i) has p components and components, be distributed according to N(jJ.(i), ::£), where
(i)=
jJ.
Zlil
Cl) ( jJ.~)' jJ.;
i
has q
=
1,2.
Find the likelihood ratio criterion (or equivalent T 2-criterion) for testing jJ.~1) = jJ.<1) given jJ.~) = jJ.l;) on the basis of a sample of Ni on Xli>, i = 1.~. [Hint: Express the likelihood in terms of the marginal density of y(l) and the conditio,laf density of Z(i) given y(i).] 5.25. Find the distribution of the criterion in the preceding problem under the null hypothesis. 5.26. (Sec. 5.5) Suppose g= 1, ... ,q.
x~g)
is an observation from N(jJ.181,::£g). a
=
1. .... Ny.
THE GENERALIZED T 2 -STATISTJC
206
(a) Show that the hypothesis ....(l) = ... = ....(q) is equivalent to t!y~i) = 0, i = I, ...• q - I, where
a=I, ... ,N1 ,
i=I, ... ,q-Ij
S; Ng • g = 2, ... , qj and (ali), ... , a~», i = 1, ... , q - 1, are linearly independent. (b) Show how to construct a T 2-test of the hypothesis using (y(l)', ... , y(q .. ) ')' yielding an F-statistic with (q - 1)p and N - (q - 1)p degrees of freedom [Anderson (1963b)j.
N)
5.27. (Sec. 5.2) Prove (25) is the density of V = x;:j( xl + xt). [Hint: In the joint density of U = xl and W = xt make the transformation u = uw(1 - u)-l, w = w and integrate out w.j
CHAPTER 6
Classification of Observations
6.1. THE PROBLEM OF CLASSIFICATION The problem of classification arises when an investigator makes a number of measurements on an individual and wishes to classify the individual into one of several categories on the basis of these measurements. The investigator cannot identify the individual with a category directly but must use these measurements. In many cases it can be assumed that there are a finite number of categories or populations from which the individual may have come and each population is characterized by a probability distribution of the measurements. Thus an individual is considered as a random observation from this population. The question is: Given an individual with certain measurements, from which population did the person arise? The problem of classification may be considered as a problem of "statistical decision functions." We have a number of hypotheses: Each hypothesis is that the distribution of the observation is a given one. We must accept one of these hypoth"!ses and reject the others. If only two populations are admitted, we have an elementary problem of testing one hypothesis of a specified distribution against another. In some instances, the categories are specified beforehand in the sense that the probability distributions of the measurements are assumed completely known. In other cases, the form of each distribution may be known, but the parameters of the distribution must be estimated from a sample from that population. Let us give an example of a problem of classification. Prospective students applying for admission into college are given a battery of tests; the "ector of
An imroductiollto Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0·471-36091-0 Copyright © 2003 John Wiley & Sons, Inc.
207
208
CLASSIFICATION OF OBSERVATIONS
scores is a set of measurements x. The prospective student may be a member of one population consisting of those students who will successfully complete college training or, rather, have potentialities for successfully completing training, or the student may be a member of the other population, those who will not complete the college course successfully. The problem is to classify a student applying for admission on the basis of his scores on the entrance examination. In this chapter we shall develop the theory of classification in general terms and then apply it to cases involving the normal distribution. In Section 6.2 the problem of classification with two populations is defined in terms of decision theory, and in Section 6.3 Bayes and admissible solutions are obtained. In Section 6.4 the theory is applied to two known normal populations, differing with respect to means, yielding the population linear discriminant function. When the parameters are unknown, they are replaced by estimates (Section 6.5). An alternative procedure is maximum likelihood. In Section 6.6 the probabilities of misclassification by the two methods are evaluated in terms of asymptotic expansions of the distributions. Then these developments are carried out for several populations. Finally, in Section 6.10 linear procedures for the two popUlations are studied when the covariance matrices are different and the parameters are known.
6.2. STANDARDS OF GOOD CLASSIFICATION 6.2.1. Preliminary Considerations In constructing a procedure of classification, it is desired to minimize the probability of misclassification, or, more specifically, it is desired to minimize on the average the bad effects of misclassification. Now let us make this notion precise. For convenience we shall now consider the case ()f only two categories. Later we shall treat the more general case. This section develops the ideas of Section 3.4 in more detail for the problem of two decisions. Suppose an individual is an observation from either population 7TI or population 7T2' The classification of an observation depends on the vector of measurements x' = (x I' ... , x p) on that individual. We set up a rule that if an individual is charactcrized by certain sets of values Jf XI"'" xp that per50n will be classified as from 7T 1 , if other values, as from 7T 2 • We can think of an observation as a point in a p-dimensional space. We divide this space into two regions. If the observation falls in R 1, we classify it as coming from population 7T 1 , and if it falls in R2 we classify it as coming from popUlation 7T 2 • In following a given classification procedure, the statistician can make two kinds of errors in classification. If the individual is actually from 7T I' the
209
6.2 STANDARDS OF GOOD CLASSIFICATION
Table 6.1
Statistician's Decision Population
o
C(2Il)
COI2)
o
statistician can classify him or her as coming from population 7TZ; if from 7Tz, the statistician can classify him or her as from 7T j • We need to know the relative undesirability of these two kinds of misclassification. Let the cost of tlJe first type of misclassification be C(21!) (> 0), and let the cost of mis· classifying an individual from 7T 2 as from 7T \ be COI2) (> 0). These costs may be measured in any kind of units. As we shall see later, it is only the ratio of the two costs that is important. The statistician may not know these costs in each case, but will often have at least a rough idea of them. Table 6.1 indicates the costs of correct and incorrect classification. Clearly. a good classification procedure is one that minimizes in some sense or other the cost of misclassification. 6.2.2. Two Cases of Two Populations We shall consider ways of defining "minimum cost" in two cases. In one casc we shall suppose that we have a priori probabilities of the two populations. Let the probability that an observation comes from population 7T\ be q\ and from population 7T2 be qz (q\ + q~ = 1). The probability properties of population 7T\ are specified by a distribution function. For convenience we shall treat only the case where the distribution has a density, although the case of discrete probabilities lends itself to almost the same treatment. Let the density of population 7T j be Pj(x) and that of 7T2 be p,(x). If we have a region R j of classification as from 7T p the probability of correctly classifying an observation. that actually is drawn from popUlation 7T j is
(1)
P(lll, R) =
f p\(x) dx. R,
where dx = dx j '" dx p , and the probability of misclassification of an observation from 7T \ is
(2)
P(211,R) =
f
pj(x)dx.
R,
Similarly, the probability of correctly c1assifying.an observation from
(3)
7T2
is
210
CLASSIFICATION OF OBSERVATIONS
and the probability of misclassifying such an observation is (4)
P(112, R) =
f P2(X) dx. R,
Since the probability of drawing an observation from 'IT I is ql' the probability of drawing an observation from 'lT 1 and correctly classifying it is q I POll, R); that is, this is the probability of the situation in the upper left-hand corner of Table 6.1. Similarly, the probability of drawing an observation from 'IT I and misclassifying it is q I P(211, R). The probability associated with the lower left-hand corner of Table 6.1 is q2P012, R), and with the lower right-hand corner is q2 P(212, R). What is t.1C average or expected loss from costs of misclassification? It is the sum of the products of costs of misclassifications with their respective probabilities of occurrence: (5)
C(211)P(211, R)ql + C(112)P(112, R)q2·
It is this average loss that we wish to minimize. That is, we want to divide our space into regions RI and R2 such that the expected loss is as small as possible. A procedure that minimizes (5) for given ql and q2 is called a Bayes procedure. In the example of admission of students, the undesirability of misclassification is, in one instance, the expense of teaching a student who will nm complete the course successfully and is, in the other instance, the undesirability of excluding from college a potentially good student. The other case we shall treat is that in which there are no known a priori probabilities. In this case the expected loss if the observation is from 'IT 1 is
(6)
C(211)P(211,R) =r(l,R);
the expected loss if the observation is from (7)
'IT 2
is
C(112)P(112, R) =r(2, R).
We do not know whether the observation is from 'lT 1 or from 'lT2' and we do not know probabilities of these two instances. A procedure R is at least as good as a procedure R* if rO, R) ~ r(1, R*) and r(2, R) s: r(2, R*); R is better than R* if at least one of these inequalities is a strict inequality. Usually there is no one procedure that is better than all other procedures or is at least as good as all other procedures. A procedure R is called admissible if there is no procedure better than R; we shall be interested in the entire class of admissible procedures. It will be shown that under certain conditions this class is the same as the class of Bayes proce-
6.3
CLASSIFICATION INTO ONE OF TWO POPULATIONS
211
dures. A class of procedures is complete if for every procedure outside the class there is one in the class which is better; a class is called essentially complete if for every procedure outside the class there is one in the class which is at least as good. A minimal complete class (if it exists) is a complete class such that no proper subset is a complete class; a similar definition holds for a minimal essentially complete class. Under certain conditions we shall show that the admissible class is minimal complete. To simplify the discussIOn we shaH consider procedures the same if they only differ on sets of probabil-. ity zero. In fact, throughout the next section we shall make statements which are meant to hold except for sets of probability zero without saying so explicitly. A principle that usually leads to a unique procedure is the mininax principle. A procedure is minimax if the maximum expected loss, r(i, R), is a minimum. From a conservative point of view, this may be consideled an optimum procedure. For a general discussion of the concepts in this section and the next see Wald (1950), Blackwell and Girshick (1954), Ferguson (1967), DeGroot (1970), and Berger (1980b).
6.3. PROCEDURES OF CLASSIFICATION INTO ONE OF TWO POPULATIONS WITH KNOWN PROBABILITY DISTRIBUTIONS 6.3.1. The Case When A Priori Probabilities Are Known We now tum to the problem of choosing regions R t and R2 so as to minimize (5) of Section 6.2. Since we have a priori probabilities, we can define joint probabilities of the population and the observed set of variables. The probability that an observation comes from 7T t and that each variate is less than the corresponding component in y is
(1) We can also define the conditional probability that an observation came from a certain popUlation given the values of the observed variates. For instance, the conditional probability of coming from population 7T t , given an observation x, is (2) Suppose for a moment that C(112) = C(21l) = 1. Then the expected loss is
(3)
212
CLASSIFICATION OF OBSERVATIONS
This is also the probability of a misclassification; hence we wish to minimize the probability of misclassification. For a given observed point x we minimize the probability of a misclassification by assigning the population that has the higher conditional probability. If
( 4)
qIPI(X) > q2P2(X) qIPI(x) +q2P2(X) - qIPI(x) +q2P2(X)'
we choose population 7T I • Otherwise we choose popUlation 7T2. Since we minimize the probability of misclassification at each point, we minimize it over the whole space. Thus the rule is
(5)
R 1: qIPI(x) ~ q2P2(X), R 2: qIPI(x)
If qIPl(X) = q2Pix), the point could be classified as either from 7TI or 7T 2 ; we have arbitrarily put it into R I . If ql PI(X) + q2P2(X) = 0 for a given x, that point also may go into either region. Now let us prove formally that (5) is the best procedure. For any procedure R* = (Rj , R!), the probability of misclass;fication is
On the right-hand side the second term is a given number; the first term is minimized if Ri includes the points x such that ql PI(X) - qzpix) < 0 and excludes the points for which ql PI(X) - q2P2(X) > O. If
(7)
i = 1,2.
then the Bayes procedure is unique except for sets of probability zero. Now we notice that mathematically the problem was: given nonnegative constants ql and q2 and nonnegative functions PI(X) and pix), choose regions Rl and R2 so as to minimize (3). The solution is (5). If we wish to minimize (5) of Section 6.2, which can be written
6.3
CLASSIFICATION INTO ONE OF TWO POPUl.ATlONS
213
we choose R j and R2 according to R j : [C(21 1)qljPI(x) ~ [C( 112)q2jPl(X),
(9)
R 2 : [C(21 1)qljPI(x) < [C(112)q2jp2(x),
since C(211)qj and C(112)q2 are nonnegative constants. Another way of writing (9) is R . pj(x) > C(112)q2 I' ['2(X) - C(211)qj ,
(10) pj(x) C(112)q2 R 2 : P2(X) < C(211)qj'
Theorem 6.3.1. If q I a, rd q2 are a priori probabilities of drawing an observation from population ~TI with density PI(X) and 7T2 with density p/x), respectively, and if the cost of misclassifying an observation from 7T I as from 7T 2 is C(21l) and an observation from 7T2 as from 7TI is C(112), then the regions of classification RI and R 2 , defined by (10), minimize the expected cost. If
(11)
i = 1,2.
then the procedure is unique except for sets of probability zero.
6.3.2. The Case When No Set of A Priori Probabilities Is Known In many instances of classification the statistician cannot assign a pnon probabilities to the two populations. In this case we shall look for the class of admissible proeedures, that is, the set of procedures that cannot be improved upon. First, let us prove that a Bayes procedure is admissible. Let R = (R 1, R 2 ) be a Bayes procedure for a given qj, q2; is there a procedure R* = (Rr, Rn such that P(112, R*) ~ p(112, R) and P(211, R*) ~ P(211, R) with at least one strict inequality? Since R is a Bayes procedure,
This inequality can be written (13)
ql[P(211,R) -P(211,R*)j ~q2[P(112,R*) -P(112,R)j.
214
CLASSIFICATION OF OBSERVATIONS
Suppose O 017T 1} = 1. If POI2,R*)=0, then Rr contains only points for which P2(X) = O. Then P(211, R*) = Pr{R! 17T!} = Pr{pc(x) > 017T 1} = 1, and R* is not better than R. Theorem 6.3.2. If Pr{pz<x) = 017T 1} = 0 and Pr{PI(X) = 017T 2} = 0, thom every Bayes procedure is admissible. Now let us prove the converse, namely, that every admissible procedure is
a Bayes procedure. We assume t (14)
I}
PI(X) Pr { P2(X) =k 7T j
=0,
i
= 1,2, 0 ~ k ~ 00.
Then for any ql the Bayes procedure is unique. Moreover, the cdf of Pl(X)/P2(x) for 7Tl and 7T2 is continuous. Let R be an admissible procedure. Then there exists a k such that (15)
P(211,R)
=pr{;~~:~ ~;kl7T!} =P(211,R*),
where R* is the Bayes procedure corresponding to qz/ql = k [Le., q! = 1/(1 + k)]. Since R is admissible, POI2, R) ~ P<112, R*). However, since by Theorem 6.3.2 R* is admissible, POI2, R) ~ P(112, R*); that is, POI2, R) = POI2, R*). Therefore, R is also a Bayes procedure; by the uniqueness of Bayes procedures R is the same as R*. Theorem 6.3.3.
If (14) holds, then every admissible procedure is a Bayes
procedllre. The proof of Theorem 6.3.3 shows that the class of Bayes procedures is complete. For if R is any procedure outside the class, we construct a Bayes procedure R* so that P(211, R) = P(211, R*). Then, since R* is admissible, P( 112. R) ~ P( 112, R*). Furthermor~, the class of Bayes procedures is minimal complete since it is identical with the class of admissible procedures.
6.4
CLASSIFICATION INTO ONE OF TWO NORMAL POPULATIONS
Theorem 6.3.4. complete.
215
If (14) holds, the class of Bayes procedures is minimal
Finally, let us consider the minimax procedure. Let PUIj, qt) = PUIj, R), where R is the Bayes procedure corresponding to qt. PUjj, qt) is a continuous function of ql' P(2j1, qt) varies from 1 to 0 as qt goes from 0 to 1; POj2, qt) varies from 0 to 1. Thus there is a value of qt' say qr, such that P(2j1, qj) = POj2, qi). This is the minimax solution, for if there were lnother proc~dure R* such that max{P(2j1, R*), POj2, R*)} .:::; P(2j1, qf) = P(lj2, qf), that would contradict the fact that every Bayes solution is admissible.
6.4. CLASSIFICATION INTO ONE OF TWO KNOWN MULTIVARIATE NORMAL POPULATIONS Now we shall use the general procedure outlined above in the case of two multivariate normal populations with equal covariance matrices, namely, N(pY),:l) and N(pP), :l), where J.l(i)' = (,Ai), ... , JL~» is the vector of means of the ith population, i = 1,2, and :l is the matrix of variances and covariances of each population. [The approach was first used by Wald (944).] Then the ith density is
The ratio of densities is
(2)
Pt( x) exp[ -!( x - J.l(t»)':l-t (x - fL(l»)] P2(X) = exp[ -!(x- J.l(2»)':l-t(x- fL(2»)]
= exp{ -
H(x -
fL(t»)':l-1 (x - fL(I»)
- (x - fL(2»)':l-t (x - J.l(2»)j). The region of classification into 7Tp R 1, is the set of x's for which (2) is greater than or equal to k (for k suitahly chosen). Sinee the logarithmic function is monotonically increasing, the inequality can be written in terms of the logarithm of (2) as
216
CLASSIFICATION OF OBSERVATIONS
The left-hand side of (3) can be expanded as
(4)
-! [x'l: -I x -
x'l: -I f.l(I) ~
f.l(I)'l: -I x
-
f.l(I)'l: -I f.l(!)
-x'l: -I x + x'l: -I f.l(2) + f.l(2)'l: -I x -
f.l(2)'l: -I f.l(2)j.
By rearrangement of the terms we obtain
The first term is the well-known discriminant function. It is a function of the components of the observation vector. The following theorem is now a direct consequence of Theorem 6.3.1. Theorem 6.4.1. If 7T; has the density (1), i = 1,2, the best regions of classification are giuen by
(6)
R 1:
x'l:-I(J.l(I) - J.l(2») - !(J.l(I)
+
J.l(2»)'l:-I(J.l(1) - J.l(2») ~
log k,
R 2 : x'l: -I (J.l(l) - J.l(2») - HJ.l(I) + J.l(2»)'l: -I (J.l(l) - J.l(2») < log k.
If a priori probabilities ql and q2 are known, then k is giuen by
(7) In the particular case of the two populations being equally likely and the costs being equal, k = 1 and log k = O. Then the region of classification into 7TI is
If we de not have a priori probabilities, we may select log k = c, say, on the basis of making the expected losses due to misclassification equal. Let X be a ranriom obs::rvation. Then we wish to find the distribution of
on the assJmption that X is distributed according to N(J.l(l), l:) and then on the assumption that X is distributed acc~rding to N(f.l(2), l:). When X is distributed according to N(J.l(l), l:), U is normally distributed with mean
(10)
G1U= J.l(l)'l:-l(J.l(l) - J.l(2») - HJ.l(l)
+
= Hf.l(l) - J.l(2»)'l:-I(J.l(I) - J.l(2»)
J.l(2»)'l:-I(J.l(1) - J.l(l»)
6.4
217
CLASSIFICATION INTO ONE OF TWO NORMAL POPULATIONS
and variance (11)
Var l ( U) = .c'1(fL(1) - fL(2»)'I. -I (X - fL(1»)( X - fL(I»), I. -·1 (fL(1) - fL(21)
= (fL(1) - fL(2»)'I.-I(fL(11 - fL(2 1). The Mahalanobis squared distance between N(fL(1), I.) and N(fL(2), I.) is
(12) say. Then U is distributed according to NqtJ"z, ~2) if X is distrihuted according to N(fL(1), I.). If X is distributed according to N(fLm, I.), then
(13)
.c'2U= fL(2)'I.-I(fL(1) -- fL(2») - HfL(l) + fL(2»)'I.-I(fL(l) - fL(21) = !(fL(2) - fLO»)'I. -I (fL(!) - fL(2»)
The variance is the same as when X is distributed according to N(fL(ll. I.) because it dep'!nds only on the second-order moments of X. Thus U is distributed according to N( _ ~~2, ~2). The probability of misclassification if the observation is from 11' I is
and the probability of misclassification if the observation is from
(15)
P(I\2) =
f
c
cc
1
I
I' ,
--e-,(z+,Il.-)-"dz=
&~
fX
-
(c+1Il.')/A
11'2
is
1 --e-'Y-dy. I
•
&
Figure 6.1 indicates the two probabilities as the shaded portions in the tails.
Figure 6.1
218
CLASSIFlCATIONOF OBSERVATIONS
For the minimax solution we choose c so that
Theorem 6.4.2. If the 71"; have densities 0), i = 1,2, the minimax regions of classification are given by (6) where c = log k is chosen by the condition (16) with CUlj) the two costs of nzisclassificatioll. It should be noted that if the costs of misclassification are equal, c = 0 and the probability of misclassification is (17)
f
x
1
_",.,
- - e ,. dy .
.'>/2&
In case the costs of misclassification are unequal, c could be determined to sufficient accuracy by a trial-and-error method with the normal tables. Both terms in (5) involve the vector (18) This is obtained as the solution of ( \9)
by an efficient computing method. The discriminant function x''O is the linear function that maximizes
(20)
[tC\(X'd) - tC'2(X'd)r Var(X'd)
for all choices of d. The numerator of (20) is
the denominator is (22)
d'tC'(X- tC'X)(X- tC'X)'d=d'Id.
We wish to maximize (21) with respect to d, holding (22) constant. If A is a Lagrange multiplier, we ask for the maximum of (23)
6.5
CLASSIFICATION WHEN THE PARAMETERS ARE ESTIMATED
219
The derivatives of (23) with respect to the components of d are set equal to zero to obtain
(24) Since (Jl,
J.l(2», d
is a scalar, say v, we can write (24) as
(25) Thus the solution is proportional to 8. We may finally note that if we have a sample of N from either 7T1 or 7T2' we use the mean of the sample and classify it as from N[J.l(1), (lIN)I] or N[J.l{2), (lIN)I]'
6.5. CLASSIFICATION INTO ONE OF TWO MULTIVARIATE NORMAL POPULATIONS WHEN THE PARAMETERS ARE ESTIMATED 6.5.1. The Criterion of Clarsification Thus far we have assumed that the two popUlations are known exactly. In most applications of this theory the populations are not known, but must be inferred from samples, one from each population. We shall now treat the case in which we have a sample from each of two normal populations and we wish to use that information in classifying another observation as. coming from one of the two populatiolls. Suppose that we have a sample x(1), ... , x~: from N(J.l(l), I) and a sample xi2), ••• , x~~ from N(J.l{2}, I). In one terminology these are "training samples." On the basis of this information we wish to classify the observation x as coming from 7T1 to 7T2' Clearly, our best estimate of J.l(1) is X(1) = I:~'X~1) IN1, of J.l(2) is X(2) = I:~2X~2) I N2 , and of I is S defined by N,
(1)
(N1 +N2 -2)S=
L,
(X~1)_X(1»)(x~l)-x(1»),
a=1
N2
+
L,
(x~)
- X(2») ( x~) - X(2»)'.
a=!
We substitute these estimates for the parameters in (5) of Section 6.4 to obtain
220
CLASSIFICATION OF OBSERVATIONS
The first term of (2) is the discriminant function based on two samples [suggested by Fisher (1936)]. It is the linear function that has greatest variance between samples relative to the variance within samples (Problem 6.12). We propose that (2) be used as the criterion of classification in the same way that (5) of Section 6.4 is used. V1hen the populations are known, we can argue that the classification criterion is the best in the sense that its use minimizes the expected loss in the case of known a priori probabilities and generates the class of admissible procedures when a priori probabilities are not known. We cannot justify the use of (2) in the same way. However, it seems intuitively reasonable that (2) should give good results. Another criterion is indicated in Section 6.5.5. Suppose we have a sample XI' ... ' XN from either 7TI or 7T 2 , and we wish to classify the sample as a whole. Then we define S by N,
(3)
(NI +N2+N-3)S=
L
(X~I)_X(I»)(X~l)-x(l»),
a~1
Nz
+
L
N
(x~)
- x(2))( X~2) - X(2») , --
a~1
L
(xa -x)( Xa - x)',
a~l
where
(4) Then the criterion is
(5) The larger N is, the smaller are the probabilities of misclassification. 6.5.2. On the Distribution of the Criterion Let
(6)
W=X'S-I(X(I) -X(2») - HX(l) +X(2»)'S-I(X(I) -X(2»)
=
[X - HX(I) +X(2»)] 'S-I(X(1) -X(2»)
for random X, X(l), X(2), and S.
6.5
221
CLASSIFICATION WHEN THE PARAMETERS ARE ESTIMATED
The distribution of W is extremely complicated. It depends on the sample sizes and the unknown 6.2 • Let
c I [ X - (NI + N z ) -I (NIX(I) + N z X(2))],
(7)
YI
(8)
Y2 = c2 ( X(I)
=
-
X(2)),
where c i = J(NI +N2 )/(NI +N2 + 1) and C z = JNI N 2 /(N I +N~). Then YI and Y2 are independently normally distributed with covariance matrix I. The expected value of Y2 is c 2 (J.l(l) - J.l(2), and the expected value of YI is c l [N2 /(N I + N 2 )](J.l(I) - J.l(Z) if X is from 7TI and -cl[NI/(N I + N~)](J.ll!) J.l(2) if X is from 7T2' Let Y= (Y I Y2 ) and
(9) Then
(10)
W=
The density of M has been given by Sitgreaves (1952). Anderson (l951a) and Wald (1944) have also studied the distribution of W. If NI = N 2 , the distribution of W for X from 7TI is the same as that of - W for X from 7T 2' Thus, If W ~ 0 is the region of classification as 7T I' then the probability of misclassifying' X when it is from 7T I is equal to the probability of misclassifying it when it is from 7T 2' 6.5.3. The Asymptotic Distribution of the Criterion In the case of large samples from N(J.l(\), I) and N(J.l(2\ I), we can apply limiting distribution theory. Since X(l) is the mean of a sample of Nl independent observations from N(J.l(\), I), we know that plim
(11)
X(l)= J.l(\).
Nl~CC
The explicit definition of (11) is as follows: Given arbitrary positive we can find N large enOugh so that for NI ~ N
(12)
-(I) Pr {Ix i
-
" I. = 1, ... , p } > 1 - s. ILi(1)1 < u,
<5
and s.
222
CLASSIFICATION OF OBSERVATIONS
(See Problem 3.23.) This can be proved by using the Tchebycheff inequality. Similarly,
X(2) =
plim
(13)
fLlZ ) ,
Nz-OC
and plim s= I
(14) as NI
--+ :X:,
N2
--+ 00
or as both N 1 , N2
--+ 00.
plim S - 1 = I
(15)
From (14) we obtain -I,
since the probability limits of sums, differences, products, and quotients of random variatles are the sums, differences, products, and quotients of their probability limits as long as the probability limit of each denominator is different from zero [Cramer (1946), p. 254]. Furthermore, plim
(16)
S-I(X(l)
-XIZ)
=
I-I (J.l11 ) - J.l(2»),
N1.Nz-X
(17) plim
(X(\) + X(2»),S -I (X(\) - X( 2») = (J.l(I) + J.l(2»)'I -1 (J.l(1) - J.l(2»).
N1·N:-.... x
It follows then that the limiting distribution of W is the distribution of U. For sufficiently large samples from 7T 1 and 7T 2 we can use the criterion as if we knew the population exactly and make only a small error. [The result was first given by Wald (1944).]
Theorem 6.5.1. Let W be given by (6) with X(\) the mean of a sample of N1 from N(J.l(\), I), X(2) the mean of a sample of N2 from N(J.l(2), I), and.s the estimate of I based on the pooled sample. The limiting distribution of W as 2 NI --+ 00 and N2 --+ 00 is N(~6.2, 6. ) if X is distributed according to N(J.l(1), I) 2 and is N( - ~6.2, 6. ) if X is distributed according to N(J.l(2), I). 6.5.4. Another Derivation of the Criterion A convenient mnemonic derivation of the criterion is the use of regression of a dummy variate [given by Fisher (1936)]. Let
(18)
(2) _
Ya
-
-N1 N1 + N
2
'
1 N ex = , ... , 2'
6.S
CLASSIFICATION WHEN THE PARAMETERS ARE ESTIMATED
Then formally find the regression on the variates minimize 2
N,
L L
(19)
x~)
[y~) - b' ( x~) - X)]2 ,
i~l a~l
where (20) The normal equations are 2
(21)
IV;
L L
2
(x~)-x)(x~")-x)'b=
i~l a~l
L
IV;
Ly~i)(X~)-x)
i-I a=1
The matrix multiplying b can be written as
(22)
2
N,
i=1
a~1
L L
(x~)-X)(i~)-i)' 2
N,
=L L
(x~)
- X(i»)( x~) - x(i»),
(x~)
- x(i») (x~) - x(i»)'
i=1 a=1
2
N,
L L i - I a=1
Thus (21) can be written as
(23)
223
by choosing b to
224
CLASSIFICATION OF OBSERVATIONS
where
(24)
A=
2
N,
i~l
,,~l
L L
(x~)-i(i»)(x~)-i(i»)'.
Since (i(\) - i(2»), b is a scalar, we see that the solution b of (23) is proportional to S-l(i(l) - i(2»). 6.5.5. The Likelihood Ratio Criterion Another criterion which can be used in classification is the likelihood ratio criterion. Consider testing the composite null hypothesis that x, xp!, ... , x~~ are drawn from N( ....(I),"I) and X\21, ... , x)J~ are drawn from N( ....(2),"I) against the composite alternative hypothesis that X\I), ... , x~~ are drrwn from N( ....(I),"I) and x, X\2l, ... , x)J~ are drawn from N( ....(2" "I), with ....(1" ....(2" and "I unspecified. Under the first hypothesis the maximum likelihood estimators of ....(1), ....(2), and "I are
(25)
Since N,
(26)
L
(x~) - fJ.\I»)(X~l) -
fJ.\ll)' + (x - M'){x - fJ.\Il)'
,,~l
N,
L ,,~1
(x~) -i(l»)(X~l)
-i(l»)' +N,(i(') -
,lV»)(i(I) - fJ.\ll)'
6.5
CLASSIFICATION WHEN THE PARAMETERS ARE ESTIMATED
we can write
(27)
±
225
as
1
1 -"1-N+N+l 1 2
i- -
[A +N+l N (x - x-(1))( x --(I)),] x, I
where A is given by (24). Under the assumptions of the alternative hypothesis we find (by considerations of symmetry) that the maximum likelihood estimators of the parameters are
(28)
N2 x(2) +x N2 + 1 '
"(2) _
""2 -
1 NI + N2 +1
i- _ -"2 -
[A + N N+z 1 ( x - x-(2))( x - x-(21)'] . z
TI-}e likelihood ratio criterion is, therefore, the (N 1 + N z + l)/2th power of
(29) A
I
+ ~(X-i(lI}(X_X(I))'I· NI + 1
This ratio can also be written (Corollary A.3.l) 1+
(30)
~(X_i(2)'A-I(X-i(2) N2 + 1
n
N, (x - x-l') + ---- )' S -I ( x N2 + 1
-'I
Xl- )
where n = NI + N2 - 2. The region of classification into 11" I consists of those points for which the ratio (30) is greater than or equal to a given number K n · It can be written
(31)
R .1l+ I·
:2:
~(x-i(2))'S-I(X-i(21) N2 + 1
Kn[n + N~~1 (X-i(I)'S-I(X-i(I))].
226
CLASSIFICATION OF
OBSERVATlON~
If Kn = 1 + 2eln and NI and N z are large, the region (31) is approximately w(x) ~C.
If we take Kn = 1, the rule is to classify as 1T1 if (30) is greater than I and as 1T 2 if (30) is less than 1. This is the maximum likelihood rule. Let (32)
Z=
HN~~l (x-i(2)'S-I(X-i(2») -
N~~l (X-i(l»'S-I(x-i(l»].
Then the maximum likelihood rule is to classify as 1T1 if Z> 0 and 1T Z if Z < O. Roughly speaking, assign x to 1T1 or 1T2 according to whether the distance to i(l) is less or greater than the distance to i(2). The difference between Wand Z is (33)
w- Z = !
[_I_(x -i(2»,S-I( x -i(2» 2 N2 + 1
-
NI~ 1 (X-i(l»'S-I(X-i(l»],
which has the probability limit 0 as N 1, N2 --> 00. The probabilities of misclassification with Ware equivalent asymptotically to those with Z for large samples. Note that for NI = N 2 , Z = [N1/(N 1+ l)]W. Then the symmetric test based on the cutoff (' = 0 is the same for Z and W. 6.5.6. Invariance
The classification problem is invariant with respect to transformations
(34)
X~I)*
= Bx~1)
+ e,
£1'=
I, ... ,N1,
x~2)*
=
BX~2)
+ e,
£1'=
1, .. . ,N2 ,
x* =Bx+e, where B is nonsingular and e is a vector. This transformation induces the following transformation on the sufficient statistics: (35)
i(l)* = Bi(l) + e,
i(2)* = Bi(2) + e,
x* =Bx+e,
S* =BSB',
with the same transformations on the parameters, ....(1), ....(2), and l'.. (Note that ",r!x = ....(1) or ....(2),) Any invariant of the parameters is a function of
6.6
227
PROBABILITIES OF MISCLASSIFICATION
6,2 = ( ....(1)
-
1-L(2»), 'I -1( ....(1)
-
....(2».
There exists a matrix B and a vector c
such that (36)
....(1)*
= B ....(1) + c = 0,
....(2)*
=B ....(2) +c = (6,,0, ... ,0)',
"I* = B'IB' = [. Therefore, 6.2 is the minimal invariant of the parameters. The elements of M defined by (9) are invariant and are the minimal invariants of the sufficient statistics. Thus invariant procedures depend on M, and the distribution of M depends only on 6,2. The statistics Wand Z are invariant.
6.6. PROBABILITIES OF MISCLASSIFICATION 6.6.1. Asymptotic Expansions of the Probabilities of Misclassification Using W
We may want to know the probabilities of misclassification before we draw the two samples for determining the classification rule, and we may want to know the (conditional) probabildes of misclassification after drawing the samples. As observed earlier, the exact distributions of Wand Z are very difficult to calculate. Therefore, we treat asymptotic expansions of their probabilities as NI and N2 increase. The background is that the limiting distribution of Wand Z is N(!6,2, 6,2) if x is from 'lT 1 and is N( - !6,2, 6,2) if x is from 'lT2' Okamoto (1963) obtained the asymptotic expansion of the distribution of W to terms of order n -2, and Siotani and Wang (1975,1977) to terms of order n- 3 • [Bowker and Sitgreaves (1961) treated the case of Nl =N2.] Let cI>O and CPO be the cdf and density of N(O,1), respectively. Theorem 6.6.1. Nl +N2 - 2),
As NI
-> 00,
N2
-> 00,
and Nil N2
->
a positive limit (n =
1
+ - - 2 [u 3 + 26.u 2 + (p - 3 + 6,2)U + (p - 2)6,] 2N2 6,
1 + 4 n [4u 3 + 46,u 2 + (6p - 6 + 6,2)U + 2(p -1)6,]) +O(n- 2 ), and Pr{ -(W + !6,2)/6,
s ul'IT 2) is
(1) with Nl and N2 interchanged.
228
CLASSIFICATION OF OBSERVATIONS
The rule using W is to assign the observation x to 1Tt if W(x) > c and to if W(x):5 c. The probabilities of miscIassification are given by Theorem 6.6.1 with u = (c - td2)/d and u = -(c + ~d2)/d, respectively. For c = 0, u = - ~d. If- N J = N 2 , this defines an exact minimax r:rocedure [Das Gupta (1965)]. 1T2
Corollary 6.6.1
(2)
pr{W:5 OI1T t , lim NNI n -+00
2
= 1}
=(-~d)+~cfJ(~d)[P~1 +~dl+o(n-I) = pr{w~ 011T2' limoo n-+
ZI = 2
1}.
Note tha"( the correction term is positive, as far as this correction goes; that is, the probability of miscIassification is greater than the value of the normal approximation._ The correction term (to order n - I) increases with p for given d and decreases with d for given p. Since d is usually unknown, it is relevant to Studentize W. The sample Mahalanobis squared distance (3) is an estimator of the population Mahalanobis squared distance d 2 • The expectation of D2 is
[2
(4)
(1 1)1 .
tCD 2 = n - pn - 1 d + P N J + N2
See Problem 6.14. If NI and N2 are large, this is approximately d 2 • Anderson (1973b) showed the following: Theorem 6.6.2.
If NJ/N2
{1
---->
a positive limit as n ---->
3
00,
3)]} +O(n -2) ,
P-1) +n1[U4+ ( P-4" u =(u)-cfJ(u) Nt (UZ--d-
6.6
229
PROBABILITIES OF MISCLASSIFICATION
1(U'2 =<1> () u -cP(u) { N2
p-l) + Ii1[U."4 + (3)]} p- 4' +0(11 _,-). 1
-
-t:,,-
U
Usually, one is interested in u :<; 0 (small probabilities of error). Then the correction term is positive; that is, the normal approximation underestimates the probability of misclassification. One may want to choose the cutoff point c so that one probability of misclassification is controlled. Let ex be the desired Pr{W < CI7T 11. Anderson (1973b, 1973c) derived the following theorem:
Let U o be such that (u o) = ex, and ief
Theorem 6.6.3.
Then as N[
N2
-> 00,
and N[/N2 -> a positive limit,
-> 00,
(8) Then c = Du + !D2 will attain the desired probability ex to within 0(11- C). We now turn to evaluating the probabilities of misclassification after the two samples have been drawn. Conditional on ill). ilcl, and S. the random variable W is normally distributed with conditional mean
(9)
S(WI'lT
j
,
i(l),
X(2l,
s) =
[ ....(i)
-
~(i(l)
+ i lc »)
r
S-1 (.i:(I)
-
il:))
= J.L(i)( x(l), X(2), S) when x is from
(10)
'lTj,
i
= 1,2, and conditional variance
'Y(Wlx(\), X(2), S) = (XII)
-
x(2), S-Il:S-[ (Xii) -
ill»
= 0' 2 ( x(l) , x(2) , S) . Note that these means and variance are functions of the samples with probability limits plim
(11)
p(i)(x(\), X(2),
S) = (_1)j-l
0' 2ex(I), X(l),
S)
N I • N 2 -··OO
plim N 1 ,N 2 -+OO
= /:1 2 .
~~c.
230
CLASSIFICATION OF OBSERVATIONS
For large NI and Nc the conditional probabilities of misclassification are close to the limiting normal probabilities (with high probability relative to Xill • xic ). and S). When c is the cutoff point, probabilities of misclassification conditional on i{l). i(~). and S are (1)( -(I)
(12)
P(211 c i(I)' i(Z) S) = c - J.L
(13)
P(112 c,i(1) i(2) S)=l- c-J.L
"
"
,
[
"
-(2)
x ,x , CT(i(1), i(2), S) (2)( -(1)
S)] , -(2)
x ,x , S)
S)] .
CT(i(1), i(2),
[
In (12) write c as DU I + ~D2. Then the argument of <1>(-) in (12) is ulD I CT + (i(1) - iCC»~' S-I (i(l) - ....(1»1 CT; the first term converges in probability to UI' the second term tends to 0 as NI -> 00, N2 -> 00, and (12) to (u j ). In (13) write c as Du, - ~Dz. Then the argument of <1>(-) in (13) is II, D I if + Crl I) - X(2 »)' S - 1-(X IZ) - ....IZ» I CT. The first term converges in probability to II c and the second term to 0; (13) converges to 1 - (u 2 ). For given i(l), XIZ>, and S the (conditional) probabilities of misclassification (12) and (13) are functions of the parameters ....(1), ....(2), 'I and can be estimated. Consider them when c = O. Then (12) and (13) converge in probability to <1>( - ~~); that suggests <1>( - ~D) as an estimator of (12) and (13). A better estimator is <1>( - ~15), where 15 2 = (n - p - l)D 2 In, which is closer to being an unbiased estimator of 1:,.2. [See (4).] McLachlan (1973, 1974a, 1974b, 1974c) gave an estimator of (12) whose bias is of order n-~; it is
(14)
+ 3in[-D 3 +4(4 P -l).JJ}.
[McLachlan gave (14) to terms of order n -I.] McLachlan explored the properties of these and other estimators, as did Lachenbruch and Mickey (1968). Now consider (12) with c = DU I + }Dz; u l might be chosen to control P(21!) conditional on i(I), x(2), S. This conditional probability as a function of iii). XIC), S is a random variable whose distribution may be approximated. McLachlan showed the following: Theorem 6.6.4. (15)
As
NI -> 00,
,- p(211, DU I + Pr vn (
N z -> 00, and NIIN2
~D', xii), Xl'>, s) - <1>(111) ,
cP(U2)[1U~+IlINlr
=
->
a positive limit,
~x
(p - l)nlNI - ~P 2- i + n/~1 )U I Inhul +nINIJ
)
ui/4] + O(n-2).
6.6
231
PROBABIUTIES OF MISCLASSIFICATION
McLachlan (1977) gave a method of selecting u! so that the probability of one misc1assification is less than a preassigned 8 with a preassigned confidence level 1 - e. 6.6.2. Asymptotic Expansions of the Probabilities of Misclassification Using Z We now tum our attention to Z defined by (32) of Section 6.5. The results are parallel to those for W. Memon and Okamoto (1971) expanded the distribution of Z to terms of order n- 2 , and Siotani and Wang (1975), (1977) to terms of order n -3. As N!
Theorem 6.6.5. limit, (16)
N2
-> 00,
-> 00,
and N!/N2 approaches a posi.'ive
I}
Pr { Z -t:.2!t:,.2 :5 u 1T!
1 + 2N t:.2 [u 3 + t:.u 2 + (p - 3 - t:.2)u - t:.3 - t:.] 2
+ 4~ [4u 3 + 4t:.u 2 + (6p - 6 + t:.2)u + 2( P - 1)t:.] } + O( n- 2), and Pr{ -(Z + ~t:.2)/ t:.:5 UI1T2} is (16) with N! and N2 interchanged.
When c = 0, then u = - !t:.. If N! = N 2, the rule with Z is identical to the rule with W, and the probability of misclassification is given by (2). Fujikoshi and Kanazawa (1976) proved Theorem 6.6.6 (17)
2 pr{ Z-JD :5UI1T!}
= <1>( u) - cfJ( u) {
2~! t:. [u 2 + t:.u -
-
2~2 t:. [u 2 + 2t:.u + p -1 + t:.2]
+
4~ [u 3 + (4p -
(p - 1)]
3)U]} + O(n- 2 ),
232
CLASSIFICATION OF OBSERVATIONS
(18)
pr{ - Z +jD2 = 4l(u) -
+
:$
UJ1T 2}
cP(U){ -
2~1/l [u 2 + 26.u +p -1 + /l2]
2~2/l [u 2 + flu -
1 3 (p -1)] + 4 n [u + (4p -
3)U]}
+ O(n-2).
Kanazawa (1979) showed the following: Theorem 6.6.7.
Let U o be such that 4l(u o) = a, and let
u = Uo + 2 ~1 D
(19)
-
[U6 + Duo -
(p - 1)]
2~ D [U6 +Du o + (p -1)
_D2]
2
Then as N1 ..... 00, N2 .....
00,
and N1 / N2 ..... a positive limit,
(20)
Now consider the probabilities of misclassification after the samples have been drawn. The conditional distribution of Z is not normal; Z is quadratic in x unless N, = N 2 • We do not have expressions equivalent to (12) and (13). Siotani (1980) showed the following: Theorem 6.6.8. (21)
P r {2
As N, .....
00,
N2 .....
00,
and Nd N z ..... a positive limit,
NIN2 P(211,0,i(l),i(2),S)-¢(-~/l) } N, +N2 cP(~/l) :$X
=4l[X-2
::~fv2 {16~,/l[4(P-l)-/l2]
+_1_[4(p_l) +36.2 ] - (P-l)/l}] +O(n-2). 16N2
4n
It is also possible to obtain a similar expression for P(2II, Du, ~D2, i(l), i(2), S) for Z and a confidence interval. See Siotani (1980).
+
6.7
CLASSIFICATION INTO ONE OF SEVERAL POPULATIONS
233
6.7. CLASSIFICATION INTO ONE OF SEVERAL POPULATIONS Let us now consider the problem of classifying an observation into one of several populations. We ~hall extend the comideration of the previous sections to the cases of more than two populations. Let Tr I' ... ,Trm be m populations with density functions PI(X),,,,, p",(x), respectively. We wish to divide the space of observations into m mutually exclusive and exhaustive regions R I , ... , R",. If an obscl vat ion falls into R i , we shall say that it comes from Tr j. Let the cost of misc\assifying an observation from Tr j as coming from Trj be C(jli). The probability of this misclassification is
(1)
PUli, R) =
f p,(x) dx. Rj
Suppose we have a priori probabilities of the populations, ql"'" qm' Then the expected loss is
(2)
(Ill
)
m j~qj j~C(jli)PUli'R) . J~'
We should like to choose R I , ••• , R", to make this a minimum. Since we have a priori probabilities for the populations. we can define the conditional probability of an observation coming from a population given th(; values of the components of the vector x. The conditional probability of th(; observation coming from Tr; is
(3) If we classify the observation as from
(4)
E
Tr j ,
};p;(x)
the expected loss is
;=1 Lk~lqkPk(X)
CUli).
;"j
We minimize the expected loss at this point if we choose j so as to minimize (4); that is, we consider
(5)
1: q;p;(x)CUli) ;=1 ;"j
234
CLASSIFICATION OF OBSERVATIONS
for all j and select that j that gives the minimum. (If two different indices give the minimum, it is irrelevant which index is selected.) This procedure assigns the point x to one of the R j • Following this procedure for each x, we define our regions R I , ••• , Rm. The classification procedure, then, is to classify an observation as coming from 7Tj if it falls in R j • Theorem 6.7.1. If qi is the a priori probability of drawing an observation from population 7Ti with density Pie x), i = 1, ... , m, and if the cost of misclassifying all observation from 7Ti as from 7Tj is C(jli), then the regions of classification, R I , ••• , R m , that minimize the expected cost are defined by assigning x to Rk if m
(6)'
L qiPi( x)C(kli) ;=1 i*k
m
<
I: qiPi( x)C(jli),
j=l, ... ,m,
j,;,k.
;=1
i,,'i
[If (6) holds for all j (j *- k) except for h indices and the inequality is replaced by equality for those indices, then this point can be assigned to any of the h + 1 7T'S,] If the probability of equality between the right-hand and left-hand sides of (6) is zero for each k and j under 7Ti (t::ach i), then the minimizing procedure is unique except for sets of probability zero. Proof We now verify this result. Let //I
(7)
hj(x) =
I: qiPi(x)C(jli). j~l
j,,'i
Then the expected loss of a procedure R is (8) where it(x\ R) = h /x) for x in R j' For the Bayes procedure R* described in the theorem, h(x\R) is h(x\R*) = minj h;Cx). Thw; the difference between the expected loss for any procedure R and for R* is (9)
J[h(xIR)-h(xIR*)]dx=I:j [hj(x)-minhj(x)]dx j
Rj
~o.
Equality can hold only if h;Cx) = minj h;Cx) for x in R j except for sets of probability zero. •
6.7
CLASSIFICATION INTO ONE OF SEVERAL POPULATIONS
235
Let us see how this method applies when C(jli) = 1 for all i and j, i *- j. Then in Rk m
(10)
m
L qiPi(X) i=1 iotk
Subtracting
"[,'['=1, iot
< L qiPi(X), i=1 iotj
k.jqiPi(X) from both sides of 00), we obtain
j*-k.
(11)
In this case the point x is in Rk if k is the index for which qiPi(X) is a maximum; that is, 7Tk is the most probable population. Now suppose that we do not have a priori probabilities. Then we cannot define an unconditional expected loss for a classificadon procedure. However, we can define an expected loss on the condition that the observation comes from a given population. The conditional expected loss if the observation is from 7Ti is m
(12)
L C(jli)P(jli,R) =r(i, R). j=1 joti
A procedure R is at least as good as R* if r(i, R) :0; r(i, R*), i = 1, ... ,m; R is better if at least one inequality is strict. R is admissible if there is no procedure R* that is better. A class of procedures is complete if for every procedure R outside the class there is a procedure R* in the class that is better. Now let us show that a Bayes procedure is admissible. Let R be a Bayes procedure; let R* be another procedure. Since R is Bayes, m
(13)
m
Lqir(i,R):o; Lqjr(i,R*). i=1
Suppose ql > 0, q2 > 0, r(2, R*) < r(2, R), and rei, R*) :0; rei, R), i = 3, ... , m. Then m
(14)
ql[r(l,R)-r(l,R*»):O; Lqi[r(i,R*)-r(i,R») <0, i=2
and r(1, R) < rO, R*). Thus R* is not better than R. Theorem 6.7.2. ble.
If qi > 0, i = 1, ... , m, then a Bayes procedure is admissi-
236
CLASSIFICATION OF OBSERVATIONS
We shall now assume that CUlj) = 1, i *- j, and Pdp/x) = 011T) = O. The latter condition implies that all p/x) are positive on the same set (except fer a set of measure 0). Suppose q; = 0 for i = 1, ... , t, and q; > 0 for i = t + 1, ... , m. Then for the Bayes solution R;, i = 1, ... , I, is empty (except for a set of probability 0), as seen from (11) [that is, Pm(x) = 0 for x in R;l. It follows that rei, R) = Ej * ;PUI i, R) = 1 - PUI i, R) = 1 for i = 1, ... , t. Then (R r + J , ••• , Rm) is a Bayes solution for the problem involving Pr+J(x), ... ,Pm(x) and qr+J, ... ,qm' It follows from Theorem 6.7.2 that no procedure R* for which P(i Ii, R* ) = 0, i = I, ... , I, can be better than the Bayes procedure. Now consider a procedure R* such that Rj includes a set of positive probability so that POll, R*) > O. For R* to be better than R, (15)
P(ili, R) =
f p;(x) dx Ri
~P(ili, R*)
=
f
p;(x) dx,
i=2, ... ,m.
Rj
In such a case a procedure R** where Rj* is empty, i = 1, ... , t, Rj* = Ri, = t + 1, ... , m - 1, and R;,,* = R;" U Rj U '" URi would give risks such that
i
(16)
P(ili; R**)
= 0,
P( iii, R**)
= P( iii, R*)
i = 1, ... , t, ~ P(ili, R),
- i=t+l, ... ,m-l,
P(mlm,R**) >P(mlm,R*) ~P(mlm,R).
Then Ri:I"'" R;,,*) would be better than (R r + I , ... , Rm) for the (m -t)decision problem, which contradicts the preceding discussion. Theorem 6.7.3.
IfCCilj) = 1, i *- j, and Pdp/x) = OI1T) = 0, then a Bayes
procedure is admissihle.
The converse is true without conditions (except that the parameter space is finite). Theorem 6.7.4.
Every admissible procedure is a Bayes procedure.
We shall not prove this theorem. It is Theorem 1 of Section 2.10 of Ferguson (1967), for example. The class of Bayes procedures is minimal complete if each Bayes procedure is unique (for the specified probabilities). The minimax procedure is the Bayes procedure for which the risks ·are equal.
6.8
CLASSIFICATION INTO ONE OF SEVERAL NORMAL POPULATIONS
237
There are available general treatments of statistical decision procedures by Wald (1950), Blackwell and Girshick (1954), Ferguson (1967), De Groot (1970), Berger (1980b), and others.
6.8. CLASSIFICATION INTO ONE OF SEVERAL MULTIVARIATE NORMAL POPULATIONS We shall now apply the theory of Section 6.7 to the case in which each population has a normal distribution. [See von Mises (1945).] We assume that the means are different and the covariance matrices are alike. Let N(tLli ), I) be the distribut ion of 7Tj • The density is given by (1) of Section 6.4. At the outset the parameters are ·assumed known. For general costs with known a priori probabilities we can form the m functions (5) of Section 6.7 and define the region R j as consisting of points x such that the jth function is minimum. In the remainder of our discussion we shall assume that the costs of misclassification are equal. Then we use the functions
If a priori probabilities are known, the region R j is defined by those x satisfying (2)
k=1, ... ,m.
k*j.
Theorem 6.8.1. If qj is the a priori probability of drawl:ng all observation from 7Tj = N(tL(i), I), i = 1, ... , m, and if the costs of miscla~sification are equal, then the regions "of classification, R l' ... , R m' that minimize the expected cost are defined by (2), where ujk(x) is given by (1).
It should be noted that each ujk(x) is the classification function related to the jth and kth populations, and Ujk(X) = -ukj(x). Since these are linear functions, the region R j is bounded by hyperplanes. If the means span an (m - I)-dimensional hyperplane (for example, if the vectors J.L(i) are linearly independent and p ~ m -1), then R j is bounded by m - 1 hyperplanes. In the case of no set of a priori probabilities known, the region R j is defined by inequalities
(3)
k=l, ... ,m,
k*i.
238
CLASSIFICATION OF OBSERV ATlONS
The constants c k can be taken nonnegative. These sets of regions form the class of admissible procedures. For the minimax procedure these constants are determined so all P(il i, R) are equal. We now show how to evaluate the probabilities of correct classification. If X is a random observation, we consider the random variables (4)
Here Vjl = - Vij" Thus we use m(m -1)/2 classification functions if the means span an (m - I)-dimensional hyperplane. If X is from 7Tj , then ~'i is distributed according to N(i~}i' ~]), where (5)
The covariance of Vjl and Vjk is (6 )
To determine the constants cj we consider the integrals
where fl is the density of
Vji,
i = 1,2, ... , m, i"* j.
Theorem 6.8.2. If 7Ti is N(Il-(i>, I) and the costs of miscLassification are equal. then the regions of classification, R l' ... , R m' that minimize the maximum conditional expected loss are defined by (3), where ujk(x) is given by (I). The constants c j are determined so that the integrals (7) are equal.
As an example consider the case of m = 3. There is no loss of generality in taking p = 2, for the density for higher p can be projected on the two-dimensional plane determined by the means of the t'\fee populations if they are not collinear (i.e., we can transform the vector x into U 12 ' u 13 , and p - 2 other coordinates, where these last p - 2 components are distributed independently of U 12 and u 13 and with zero means). The regions R j are determined by three half lines as sho\\'n in Figure 6.2. If this procedure is minimax, we cannot move the line between R \ and R 2 rearer ( J.I-?), J.I-~l», the line between R2 and R3 nearer (J.I-\21, J.l-j21), and the line between R) and R\ nearer (J.I-\31, J.I-~» and still retain the equality POll, R) = P(212, R) = P(313, R) without leaving a triangle that is not included in any region. Thus, since the regions must exhaust the space, the lines must meet in a point, and the equality of probabilities determines ci - cj uniquely.
6.8
CLASSIFICATION INTO ONE OF SEVERAL NORMAL POPULATIONS
239
----------------~~~--~---------------%1
R3
Figure 6.2. Classification regions.
To do this in a specific case in which we havc numerical values for the components of the vectors ....(1), ....(2), ....(3), and the maUx I, we would consider the three (:5.p + 1) joint distributions, each of two ll;/s (j "* n. We could try the values of c j = 0 and, using tables [Pearson (1931)] of the bivariate normal distribution, compute POli, R). By a trial-and-error method we could obtain c j to approximate the above condition. The preceding theory has been given on the assumption that the parameters are known. If they are not known and if a sample from each population is available, the estimators of the parameters can be substituted in the definition of uij(x), Let the observations be Xli), ... , X~~ from N( ....(i), I), i = 1, ... ,m. We estimate ....(i) by
(8) and I by S defined by
(9)
C~N;-m)s= i~ "~1 (X~)-i(i»)(X~)-i(i»),.
Then, the analog of uij(x) is (10)
wij(x)
=
[x- !(i(i) +i U»]' S-I(X(i)
-xU».
If the variables 'above are random, the distributions are different from those of Uij . However, as Ni -> 00, the joint distributions approach those of Uij . Hence, for sufficiently large sa'l1ples one can use the theory given above.
240
CLASSIFICATION OF OBSERVATIONS
Table 6.2 Mean Measurement Stature (x I) Sitting height (X2) Nasal depth (X3) Nasal height (x 4 )
164.51 86.43 25.49 51.24
160.53 81.47 23.84 48.62
158.17 81.16 21.44 46.72
6.9. AN EXAMPLE OF CLASSIFICATION INTO ONE OF SEVERAL MULTIVARIATE NORMAL POPULATIONS Rao (1948a) considers three populations consisting of the Brahmin caste ('lT l ), the Artisan· caste ('lT2)' and the KOIwa caste ('lT3) of India. The measurements for each individual of a caste are stature (Xl)' sitting height (x 2), nasal depth (X3)' and nasal height (x 4). The means of these variables in
the three popUlations are given in Table 6.2. The matrix of correlations for all the ):opulations is 1.0000 0.5849 [ 0.1774 0.1974
(1)
0.5849 1.0000 0.2094 0.2170
0.1774 0.2094 1.0000 0.2910
0.1974] 0.2170 0.2910 . 1.0000
The standard deviations are CT1 = 5.74, CT2 = 3.20, CT3 = 1.75, (14 = 3.50. We assume that each population is normal. Our problem is to divide the space of the four variables XI' x 2 , x 3 , X 4 into three regions of classification. We assume that the costs of misclassification are equal. We shall find 0) a set of regions under the assumption that drawing a new observation from each population is equally likely (ql = q2 = q3 = t), and (ij) a set of regions such that the largest probability of misclassification is minimized (the minimax solution). We first compute the coefficients of "I-I (J.L(I) - J.L(2» and "I-I (J.L(I) - J.L(3». Then "I-I (J.L(2) - J.L(3» = "I-l(J.L(I) - J.L(3» - "I-l(J.L(I) - J.L(2». Then we calculate ~(J.L(i) + J.L(j»'"I -I (J.L(i) - J.L(j». We obtain the discriminant functions t unC x) = - 0.0708x 1 + 0.4990x 2 + 0.3373x 3 + 0.0887x 4
(2)
u 13 (x) =
0.0003x l + 0.3550x 2 + 1.1063x 3
u 23 ( x) =
0.0711x l
-
-
43.13,
+ 0.1375x 4 -
62.49,
0.1440x 2 + 0.7690x 3 + 0.0488x 4 - 19.36.
tOue to an error in computations, Rao's discriminant functions are incorrect. I am indebted to Mr. Peter Frank for assistance in the computations.
241
6.9 AN EXAMPLE OF CLASSIFICATION
Table 6.3
Population of r
U
Means
Standard Deviation
'lT l
u l2
1.491 3.487
1.727 2.641
0.8658
1.491 1.031
1.727 1.436
-0.3894
3.487 1.031
2.MI 1.436
0.7983
u I3 'lT2
U 21 U2.1
'IT)
U)l
u32
Correlation
~lhe
other three functions are UZl(x) = -u l2 (x), U3l (X) = -un(x), and = -U 23 (X). If there are a priori probabilities and they are equal, the best set of regions of classification are R l : u l2 (x):?: 0, Lln(x):?: 0; R 2: u 21 (x) ~ 0, u 23 (X):?: 0; and R3: u 3j (x):?: 0, U32(X):?: O. For example, if we obtain an individual with measurements x such that u l2 (x):?: 0 and u l3 (x):?: 0, we classify him as a Brahmin. To find the probabilities of misclassification when an individual is drawn from population 'lTg we need the means, variances, and covariances of the proper pairs of u's. They are given in Table 6.3. t The probabilities of misc1assification are then obtained by use of the tables for the bivariate normal distribution. These probabilities are 0.21 for 'lT1, 0.42 for 'lT 2 , and 0.25 for 'lT3' For example, if measurements are made on a Brahmin, the probability that he is classified as an Artisan or Korwa is 0.21. The minimax solution is obtained by finding the constants c l ' c 2 • and c, for (3) of Section 6.8 so that the probabilities of misclassification are equal. The regions of classification are
U3zCX)
R'j: u l2 (x):?:
(3)
0.54,
Ul3 (X):?:
0.29;
R'2: U21 (X):?: -0.54,
U23 (X):?: -0.25;
R~: U31 (X):?:
U32 (x):?:
--0.29,
0.25.
The common probability of misc1assification (to two decimal places) is 0.30. Thus the maximum probability of misclassification has been reduced from 0.42 to 0.30.
tSome numerical errors in Anderson (1951a) are corrected in Table 6.3 and (3).
242
CLASSIFICATION OF Ol:lSERVATlONS
6.10. CLASSIFICATION INTO ONE OF TWO KNOWN MULTIVARIATE NORl\-LU POPULATIONS WITH UNEQUAL COVARIANCE MATRICES
6.10.1. Likelihood Procedures
Let 7TI and 7T2 be N(ILl !), II) and N(ILl2 ),I 2hvith IL(I) When the parameters are known, the likelihood ratio is
(1)
* IL(2) and II * 1 2,
PI(X) II21!exp[ -Hx- IL(l))'I1I(x- IL(I))] P2(X) = IIII~exp[ -~(x- IL(2))'I 21(x- IL(2))]
= II21:lIIII-texp[Hx- IL(2))'I 2i (x- IL(2))
-Hx- IL(I))'I1I(x- IL(I))]. The logarithm of (1) is quadratic in x. The probabilities of misclassification are difficult to compute. [One can make a linear transformation of x so that its covariance matrix is I and the matrix of the quadratic form is diagonal; then the logarithm of (J) has the distribution of a linear combination of noncentral X 2-variables plus a constant.] When the parameters are unknown, we consider the problem as testing the hypothesis that x, X\I" ... , x~~ are observations from N(IL(1), II) and X\21, ...• xJJ! are observations from N(IL(2),I 2) against the alternative that (Ii are 0 bserva t'Ions f rom N( IL(I) '''I '<') an d X,XI(2) , .•. ,XN2 (2) XI(II •..•. x"', are 0 bservations from N(ILI2 l, I). Under the first hypothesis the maximum likelihood estimators are fi\11 = (NIX(I) +x)/(NI + 1), fi\2) =X(2),
~(l) -_NI _l_[A + 1 I + ~(_-(I»)( NI + 1 x x x
"I
_(1)'] x ,
(2)
where Ai = ~~:'I(X:,i) -i(i»(X~i) -xU»', i = 1,2. (See Section 6.5.5.) Under the second hypothesis the maximum likelihood estimators are fi~) = x(l), fJ..~) = (N-c iI2 ) +x)/(N2 + 1),
(3)
~
"2
(2) -
1
N2 + 1
[A
2
z + N zN + 1 ( x - x-(2»( x - x
-(2),] .
6.10
POPULATIONS WITH UNEQUAL COVARIANCE MATRICES
243
The likelihood ratio criterion is
(4)
[1 + (x - x(2»),A zl (x - X(2»] t(N,+1) [1 + (x -x(1»'Ajl(x _X(l»]i
111(2) 1tN'1 12(2) 1t(N,+ I) 11 1(1) 1teN, +])1 1 2( 1) 1tN,
(N1 + 1) t(N, +l)p NiN'PIA 2 1~ NiN,P(N2 + 1)t(N,+I)PIA11~·
The observation xis classified into 7T1 if (4) is greater than 1 and into 7T2 if (4) is less than 1. An alternative criterion is to plug estimates into the logarithm of (1). Use
to classify into 7T I if (5) is large and into 7T 2 if (5) is small. Again it is difficult to evaluate the probabilities of misclassification. 6.10.2. Linear Procedures
'*
The best procedures when :I I :I 2 are not linear; when the parameters are known, the best procedures are based on a quadratic function of the vector observation x. The procedure depends very much on the assumed normality. For example, in the case of p = 1, the region for classification with one popUlation is an interval and for the other is the complement of the interval -that is, where the observation is sufficiently large or sufficiently small. In the bivariate case the regions are defined by conic sections; for exam1)le, tpe region of classification into one population might be the interior of an ellipse or the region between two hyperbolas. In general, the regions are defined by means of a quadratic function of the observations which is not necessarily a positive definite quadratic form. These procedures depend very much on the assumption of 1I0rmaiity and especially on the shape of the normal distribution far from its center. For instance, in the univariate case cited above the region of classification into the first population is a finite interval because the density of the first population falls off in either direction more rapidly than the density of the second because its standard deviation is smaller. One may want to use a classification procedure in a situation where the two popUlations are centered around different points and have different patterns of scatter, and where one considers multivariate normal distributions to be reasonably good approximations for these two populations near their centers and between their two centers (though not far from the centers, where the densities are small). In such a case one may want to divide the
244
CLASSIFICATION OF OBSERVATIONS
sample space into the two regions of classification by some simple CUIVe or surface. The simplest is a Iip.e or hyperplane; the procedure may then be termed linear. Let b ("* 0) be a vector (of p components) and c a scalar. An obseIVation x is classified as from the first population if b' x ~ c and as from the second if b' x < c. We are primarily interested in situations where the important difference between the two populations is the difference between the centers; we assume /-l(I)"* /-l(2) as well as 1: 1 "* I2' and that II and I2 are nonsingular. When sampling from the ith population, b' x has a univariate normal distribution with mean G(b' XI1T) = b' /-l(i) and variance
The probability of misclassifying an obseIVation when it comes from the first population is
(7)
The probability of misclassifying an obseIVation when it come3 from the second population is
(8)
It is desired to make these probabilities small or, equivalently, to make the arguments
(9)
b'/-l(1) -
C
Y\ = (b'llb)t'
c - b'/-l(2) Y2 = (b'12b)t
large. We shall consider making Y\ large for given Y2' When we eliminate c from (9), we obtain
(10)
245
dO POPULATIONS WITH UNEQUAL COVARIANCE MATRICES
"here "{ = /-l(I) - /-l(2). To maximize YI for given Y2 we differentiate YI with :espeet to b to obtain
~~I = ["{.-y~(b''I~b)-~'Izbl(b'''Ilb)-~
:11)
Y2(b''I~b)~l(b''Ilb) -~'Ilb.
- [b'''{ -
[f we let (12)
11 =
(13)
t
b'''{ - Y2(b''I 2 b)! b''2.\b
v
=
Y2
b ''I 2 b''
2
then (11) set equal to 0 is
(14) Note that (13) and (14) imply (12). If there is a pair tl' t z , and a vector b satisfying (12) and (13), then c is obtained from (9) as
(15) Then from (9), (12), and (13) b'/-l(l) - (tzb''I~b
(16)
~
YI=
+ b'/-l!2I)
,
=tl~·
Now consider (14) as a function of t(O ::; t ::;1). Let t I = t and t ~ = 1 - (; then b=(tl'Il +t 2 'I 2 )-I,,{. Define VI =tlvb''Ilb and v~=t~Vb''Izb. The derivative of vf with respect to t is (17)
t)'I~]-I'Y
:tt 2 "{'[t'Il + (1- t)'IZ]-1 '1 1[1'1 1 + (1 -
l 'Il[t'Il + (1 - t)'I 2r "{ l l - t 2 "{'[t'Il + (1- t)'I 2r ('I1 - 'I~)[t'Il + (1 - t)'Izr l
= 2t"{'[t'Il + (1 - t)'I 2r
·'II[t'Il + (1- t)'IJ
-1"1
2
- t "{'[t'Il + (1- t)'I 21-1'II[t'Il + (1- t)'I 21-1
.('1 1 - 'I 2)[t'Il + (1-t)'I zl··I"{ =t"{'[t'Il + (1 -t)'I z r
l
{'I 2 [t'Il + (1 -t)'Izrl'Il
+'II[t'Il + (1-t)'I 2 r by the following lemma.
l
I
'I Z }[t'Il + (l-t)'IJ- 'Y
2.t6
CLASSIFICATION OF OBSERVATIONS
Lemma 6.1 11.1.
If'll and
'1 2
are positive definite and t I > 0,1 2 > 0, then
is posiriul! definire. Proof The matrix (1S) is
•
(19)
Similarly dvildt < 0. Since VI :2: 0, Vz :2: 0, we see that VI increases with I from at (= to V-y''lll-y at 1=1 and V 2 decreases from V-Y''l"2 1 -y at I = to at I = 1. The coordinates VI and V z are continuous functions of t. For given y" 0:$)"2:$ V-Y''l2 1 -y, there is a t such that Y2 =vz =t2Vb''l2b and b satisfies (14) for (I = t and 12 = 1 - t. Then' YI = VI = 11..[b'l;b maxi-
° ° ° °
°
mizes )"1 for tLat value of Y2. Similarly given Yi' :$YI:$ V-Y''lll-y, there is a ( such that .h = VI = IIVb''l Ib and b satisfies (14) for tl = t and t2 = 1 - t,
°
and Y2 = v, = Idb''l2b maximizes Ye. Note that YI :2: 0, Y2 :2: implies the errors of misclassification are not greater than ~. We now argue that the set of YI' Y2 defined this way correspond to admissible linear procedures. Let XI' X2 be in this set, and suppose another proceJun: defined by ZI,Z2 were better than xI,X Z , that is, XI :$ZI' X2 :$Z2 with at least one strict inequality. For YI = ZI let y~ be the maximum Y2 among linear procedures; then ZI = YI' Z2 :$Y~ and hence Xl :$YI' X2 :$Y~· However, this is possible only if XI = YI' x 2 = yi, because dYlldY2 < 0. Now we have a contraJiction to the assumption that ZI' Z2 was better than XI' x 2 • Thus x I' x: corresponds to an admissible linear procedure.
Use of Admissible Linear Procedures Given t I and (2 such that (1'1 I + (2'1 2 is positive definite, one would compute the optimum b by solving the linear equations (15) and then compute c by one of (9). U~ually (I and t2 are not given, but a desired solution is specified in another way. We consider three ways. Minimization of One Probability of Misciassijication for a Specijied Probability of the Other Suppose we arc given Y2 (or, equivalently, the probability of misclassification when sampling from the second distribution) and we want to maximize YI (or. equivalently, minimize the probability of misclassification when sampling from the first distribution). Suppose Y2 > (i.e., the given probability of misclassification is less than ~). Then if the maximum YI :2: 0, we want to find (2 = 1 - (I such that Y2 = (Z(b''l2b)~, where b = [(I'll + t 2 'lzl-1-y. The solu-
°
6.10
POPULATIONS WITH UNEQUAL COVARIANCE MATRICES
247
tion can be approximated by trial and error, since Y2 i~ an incre~sing function of t 2, For t2 = 0, Y2 = 0; and for I~ = 1, Y2 = (b'12b)1 = (b''Y)! = ('Y ':£'2 1 'Y), where I2b = 'Y. One could try other values of t2 successively by solving (14) and inserting in b'I2b until 12(b'12b)1 agreed closely enough with the desired Y2' [YI > 0 if the specified Y2 < ('Y'12 1 'Y)!.] The Minimax Procedure The minimax procedure is the admissible procedure for which YI = Yz. Since for this procedure both probabilities of correct classification are greater than ~, we have YI =Y2 > 0 and II> 0,1 2 > O. We want to find t (=t l = I-t 2) so that
(20)
o= Y~ -
Y~ = t 2b'Ilb - (1- t)2b'12b
=b'[t 2 1
1-
(l-t)2 I2 ]b.
Since Y~ increases with t and Y~ decreases with increasing t, there is one and only one solution to (20), and this can be approximated by trial and error by guessing a value of t (0 < t < 1), solving (14) for b, and computing the quadratic form on the right of (20). Then another I can be tried. An alternative approach is to set Yl = Y2 in (9) and solve for c, Thf;n the common value of YI = Y2 is (21)
b''Y
and we want to find b to maximize this, where b is of the form (22) with 0 < t < 1. When II = I2, twice the maximum of (21) is the squared Mahalanobis distance between the populations. This suggests that when II may be unequal to I2' twice the maximum of (21) might be called the distance between 'the populations. Welch and Wimpress (1961) have programmed the minimax procedure and applied it to the recognition of spoken sounds. Case of A Priori Probabilities
Suppose we are given a priori probabilities, ql and q2, of the first and second populations, respectively. Then the probability of a misclassification is
248
CLASSIFICATION OF OBSERVATIONS
which we want to minimize. The solution will be an admissible linear procedure. If we know it involves Yl ~ 0 and Y2 ~ 0, we can substitute 1 Yl = t(b'I.1b)t and Y2 = (1- t)(b'I. 2b)t, where b = [tI. 1 + (1- t)I. 2 1- y, into (23) and set the derivative of (23) with respect to t equal to 0, obtaining dYI dyz QI4>(Yl) dt + Qz 4>(Yz) dt = 0,
(24)
where 4>(u) = (21T)- te- til'. There does not seem to be any easy or direct way of solving (24) for t. The left-hand side of (24) is not necessarily monotonic. In fact, there may be several roots to (24). If there are, the absolute minimum will bc found by putting the solution into (23). (We remind the reader that the curve of admissible error probabilities is not necessary convex.) Anderson and Bahadur (1962) studied these linear procedures in general, induding Yl < 0 and Yz < O. Clunies-Ross and Riffenburgh (1960) approached the problem from a more geometric point of view.
PROBLEMS 6.1. (Sec. 6.3) Let 1Ti be N(IL, I.), i = 1,2. Find the form of the admissible dassification procedures. 6.2. (Sec. 6.3) Prove that every complete class of procedures includes the class of admissible procedures. 6.3.
~Sec. 6.3) Prove that if the class of admissible procedures is complete, it is minimal complete.
6.4. (Sec. 6.3) The Neymull-Peursoll!ulldumelllullemmu states that of all tests at a given significance level of the null hvpothesis that x is drawn from Pl(X) agaimt alternative that it is drawn from P2(X) the most powerful test has the critical region Pl(x)/pix) < k. Show that the discussion in Section 6.3 proves this result. 6.5. (Sec. 6.3) When p(x) = n(xllL, ~) find the best test of IL = 0 against IL = IL* at significance level 8. Show that this test is uniformly most powerful against all alternatives fL = CfL*, C > O. Prove that there is no uniformly most powerful test against fL = fL(!) and fL = fL(2) unless fLO) = qJ.(2) for some C > O. 6.6. (Sec. 6.4) Let P(21!) and POI2) be defined by (14) and (5). Prove if - ~~2 < C < ~~z, then P(21!) and POI2) are decreasing functions of ~. 6.7. (Sec. 6.4) Let x' = (x(1)', X(2),). Using Problem S.23 and Problem 6.6, prove that the class of classification procedures based on x is uniformly as good as the class of procedures based on x(l).
249
PROBLEMS
6.S. (Sec. 6.5.1) Find the criterion for classifying irises as Iris selosa or Iris versicolor on the basis of data given in Section 5.3.4. Classify a random sample of 5 Iris virginica in Table 3.4. 6.9. (Sec. 6.5.1) Let W(x) be the classification criterion given by (2). Show that the T 2-criterion for testing N(fL(I), l:) = N(fL(2),l:) is proportional to W(ill,) and W(i(2».
6.10. (Sec. 6.5.1) Show that the probabilities of misclassification of assumed to be from either 7T I or 7T 2) decrease as N increases.
Xl •... '
x'"
(all
6.11. (Sec. 6.5) Show that the elements of M are invariant under the transformation (34) and that any function of the sufficient statistics that is invariant is a function of M. 6.12. (Sec. 6.5)
Consider d'x(i). Prove that the ratio
N]
L: a=l
6.13. (Sec. 6.6)
N~
"'J
L:
(d'x~I)-d'i(l)r+
.,
(d'x~2)-d'i(2)r
a= I
Show that the derivative of (2) to terms of order
11- I
is
{I
I p-2 p ']} . -t/>(ztl) "2+n1[P-1 ~+-4--81l-
6.14. (Sec. 6.6) Show IC D2 is (4). [Hint: Let l: = I and show that IC(S -11:l = I) = [n/(n - p - 1)]1.] 6.15. (Sec. 6.6.2)
Show 2
Z_lD Pr { -y}---:S U =
I} 7T 1
2
-
Pr {Z-{tl --Il---:S U
t/>(1l){_1_2 [1l3 2Nltl
I} 7T I
+ (p - 3)l/- 1l2 l/ + ptll
+ _1_2 [u 3 + 2tl1l2 + (p - 3 + 1l2)1l- tl3 +ptll 2N26.
+
4~ [3u 3 + 4tl1l2 + (2p -
3 + 1l2 ) l/ + 2(p - I )Ill } + O(n -2).
be N(fL(i),l:), i = L ... , m. If the fL(i) arc on a line (i.e .. show that for admissible procedures the Ri are defined by parallel planes. Thus show that only one discriminant function llrk(X) need be used. .
6.16. (Sec. 6.8) fLU) = fL
Let
+ v i l3),
7Ti
250
CLASSIFICATION OF OBSERVATIONS
6.17. (Sec. 6.8) In Section 8.8 data are given on samples from four populations of skulls. Consider the first two measurements and the first three sample>. Construct the classification functions uij(x), Find the procedure for qi = Nj( N, + N z + N,). Find the minimax procedure. 6.18. (Sec. 6.10) Show that b' x = c is the equation of a plane that is tangent to an ellipsoid of constant density of 7T, and to an ellipsoid of constant density of 7T2 at a common point. 6.19. (Sec. 6.8) Let x\j), ... ,x~) be observations from NCJL(i),I.), i= 1,2,3, and let x be an observation to b~ classified. Give explicitly the maximum likelihood rule. 6.20. (Sec. 6.5)
Verify (33).
CHAPTER 7
The Distribution of the Sample Covariance Matrix and the Sample Generalized Variance
7.1. INTRODUCTION The sample cOvariance matrix, S = [lj(N -l)n:::=l(X" - i)(x" - i)', is an unbiased estimator of the population covariance matrix l:. In Section 4.2 we found the density of A = (N - 1)S in the case of a 2 X 2 matrix. In Section 7.2 this result will be generalized to the case of a matrix A of any order. When l: =1, this distribution is in a sense a generalization of the X2-distribution. The distribution of A (or S), often called the Wishart distribution, is fundamental to multivariate statistical analysis. In Sections 7.3 and 7.4 we discuss some properties of tJoe Wishart distribution. The generalized variance of the sample is defined as ISI in Section 7.5; it is a measure of the scatter of the sample. Its distribution is characterized. The density of the set of all correlation coefficients when the components of the observed vector are independent is obtained in Section 7.6. The inverted Wishart distribution is introduced in Section 7.7 and is used as an a priori distribution of l: to obtain a Bayes estimator of the covariance matrix. In Section 7.8 we consider improving on S as an estimator of l: with respect to two loss functions. Section 7.9 treats the distributions for sampling from elliptically contoured distributions.
An Introduction to Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0-47\-3609\-0 Copyright © 2003 John Wiley & Sons. Inc.
251
252
COVARIANCE MATRIX mS1RlAUTION; GENERALIZED VARIANCE
7.1. THE WISHART DISTRIBUTION
We shall obtain the distribution of A = I:~_I(Xa - XXXa - X)', where X I' ... , X N (N > p) are independent, each with the distribution N(p., l:). As was shown in Section 3.3, A is distributed as I::_I Za Z~, where n = N - 1 and Zp ... , Zn are independent, each with the distribution N(O, l:). We shall show that the density of A for A positive definite is
(1)
IAI Hn-p-I) exp( - ~tr l: -IA)
We shall first consider the case of l: = 1. Let
(2)
Then the elements of A = (a i ) are inner products of these n-component vectors, aij = V;Vj' The vectors VI"'" vp are independently distributed, each according to N(O, In). It will be convenient to transform to new coordinates according to the Gram-Schmidt orthogonalization. Let WI = VI'
i = 2, ... ,p.
(3)
We prove by induction that Wk is orthogonal to Wi' k < i. Assume WkWh = 0, k h, k, h = 1, ... , i-I; then take the inner product of W k and (3) to obtain wi Wi = 0, k = 1, ... , i - 1. (Note that Pr{lIwili = O} = 0.) Define tii = II Wi II = ';W;Wi' i = 1, ... , p, and tij = v;w/llwjll, f= 1, ... , i-I, i = 2, ... , p. Since Vi = I:~_I(ti/llwjll)wj'
"*
min(h. i)
(4)
a hi = VhV; =
E .1'-1
t h/ ij ·
If we define the lower triangular matrix T = (tij) with ti; > 0, i = 1, ... ,p, and tij = 0, i <j, then
(5)
A=TT'.
Note that tii' j = 1, ... , i-I, are the first i - I coordinates of Vi in the coordinate system with WI"'" Wi - 1 as the first i -1 coordinate axes. (See Figure 7.1.) The sum of the other n - i + 1 coordinates squared is IIv;1I2_. I:;.:\tt=ti~=IIWiIl2; Wi is the vector from v; to its projection on w1,,,,,w;_1 (or equivalently on VI"'" Vi-I)'
253
7.2 THE WISHART DISTRIBUTION
2
Figur~
7.1. Transformation of cOf'rdinates.
Lemma 7.2.1. Conditional on wl"",W i _ 1 (or equivalently on vl, ... ,Vi _ I ), til"'" ti,i-I and ti~ are independently distributed; t ij is distributed according to N(O,1), i > j; and ti~ has the .(2-distribution with n - i + 1 degrees of freedom.
Proof The coordinates of Vj referred to the new orthogonal coordinates with VI"'" Vi -I defining the first coordinate axes are independently normally distributed with means I) and variances 1 (Theorem 3.3.1). ti~ is the sum • of th.: coordinates squared omitting the first i - 1.
Since the conditional distribution of til"'" tii does not depend on Vi - I ' they are distributed independently of til' t 21> t 22' ... , ti _ I. i - I '
VI' ... ,
Corollary 7.2.1. Let ZI"'" Zn (n ~p) be independently distributed, each according to N(O, I); let A = I::~IZaZ~ = IT', ""here t ij = 0, i <j, and tii > 0, i = 1, ... , p. Then til' t 21 , . .. , t pp are independently distributed; t ij is distributed according to N(O, 1), i > j; and t,0 has the X 2-distribution with n - i + 1 degrees offreedom. Since tii has density 2 - i(n -i-I) t n - i e- l,2/ fl !(n + 1 - i)], the joint density of tji' j = 1, ... ,.i, i = 1, ... , p, is
n p
(6)
H
n-i
,.tii
,exp
1T,(,-J)2,n-1
(_.!."i
2'-j~1
t2) ij
r[ ±en + 1 -
i) 1
254
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCS
Let C be a lower triangular matrix (c;j = 0, i <j) such that I = CC ' and c;; > O. The linear transformation T* = CT, that is, i
( 7)
t*'J
=
L C;ktkj'
i "2j,
k=j
i <j,
=0, can be written
x
0 C22 x
f*/II
x
t;p
X
til til tiz
C ll
X
0 0
0 0 0
tll
C Z2
0 0 0
x
X
c pp
0
tpl
X
X
X
cpp
tpp
t21 t22
(8)
where x denotes an element, possibly nonzero. Since the matrix of the transformation is triangular, its determinant is the product of the diagonal elements, namely, nf= I cl;. The Jacobian of the transformation from T to T* is the reciprocal of the determinant. The density of T* is obtained by substituting into (6) t;; = t~/c;; and P
(9)
i
L L
t,~.
= tr IT'
;= I j= I
= tr T* T* 'C' -I C- I = tr T* T* 'l-I = tr T* 'l- IT* , and using nf=lc~ = ICIIC'I = Ill. Theorem 7.2.1. Let ZI' ... ' Zn (n "2p) be independently distributed, each according to N(O, I); let A = I Z" Z~; and let A = T* T* " where t7J = 0, i <j, and t~ > o. Then the density of T* is
I::=
(10)
7.2
255
THE WISHART DISTRIBUTION
We can write (4) as ahi = E~~l tt/'!} for h ~ i. Then
(11)
aa hi at*kl = 0, =0,
k>h, k=h,
I>i;
that is, aa h ;/ atkl = 0 if k, I is beyond h, i in the lexicographic ordering. The Jacobian of the transformation from A to T* is the determinant of the lower triangular matrix with diagonal elements
(12)
(13)
h
> i,
The Jacobiar. is therefore 2pnf:'lt~P+I-i. The Jacobian of the transfurmation from T* to A is the reciprocal. Theorem .7.2.2. Let Zl"'" Zn be independently distributed, each according to N(O, I,). The density of A = E~~l ZaZ~ is
(14) for A positive definite, and 0 otherwise.
Corollary 7.2.2. Let Xl"'" X N (N > p) be independently distributed, each according to N(p., I,). Then the density of A = E~~ I(Xa - XXXa - X)' is (14) for n = N-1. The density (14) will be denoted by w(AI I" n), and the associated distribution will be termed WeI, n). If n < p, then A does not have a density, but its distribution is nevertheless defined, and we shall refer to it as WeI, n). Corollary 7.2.3. Let Xl" .. , X N (N > p) be independently distributed, each according to N(p., I). The distribution of S = (1/n)E~~ l(Xa - X)(X a - X)' is W[(1/n)I, n], where n = N - l. Proof S has the distribution of E~~I[(1/rn)Za][(1/rn)Za]', where (1/ rn)ZI"'" (1/ rn)ZN are independently distributed, each according to N(O,(1/n)I). Theorem 7.2.2 implies this corollary. •
256
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
The Wishart distribution for p = 2 as given in Section 4.2.1 was derived by Fisher (1915). The distribution for arbitrary p was obtained by Wishart (1928) by a geometric argument using Vi •...• vp defined above. As noted in Section 3.2. the ith diagonal element of A is the squared length of the ith vector. a ii = V; Vi = IIvi 11 2 • and the i.jth off-diagonal element of A is the product of the lengths of Vi and vi and the cosine of the angle between them. The matrix A specifies the lengths and configuration of the vectOrS. We shall give a geometric interpretation t of the derivation of the density of the rectangular coordinates tii' i "2:.j. when I = I. The probability element of t11 is approximately the probability that IIvllllies in the intervrl t11 < IIvIIi < t 11 + dt 11' This is tile probability that VI falls in a sphericdl shell in n dimensions with inner radius t II and thickness dt 11' In this region. the density (2'7T)- in exp( - tV/IV I) is approximately constant. namely. (2'7T)- in exp( - ttll ). The surface area of the unit sphere in n dimensions is C(n) '7 2'7T in /ntn) (Problems 7.1-7.3). and the volume of the spherical shell is :lpproximately C(n)t~I-1 dtll' The probability element is the product of the volume and approximate density. namely. (15) The probability element of ti\ •...• li,i-I.IU given vI ..... Vi_ 1 (Le .• given 11'1 •••• ' Wi-I) is approximately the probability that Vi falls in the region for which t;1
(16)
l)
t ) '7T -W-I)tn-i ii exp (_l"i 2£"i-1 ii r[Hn+l-i)] dti\· .. dt ii · 2
Then the product of (15) and (16) for i = 2•...• P is (6) times dt 11 ... dt pp . This ar alysis, which exactly parallels the geometric derivation by Wishart [and later by Mahalanobis. Bose. and Roy (1937)], was given by Sverdrup In the first edition of this book, the derivation of the Wishart distribution and its geometric interpretation were in terms of the nonorthogonal vectors Vi' .... vp-
t
257
7.2 THE WISHART DISTRIBUTION
(1947) [and by Fog (1948) for p = 3). Another method was used by Madow (1938), who drew on the distribution or corrdation coefficients (for! = [) obtained by Hotelling by considering certain partial correlation coefficients. Hsu (1939b) gave an inductive proof, and Rasch (1948) gave a method involving the use of a functional equation. A dilferent method is to obtain the characteristic function and invcrt it, as was done by Ingham (1933) and by Wishart and Bartlett (1933). Cramer (1946) verified that the Wishart distribution has the characteristic function of A. By means of alternative matrix transformations Elfving (1947). Mauldon (1955), and Olkin and Roy (1954) derived the Wishart distribution via the Bartlett decomposition; Kshirsagar (1959) based his derivation on random orthogonal transformations. Narain (1948), (J 950) and Ogawa (1953) us\.;d a regression approach. James (1954), Khatri and Ramachandran (1958), and Khatri (1963) applied different methods. Giri (J 977) used invariance. Wishart (1948) surveyed the derivations up to that date. Some of these methods are indicated in the problems. The relation A = TT' is known as the Bart/ell decomposition [Bartlett (1939»), and the (nonzero) elements of T were termed rectangular coordinates by Mahalanobis, Bose, and Roy (1937). Corollary 7.2.4
(17)
I· -f IBI(-~(p+l)e-LrB
p
dB =
,,1'(1'-11/.
8>0
n r[t -
~(i - I)].
(~l
Proof Here B > 0 denotes B positive definite. Since (14) is a density. its integral for A > 0 ;s 1. Let I = I, A = 2B (dA = 2 dB), and 11 = 2t. Then the fact that the integral is I is identical to (17) for t a half integer. However. if we derive (14) from (6), we can let n be any real number greater than p .- 1. In fact (17) holds for complex t such that .JJft > P - 1. (/lIt means the real part of t.) • Definition 7.2.1.
The multivariate gamma function is I'
(18)
rp(t) = "p(p-Ilj. nr[t- ~(i-I)]. i~1
The Wishart density can be written
(19)
_ ...'0
........... \ . " . . . . . , . . . . . . . . . . . . . . . " • • •
~'~"
................ • -....1 •• ,
........... ___ " ....... _
7.3. SOME PROPERTIES OF THE WISHART DISTRIBUTION 7.3.1. The Characteristic Function The characteristic function of the Wishart distribution can be obtained directly from the distribution of the obseIVations. Suppose ZI"'" Zn are distributed independently, each with density (1) Let n
(2)
A=
L
ZaZ~.
a~l
Introduce the p X P matrix 0 = «(}i) with (}ij = (}ji' The characteristic function of All' A 22 ,· •• , App,2AI22AI3, ... ,2Ap_l,p is (3)
tC exp[ i tr( A0) 1= tC exp ( i tr = tC exp(i tr
a~l ZaZ~0 )
t
,,~l
= tCexp(i
Z~0Za)
"t Z~0Za).
It follows from Lemma 2.6.1 that
where Z has the density (1). For 0 real, there is a real nonsingular matrix B such that
(5) (6)
B,!,-IB=I, B'0B =D,
where D is a real diagonal matrix (Theorem A.:.2 of the Appendix). If we let z =By, then (7)
.c'exp(iZ'0Z) = tCexp(iY'DY) p
= tC
n exp( id
j~l
jJ
~2)
P
=
n tC exp(idl?)
j~l
II
7.3
259
SOME PROPERTIES OF THE WISHART DISTRIBUTION
by Lemma 2.6.2. Th~ jth factor in the second product is tC exp(idjjY/), ",:,here ~. has the distribution N(O, 1); this is the characteristic fun~tion of the x 2-distribution with one degree of freedom, namely (1 - 2idjj )- ., [as can be proved by expanding exp(idjjy}) in a power series and integrating term by terml. Thus
(8)
tC exp( iZ'0Z) =
nP (1 - 2id
_!
jj )
I
'= II - 2iDl-;:
j=l
since 1- 2iD is a diagonal matrix. From (5) and (6) we see that
(9)
II-2iDl =IB'r I B-2iB'0BI
= IB'(l:-1 - 2i0)BI
= IB'I'Il:-1 -2i01·IBI 2 = IBI '1l:- 1- 2i01, IB'I ·Il: -II 'IBI = III = 1, and IBI2 = 1/ Il: -II. Combining the above resuits, we obtain
(10) It can be shown that the result is valid provided (0'l(U ik - 2iOjk » is positive definite. In particular, it is true for all real 0. It also holds for l: singular.
Theorem 7.3.1. If ZI"'" Zn are independent, each with distribution N(O, l:), then the characteristic function of All"'" A pp ' 2A I2 , ... , 2Ap-I. p' where (A ij ) =A = E~=IZaZ~, is given by (10). 7.3.2. The Sum of Wishart Matrices Suppose the Ai' i = 1,2, are distributed independently according to W(l:, n), respectively. Then Al is distribute.d as E~'= I Za Z~, and A2 is distributed as E~'::,2+IZaZ~, where Zp ... ,Zn,+1I2 are independent, each with distribution N(O,l:). Then A=AI +A2 is distributed as E~=lZaZ~, where n=n l +n2' Thus A is distributed according to W(l:, n). Similarly, the sum of q matrices distributed independently, each according to a Wishart distribution with covariance l:, has a Wishart distribution with covariance matrix l: and number of degrees of freedom equal to the sum of the numbers of degrees of freedom of the component matrices.
260
COY ARIANCE MATRIX DISTRIBUTION; :;ENERALIZEO VARIANCE
Theorem 7.3.2. If AI>"" Aq are independently distributed with Ai dis-·~ tributed according to W(I, n), then
(11) is distributed according to W( I,
r:,
= 1 n).
7.3.3. A Certain Linear Transformation We shah frequently make the transformation (12)
A=CBC',
where C is a nonsingular p X P matrix. If A is distributed according to W(I, n), then B is distributed according to W( ~, n) where
(13) This is proved by the following argument: Let A = r::=!ZaZ~, where Z!, ... , Zn are independently distributed, each according to N(O, I). Then Ya = C- I Zu is distributed according to N(O, q-.). However, n
(14)
B=
L:
n
YaY~=C-1
a=!
is distributed according to the transformation (12), is
(15)
L: ZaZ~C'-1 =C-1AC- 1 a=!
W(~,
I
n). Finally, Ia(A)/ a(B)I, the Jacobian of
a(A) = w(B,~,n) = IBI~(n-p-J)I:lI-in
Ia(B)
w(A,I,n)
I
I
IAI;:(n-p-I)I~I;:n
= modIClp+l.
Theorem 7.3.3. Thelacobian of the transformation (12) from A to B, where A and B are symmetric, is modi CI P +1 • 7.3.4. Marginal Distribntions If A is distributed according to W(I, n), the marginal distribution of any arbitrary set of the elements of A may be awkward to obtain. However, the marginal distribution of some sets of elements can be found easily. We give some of these in the following two theorems.
7.3
SOME PROPERTIES OF THE WISHART DISTRIBUTION
Theorem 7.3.4. columns,
(16)
261
Let A and 'I be partitioned into q and p - q rows and
An)
A22 '
If A is distributed according to W('I,I1), then W('III' n).
All
is distribllted according to
Proof A is distributed as L:_IZaZ~, where the Za are independent, each with the distribution N(O, 'I). Partition Za into subvectors of q and p - q components, Za = (Z~I)I, Z~2)1)1. Then ZP), ... , Z!,l) are independent, each with the distribution N(O, 'Ill), and All is distributed as L:_IZ~I)Z~I)', which has the distribution W('I II' n). •
Theorem 7.3.5. Let A and 'I be partitioned into PPP2, ... ,Pq rows and columns (PI + ... +P q = p),
( 17)
If 'I ij = 0 for i"* j and if A is distn'buted according (0 W('I, n), thell All' A 22 , ... , Aqq are independently distributed and Ajj is distributed according (0 W('I ji , n). Proof A is distributed as L:_IZaZ~, where ZI"",Zn are independently distributed, each according to N(O, 'I). Let Za be partitioned
(18)
as A and 'I have been partitioned. Since 'Iij = 0, the sets Z\ll, ... , Z~I), ... , Z\q), ... , z~q) are independent. Then A II = L:_ I Z~})Z~\)I, .... Aqq = z~q)Z~'1)1 are independent. The rest ol Theorem 7.3.5 follow'S from I • Theorem 7.3.4.
L:_
262
COVARIANCE MATRIX DISTRIBUTiON; GENERALIZED VARIANCE
7.3.5. Conditional Distributions
In Section 4.3 we considered estimation of the parameters of the conditional distribution of X( I) given X(2) = X(2). Application of Theorem 7.2.2 to Theorem 4.3.3 yields the following theorem:
row;
Theorem 7.3.6. Let A and "I be partitioned into q and p _. q and columns as in (16). If A is distributed according t(' W("I, n), the distribution ,Jf A 11 . 2 = All -A12A221A21 is W("I ll . 2 , n - p + q), n ?.p - q.
Note that Theorem 7.3.6 implies that A ll . 2 is independent of
A22
and
A 12 A 221 regardless of "I.
7.4. COCHRAN'S THEOREM Cochran's theorem [Cochran (1934)] is useful in proving that certain vector quadraric j"onlls are distributed as sums of uector squares. It is a statistical statement of an algebraic theorem, which we shall give as a lemma. Lemma 7.4.1.
If the N X N symmetric matrix Ci has rank ri , i = 1, ... , m,
and
( 1) then
(2)
is a necessary and sufficient condition for there to exist an N X N orthogonal maniT P such that for i = 1, ... , m
(3)
PCiP'
=
(i
o I
o
I::·:h
where 1 is of order r" the upper left-hand 0 is square of order (which is vacuous for i = 0, and the lower-right hand 0 is square of order I:j~i+l rj (which is vacuous for i = m). Proof The necessity follows from the fact that (I) implies that the sum of (3) over i = 1, ... , m is IN' Now let us prove the sufficiency; we assume (2).
7.4
263
COCHRAN'S THEOREM
There exists an orthogonal matrix If such that IfCjPj is diagonal with diagonal elements the characteristic roots of C j • The number of nonzero roots is 'j' the rank of C j , and the number of 0 roots is N -Ii- We write
o I1 j
(4)
o where the partitioning is according to (3), and 11 j is diagonal of order is possible in view of (2). Then
rj'
This
(5)
Since the rank of (5) is not greater than L:7~ I r,. - rj = N - r j , which is the sum of the orders of the upper left-hand and lower right-hand I's in (5), the rank of 1- 11 j is 0 and 11 j = I. (Thus the rj nonzero roots of Cj are 1, and Cj is positive semidefinite.) From (4) we obtain
o (6)
I
o where B j con~ists of the we obtain
'j
columns of P; corresponding to I in (6). From (1)
III
(7)
1=
L
=P'P,
BjB; = (Bl> B 2 , .. ·, Bm)
j~l
B'm
We now state a multivariate analog to Cochran's theorem. Theorem 7.4.1. Suppose Y1, ••• , YN are independently distributed, each according to N(O, l:). Suppose the matrix (c~f3) = C; used in forming N
(8)
Q,. =
L "./3~1
c~f3 Ya Y~ ,
i=l, ... ,m,
264
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
is of rank ri , and suppose m
(9)
N
E YaY~.
EQi= i~1
a~l
Then (2) is a necessary and sufficient condition for QI"'" Qm to be independently distributed with Qi having the distribution W(l:, rJ It follows from (3) that Ci is idempotent. See Section A.2 of the Appendix. This theorem is useful in generalizing results from the univariate analysis of variance. (See Chapter 8.) As an example of the use of this theorem, let us prove that the mean of a sample of size N times its transpose and a multiple of the sample covariance matrix are independently distributed with a singular and a nonsingular Wishart distribution, respectively. Let YJ , ••• , YN be independently distributed, each according to N(O, l:). We ~hall use the matrices CJ = (c~JJ) = (lIN) and C2 = (C~2J) = [oa/3 - (lIN)]. Then N
(10)
QJ =
E
1
__
NYaY~ =NYY/,
a,/3-1
(11)
Q2 =
E (oa/3 - ~ )
Y" If;
",/3~1
N
E YaY~ -NIT/ a~1
N
=
E (Y" - r)(Y" - r)/, a~l
and (9) is satisfied. The matrix C J is of rank 1; the matrix C2 is of rank N - 1 (since the rank of the sum of two matrices is less than or equal to the sum of the ranks of the matrices and the rank of the second matrix is less than N). The conditions of the theorem are satisfied; therefore QJ is distributed as ZZ/, where Z is distributed according to N(O, l:), and Q2 is distributed independently according to W(l:, N - 1). Anderson and Styan (l982) have given a survey of pfLlofs and extensions of Cochran's theorem.
7.5. THE GENERALIZED VARIANCE 7.5.1. Definition of the Generalized Variance One multivariate analog of the variance a 2 of a univariate distribution is the covariance matrix l:. Another multivariate analog is the scalar 1l:1, which is
7.5
265
THE GENERALIZED VARIANCE
called the generalized variance of the multivariate distribution [Wilks (932); see also Frisch (1929)]. Simillrly, the generalized variance of the sample of vectors Xl"'" XN is
(1)
lSI
=\N~l
E(Xa-X)(Xa-Xr\.
a~1
In some sense each of these is a measure of spread. We consider them here· because the sample generalized variance will recur in many likelihllod ratio criteria for testing hypotheses. A geometric interpretation of the sample generalized variance comes from considering the p rows of X = (Xl>"" x N ) as P vectors in N-dimensional space. In Section 3.2 it was shown that the rows of
(2)
(Xl
-x, ... , x N -x) =X -xe',
where e = (1, ... ,1)', are orthogonal to the equiangular line (through the origin and e); see Figure 3.2. Then the entries of
(3)
A = (X-xe')(X -xe')'
are the inner products of rows of X - xe ' . We now define a parallelotope determined by p vectors VI"'" vp in an n-dimensional space (n ~p). If P = 1, the parallelotope is the line segment VI' If P = 2, the parallelotope is the parallelogram with VI and v 2 as principal edges; that is, its sides are VI' V2, VI translated so its initial endpoint is at v~. and v 2 translated so its initial endpoint is at ~'I' See Figure 7.2. If p = 3. the parallelotope is the conventional parallelepided with VI' II", and V) as
Figure 7.2. A parallelogram.
266
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
principal edges. In general, the parallelotope is the figure defined by the principal edges L'I .... ' £'1" It is cut out by p pairs of parallel (p - 1)dimensional hyperplanes, one hyperplane of a pair being spanned by p - lof L'I •..•• L'p and the other hyperplane going through the endpoint of the n:maining vector. Theorem 7.5.1.
If V =
(VI"'"
va/lime of [he parallelotope with
1,'1'),
VI"",
thell the square of the p-dimensional (IS principal edges is IV'V!.
vI'
Proot: If p= 1, then Iv'V! = 1,,'1 V I =lIvI1I2, which is the square of the one-dimensional volume of V I' If two k-dimensional parallelotopes have bases consisting of (k - I)-dimensional parallelotopes of equal (k - 1)dimensional volumes and equal altitudes, their k-dimensional volumes are equal lsi nee the k-dimensional volume is the integral of the (k - 1)dimensional volumesl. In particular, the volume of a k-dimensional parallelotope is equal to the volume of a parallelotope with the same hase (in k - 1 dimensions) and same altitude with sides in the kth direction orthogonal to the first k - I dirt:ctions. Thus the volume of the parallelotope with principal edges L' I' ... , I· k • say Pk , is equal to the volume of the parallelotope with principal edges /'1.·· .• £'k - I ' say Pk - I ' times the altitude of Pk over Pk'-l; that is, (4) It follows (by induction) that
By tht: construction in Section 7.2 the altitude of Pk over Pk ' l is tkk = IIwk II; thaI is. II.! is the distance of 1', from the (k - n-dimensional space spanned by £'I . . . . '£'k-I (or wl' ... 'W k _ I ). Hence Vol(J~,)=llL'~llkk' Since IV'V! = I TT'I = nf'~ I t~, the theorem is proved. • We now apply this theorem to the parallelotope having the rows of (2) as principal edges. The dimensionality in Theorem 7.5.1 is arbitrary (but at least p). Corollary 7.5.1. The square of the p-dimensional volume of the parallelotope with the rows of (2) as principal edges is IAI, where A is given by (3). Wc shall see later that many multivariate statistics can be given an interpretation in terms of these volumes. These volumes arc analogous to distances that arise in special cases when p = 1.
75
267
THE GENERALIZED V ARt \NCE
We now consider a geometric interpretation of IAI in terms of N points in p-space. Let the columns of the matrix (2) be YI"'" YN' representing N points in p-space. When p = 1, IAI = LaYta' which is the sum of 3quarcs of the distances from the points to the origin. In general IAI is the sum of squares of the volumes of all parallelotopes formed by taking as principal edges p vectors from the set YI"'" YN' We see that Lyfa
IAI = LYp-I,aYla
(6)
LY""Yla
LY~a
LYlaYp-I,a
LYI/3Yp/3 f3
LY;-I,a
LYp-I,/3Yp/3 f3
LYp"Y,,-I.u
LY;/3 f3
LYlaYp-I,a
YI/3Y p/3
LY;-I,a
Yp-I,/3Y p/3
LYpaYp-I,a
Y;f3
=L LYp-I,aYla f3 a LYpaYla
by the rule for expanding determinants. [See (24) of Section A.l of the Appendix.] In (6) the matrix A has been partitioned into p - 1 and 1 columns. Applying the rule successively to the columns, we find N
(7)
IAI
=
L al •...•
IY;(t}Yjll}
I·
a p =l
By Theorem 7.5.1 the square of the volume of the parallelotope with < ... < 11" as principal edges is
Y"fl'"'' Y-Yr' 1'1
(8) where the sum on f3 is over (1'1"'" 1'p), If we now expand this determinant in the manner used for IAI, we obtain
(9)
268
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
7.~
where the sum is for each f3j over the range ()'I"'" )'p)' Summing (9) over all different sets ()'I < .. ' )'p)' we obtain (7). (lYi/3jYj/3jl = 0 if two or more f3j are equal.) Thus IA I is the sum of volumes squared of all different parallelotopes formed by sets of p of the vectors y" as principal edges. If we replace y" by x" - i, we can state the following theorem:
tt 7 c;
Theorem 7.5.2. Let lSI be defined by 0), where X"""XN are the N vectors of a sample. Then lSI is proportional to the sum of squares of the volumes of all the different parallelotopes formed by using as principal edges p vectors with p of XI"'" X N as one set of endpoints and i as the other, and the factor of proportionality is 1j( N - l)p. The population analog of lSI is II.I, which can also be given a geometric interpretation. From Section 3.3 we know that
(10) if X is distributed according to N(O, I.); that is, the probability is 1 - a that X fall hsid..: the ellipsoid
(11) The v'Jlume of this ellipsoid is C(p)1 I.11[ x;(a)]W jp, where C(p) is defined in Problem 7.3. 7.5.2. Distribution of the Sample Generalized Variance The distribution of lSi is the same as the distribution of IAI j(N - l)P, where A = E:= I Z" Z~ and ZI"'" Zn are distributed independently, each according to N(O, I.), and n = N - 1. Let Z" = CY", a = 1, ... , n, where CC' = I.. Then YI , ••• , Yn are independently distributed, each with distribution N(O, I). Let (12)
B=
n
n
,,=1
a=1
E Y"Y~= E C-IZ"Z~(C-I)' =C-IA(C- I );
then IAI = ICI·IBI·IC'I = IBI·II.I. By the development in Section 7.2 we see that IBI has the distribution of nf=lti~ and that t;l, ... ,t;p are independently distributed with X 2-distributions. The distribution of the generalized variance ISI of a sample I.) is the same as the distribution of II.I j(N -l)p times the product of p independent factors, the distribution of the ith factor being the X2-distribution with N - i degrees offreedom.
Theorem 7.5.3.
Xl"'" X N from N(p.,
i
269
THE GENERALIZED VARIANCE
lSI has the distribution of 1'lI·x~_I/(N-1). If p=2, lSI has I'll X~-l . X~_zl(N - 1)2. It follows from Problem 7.15 or .37 that when p = 2, lSi has the distribution of l'll( XiN_4)2 /(2N - 2f. We If p= 1,
Ie distribution of
an write
I'll XX~_1 XX~_2 x .. · XX~_p.
13)
IAI
f p
= 2r, then IAI is distributed as
=
14) ii01ce the hth moment ofax 2-variable with m degrees of freedom is + h)/fC!m) and the moment of a product of independent variables .s the product of the moments of the variables, the hth moment of 1041 is ~hf(!m
(15) l'llhn {2J[H~-i) .+hl} i~l
rh(N -I)]
.+hl
=2hpl'llh f1f_lr[H,N-i) f1f_lrh( N -I)]
r
[!(N- 1) + h]
- 2" PI'll h -,-,P_2....,.-_ _-.-"rp[HN - 1)]
Thus p
(16)
tCIAI=I'lIO(N-i). i-I
where
r(iAI) is the variance of IAI.
7.5.3. The Asymptotic Distribution of the Sample Generalized Variance Let IHI /n P = Vl(n) X V 2 (n) X ... X ~(n), where the V's are independently distributed and nV;(n)=:tn2_p+j. Since Xn2_p+1 is distributed as L~:~+jwa2. wfiere the Wa are independent, each with distribution N(O, 1), the central limit theorem (applied to Wa2 ) states that
(13)
p-i nV;(n)-(n--p+i) =rnV;(n)-l+nJ2(n-pf-i)
fiV1-P:i
is asymptotically distributed according to N(O, 1). Then rn [V;( /I) - 1] is asymptotically distributed according to N(O, 2). We now apply Theorem 4.2.3.
270
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
We have
Vl(n)] U(n) =
(19)
IBI /n P = w = f(u 1, ••• , up) =
:
( ~(n)
U 1U 2 ...
,
up, T = 21, af! auil"ab = 1, and rjJ~TrjJb
= 2p. Ihus
.
(20) is asymptotically distributed according to N(O, 2 p).
Theorem 7.5.4. Let S be a p X P sample covariance matrix with n degrees of freedom. Then (I sl / Il: I - 1) is asymptotically normally distributed with mean 0 and variance 2p.
rn
7.6. DISTRIBUTION OF THE SET OF CORRELATION COEFFICIENTS WHEN THE POPULATION COVARIANCE MATRIX IS DIAGONALIn Section 4.2.1 we found the distribution of a single sample correlation when the corresponding population correlation was zero. Here we shall find the density of the set rij , i <j, i, j = 1, ... , p, when Pij = 0, i <j. We start with the distribution of A when ~ is qiagonal; that is, W[(O"jj/)j), nJ. The density of A is
(1) since
(2)
Il:l =
0"11
0
0
0"22
0 0
P
=nuiio i=1
0
0
O"pp
We make the transformation
(3)
( 4)
i +j,
•
271
7.6 DISTRIBUTION OF SET OF CORRElATION COEFFICIENTS
The Jacobian is the product of the Jacobian of (4) and that of (3) for aii fixed. The Jacobian of (3) is the determinant of a p(p -l)/2-order diagonal Since each particular subscript k, matrix with diagonal elements say~ appearf in the set rij (i <j) p - 1 times, the Jacobian is
,;a:; va;;. p
(5)
J=
naj/p-O. jQ1
IT we substitute from (3) and (4) into w[A!(uii 8ij ), n] and multiply by (5), we
obtain as the joint density of {aii} and {rij }
!
I ra:ra::r.. I VUii VUjj
(6)
IJ
P
I exp(--2EP1- la··/a:··) P !
n aii
. crti!nrp (!2 n )
2~npn
i=l
i-I
n
= !r ij !!
I
,
since
(7) where rii = 1. In the ith term of the product on the right-hand side of (6), let aii /(2ui ) = ui ; then the integral of this term is
(8)
ooa tn - I exp(-!a .. /a: .. ) 00 (ii 2" " d .. = ( !n-I -U'd., = f(! n ) 10 2tnut!' a" 10 u i e U 2 "
by definition of the gamma function (or by the fact that aiil U ii has the x2-density'vith n degrees of freedom). Hence the density of rij is
(9) Theorem 7.6.1.
fP(!n )!rij! t
N[P.,(Uii8ij)], then the density of the sample correlation coefficients is given by
(9) where n =, N - 1.
272
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
7.7. THE INVERTED WISHART DISTRIBUTION AND BAYES ESTIMATION OF THE COVARIANCE MATRIX 7.7.1. The Inverted Wishart Distribution As indicated in Section 3.4.2, Bayes estimators are usually admissible. T!le
calculation of Bayes estimators is facilitated when the prior distributions of the parameter is chosen conveniently. When there is a sufficient statistic, there will exist a family of prior distributions for the parameter such that the posterior distribution is a member of this family; such a family is called a conjugate family of distributions. In Section 3.4.2 we saw that the normal family of priors is conjugate to the normal family of distributions when the covariance matrix is given. In this section we shall consider Bayesian estimation of the covariance matrix and estimation of the mean vector and the covariance matrix. Theorem 7.7.1. density
If A has the distrihution W( I, m), then B 1"'I~mIBI- !t",+p+l) e-1lr
(1)
'1'8-
=
A - ~ has the
1
2impfp(tm)
for B positive definite and 0 elsewhere, where 'IT
=I
-I.
Proof By Theorem AA.6 of the Appendix, the Jacobian of the transformation A = B- 1 is IBI-( p+ I). Substitution of B- 1 for A in (6) of Section 7.2 • and multiplication by IBI-(p+l) yields 0). We shall call (1) the density of the inverted Wishart distribution with m degrees of freedom t and denote the distribution by W-I('II,m) and the density by w-I(BI'II,m). We shall call 'IT the precision matrix or concentration matrix. 7.7.2. Bayes Estimation of the Covariance Matrix The covariance matrix of a sample of size N from N(p., I) has the distribution of O/n)A, where A has the distribution W(I,I1) and 11 = N - I. We shall now show that if I is assigned an inverted Wishart distribution, then the conditional distribution of I given A is an inverted Wishart distribution. In other words, the family of inverted Wishart distributions for I is conjugate to the family of Wishart distributions. tThe definition of the number of degrees of freedom differs from that of Giri (1977), p. 104, and Muirhead (1982), p. 113.
7
.7
273
INVERTED WISHART DISTRIBUTION AND BAYES ESTIMATION
Theorem 7.7.2. If A has the distribution weI"~ n) and I, has the a priori listribution w- I('IT, m), then the conditional distribution of I, is W- I (A + V,n +m). Proof The joint density of A and I, is I 'IT I tmII,I-t(n+m+p+I )IAI !(n-p-I) e-ttr(A+",n:-
1
2 t (n +m)p rp( ~n) rpn m )
for A and I, positive definite. The marginal density of A is the integral of (2) lJver the set of I, positive definite. Since the integral of 0) with respect to B is 1 identically in 'IT, the integ ral of (2) with respect to I, is fp[Hn + m)] 1'1'1 tmlAI t(n-p-I )IA +
(3)
'lT1- t(n+m)
fpnn )rp(tm)
for A positive definite. The conditional density of I, given A is the ratio of (2) to (3), namely,
IA + 'ltl ~(n +m )11:\ - !(" + m +p + I ) e - ~1r(.4 +"')I. 2t (n+m)prAHn + m)]
(4)
which is w-I(I,IA+'IT,n+m).
-1
_
Corollary 7.7.1. If nS has the distribution weI"~ n) and I, has the a priori distribution W- I ('IT, m), then the conditional distribution of I, given S is W- I(nS + 'IT, n + m). Corollary 7.7.2. If nS has the distribution WeI"~ n), I, has the a priori distribution W-I('IT, m), and the loss function is tr(D - I,)G(D - I,)H, where G and H are positive definite, then the Bayes estimator for I, is 1
(5)
n + m-p- 1 (nS + 'IT).
Proof It follows from Section 3.4.2 that the Bayes estimator for I, is 0"(I,IS). From Theorem 7.7.2 we see that I.-I has the a posteriori distribu-
tion WrenS
+ 'IT)-I, n + m]. The theorem results from the following lemma.
Lemma 7.7.1.
(6)
If A has the distribution iCA-I =
WeI"~
n), then
1 I.-I. n-p -1
-
274
COy ARIANCE MATRIX DlSTRIBUTI' )N; GENERALIZED YARlANCE.
Proof If C is a nonsingular matrix such that I = CC', then A has the distribution of CBe', where B has the distribution W(/, n), and $A- 1 = (C') -I ( $ B- 1 )C- I . By symmetry the diagonal elem~nts of $ B- 1 are the same and the off-diagonal elements are the same; that is, $ B- 1 = k 11+ k ~ ss'. For every orthogonal matrix Q, QBQ' has the distribution W(J, n) and hence $(QBQ')-I = Q$B-IQ' = $B- 1 • Thus k2 = O. A diagonal element of B- 1 has the distribution of (Xn2_p+I)-I. (See, e.g., the proof of Theorem 5.2.2.) Since $( X}-P+ 1)-1 = (n - p - 1)-1, $B- I = (n - p _1)-1 I. Then (6) • follows.
We note that (n-p-l)A- I =[(n-p-l)/(n-1)]S-1 is an unbiased estimator of the precision I -I. If J.l is known, the unbiased estimator of I is (1IN)I:~=I(xa - J.l)(x a J.l)'. The above can be applied with n replaced by N. Note that if n (or N) is large. (5) is approximately S. Theorem 7.7.3. Let XI"'" x N be observations from N(J.l, I). Suppose J.l and I have the a priori density n(J.l1 v, (11 K)I) X w-I(II 'IT, m). Then the a posteriori density of J.l and I given x = (1IN)I:~=lxa' and S = (1ln)I:~=l(xa - xXxa - x)' is (7)
n(J.lIN:K(Ni+Kv)'N:KI)
'w- I (I I'IT + nS + N:KK (x- v)(x - v)', N +m). Proof Since x and nS = A are a sufficient set of statistics, we can consider the joint density of x, A, J.l, and I, which is
( 8)
K!P NY> I'IT I tml II- t(N+m +p+2 )IAI t(N-p-2) 2 t(N+m+1 ·exp{ -
)P'7T P f
p[ HN - 1) 1fpo-m)
HN(x -
J.l),I-1 (x - J.l) + tr AI- 1 +K(J.l- V)'I-I(J.l- v) + tr 'lTI-1]}.
The marginal density of x and A is the integral of (8) with respect to J.l and I. The exponential in (8) is - ~ times (9)
(N+K)J.l'I-IJ.l-2(NX+Kv)I-IJ.l
+ NX'I-I X +Kv'I-1 v + tr(A + 'IT)I-I =(N+K)[J.l-
N:K(Ni+K~)rI-I[J.l-
N:K(NX+KV)]
+ ~KK ( x - v)' I-I ex - v) + tr( A + 'IT) I
-1 .
7.7
INVERTED WISHART DISTRIBUTION AND BAYES ESTIMATION
275
The integral of (8) with respect to JL is
KWNWI'lT1 tmlll-t(N+m+p+l )IAI t(N-p-2) (10)
(N +K)W2t (N+m)p7TWr p[t(N - 1)]rA~m) .exp { -
4[tr AI-l + ::K(X - v)'I-l(X - v) + tr 'lTI- l n·
In turn, the integral of (10) with respect to I is (11)
wrA HN - 1)]rp(~m)(N + K)W
7T
·IAI t(N-p-2)1'lT1 tml'IT +A + Nr:.KK(X - v)(X - v)/I-t(N+m). The conditional density of JL and I given i and A is the ratio of (8) to (11), namely,
(N +K)WIII- t(N+m+p+2 )1 'IT +A + ~(i - v)(i - v~'I.t(N+,") (12)
·exp { -
2t (N+m+l )P7TWr p[~(N + m)]
4[(N + K) [
JL - N
~ K ( NX + K v )
r
I
-1 [
JL - N
~ K ( NX + K v) 1
+tr[ 'IT +A + ::K(X - v)(i - v)/l I Then (12) can be written as (7).
l
n·
•
Corollary 7.7.3. If Xl' ••. ' XN are observations from N(JL, I), if JL and I have the a priori density n[JLlv,(1/K)l:]xw- 1 Ctl'IT,m), and if the loss function is (d - JL)'](d - JL) - tr(D - I)G(D - I)H, then the Bayes estimators of JL and I are
(13)
and (14)
N+m~p-1 [ns+'IT+ ::K(X-v)(X-v)/l,
respectively. The estimator of JL is a weighted average of the sample mean i and the a priori mean v. If N is large, the a priori mean has relatively little weight.
276
COVARIANCE MATRIX DISTRIBUTION; GENERAUZED VARIANCE
The estimator of I is a weighted average of the sample covariances S, 'II, and a term deriving from the difference between the sample mean and the a priori mean. If N is large, the estimator is close to the sample covariance matrix. Theorem 7.7.4. If XI"'" x N are observations from N(p., I) and if p. and I have the a priori density n[p.1 v, (ljK)Il X w-I(II 'IT, m), then the marginal a posteriori density of p. given x and S is
(15)
Wr[!( N + m + 1 - p)] [1 + (N + K)(p. - p.*)' B-1 (p. _ p.*)] t(N+m-l)
7T
,
where p.* is (13) and B is N + m - p - 1 times (14). Proof The exponent in (12) is -
!
times
tr[ B + (N + K) (p. - p.* ) (p. - p.* ),] I
(16)
-I.
Then the integral of (12) with respect to I is
(17) Since
IB +;.x'i
=
IBI(1 +x'B-Ix) (Corollary A.3.1), (15) follows.
The denr,ity (15) is the multivariate t-distribution with N degrees of freedom. See Section 2.7.5, ExaIT'ples.
•
+m + 1 -
P
7.8. IMPROVED ESTIMATION OF THE COVARIANCE MATRIX Just as the sample mean i can be improved on as an estimator of the population mean p. when the loss function is quadratic, so can the sample covariance S be improved on as an estimator of the population covariance I for certain loss functions. The loss function for estimation of the location parameter p. was invariant with respect to translation (x -> X + a, p. -> p. + a), and the risk of the sample mean (which is the unique unbiased function of the sufficient statistic when I is known) does not depend on the parameter value. The natural group of transformations of covariance matrices is multiplication on the left by a nonsingular matrix and on the right by its transpose
7
.8
IMPROVED ESTIMATION OF THE COVARIANCE MATRIX
277
x ..... ex, S ..... CSC/, l: ..... Cl:C/). We consider two loss functions which arc nvariant with respect to such transformations. One loss function is quadratic:
))
Lq(l:,G) = tr(G _l:)l:-I(G -l:)l:-I = tr( G l: -I
-
I) 2,
",here G is a positive definite matrix. The other is based on the form of the lik.elihood function: (2) (See Lemma 3.2.2 and alternative proofs in Problems 3.4, 3.8, and 3.12.) Each of these is 0 when G = l: and is positive when G"* l:. The second loss function approaches 00 as G approaches a singular matrix or when one or more elements (or one or more characteristic roots) of G approaches x. (See proof of Lemma ~.2.2.) Each is invariant with respect to transformations G* = CGC/, l:* = Cl:C/. We can see some properties of the loss functions from L/I, D) = Ef=l(dji - 1)2 and L,= $(s-O')(s-O')/, then L/l:G) is a constant multiple of (g-O')/-I(g-O'). (See Problem 7.33.) The maximum likelihood estimator i and the unbiased estimator S are of the form aA, where A has the distribution W(l:, n) and n = N - 1. Theorem 7.8.1. The quadratic risk of aA is minimized at a = 1 /(n + p + 1). and its value is pep + 1)/(n + p + 1). The likelihood risk of aA is minimized at a = l/n (i.e., aA = S), and its value of p log n - Ef~ 1 tC log X;+ 1-;' Proof. By the invariance of the loss function
(3)
tCJ:Lq(l:, aA) = $[Lq(I, aA*)
= $[tr(aA* _1)2
= $[ (a 2
t
i.j~
a7/ - 2a I
t
a7; + p)
i~1
= a 2 [(2n + n 2 )p + np(p - 1))- 2allp + p = P [ n( n + P + 1) a2 - 2na + I).
278
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
which has its minimum at a = 1/(n + p + 1). Similarly
(4)
.F1L,(:t,aA) = Lfi'rL,(I,aA*)
= #r{a tr A* -logIA*1 - p log a - p} = p[na -log a -1] -
•
Although the minimum risk of the estimator of the form aA is constant for its loss function, the estimator is not minimax. We shall now consider estimators G(A) such that
G(HAll') =HG(A)H'
(5)
for lower triangular matrices H. The two loss functions are invariant with respect to transformations G* =HGH',:t* =H:tH'. Let A = 1 and H be the diagonal matrix Di with -1 as the ith diagonal element and 1 as each other diagonal element. Then HAll' = I, and the i, jth component of (5) is ( 6) Hence, gill) = 0, i *- j, and G(l) is diagonal, say D. Since A = IT' for T lower triangular, we have
G(A) = G(TIT')
(7)
= TG(I)T' =TDT', where D is a diagonal matrix not depending on A. We note in passing that if (5) holds for all nonsingular H, then D = al for some a. (H can be taken as a permutation matrix.) If :t = KK', where K is lower triangular, then
(8)
f L[:t, G( A)]C( p, n)1 11- tnlAI = f L[KK', G(A)]C( p, n) IKK'I- iniAl
g I L[ I, G( A)] =
!(n-p-l)
e-!tr I-'A dA
i(n-p-l)
7.8
279
IMPROVED ESTIMATION OF THE COVARIANCE MATRIX
J
= L[KK', G(KA* K')]C( p, n)IA*1 t(n-p-i) e- !trA* dA.* ,;" cC'JL[KK',KG(A*)K/] = cC'JL[I,G(A*)] by invariance of the loss function. The risk does not depend onl:. For the quadratic loss function we calculate
= cC'J tr(TDT' _1)2 = ct.'J tr(TDT'TDT' - 2TDT' + I) p
= cf.'1
E
p
tjjd/k/k,d,tj, - 2 cC'1
i,j,k,/~1
E
tj~dj + p.
j,j~l
The expectations can be evaluated by using the fact that the (nonzero) elements of T are independent, ti~ has the X2-distribution with n + 1 - i degrees of freedom, and tij' i > j, has the distribution N(O,1). Then
(10) where F = (tij), f = (t),
(11)
fii=(n+p-2i+l)(n+p-2i+3), fjj = n + p - 2j + 1,
i <j,
fi = n + p + 2i + 1, and d=(dj, ... ,dp )" Since d'Fd=cC'tr(TDT,)2>0, F is positive definite and (0) has a unique minimum. It is attained at d = F-1f, and the minimum
is p - rP-lf. Theorem 7.8.2. With respect to the quadratic loss function the best estimator invariant with respect to linear transformations l: -+ BIB', A -+ HAH/, where B is lower triangular, is G(A) = TDT', where D is the diagonal matrix whose diagonal elements compose d = F-1f, F andf are defined by (11), and A = IT/ with T lower triangular.
280
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
Since d=F-lf is not proportional to E =(1, ... ,1)/, that is, FE is not proportional to f (see Problem 7.28), this estimator has a smaller (quadratic) loss than any estimator of the form aA (which is the only type of estimator invariant under the full linear group). Kiefer (1957) showed that if an estimator is minimax in the class of estimators invariant with respect to a group of transformations satisfying certain conditions,t then it is minimax with respect to all estimators. In this problem the group of trianr,ular linear transformations satisfies the conditions, while the group of all linear transformations does not. The definition of this estimator depends on the coordinate system and on the numbering of the coordinates. These properties are intuitively unappealing. Theorem 7.8.3. The estimator G(A) defined in Theorem 7.8.2 is minimax with respect to the quadratic loss function. In the case of p = 2 (12)
d = 1
(n-l-l)2-(n-l) (n-l-l)\n-l-3)-(n-l)'
(n -I- 1) (n -I- 2) 2- (n-l-l)2(n-l-3)-(n-l)
d _
The risk is 2
(13)
n3
3n 2 -I- 5n -I- 4 -I- 5n 2 -I- 6n -I- 4 .
The difference between the risks of the best estimator aA and the best estimator TDT' is
(14)
6 n
-I-
6n 2 -I- lOn -I- 8 3 - n 3 -I- 5n 2 -I- 6n -I- 4
2n(n-l)
tr
The difference is is for n = 2 (relative to %), and for n = 3 (rela~ive to 1); it is of the order 2jn 2; the improvement due to using the estimator TDT' is not great, at least for p = 2. For the likelihood loss function we calculate (15)
cC'/L,[I,G(A)]
= cC'/L,[I, TDT'] = ,CAtr TDT' -iogi TDT'I - p] t
The essential condition is that the group is solvable. See Kiefer (1966) and Kudo (1955).
'.8
281
IMPROVED ESTIMATION OF THE COVARIANCE MATRIX
p
p
= E(n+p-2j+1)d j j-I
-
p
Elogd j
-
j~1
EJ"logX,;+I_j-P, j~I
fhe minimum of (15) occurs at dj = 1/(n + p - 2j + 1), j = 1, ... , p.
Theorem 7.8.4. With respect to the likelihood loss function, the best estima'or invariant with respect to linear transformations ::£ -> H ::£ H ' , A -> BAH' , ",here H is lower triangular, is G(A) = TDT', where the jth diagonal element of 'he diagonal matrix Dis l/(n + p - 2j + 0, j = 1, ... , p, and A = TT', with T 'ower triangular. The minimum risk is p
(16)
J'IL[::£,G(A)]
=
p
E log(n + p -
L
2j + l) -
j-I
cf;
log X,;+I-j'
j-I
Theorem 7.8.5. The estimalOr G(A) defined in Theorem 7.8.4 is minimax with respect to the likelihood loss function. James and Stein (1961) gave this estimator. Note that the reciprocals of the weights l/(n + p - 1), l/(n + p - 3), .... l/(n - P + 1) are symmetrically distributed about the reciprocal of lin. If p = 2,
1
G(A) =--A+ n+1
(17)
(18)
(00 2 (00
n .c'G(A) = n+1::£+ n+1
The difference between the risks of the best estimator aA and the best estimator TDT' is
(19)
p
p
j-I
j-I
(
plogn- Elog(n+p-2j+1)= - Elog 1+ P
--"+1)
~J
.
282
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
If p = 2. the improvement is
(20)
-log(I+~)-log(I-*)=
-IOg(l-
1
:2)
1
1
= n 2 + 2n 4 + 3n 6 + ... ,
which is 0.288 for n = 2, 0.118 for n = 3, 0.065 foOr n = 4, etc. The risk (19) is 0(1/n 2 ) for any p. (See Problem 7.31.) An obvious disadvantage of these estimaton is that they depend on the coordinate system. Let P; be the ith permutatioll matrix, i = 1, ... , p!, and iet P;AP; = T;T;, where T; is lower triangular and t ji > 0, j = 1, ... , p. Then a randomized estimator that does not depend on the numbering of coordinates is to let the estimator be P;T;DT;' P; with probability l/p!; this estimator has the same risk as the estimaLor for the original nnmbering of coordinates. Since the loss functions are convex, O/p!)L;P;T;DT/P; will have at least as good a risk function; in this case the risk will depend on :t. Haff (1980) has shown that G(A) = [l/(n + p + l)](A + yuC), where y is constant, 0 ~ y ~ 2(p - l)/(n - p + 3), u = l/tr(A -IC) and C is an arbitrary positive definite matrix, has a smaller quadratic risk than [1/(11 + P + I)]A. The estimator G(A) = (l/n)[A + ut(u)C], where t(u) is an absolutely continuous, nonincreasing function, 0 ~ t(u) ~ 2(p - l)/n, has a smaller likelihood risk than S.
7.9. ELLIPTICALLY CONTOURED DISTRIBUTIONS 7.9.1. Observations Elliptically Contoured Consider
XI"'"
Xv
observations on a random vector X with density
(\)
Let A=L~~I(x,,=iXx,,-i)', n=N-l, S=(1/n)A. Then S~:t as --> 'X). The limiting normal distribution of IN vec(S -:t) was given in Theorem 3.6.2. The lower triangular matrix T, satisfying A = IT', was used in Section 7.2 in deriving the distribution of A and hence of S. Define the lower triangular matrix f by S = ff', l;; ~ 0, i = 1, ... , p. Then f = 0/ /n)T. If :t = I, then N
7.9
ELLIPTICALLY CONTOURED DISTRIBUTIONS
283
s~/ and T~I, /N(S -I) and /N(T-I) have limiting normal distributions, and
/N(S -I) = /N(T-I) + /N(T-I)' + Opel).
(2)
That is, /N (s;; - 1) = 2/N (i;; - 1) + Op(1), and /N s;i = /N'i'i + 0/1), i > j. When :t =/, the set /N(sn -1), ... , /N(spp - 1) and the set /Ns;i' i > j, are asymptotically independent; /NSI2, ... ,/NSp_l,p are mutually asymptotically independent, each with variance 1 + K; the limiting variance of {jij (Sjj - 1) is 3 K + 2; and the limiting covariance of {jij (Sjj - 1) and /N (Sjj - 1), i *" j, is K. Theorem 7.9.1. If:t = Ip, the limiting distribution of /N (T -Ip) is normal with mean O. The variance of a diagvnal element is (3 K + 2) /4; the covariance of two diagonal elements is K/4; the variance of an off-diagonal element is K + 1; the off-diagonal elements are uncorrelated and are uncorrelated with the diagonal elements. Let X = v + CY, where Y has the density g(y' y), A = CC', and :t = tC(X = (tCR 2 /p)A = ff', and C and f are lower triangular. Let S be the sample covariance of a sample of Non X. Let S = IT'. Then S ~:t, T~ f, and - v)(X - v)'
(3)
/N(S -:t) = /N(T- f)f' + f/N(i- r)' + Op(l).
The limiting distribution of If(T - f) is normal, and the covariance can be calculated from (3) and the covariances of the elements of /N (S - :t). Since the primary interest in T is to find the distribution of S, we do not pursue this further here. 7.9.2. Elliptically Contoured Matri); Distributions Let X (NXp) have the density
(4)
ICI-Ng[ C-1(X - ENV')'(X- ENV,)(C')-lj
based on the left spherical density g(Y'Y). Theorem 7.9.2. Define T = (t i) by Y'y = IT', tii = 0, i <j, and tii ~ O. If the density of Y is g(Y'Y), then the density of T is
284
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
Proof Let Y=(vl, ... ,vp )' Define W; and w;'recursivelyby WI=V I' u I =
wl/llwllI, (6) and u j =' wJIIWj II. Then W;Wj = 0, U'jU j = 0, i"* j, and U'jU; = 1. Conditional on (that is, wl"",W;_I)' let Q; be an orthogonal m.itrix with U'I"'" U~_I as the first i - 1 rows; that is, vl"",V;_1
(7) (See Lemma A.4.2.) Define
(8)
Z,
= Q,v, =
td
:
t",_1 zi
1 .
r
This transformation of Vj is linear and has Jacobian 1. The vector N + 1 - i components. Note that liz; 112 = IIw j ll 2, i-I
(9)
z;
has
i-I
Vj = LtijUj+Wj= LtijUj+Q;'z;, i-I
j=1 i-l
(10)
i
V;V;= Ltj~+z;'zi= Lt;~, j= I
(11)
l!}"'i =
L
j-I
tjktik'
j
k=1
The transformation from Y = (VI"'" v p) to Zl"'" zp has Jacot-ian 1. To obtain the density of T convert zi to polar coordinates and integrate • with respect to the angular coordinates. (See Section 2.7.1.) The above proof follows the lines of the proof of (6) in Section 7.2, but does not use information about the normal distribution, such as tj~ g, X~+I-j' See also Fang and Zhang (1990), Theorem 3.4.1. Let C be a lower triangular matrix such that A = CC'. Define X = YC'. Theorem 7.9.3.
(12)
If X (N X p) has the density
'ROBLEMS
285 hen the Lower trianguLar matrix T* satisfying X' X = T* T*' and tensity 2P
(13)
~Np
P
fp(tN)IAlt
Theorem 7.9.4.
0 has the
Ot*:N-1 [C-IT*T*'(C,)-I).
1T-
Let A = X' X = T* T*
I jj ~
N
;=1
/I
g
I.
If X has the density (12), then A = X' X has the density
(14)
The class of densities g(tr Y'Y) is a subclass of densities g(Y'Y). Let X = ENV' + YC'. Then the density of X is
(15) A stochastic representation of X is vec X !l: R(C ® IN) vee U + v ® EN' Theorems 7.9.3 and 7.9.4 can be specialized to this form. Then Theorem 3.6.5 holds. Theorem 7.9.5. Let X have the density (12) where A is diagonal. Let S = (N _I)-I(X - £Nil)'(X - ENi') and R = (diag S)- !S(diag S)-!. Then the density of R is (9) of Section 7.6.
PROBLEMS 7.1. (Sec. 7.2)
A transfonnation from rectangular to polar coordinates is
YI=wsin/l l ,
Y"_I Y" where
w
= W =
cos /II cos /1 2 ... cos /1"-2 sin /1""1'
w cos /II cos /1 2 ... cos /1"-2 cos /1"_1'
-t1T
-1T
and
0::;
286
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
(a) Prove w 2 = r.y~. [Hint: Compute in turn Y; + Y;_I'(Y;+Y;_I) +Y;-2' and so forth.] (b) Show that the Jacobian is w" -I cos"- 2 /II cos" -3 /1 2 .. , cos /I" -2' [Hint: Prove cos
I
r)(YI" ...
y,,)
iI(AI.···,/I,,_I,w)
I
/II
0
0
x
0 0
cos /1"-1
0
0 (II
w sin
(12
w sin /1,,-1
X
X
X
X
0
wcos
0
0
w cos
0
0
0
(II
0 0
.
wsin II'
0
cos /1 2
/II
cos /1,,-2
X
cos /11
cos /1,,-1
where x denotes elements whose explicit values are not needed.] 7.2. (Sec. 7.2)
Prove that
[Hint: Let cos 2 (I = u, and use the definition of B(p, q).] 7.3. (Sec. 7.2) Use Problems 7.1 and 7.2 to prove that the swface area unit radius in n dimensions is
of a sphere of
7.4. (Sec. 7.2) Use Problems 7.1, 7.2, and 7.3 to prove that if the density of y' =yp".,y,,) is f(y'y), then the density of u = y'y is ~C(n)f(u)ul"-I. 7.5. (Sec. 7.2) X 2-distribution. Use Problem 7.4 to show that if Yl'"'' Yn are independe?tly dist.ribut~d, each according to N(O, 1), then U = 1Y; has the density U,"-I e- ,Uj[2,nntn)], which is the x2-density with n degrees of freedom.
r.:=
7.6. (Sec. 7.2)
Use (9) of Section 7.6 to derive the distribution of A.
7.7. (Sec. 7.2)
Use the proof of Theorem 7.2.1 to demonstrate Pr{IAI = O} = O.
PROBLEMS
287 7.S. (Sec. 7.2) Independence of estimators of the parameters of the complex normal distribution. Let ZI' ... ' z,v be N obseIVations from the complex normal distribution with mean IJ and covariance matrix P. (See Problem 2.64.) Show that Z and A = Z~_I(Za - Z)(Za - Z)* are independently distributed, and show that A has the distribution of [:_IWaWa*' where WI, ... ,Wn are independently distributed, each according to the complex normal distribution with mean () and covariance matrix P. 7.9. (Sec. 7.2) distributed, covariance [:_1 Wa Wa*
7.10. (Sec. 7.3)
The complex Wishart distribution. Let WI' ... ' w" be independently each according to the complex normal distribution with mean 0 and matrix P. (See Problem 2.64.) Show that the density of B = is
Find the characteristic function of A from WeI, n). [Hint: From < , one derives
[w(AI I, n) dA =
as an identity in ~.] Note that comparison of this result with that of Section 7.3.1 is a proof of the Wishart distribution. 7.11. (Sec. 7.3.2)
Prove Theorem 7.3.2 by use of characteristic functions.
7.12. (Sec. 7.3.1) Find the first two moments of the elements of A by differentiating the characteristic function (11).
"'.13. (Sec. 7.3) Let ZI' ... ' Z" be independently distributed, each according to N(O, I). Let W = [:. Il- I ball Za Z~. Prove that if a' Wa = X~ for all a such that a' a = 1, then W is distributed according to W{/, m). [Hint: Use the characteristic function of a'Wa.] 7.14. (Sec. 7.4) Let Xu be an obseIVation from NCJ3z,,, I), a = I, ... , N, where za is a seal,,£. Let b = LaZaXa/LaZ';. Use Theorem 7.4.1 to show that [axax~ bb'["z,; and bb' are independent. 7.15. (Sec. 7.4)
Show that
2)" = tC (2 2 XN-2 tC ( XN-I X2N.,4/4 )" ,
h
~O,
by use of the duplication formula for the gamma function; X~ _I and X~- 2 are independent. Hence show that the distribution of X~- J X~-2 is the distribution of xlN-4/4.
p
288
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
7.16. (Sec. 7.4)
Verify that Theorem 7.4.1 follows from Lemma 7.4.1. [Hint: Prove that Qj having the distribution W(I.,,) implies the existence of (6) where 1 is of order 'j and that the independence of the Q;'s implies that the I's in (6) do not overlap.]
7.17. (Sec.7.5)
7
Find tS'IAlh directly from W(I.,n). [Hint: The fact that 7 jw(AII.,n)dA == 1
shows
as an identity in n.] 7.1S. (Sec. 7.5)
Consider the confidence region for .... given by N(i- Ir-*)'S-I(i_"*):::;(N-l)PF rN _ P p, N-p (e)' ,
where i and S are based on a sample of N from N( .... , I.). Find the expected value of the volume of the confidence region. 7.19. (Sec. 7.6)
Prove that if I.
'Ip"""p-I,p
=
1, the joint density of
'ij.p'
i, j
=
1, ... , P - 1, and
is
7.20. (Sec. 7.6) Prove that the joint density of 'Ip"""p-I.p is
'12·3 ..... p"13.4 •..• p"2J.4 .... ,p'
r{ Un - (p - 2)]} (1- 2 )±In-
2 .n .
1-17T'
... n
r [Iz(n - 1)] (1 j-I 7T1r[!(n - 2)]
p-2
.p-I n
r (Izn )
j-I 7T,r [Iz(n - 1) ] !
(1 _
2 t
_,2
)±
I.p-I·p
2
"P
)1
[Hint: Use the result of Problem 7.19 inductivity.]
.
... ,
ROBLEMS
289 .21. (Sec. 7.6) Prove (without the use of Problem 7.20) that if I = I, then 'Ip"'" 'p-l, P are independently distributed. [Hint: 'jp = ajp/(,;a;; Prove that the pairs (alp' all)' ... ,(ap-l,p, ap-l. p _ l ) are independent when (Zip"", z"p) are fixed, and note from Section 4.2.1 that the marginal distribution of 'jp, conditional on zap' does not depend on zap']
,;a;;).
'.22. (Sec. 7.6) Prove (without the use of Problems 7.19 and 7.20) that if I = I. then the set 'II', ... ,',,-I,p is independent of the set 'ji'p' i,j = l. .... p _. I. [Hillt: From Section 4.3.2 a pp ' and (u jp ) arc independent of (u'i'pl. Prow thaI app,(a j ) , and a jj , i=l,oo.,p-l, are independent of ('jj-,,) hy proving that ajj .p are independent of ('ji'P)' See Problem 4.21.] r.23. (Sec. 7.6) Prove the conclusion of Problem 7.20 by using Problems 7.21 and 7.22. 1.24. (Sec. 7.6) Reverse the steps in Problem 7.20 to derive (9) of Section 7.6. 1.25. (Sec. 7.6) Show that when p mutually independent.
=
3 and I
is diagonal
7.26. (Sec. 7.6) Show that when I is diagonal the set
'ij
'12' '13' '23
are not
are pairwise independent.
7.27. (Sec. 7.7) Multivariate t-distributioll. Let y and u be independently distributed according to N(O, I) and the x;-distribution, respectively, and let .;nTriy = xJ.l..
(a) Show that the density of x is
r[t(n +p)]
(b) Show that Gx = J.l. and
7.28. (Sec. 7.8)
Prove that Fe is not proportional to f by calculating FE.
7.29. (Sec. 7.8)
Prove for p
=
2
7.30. (Sec. 7.8) Verify (17) and (18). [Hillt: To verify (18) let I = [([(', A"' [(A*[(', and A* = T* T*, where K and T* are lower triangular.]
290
COVARIANCE MATRIX DISTRIBUTION; GENERALIZED VARIANCE
7.31. (Sec. 7.8)
Prove for optimal D
P even,
~lp -I)
L
[
log 1-
(_
P
2i
n
+ 1 )2)
'
P odd.
i= 1
7.32. (Sec. 7.8) Prove LiI., G) and LI(I., G) are invariant with respect to transformations G* = CGe', I.* = CI.C' for C nonsingular. 7.33. (Sec. 7.R) Prove L,/I.,G) is a multiple of (g_(J"Yct>-I(g-U). Hint: Transform so ~ = I. Then show
<1>=
.~(2I tI
7.34. (Sec. 7.8)
0
0)I '
Verify (II).
7.35. Let the density of Y he I(y) = K for y' y 5, P + 2 and 0 elsewhere. Prove that K = r<1P + I )/[(p + 2hT l~P, and show that (~'Y = 0 and @YY' = 1. Dirichlet distribution. Let YI , •.• , Y,n be independently distributed as ,\'~-variables with Pi' ... ' Pm degrees of freedom, respectively. Define Z; = YjL7~1}j, i = [, ... ,m. Show that the density of ZI"",Zm_1 is
7.36. (Sec. 7.2)
for
Zi ~
0, i
=
1, .... m.
7.37. (Sec. 7.5) Show that if X~-I and X~-z are independently distributed, then x~· I X~ z is distrihuted ~s (:d'l ,l" /4. [Hint: In the joint density of x = X~-I and y = x.~ _z substitute z = 2/;;, x = x, and CXPI ess the marginal density of z as z'\" .1h(Z)' where h(z) is an integral with respect to x. Find h'(z), and solve the difflon:ntial L'quation. Sec Srivastava and Khatri (J 979), Chapter 3.]
CHAPTER 8
Testing the General Linear Hypothesis; Multivariate Analysis of Variance
S.l. INTRODUCTION In this chapter we generalize the univariate least squares theory (Le., regression analysis) and the analysis of variance to vector variates. The algebra of the multivariate case is essentially the same as that of the univariate case. This leads to distribution theory that is analogous to that of the univariate case and to test criteria that are analogs of F-statistics. In fact, given a univariate test, we shall be able to write down immediately a corresponding multivariate test. Since the analysis of variance based on the model of fixed effects can be obtained from least squares theory, we obtain directly a theory of multivariate analysis of variance. However, in the multivariate case there is more latitude in the choice of tests of significance. In univariate least squares we consider scalar dependent variates XI' •• ·' x N drawn from populations with expected values P' Z I' ... , P' Z N' respectively, where Jl is a column vector of q components and each of the z" is a column vector of q known components. Under the assumption that the variances in the populations are the same, the least squares estimator of P' is
(1)
An Introductwn to Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0·471·36091·0 Copyright © 2003 John Wiley & Sons, Inc.
291
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
g
If the populations are normal, the vector is the maximum likelihood estimator of p. The unbiased estimaior of the common variance a 2 is
v
292
N
(2)
S2=
E
(x a ';"b'za)2 j (N-q),
a=J
and under the assumption of normality, the maximum likelihood estimator of a 2 is u2 =(N-q)s2jN. In the multivariate case xa is a vector, W is replaced by a matrix p, and a 2 is replaced by a covariance matrix l:. The estimators of p and l:, given in Section 8.2, are matric analogs of (1) and (2). To test a hypothesis concerning p, say the hypothesis p = 0, we use an F-test. A criterion equivalent to the F-ratio is
(3)
[qj{N-q)]F+ 1
where u02 is the maximum likelihood estimator of a 2 under the null hypothesis. We shall find that the likelihood ratio criterion for the corresponding multivariate hypothesis, say p = 0, is the above with the variances replaced by generalized variances. The distribution of the likelihood ratic criterion under the null hypothesis is characterized, the moments are found, and some specific distributions obtained. Satisfactory approximations are given as well as tables of significance points (Appendix B). The hypothesis testing problem is invariant under several groups of linear transformations. Other invariant criteria are treated, including the Lawley-Hotelling trace, the Bartlett-Nanda-Pillai trace, and the Roy maximum root criteria. Some comparison of power is made. Confidence regions or simultaneous confidence intervals for elements of p can be based on the likelihood ratio test, the Lawley-Hotelling trace test, and the Roy maximum root test. Procedures are given explicitly for several problems of the analysis of variance. Optimal properties of admissibility, unbiased ness, and monotonicity of power functions an~ studied. Finally, the theory and methods are extended to elliptically contoured distributions.
S.2. ESTIMATORS OF PARAMETERS IN MULTIVARIATE LINEAR REGRESSION S.2.1. Maximum Likelihood Estimators; Least Squares Estimators Suppose Xl' •.. ' X N are a set of N independent observations, xa being drawn from N(Pz a , l:). Ordinarily the vectors Za (with q components) are known
293
.2 ESTIMATORS OF PARAMETERS IN LINEAR REGRESSION
'ectors, and the p x p matrix £ and the p x q matrix lssume N ~ P + q and the rank of
s q. We shall estimate ikelihood function is
I
and
P by the
P are
unknown. We
method of maximum likelihood, The
In (2) the elements of I* and P* are indeterminates. The method of maximum likelihood specifies the estimators of I and P based on the given sample xpz1,,,,,XN,ZN as the I* and P* that maximize (2). It is convenit:nt to use the following lemma. Lemma 8.2.1.
Let
(3) Then for any p
X
q matrix F
N
(4)
N
L (x,,-Fz,,)(xa-Fza)' = L (x,,-Bz,,)(x,,-Bz,,)' N
+(B-F) LZaz~(B-F)'. a=\
Proof The left-hand side of (4) is N
(5)
L [(xa-Bza)+(B-F)z,,][(xa-Bza)+(B-F)z,,]', a~l
which is equal to the right-hand side of (4) because N
(6)
L za(x" - BZa)' = 0 a=1
by virtue of (3).
•
294
TESTING THE GENERAL LiNEAR HYPOTHESIS; MAN OVA
t
Th(' exponential in L is -
times
(7) N
tr ~*-.I
L
N
(xC( - ~zn )(x n
-
p*zu)'
=
tr l:*-l
L
(xa -Bza)(xa -Bz a )'
n=!
+ tr I * - I ( B - p* ) A( B - 13*)' , where
(8) The likelihood is maximized with respect to inl?). Lemma 8.2.2. Proof Let A
(9)
by minimizing the last term
If A and G are positive definite, trFAF'G
> 0 for F"* O.
= HH', G = J(]('. Then tr FAF'G = tr FHH'F'J(](' = tr K'FHH'F'K =
tr(K'FH)(K'FH)'
for F"* 0 because then K' FH
by
13'
"* 0 since
>0
Hand K are nonsingular.
•
It follows from (7) and the lemma that L is maximized with respect to p* 13* = B. that is.
( 10)
where
(11 ) Then by Lemma 3.2.2, L is maximized with respect to l:* at ( 12)
i
=
~ ~ (x" - PZa)(x" - Pz,,)'. u=i
This is the multivariate analog of Section 8.1.
a- 2 =(N-q)S2/N
defined by (2) vf
Theorem 8.2.1. ffx" is an observation from N(Pz", l:), a = 1, ... , N, with (z I" .,. z,v) of rallk q, the maximum likelihood estimator of 13 is given by (10), where C = L"X" z;, and A = L" z" z;,. The maximum likelihood estimator of I is givel1 by ([2).
0...
c')l1MA1U!<.:' Ui;' t'AKAMt::U::;l{S IN LINEAR REGRESSION
295
A useful algebraic result follows from (12) and (4) with F = 0:
Now let us consider a geometric interpretation of the estimation procedure. Let the ith row of (XI" ., xN ) be xi (with N components) and the ith row of (ZI"'" ZN) be zi (with N components). Then Lj ~ijZj, being a linear combination of the vectors zj, ... , z~, is a vector in the q-space spanned by zj, ... , z~, and is in fact, of all such vectors, the one near~st to xi; hence, it is the projection of xi on the q-space. Thus xi - Lj f3ijzj is the vector orthogonal to the q-space going from the projection of xi on the q-space to xi. Translate this vector so that one endpoint is at the origin. Then the set of . . . f3 Z*.~s a set of vectors emanatmg . from p vectors Xl* - l.jf3ljZj*""'AXp* - L. j pj j the origin. NUji = (xi - Lj f3ijzj Xxi - Lj f3ijz'/')' is the square of the length of the ith such vector, and Na;j = (xi - Lh ~ihZn(xj - Lg ~jgZ;)' is the product of the length of the ith vector, the length of the jth vector, and the cosine of the angle between them. The equations defining the maximum likelihood estimator of p, namely, AB' = C', consist of p sets of q linear equations in q unknowns. Each set can be solved by the method of pivotal condensation or successive elimination (Section A.S of the Appendix). The forward solutions are the same (except the right-hand sides) for all sets. Use of (13) to compute N I involves an efficient computation of Let X", = (x]"', ... , x p ) ' , B = (b l , ... , bp )', and p = (PI"'" Pp )'. Then tCx;", = WiZ",. and bi is the least squares estimator of Pi' If G is a positive definite matrix, then tr G L:~I(X", - Fz",)(x", - Fz",)' is minimized by F = B. This is another sense in which B is the least squares estimator. A
A
pAp'.
8.2.2. Distribution of
P and I
Now let us find the joint distribution of ~ig (i = 1, ... , p, g = 1, ... , q). The joint distribution is normal since the ~jg are linear combinations of the Xj",. From (0) we see that N
(14)
tCp =,g L X",Z~£l ",~l
N
L ,,~l
=p.
PZ",Z~A-l =
PAA- 1
296
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
P
Thus is an unbiased estimator of two rows of ii3 , is
p. The covariance between
P'i and Pi,
(15)
tff(~i - Pi)(~j -
pJ =A
N -I
N
L
tff
(Xi" - (g'X;a )Za
a~1
L
(Xjy - (,0/ Xjy )Z~A-I
y~1
N
=
L
A -I a,
y~
c8'(Xia - cg'Xia)(Xjy
-
tffXjy )ZaZ~[1
1
N
L
=A- 1 (X,
8ayUijZaZ~A-I
y= 1
N
L
=A- 1
UijZaZ~A-I
{r= 1
To summarize, the vector of pq components (P'I"'" P~)' = vec P' is normally distributed with mean (WI"'" p~)' = vec 13' and covariance matrix
(16)
Ul1 A - 1
ui2 A - 1
U21 A- 1
U22 A - 1
uplA- 1
Up2 A -
1
u1pA- 1 u 2p A- 1 uppA- 1
The matrix (16) is thc Kroncckcr (or dircct) product of thc matrices I and A -I, denoted by I ® A - I. From Theorem 4.3.3 it follows that NI = L~~IXnX~ - PA~' is distributed according to W( I, N - q). From this we see that an unbiased estimator of I is S = [N I(N - q)]I.
P
Theorem 8.2.2. The maximum likelihcod estimator based on a set of N observations, the ath from N(Pza' I), is normally distributed with mean P, and the couar:ance matrix of the ith andjth rows of~ i,: uijA -I, where A = La Za z~. The maximum likelihood estimator I multiplied by N is independently distributed according to W(I, N - q), where q is the number of components of za'
8.2
297
ESTIMATORS OF PARAMETERS IN LINEAR REGRESSION
The density then can be written [by virtue of (4)]
(17)
,IN' exp(--hr{I-l[(P -P).t(P -p)'+Nil}).
(21r)'P III,N
-
This proves the following: Corollary 8.2.1.
P and i
form a sufficient set of statistics for
p alld 1.
A useful theorem is the following. 'fheorem 8.2.3. Let X" be distributed according to N(Pz", I), a and suppose Xl' ... , X N are independent.
1..... N,
=
and r = PH- l , (/Jell Xa is distributed according to N(rw", I). (b) The maximum likelihood estimator of r based on observations x" on X"' l a = 1, ... , N, is t = is the maximum likelihood estima, where tor of p. (c) tn::aw"w~)t'=PAP', where A=LaZaZ~, and the ma:imum likelihp01 estimator of NI is Ni = L"X"X~ - tcL"waW~)t' = Lo x" x:, -
(a) If w" = HZa
PH-
P
pAp'.
(d) (e)
t t
and I are independently distributed. is normally distributed with mean r and the covariance matrix of the ith andjth rows of is U'i/HAH,)-l = U'ijH'-lA-lH-- l .
t
The proof is left to the reader. An estimator F is a linear estimator of f3i~ if F = L~'~ 1 f~ xo' It is a linear unbiased estimator of f3ig if . N
P8)
f3ig = ,P,F =
,t
N
L
f~x" =
n=i
is an identity in
L
N
f~pzu
n=i
=
p
q
L L Lf 11'=1 j=i
j"
f3 j h Z h"
},=\
p, that is, if N
(19)
L
fjazha = 1,
j=i,
h=g,
a=\
=0,
otherwise.
A linear unbiased estimator is best if it has minimum variance over all linear unbiased estimators; that is, if J;'(F - f3igf :c;; ,r(G - f3igr for G = L:~~ 1 g~x" and ,cG = f3ig'
298
TESTING THE GENERAL LINEAR HYPOTHESIS; MAN OVA
Theorem 8.2.4. estimator of f3ig'
The least squares estimator is the best linear unbiased
Proof Let ~ig = L~= I Lf= I fia x ia be an arbitrary unbiased estimator of f3;g' and let f3;g = L~= I Lh= I X ia Zhaahg be the least squares estimator, where A = L~=IZaZ~. Then A
(20) ,f (
- - f3 ig )" = ,if [A - - f3;g A f3;g f3;g - f3;g + ( f3;g =
A
rf ( f3ig - f3ig
)]2
)2 + 2
f3ig
)( _ A) f3ig - f3ig
A)2 . + Iff ( f3_ig - f3ig
Because ~ig and Pig are unbiased, ~ig - f3;g = L~= I LJ= I fj"u ia , P;g - f3ig = L:=IL7.=lu;azhaahg, and
where 8;; = 1 and 8;i = 0, i *- j. Then
q
=
(Tii agg - (Tii
q
L L
ahh,ahgah'g
h=1 h'=1
=0.
8.3. LIKELIHOOD RATIO CRITERIA FOR TESTING LINEAR HYPOTHESES ABOUT REGRESSION COEFFICIENTS 8.3.1., Likelihood Ratio Criteria Suppose we partition (1)
lU
L1Kl:iLlHOOD RATIO CRITERIA FOR REGRESSION COEFFICIENTS
299
so that PI has ql columns and P2 has q2 columns. We shall derive the likelihood ratio criterion for testing the hypothesis
(2) where ~ is a given matrix. The maximum of the likelihood function L for the sample XI"'" X N is
maxL = (21T) - y,N
(3)
1
~.I
In l - tN e- Y,N,
where In is given by (12) or (13) of Section 8.2. To find the maximum of the likelihood function for· the parameters restricted to w defined by (2) we let y" =x" - ~z~l),
(4)
a=l, ... ,N,
where
( Z(I»)
(5)
Z,,=
z~)
a=l, ... ,N,
,
is partitioned in a manner corresponding to the partitioning of p. Then y" can be considered as an observation from N(P2Z~2), l:). The estimator of P2 is obtained by the procedure of Section 8.2 as N
(6)
A ..... 2w
N
="i...J Y z(2)'A- = "i...J (x a
a
22
1
a=1
a
-
1 <>*z(I»)z(2)'A11-'1 a a 22
,,=1
= (C 2 - P'rAI2)A221 with C and A partitioned in the manner corresponding to the partitioning of P and za'
(7) (8) The estimator of I is given by N
(9)
NIw=
L
(Yu-P2wZ~;»)(Ya-P2(.Z~;»)'
a=J N
=
L
a=J
y"y~ - P2w A22P;w
300
TESTING THE GENERAL LINEAP HYPOTHESIS; MANOVA
Thus the maximum of the likelihood function over w is
(10) The likelihood ratio criterion for testing H is (10) divided by (3), namely,
(11) In testing H, one rejects the hypothesis if A < 11.0' where 11.0 is a suitably chosen number. A special case of this problem led to Hotelling's T 2-criterion. If q = ql = 1 (q2 = 0), za = 1, a = 1, ... , N, and P = PI = .... , then the T 2-criterion for testing the hypothesis .... = .... 0 is a monotonic function of (11) for ~ = .... 0. The hypothesis .... = 0 and the T 2-statistic are invariant with respect to the transformations X* = DX and x: = Dx a , a = 1, ... , N, for nonsingular D. Similarly, in this problem the null hypothesis PI = 0 and the lik~lihood ratio criterion- for testing it are invariant with respect to nonsingular linear transformations. Theorem 8.3.1. The likelihood ratio criterion (11) for testing the null hypothesis PI = 0 is invariant with respect to transformations x~ = Dx a , a = 1, ... , N, for nonsingular D.
Proof The estimators in terms of x: are
(12)
p* = DCA-I = DP
(13)
'lil = N
~
1
N
L
, ~
~
~
(Dxa - DPza)( DXa - DPza)' = D'ln D ',
a~l
(14)
P'L = DC2A 22
(15)
~ * _ 1 ;, ( ~ (2»)( ~ (2»), -_ D'lw ~ D ,. 'lw - N L.... Dx" - DP2wZa DXa - DP2wZa
1
=DP2w'
a~1
•
8.3.2. Geometric Interpretation An insight into the algebra developed here can be given in terms of a geometric interpretation. It will be convenient to use the following lemma: Lemma 8.3.1. (16)
8.3
301
LIKELIHOOD RATIO CRITERIA FOR REGRESSION COEFFICIENTS
Proof. The normal equation PoA = C is written in partitioned form (17)
(PWAlJ +PWA21,PWAI2 + PWA22) = (C I ,C2 ).
Thus P w = C 2A z21 - pwA!2Azr The lemma follows by comparison 'vvith (6). • We can now write (18)
X - PZ = (X - PIlZ) + (P211 - PJ Z2 + (Pili - pnZI
= (X- PoZ) +
-(P2w =
(X -
(P2w -
P2)Z2
P211)Z2 + (P!l1 - ~)ZI
PoZ) +
(P2w -
P2)Z2
+(Pw - pn(ZI -AI2Az2IZ2) as an identity; here X = (XI"'" x N ), ZI = (Z\I), ... , z~), and Z2 = (Z\2), ... , z~). The rows of Z = (Z;, Z~)' span a q-dimensional subspace in N-space. Each row of PZ is a vector in the q-space, and hence each row of X - PZ is a vector from a vectOr in the q-space to the corresponding row vector of X. Each row vector of X - PZ is expressed above as the sum of three row vectors. The first matrix on the right of (18) has as its ith row a vector orthogonal to the q-space and leading to the ith row vector of X (as shown in the preceding se;tion). The row vectors of (P2w - P2)Z2 are vectors in the q2-space spanned by the rows of Z2 (since they are linear combinations of the rows of Z2)' The row vectors of (Pili - ~XZI -A\2A /Z 2 ) are vectors in the ql-space of ZI -AI2Az21Z2' and this space is in the q-space of Z, but orthogonal to the q2-space of Z2 [since (ZI -AI2A22IZ2)Z~ = 0]. Thus each row of X - PZ is indica; ed in Figure 8.1 as the sum of three orthogonal vectors: one vector is in the space orthogonal to Z, one is in the space of Z2' and one is in the subspace of Z that is orthogonal to Z2'
z
x
Figure 8.1
302
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOV A
From the orthogonality relations we have (19)
(X- PZ)(X= (X -
PZ)'
PnZ)( X - PnZ)' + (P2w -
P2)Z2 Z Z(P2w -
P2)'
+(PIH - P7)(ZI -A12A22IZ2)(ZI -A12A22IZd'(PIn - P7)' = Nin + (P2w -
P2)A 22 (P2w -
P2)'
+(PUl- P7)(All -A12Az2IA21)(PIn - P7)'. If we subtract
(P2w - P2 )Z2
from both sides of (18), we have
From this we obtain (21)
Niw=(x-p~ZI-P2wZ2)(X-P~ZI-P2wZ2)' .= (
X-
PI! Z) ( X - PI! Z),
+ (Pw - ~)( ZI -
A I2 A zlz 2)( ZI
= Nin + (Pin - ~)( All -
-
A12AzlZ2)'(PIfi - ~)'
A12Az2IA2l )(PIn
- pr)'.
The determinant linl =(1/W)I(X:"'PnZ),X-PnZ)'1 is proportional to the volume squared of the parallelotope spanned by the row vectors of X - PnZ (translated to the origin). The determinant liJ = (l/NP)I(xP7z1 - PZwZ2XX - ~ZI - P2wZ2)'1 is proportional to the volume squared of the parallelotope spanned by the row vectors of X - ~ Zl - P2wZ2 (translated to the origin); each of these vectors is the part of the vector of X- P7Z1 that is orthogonal io Z2. Thus the test based on the likelihood ratio criterion depends on the ratio of volumes of parallelotopes. One parallelotope involves vectors orthogonal to Z, and the other involves vectors orthogonal to ZC. From (5) we see that the density of xP .•. ,XN can be written as (22)
Thus,
i, Pin,
and
P2w
form a sufficient set of statistics for I,
PI'
and
P2.
8.3
LIKELIHOOD RATIO CRI1ERIA FOR REGRESSION COEFFICIENTS
303
Wilks (1932) first gave the likelihood ratio criterion for testing the equality of mean vectors from several populations (Section 8.8). Wilks (1934) and Bartlett (1934) extended its use to regression coefficients. 8.3.3. The Canonical Form In studying the distributions of criteria it will be convenient to put the distribution of the observations in canonical form. This amounts to picking a coordinate system in the N-dimensional space so that the first ql coordinate axes are in the space of Z that is orthogonal to Z2' the next q2 coordinate axes are in the space of Z2' and the last n (= N - q) coordinate axes are orthogonal to the Z-space. Let P2 be a q2 X q2 matrix such that (23)
Then define the N x N orthogonal matrix Q as
(25)
where Q3 is any n
X
N matrix making Q orthogonal. Then the columns of
are independently normally distributed with covariance matrix I (Theorem 3.3.1). Then
(28)
tffW2 = tff XQ~ = (PIZ I + P2Z2)ZZP;
= (PI AI2 + P2 A 22)P;, (29)
.cW3 = tff XQ3 = PZQ 3 = o.
304
TESTING THE GENERAL LINEAR HYPOTHESIS; MAN OVA
Let
(31)
r l = ('YI'·.·' 'Yq) = PIAu·zPi = PIPI- I , r z = ('Yq,+I'·.·' 'Yq) = (PI A 12 + pzAzz)Pz ,
(32)
W=(WI
(30)
Wz
W3)=(WI, ... ,Wq"Wq,+I, ... ,Wq,Wq+I, ... ,WN).
Then WI' ... ' WN are independently normally distributed with covariance matrix I. and $w",='Y"" a= 1, ... ,q, and $w",=O, a=q + 1, ... ,N. The hypothesis PI = P~ can be transformed to PI = 0 by subtraction, that is, by letting x", - P~ z~1) = y"" as in Section 8.3.1. In canonical form then, the hypothesis is r l = O. We can study problems in the canonical form, if we wish, and transform solutions back to terms of X and Z. In (17), which is the partitioned form of Pn A = C, eliminate to obtain
Pw
Pln(A 11
(33)
-AIZA2ZIAzI)
= C1 =
that is, WI = PlnAI1.zP; = PWPI- I and obtain
CZA2ZlAzi
X( Z; - Z; A2ZIAzI)
r l = PIPI-1.
Similarly, from (6) we
(34) that is, Wz = (pzwAzz + P~Alz)Pz = Pzwpz-I + P~AI2P21 and I
r z = PZPZ I +
PIAIZPZ .
8.4. THE DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION WHEN THE HYPOTHESIS IS TRUE 8.4.1. Characterization of the Distribution The likelihood ratio criterion is the ~Nth power of
(1) where A II . z =AII -AlzA2ZIAzl. We shall study the distribution and the moments of U when PI = ~. It has been shown in Section 8.2 that Nin is distributed according to W(I., n), where n = N - q, and the elements of Pn - P have a joint normal distribution independent of Ni n .
8.'1
305
DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION
From (33) of Section 8.3, we have
(2)
(Pw - ~)AJI'2(pjn -l:5n' = (WI
- rl)PIAJI.2P;(WI -
r l )'
= (Wj - rj)(Wj - rj)', by (24) of Section 8.3; the columns of WI each according to N(O, 'l). Lemma 8.4.1. W('l, qj)' Lemma 8.4.2.
(Pw -
rl
are independently distributed,
1:5~)AII'2(Pln - 1:5~)' is distributed according to
The criterion U has the distrib7ltion of
(3) where G is distributed according to W('l, n), H is distributed according to W('l, m), where m = ql, and G and H are independent.
Let
(4)
G=Ni n =XX' -Xl'(ll') -I ZX',
(5)
G+ H = Nil! + (Pill -l:5nAII'2(PIl! - ~)' = Niw = YY' - Yl~(l2l~)-1 l2Y"
where Y = X - 1:5~ II = X - (1:5~ O)l. Then
(6)
G = YY' - Yl'(ll')
-I
lY'.
We shall denote this criterion as Up. m. n' where p is the dimensionality, m = qj is the lUmber of columns of 1:5 1, and n = N - q is the number of cegrees of freedom of G. We now proceed to characterize the distribution of U as the product of ~eta variables (Section 5.2). Write the criterion U as
(7)
(8)
i=2 ..... p,
306
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
and G; and Hi are the submatrices of G and H, respectively, of the first i rows and columns. Correspondingly, let y~) consin of the first icomponents of Ya = xa - P': z~), a = 1, ... , N. We shall show that II; is the length squared of the vector from yt =(Yn""'YiN) to its projection on Z and Y;-I = (J\i-I), ... , y~-I» divided by the length squared of the vector from yt to its projection on Z2 and 1'; -I' Lemma 8.4.3. Let y be an N-component row vector and U an r X N matrix. Then the sum of squares of the residuals of y from its regression on U is
YU'I
yy'
IUy'
(9)
uU'
\UU'\
Proof By Corollary A.3.1 of the Appendix, (9) is yy' -yU'(UU,)-IUy', which is the sum of squares of residuals as indicated in (13) of Section 8.2 .
•
Lemma 8.4.4. V; defined by (8) is the ratio of the sum of squares of the residuals ofYil'''''Y,N from their regression on Yli-I), ... ,y~-I) and Z to the sum of squares of residuals ofYil' ... , y, N from their regression on yli -I) , ... , y~ -I) and Z2'
Proof The numerator of II; can be written [from (13) of Section 8.2]
(10)
\G\i
-ra;-:-J =
\Y,Y,'-Y"Z'(ZZ,)-IZY,'\ - 1';-1 Z'( ZZ') 1ZY;'_I\
\ 1';-1 1';'-1
1';1';' 1
ZY/
YZ'II ~Z' \ZZ'\
Yi-I1';'-1 Z1';'_1
Y'~~,Z'II \ZZ'\
1';-1 Yi'-I
1';yt' ytyt' Zyt '
I
yt1';'-1 ZYi'_1
Yi-I Y;'-I I ZY;'_I
_
1';-1 z'
ytz' zz'
1';-I Z '1
zz'
8.4
DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION
YfYf'
Yf[lj'-I
307
Z,]
JY~I]Yf [lj~I][lj'_1 Z,]
I[lj~I ] [lj'-I = Yf Yf' -
Z'] I
lj'-I yf( lj'-I z' ) [lj-I Zlj'_1
lj_IZ,]-I(lj_l) *, ZZ' Z y,
by Corollary A.3.1. Application of Lemma 8.4.3 shows that the right-hand side of (10) is the sum of squares of the residuals of on lj_1 and Z. The denominator is evaluated similarly with Z replaced by Zz. •
yt
The ratio V; is the 2/ Nth power of the likelihood ratio criterion for testing the hypothesis that the regression of ~rIZI on ZI is 0 (in the presence ofregression on lj-I and Zz); here ~rl is the ith row of !=It. For i = 1, gll is the sum of squares of the residuals of yi = (Yll"'" YIN) from its regression on Z, and gil + hll is th~ sum of squares of the residuals from Zz. The ratio VI = g 11 /(g 11 + h 11)' which is approximate to test the hypothesis that regression of yi on Z\ is 0, is distributed as X.z/( x.z + X~) (by Lemma 8.4.2) and has the beta distribution f3(v; ~n,~m). (See Section 5.2, for example.) Thus V; has the beta density
yt =x; -
(11)
f3[u;hn+l-i),4m]
r[Hn+m+l-i)] l(.+I-il-1 !m-I r[h n + 1 - i )]r(!m)v (I-v), 2
for 0 :<::; v ::;; 1 and 0 for v outside this interval. Since this distribution does not depend on lj -I' we see that the ratio V; is independent of lj -I' and hence independent of IVI' ... , V; - I ' Then Vi"'" Vp are independent. Theorem 8.4.1. The distribution of U defined by (3) is the distribution of the product nf~ IV;, where VI"'" ~ are independent and V; has the density (11).
The cdf of U can be found by integrating the joint density of VI"'" over the range
(12)
~
308
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
We shall now show that for given N - qz the indices p and ql can be interchanged; that is, the distributions of Up,Q"N-Q2-q, = Up,m,n and of Uq"P,N-q,-P = Um,p,n+m-p are the same. The joint density of G and WI defined in Section 8.2 when 'l = 1 and PI = 0 is
I GI t(n -p-l) e- ttr G- ttr w,w;
(13)
Let G + WIW{ =J= CC' and let WI = CV. Then
(14)
IGI
Up,m,n = IG
+ W1W{1
=
IcC' - CVV'C'I ICC'I
V'\ Ip = 11m -
, = IIp - VV
I
v'vl;
the fourth and sixth equalities follow from Theorem A.3.2 of the Appendix, and thf fifth from permutation of rows and columns. Since the Jacobian of WI = CV is model CI m = IJI tm, the joint density of J and V is
(15) . p
J]
{r[!(n+m+l-i j ]} IIp-VV'lt(n-p-O 7Ttmp
r[!(n+l-i)]
for J and Ip - VV' positive definite, and 0 otherwise, Thus J and V are independently distributed; the density of J is the first term in (15), namely, w(JIIp, n + m), and the density of V is the second term, namely, of the form (16)
KIIp - VV'I t(n-p-l)
for Ip - VV' positive definite, and 0 otherwise. Let ,) * = V', p* = m, m* = p, and n* = n + m - p. Then the density of V * is
(17)
KII - V' V It(n-p-l) p
* *
for Ip - V~ V * positive definite, and 0 otherwise. By (14), IIp - V~ V * I = 11m - V* V~ I, and hence the density of U* is
(18)
KII
p*
- V V'lt(n*-p*-I)
* *
,
8.4 DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION
309
which is of the form of (16) with p replaced by p* = m, m replaced by m* = p, and n - p - 1 replaced by n* - p* - 1 = n - p - 1. Finally we note that u".m.n given by (14) is 11m - U*U~I = Um.p.n+m-p· Theorem 8.4.2. When the hypothesis is true, the distribution of Up • q \. N-q I _q, is the same as that of Uq\.P. N-p-q, (i.e., that of U p. m . n is that of Um.p.lI+m--p). 8.4.2. Moments Since (l1)\s a density and hence integrates to 1, by change of notation
From this fact we see that the hth moment of V; is
(20)
r[t(n+m+1-i)] v t
tffV h = (1 I
r[t(n+l-i) +h]r[Hn +m+l-i)]
= r[t(n+1-i)]r[t(n+m+l-i)+h)' Since VI"'" Vp are independent, tffU h = tffnf~ I V;h = nf~ I tffV/,. We obtain the following theorem: The hth moment of U[if II > - ten + 1 - p)] is
Theorem 8.4.3. h _
(21)
tffU
n
-i=l
=
r[ Hn + 1 - i) + h] r[ H n + m + 1- i)] r [t(n+1-i)]r[Hn+m+1-i)+h]
nH i=l
r[ N - ql - q2 + 1 - i) + h] r[ HN - q2 + 1 - i)] r[HN-ql-qz+1-i)]r[±(N-q2+ 1 - i )+h]'
In the first expression p can be replaced by m, m by p, and n by . n +m -p. Suppose p is even, that is, p = 2r. We use the duplication fonnula
(22)
r( a + })r( a + 1) =
{;r~~~ + 1)
310
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Then the hth moment of U2r . m ." is (23)
CU h
_Or {r[Hm+n+2)-J]
2r."',"-
j~1
r[Hm+n+l)-J]
r[~(m+n+2)-J+h] r[Hm+n+l)-J+h]
. r [ Hn + 2) - J + h] r [ H n 1- 1) - J + h] } r[i(n+2)-J]r[Hn+l)-J]
D r
=
{r c m+n+1-2j)f(n+1-2 J +2h)} f(m+n-1-2J+2h)f(n+1-2j) .
It is clear from the definition of the beta function that (23) is
(24)
fIl!1 f(m + n + ~ - 2j) j~l \ 0 r(n + 1- 2J)f(m)
y
d } Y
where the ~. are independent and "lj has density {3(y; n + 1 - 2J, m). Suppose p is odd; that is, p = 2s + 1. Then (25)
Co'
h
_
'" U2s+1.m.n -
S
«,
(n
(
2
OZi Zs+1 ,~l
)
h
, .
where the Zi are independent and Zi has density {3(z; n + 1 - 2i, m) for i = 1, ... , sand Zs+ 1 is distributed with density f3[z; (n + 1 - p)j2, mj2l. Theorem 8.4.4. U2r.~'.n is distributed as n~~lr?, where Y1'''''Yr are independelll and ~ has density {3(y; n + 1 - 2i, m); U2 s+ 1 m n is distributed as n;o.IZi2Z,+I' where the Zi' i= 1, ... ,s, are independent'a',w Zi has density {3(z; II + 1 _. 2i, m), and Zs+ 1 is independently distributed with density {3[z;i(n 1- 1- pHIII]. 8.4.3. Some Special Distributions p
=1
From the preceding characterization we see that the density of U1, m, n is (26)
r[l(n + m)] ,n , - 1 (1 _ ) ,m , -1 2 r(i.n)r(!m) u u.
8.4
311
DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION
Another way of writing UI , m, n is
(27)
1 + (m/n)Fm,n '
where gil is the one element of G = Nin and Fm,n is an F·statistic. Thus 1 - VI, m U1.m,n
(28)
n •
!!... = F
m,n'
m
Theorem 8.4.5. The distribution of [(1- UI,m,n)/UI,m,n]'n/m is the F·distribution with m and n degrees of freedom; the distribution of [(1-Up,l,n/Up,l,n]-(n+l-p)/p is the F·distribution with p and n+l-p degrees of freedom.
p=2
From Theorem 8.4.4, we see that the density of (29)
VU
2 , m, n
is
r(n + m -1) x n- 2(1_x)m-1 r(ll- l)f(m) ,
and thus the density of Uz, m," is
(30)
f(n + m - 1) ~(n-3)(1 2r(n - ])f(m) u
_
C)m-I yu.
From (29) it follows that
(31)
l-ru;::: ~
VU
n-]
·-----,:;;-=FZm,Z(n-I)·
2,m,n
Theorem 8.4.6. The distribution of [(1- VUZ,m,n)/ VUz,m,n]-(n -l)/m is the F·distribution with 2m and 2(n - 1) degrees of freedom; the distribution of [(1- ";Up,Z,n)/ ";Up,Z,/I Hn + 1 - p)/p is the F-distribution with 2p and 2(n + 1 - p) degrees of freedom,
pEven
Wald and Brookner (1941) gave a method for finding the distribution of Up,m,n for p or m even. We shall present the method of Schatzoff (1966a). It will be convenient first to consider Up,m,n for m = 2r. We can write the event nf~l~ ~ u as
(32)
YI + ... + y;,
~
-log u,
312
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
where Y1 , •.• , r;, are independent and
Yi =
-log V; has the density
for 0 ::; Y < CXJ and 0 otherwise, and
(34)
r[1(n+l-i)+r] = __I_'ilfl+1-i+2j (r-l)!j=o 2
K i =r[1(n+l-i)]r(r)
The joint density of Y1, ••. , r;, is then a linear combination of terms exp[ - r.f= 1a i Yi J. The density of U) = r.{ = 1Yi can be obtained inductively from the density of U)-l = r.{::l Yi and l-j, j = 2, ... , p, which is a linear combination of terms w/- 1 eCWj - 1 +a jYj. The density of U-j consists of linear combinations of
(35)
if
+(
_l)k+l
ea.w J J
k!
(c - a j )
aj
= c,
k+l
The evaluation involves integration by parts. Theorem 8.4.7. If P is even or if m is even, the density of Up,m,n can be expressed as a linear combination of terms ( -log U)k u l , where k is an integer and I is a half integer. From (35) we see that the cumulative distribution function of -log U is a linear combination of terms w k e- 1w and hence the cumulative distribution function of U is a linear combination of terms (-log U)kU I• The values of k and I and the coefficients depend on p, m, and n, They can be obtained by inductively canying out the procedure leading to Theorem 8.4.7. Pillai and Gupta (1969) used Theorem 8.4.3 for obtaining distributions.
8.4
DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION
313
An alternative approach is to use Theorem 8.4.4. The complement to the cumulative distribution function U2r , m." is
(36)
Pr{U2r ,m,,,
~ u} =
pr{tp> ru}
=JI" J- "'J \
\
_
.jii
I
r
·"dv,dy. .fii.nf3(y\n+1-2i I 'm)dv .,r . _ 1
--,··1
Y,
Ili~,' Y,
In the density, (1 - y)m -I can be expanded by the binomial theorem. Then all integrations are expressed as integrations of powers of the variables. As an example, consider r = 2. The density of Y\ and Yz is
_
[m-\
-c i,j=O L:
1) !]2( _1)i+1 ,,-2+; ,,-4+i (m --'-1)1( Y2 , i . m _·_l)I.,.,Yl J .I.J. [( m
-
where
(38)
c=
f(n+m-1)f(n+m-3) f(n - l)r(n - 3)r2(m)
The complement to the cdf of U4• m," is m-\ [(m -1) !]2( _1)i+1 (39) Pr{U4,m,,,~U}=C i,j=O L: (m -'-1)1( i . m -'-1)1"" J .I.J.
m-l
=Ci'~O
[(m -1)!]\ _l);+i (m-i-1)!(m-j-1)!i!j!(n-3+j)
The last step of the integration yields powers of of and log u (for 1 + i - j = -1).
ru
ru and products of powers
314
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Particular Values Wilks (1935) gives explicitly the distrihutions of U for p = 1, p = 2, P = 3 with m = 3; p = 3 with In = 4; and p = 4 with m = 4. Wilks's formula for p =0 3 with m = 4 appears to be incorrect; see the first edition of this book. Consul (1966) givl.!s many disIrihulivns for spl.!cial cas\,;s. See also Mathai (1971).
8.4.4. The Likelihood Ratio Procedure Let up.m.n(a) be the a significance point for Up. m.,,; that is, ( 40)
Pr{Up. m. n ~up.m.,,(a)IH true} = a.
It is shown in Section 8.5 that -[n - i(p -m + l)]logUp • m . n has a limiting XC-distribution with pm degrees of freedom. Let X;II,(a) denote the a significance point of Xp2m, and let (41 )
- [II - ~ ( P Cp.lIl.n-p+1 (a) =
/II
+ I) ling LI fl. 111,..( a) "( ) .
XPIII a
Table B.l [from Pearson and Hartley (1972)] gives value of Cp.m.M(a) for a = 0.10 and 0.05. p = 1(1 )10. various even values of m, and M = II - P + 1 = 1(1)10(2)20.24,30,40,60,120. To test a null hypothesis one computes Up. "'./. and rejects the null hypothesis at significance level a if
Since Cp .",. n(a) > 1, the hypothesis is accepted if the left-hand side of (42) is less than xim(a). The purpose of tabulating Cpo m. M(a) is that linear interpolation is reasonably accurate because the entries decrease monotonically and smoothly to 1 as M increases. Schatzoff (l966a) has recommended interpolation for odd p by using adjacent even values of p and displays some examples. The table also indicatl.!s how accurate the X "-approximation is. The table has been extended by Pillai and Gupta (1969).
8.4.5. A Step-down Procedure The criterion U has been expressed in (7) as the product of independent beta variables VI' V2 , ... , Vp. The ratio V; is a least squares criterion for testing the null hypothesis that in the regression of xi - PilZI on Z = (Z; Zz)' and
8.4
315
D1STRIBUTlO'" OF THE LIKELIHOOD RATIO CRITERION
Xj _ 1 the coefficient of ZI is O. The null hypothesis that the regression of X on ZI is ~, which is equivalent to the hypothesis that the regression of X - ~ Z I on Z I is 0, is composed of the hypotheses that the regression of xi - ~ilZI on ZI is 0, i = 1, ... , p. Hence the null hypothesis PI = ~ can be tested by use of VI"'" Since ~ has the beta density (11) under the hypothesis ~il = ~il'
v;,.
(43)
1-V;n-i+1 Vi m
has the F·distribution with m and n - i + 1 degrees of freedom. The stepdown testing procedure is to compare (43) for i = 1 with the significance point Fm,n(e l ); if (43) for i = 1 is larger, reject the null hypothesis that the regression of xi - ~ilZI on ZI is 0 and hence reject the null hypothesis that PI = ~, If this first component null hypothesis is accepted, compare (43) for i = 2 with Fm,n-l(eZ)' In sequence, the component null hypotheses are tested. If one is rejected, the sequence is stopped and the hypothesis PI = ~ is rejected. If all component null hypotheses are accepted, the composite hypothesis is accepted. When the hypothesis PI = ~ is true, the probability of accepting it is nr~l(1 - e). Hence the significance level of the step-down test is 1 - nr_l(1 - e). In the step-down pr.ocedure tt.e investigator usually has a choice of the ordering of the variablest (i.e., the numbering of the components of X) and a selection of component significance levels. It seems reasonable to order the variables in descending order of importance. The choice of significance levels will affect the rower. If e j is a very small number, it will take a correspondingly large deviation from the ith null hypothesis to lead to rejection. In the absence of any other reason, the component significance levels can be taken equal. This procedure, of course, is not invariant with respect to linear transformation of the dependtnt vector variable. However, before cafrying out a step-down procedure, a linear transformation can be used to determine the p variables. The factors can be grouped. For example, group XI"'" x k into one ~et and xk+P, .. ,xp into another set. Then Uk,m,n =n7~IVi can be used to test the null hypothesis that the first k rows of PI are the first k rows of ~, Subsequently nr~k+ I Vi is used to test the hypothesis that the last p - k lOWS of PI are those of ~; this latter criterion has the distribution under the null hypothesis of Up-k,ln,n-k' tIn some cases the ordering of variables may be imposed; for example, observation at the first time point, X2 at the second time point, and so on.
XI
might be an
316
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
The investigator may test the null hYP?thesis PI = ~ by the likelihood ratio procedure. If the hypot\1esis is rejected, he may look at the factors VI"'" to try to determine which rows of PI might be different from ~. The factors can also I)e used to obtain confidence regions for !J11"'" ~pl' Let vi(e) be defined by
v;,
Vie e;) Vie e;)
1-
(44)
Then a confidence region for
(45)
n - i +1 m ~ i1
of confidence 1 - e i is
xjxj'
xiX;_1
xjZ'
Xi_lxi' lxi'
Xi_IX;_1
Xi_IZ' ZZ'
(xi - ~ilZI)(xi - ~ilZd' Xi_l(xi - PilZI)' Zz( xi - Pilzd'
ZXI_ I
(xi - ~iIZI)X;_1 Xi-lXI-I
(xi - ~ilZI)Z; Xi_IZ;
ZzXI_ I
ZzZ;
Xi_IX;_1 \ ZzXI_ I Xi_IX:_ I \ ZX;_I
8.5. AN ASYMPTOTIC EXPANSION OF THE DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION 8.5.1.
Gen~ral
Theory of Asymptotic Expansions
In this seeton we develop a large-sample distribution theory for the criterion studiea in this chapter. First we develop a general asymptotic expansion of the distribution of a random variable whose moments are certain functions of gamma functions [Box (949)]. Then we apply it to the case of the likelihood ratio criteIion for the linear hypothesis. We consider a random variable W (0 ::;; W::;; ]) with hth moment t
h=O,l, ... ,
(1) tIn all cases where we apply this result, the parameters is a distribution with such moments.
Xk'
~k> Yi' and 1Jj will be such that there
8.5
ASYMPTOTIC EXPANSION OF DISTRIBUTION OF CRITERION
317
where K is a constant such that $Wo = 1 and a
(2)
L,
b Xk
k= I
= L, Yj· j= \
It will be observed that the hth moment of A = uj~" n is of this form where x k = ~N = Yj' I;k = ~(-q + 1 - k), TJj = ~(-q2 + 1 - j), a = b = p. We treat a more general case here because applications later in this book require it. If we let
(3)
M= -210gW,
the characteristic function of pM (0 s p < 1) is
(4)
cjJ(t)=tCe i1pM
= tCW- 2i1 p
Here P is arbitrary; later it will depend on N. If a = b, X k = Yk, /;k S TJk' then of powers of variables with beta h for which the gamma functions We shall assume here that (4) holds apply the result we shall verify this assumption. Let
(1) is the hth moment of the product distributions, and then (1) holds for all exist. In this case (4) is valid for all real t. for all real t, and in each case where we
(5)
<1> ( t) = log cjJ ( t) = g ( t) - g ( 0) •
where
a
+ L, log
r[
pxk(l - 2it) + 13k + I;k]
k=\ b
- L,logr[Py/1-2it)+ej +TJj], j=l
where 13k = (1 - p)x k and ej = (1- p)Yj' The form get) - g(O) makes <1>(0) = 0, which agrees with the fact that K is such that cjJ(O) = l. We make use of an
318
TESTING THE GENERAL LINEAR HYPOTHESIS; MAN OVA
expansion formula for the gamma function [Barnes (1899), p. 64] which is asymptotic in x for bounded h: log r( x + h) = logfu + (x + h - ~) logx - x
( 6)
~
(-1)
_. L...
,~1
, B,+,(h) ) ( 1)' +Rm+1(x , r r+ x
t
where Rm +1(X) = O(x-(",+1)) and B/h) is the Bernoulli polynomial of degree,. and order unity defined by; Te'"
(7)
-T-1 = e -
x
Tr
L -, r. B,(h).
,~O
The first three polynomials are [Bo(h)
1]
=
Bl(h)=h-~, B~(h)=h2_h+L
(8)
B 3 ( h) = h 3
~h2
-
+ ~h.
13k + gb
Taking x = px k (1 - 2it), PYj(1- 2it) and h = obtain (9)
(t) = Q - g(O) -
!f log(1- 2it)
m
+
L
E:j + 7]j in turn, we
h
a
w,(1- 2itf' +
,~1
L
O(x;(m+l))
+
k~l
L
O(Yj-(m+l)),
j~l
where (10)
( 11)
( 12)
f= -
2{ ~ gk -17]j -
_ (- 1) ,+ 1 r(r+1)
(u,-
Q=
{L k
B,+ I (
ha-
b) },
13k + gk) _
(px k
L j
)'
B,+, (E:j +, 7]j) }, (PYj)
h a - b) log 2'7T - tfiog P + L (Xk + gk - ~)log x k - L (Yj + 7]j k
t)log Yj'
j
means Ixm+IRm+l(x)1 is bounded as Ixl-->oo. IThis definition differs slightly from that of Whittaker and Watson [(1943), p. 126], who expand h rk '-I)/(e'·-1l.1f 8:(") is Ihis second Iype of polynomial, 8,(hl=Bf(")-t, B2 /iJ)= B2_(") + (- [)" 'B" where B, is the rlh Bernoulli number. and B2,+ ,(") = Bi,+ 1("). 'Rm.I(X)=O(X-Im+I)
8.5
319
ASYMPTOTIC EXPANSION OF DISTRIBUTION OF CRITERION
One resulting form for cf>(t) (which we shall not use here) is
(13)
cf>(t) = C(I) = eQ- g (O)(l - 2itf tf
Eau(1- 2itfU + R:':,+l'
v=o
where L~_oauz-u is the sum of the first m + 1 terms in the series expansion of exp( - L;'~o wrz- r ), and R:':,+. is a remainder term. Alternatively,
(14)
(t)=-iflog(1-2it)+ Ewr[(1-2it)-r- 1]+R:"+ 1, r-I
where R:"+l = EO(x;(m+l»)
(15)
k
+
EO(Yj-(m+l»). j
In (14) we have expanded g(O) in the same way we expanded g(t) and have collected similar terms. Then (16)
cf>( t) = e(t) =(1-2itfli
expL~1 wr(1-2it)-r - r~1 Wr+R'm+l)
= (1- 2it) -li
{fl
X
fl (1 -
wr +
[1 + wr(l- 2it) -r +
i! w;(1- 2it) -Zr ... ]
i! wrz - ... ) + R':"+l}
= (1- 2it) -li[l + T1(t) + Tz(t) + ... + Tm(t) +R'~+I], where Tr(t) is the term in the expansion with terms example,
~
wi' ... w:',
Lis . = r; for
In most applications, we will have x k = Ck8 and Yj = d j 8, where c k and d j will be constant and 8 will vary (i.e., will grow with the sample size). In this case if p is chosen so (1- p)x k and (1- p)Yj . have limits, then R'::.+l is O(8-(m+l». We collect in (16) all terms wi' ... Lisi = r, because these terms are O( 8- r ).
w:',
320
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
It will be observed that T,.(t) is a polynomial of degree r in (1 - 2it)-1 and each term of (1- 2it)- ti1',(t) is a constant tir.les (I - 2it)- ~u for an integral v. We know that (1 - 2it)- i .. is the characteristic function of the xZ-density with v degrees of freedom; that is,
(19)
Let S, ( z) = {YOoo 2~ (1 - 2 it) -
tr 1', ( t ) e - i IZ dt,
(20) iv R 111+1
--foo -
-00
1(1
21T
-
2--)-tr R ", It
111+1 ('
-ilzd
t.
Then the density of pM is
=gr(Z) + w1[gr+z(z) -gr(z)]
+ {wz[gr+4(Z) -gr(z)]
Let
(22)
The cdf of M is written in terms of the cdf of pM, which is the integral of
~.5
321
ASYMPTOTIC EXPANSION OF DISTRIIlUTION OF CRITERION
the density, namely, (23)
Pr{MsMo}
= Pr( pM s pMol m
=
L
~(pMo) + R~+I
r~O
= Pr{
xl s pMo} + wo(Pr{ Xi+2 s pMo} -
Pr{ xl
+ [wz(pr{ x/+4 s pMo} - Pr( xl s pMo}) + - 2 Pr{ xl+2
s pMo})
~~ (Pr{ xk. s pMl,j
s pMo} + Pr{ xl s PM o})]
The remainder R~+I i~ O(lr(m+I»; this last statement can be verified by following the remainder terms along. (In fact, to make the proof rigorous one needs to verify that eal:h remainder is of th~ proper order in a uniform sense.) In many cases it is desirable to choose p so that WI = O. In such a case using only the first term of (23) gives an error of order (r 2 • Further details of the eXI,ansion can be found in Box's paper (1949).
Theorem 8.5.1. Suppose that GW h is given by (1) for all pure(v imaginary h, with (2) holding. Then the cdf of - 2p log W is given by (23). The error, R;;'+l' is O(II-(m+l» if Xk ~~ ckll, Yj ~ djll (C k > 0, d j > 0), and if (1 - p)X k • (1- p)Yj have limits, where p may depend on II. Box also considers approximating the distribution of - 2 p log IV by an F-distribution. He finds that the error in this approximation can be made to be of order 11- 3•
8.5.2. Asymptotic Distribution of the Likelihood Ratio Criterion We now apply Theorem 8.5.1 to the distribution of - 2 log A, the likelihood ratio criterion developed in Section 8.3. We let W = A. The 11th moment of A is
(24)
(~'Ah =KOC=lf[
(N -q + 1- k +Nh)] O[=lf[ (N-q2+ I -j+Nh)] '
322
TESTING THE GENERAL LINEAR HYPOTHESIS; Iv' ANOV A
and this holds for all h for which the gamma functions exist, including purely imaginary h. We let a = b = p, .\k
= ~N,
(25)
Jj
=
/;k=h-q+l-k),
~N,
TJj =
13k =
H -qz + 1- j),
6j
!(l·- p)N,
= 1(1- p)N.
We observe that ( 26) \
2wI
=
P
1: k~1
{{
H(1 -
p) N - q + 1 - k ]}
2 -
I N zp
_ {~[ (1 - p) N - qz + 1 - k 1}
H(1 -
p) N - q + 1 - k
~ [( 1 - p) N - qz + 1 - k] }
2 -
~pN
=
i:kr [-
2(1 - p) N + 2qz - 2 + (p + 1) + ql + 2].
To make this zero, we require that
(27)
p=
N-qz-!(p+ql+l) N
Then
(28)
pr{-2~IOgA::;Z} =Pr{-klogUp.q\.N_c =
sz}
Pr{ X;q\ S z}
+ k\ [ Y4(Pr{ X;q\ S z} - Pr{ X;q\ Sz}) +8
-yf(Pr{X;q\+4
sz} -
1
Pr{X;q\
sz})] +R~,
8.5
323
ASYMPTOTIC EXPANSION OF DISTRIBUTION OF CRITERION
where
k = pN = N - qz -
(29) (30)
"Yz =
pql(pZ + qf 48
Hp + q
I
+ 1) = n -
tc p - q
I
+ 1),
5) '
Z
(31)
"Y4 =
"Y~ + {g~0 [3p4 + 3qi + lOpZqf - 50(pZ + qO + 159].
Since A = Up~~I' n' where n = N - q, (28) gives Pr{ - k log Up,q" n ~ z}. Theorem 8.5.2. The cdf of - k log Up,q" n is given by (28) with k = n - !(p - ql + 1), and "Yz and "Y4 given by (30) and (31), respectively. The remainder term ~ O(N- 6 ). The coefficient k = n - !(p - ql + 1) is known as the Bartlett correction. If the first tenn of (28) is used, the error is of the order N- z ; if the second, N- 4 ; and if the third,t N- 6 • The second term is always negative and is numerically maximum for z = V(pql + 2)(pql) (= pql + 1, approximately). For p ~ 3, ql ~ 3, we have "Y2/kz ~ [(pZ + qf)/kF /96, and the contribution of the second term lies between -0.005[(pZ + qf)/kF and O. For p ~ 3, ql ~ 3, we have "Y4 ~ "Yi, and the contribution of the third tenn is numerically less than ("Yz/kZ)z. A rough rule that may be followed is that use of the first term is accurate to three decimal places if pZ + ~ k/3. As an example of the ca\cul..ltion, consider the case of p = 3, ql = 6, N - qz = 24, and z = 26.0 (the 10% significance point XIZB)' In this case "Yz/k2 = 0.048 and the second term is - 0.007: "Y4/k4 = 0.0015 and the third term is - 0.0001. Thus the probability of -1910g U3,6, 18 ~ 26.0 is 0.893 to three decimal places. Since
qt
(32)
-
[n -
Hp - m+ l)]log Up,m,nC a) = Cp,m,n-p+1 (a)X;m( a),
the proportional error in approximating the left-hand side by X;m(a) is Cp,m,n_p+1 - 1. The proportional error increases slowly with p and m. 8.5.3. A Normal Approximation Mudholkar and Trivedi (1980), (1981) developed a normal approximation to the distribution of -log Up,m,n which is asymptotic as p and/or m -+ 00. It is related to the Wilson-Hilferty normal approximation for the XZ-distribution. Box has shown th.lt the term of order N- 5 is 0 and gives the' coefficients to be used in the term of order N- 6 •
t
324
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
First, we give the background of the approximation. Suppose {Yk } is a sequence of nonnegative random variables such that (Yk - ILk) / O"k ~ N(O, 1) as k -> 00, where .cYk = ILk and 'Y(Yk) = O"t Suppose also that ILk -> 00 and O"l/ ILk is bounded as k -> 00. Let Zk = (Yk/ ILk)h. Then (33)
by Theorem 4.2.3. The approach to normality may be accelerated by choosing h to make the distribution of Zk nearly symmetric as measured by its third cumulant. The normal distribution is to be used as an approximation and is justified by its accuracy in practice. However, it will be convenient to develop the ideas in terms of limits, although rigor is not necessary. By a Taylor expansion we express the hth moment of Yd ILk as (34)
Yk)h .cZk =.c ( ILk 2
_ 1 + h( h - 1) O"k 2 ILk
+
h(h-l)(h-2) 44>k- 3(h-3)(0"//ILk/
24
2
ILk
+
O( -3) ILk
,
where 4>k = .c(Yk - ILk)3/ILk' assumed bounded. The rth moment of Zk is expressed by replacment of h by rh in (34). The central moments of Zkare
To make the third moment approximately 0 we take h to be
(37) Then Zk = (Yd ILk)ho is treated as normally distributed with mean and variance given by (34) and (35), respectively, with h = h o.
8.5
325
ASYMPTOTIC EXPANSION OF DISTRIBUTION OF CRITERION
Now we consider -log ~" m, n = - 'f.f= I log V;, where VI"'" VI' are independent and V; has the density f3(x;(n + 1 - 0/2, m/2), i = 1, .... p, As n --> 00 and m ...... 00, -log 1'-; tends to normality. If V has the density f3(x; a /2, b /2), the moment generating function of -log V is .:b"e-tlogV=
(38)
r[(a+b)/2]f(a/2-t) r(a/2)r[(a + b)/2 - t]
.
Its logarithm is the cumulant generating function, Differentiation of the last yields as the rth cumulant of V
r=1,2, ....
(39)
where t/J(w) = d log r(w)/dw, [See Abramovitz and Stegun (1972), p, 258, for e~ample,] From r(w + 1) = wr(w) we obtain the recursion relation ljJ(w + 1) = t/J(w) + l/w, This yields for s = 0 and 1 an integer (40)
The validity of (40) for s = 1,2" .. is verified by differentiation, [The expression for t/J '(Z) in the first line of page 223 of Mudholkar and Trivedi (1981) is incorrect.] Thus for b = 21 1
,.- I
Cr = (r - I)! L.
( 41)
j=O
(a/2+j)
,.
From these results we obtain as the rth cumulant of -log Up . 21 . n P
(42) As I ......
/-1
K r (-logUp ,21,n)=2'(r-1)!L. L. i=lj=O(n
00
_. ,1 IT1
_?
"
-])
the series diverges for r = 1 and converges for r = 2,3, and hence The same is true as p ...... 00 (if n /p approaches a positive
Kr/ K J ...... 0, r = 2,3,
constant), Given n, p, and I, the first three cumulants are calculated from (42). Then ho is determined from (37), and (-logUp . 21 .,,)"" is treated as approximately normally distributed with mean and variance calculated from (34) and (35) for h = h(J' Mudholkar and Trivedi (1980) calculated the error of approximation for significance levels of 0,01 and 0.05 for n from 4 to 66, P = 3,7, and
326
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOY A
q = 2.6, 10. The maximum error is less than 0.0007; in most cases the error is considerably less. The error for the X 2-approximation is much larger, especially for small values of n. In case of m odd the rth cumulant can be approximated by p
(43)
2'(r-I)!L
[~(m - 3)
i~1
L j~O
1
1
1
]
"+2' ,. (1I-/+1-2J) (n-l+m)
Davis (1933. 1935) gave tables of t/J( w) and its derivatives. 8.5.4. An F-Approximation
Rao (1951) has used the expansion of Section 8.5.2 to develop an expansion of the distribution of another function of Up . m . in terms of beta distributions. The constants can he adjusted so that the term aftcr the leading one is of order m .. 4 • A good approximation is to consider 1I
I-U I / U I/ s
(44)
s
ks-r pm
as F with pm and ks - r degrees of freedom, where
(45)
s=
p 2m 2 p2
-
4
+ m2 - 5 '
pm
r= 2
-1,
and k is 11- ~(p - m - 1). For p = 1 or 2 or m = 1 or 2 the F-distribution is exactly as given in Section 8.4. If ks - r is not an integer, interpolation he tween tv.·o integer values can be used. For smaller values of m this approximation is mure accurate than thc x2-approximation.
M.6. OTHER CRITERIA FOR TESTING TilE LINEAR HYPOTHESIS
8.6.1. Functions of Roots
Thus far the only test of the linear hypothesis we have considered is the likelihood ratio test. In this section we consider other test procedures. and tl2 w be the estimates of the parameters in N(Pz,:I), Let in, based on a sample of N observations. These are a sufficient set of statistics, and we shall base test procedures on them. As was shown in Section 8.3, if the hypothesis is PI = P~, one can reformulate the hypothesis as PI = 0 (by
tlw,
8.6
327
OTHER CRITERIA FOR TESTING THE LINEAR HYPOTHESIS
replacing
(1)
Xa
by
Xa -
PZa
pi z~1). Moreover,
= PIZ~) + P2Z~) = PI(Z~I) -A12A22IZ~») + (P 2 + PIAI2A2i1 )Z~2)
= PI z!(1) + Pi Z~2), where ~az!(1)z~)'
=0
and L a Z:(1)Z!(I),
=A ll .2 .
Then
131 =!3lfi
and
Pi =
P2w'
We shall use the principle of invariance to reduce the set of tests to be considered. First, if we make the transformation X: =Xa + rz~), we leave the null hypothesis invariant, since ,ffX: = PIZ!(I) + (P~ + r)Z~2) and Pi + r is unspecified. The only invariants of the sufficient statistics are i and PI (since for each there is a r that transforms it to 0, that is, I Second, the nJlI hypothesis is invariant under the transformation Z:*(I) = Cz!(I) (C nonsingular); the transformation carries PI to PiC-I. Under this transformation and !3IA11.2!3; are invariant; we consider A 11 .2 as information relevant to inference. However, these are the only invariants. Fur consider a function of PI and A ll .2 , say !(!31' A 11 .2 ). Then there is a C* that carries this into !(!3IC*-I, I), and a further orthogonal transformation carries this into !(T,1), where t i " = 0, i < V, tii ~ O. (If each row of T is considered a vector in ql-space, the rotation of coordinate axes can b{; done so the first vector is along the first coordinate axis, the second vector is in the plane determined by the first two coordinate axes, and so forth). But T is a function of IT' = PI A ll . 2 that is, the elements of T are uniquely determined by this equation and the preceding restrictions. Thus our tests will depend on i and !3IAll'2!3;. Let Ni = G and !3IAll.2!3; =H. Third, the null hypothesis is invariant when Xa is replaced by Kx a , for :I and Pi are unspecified. This transforms G to KGK' and H to KHK'. The only invariants of G and H under such transformations are the roots of
Pi,
Pi).
±
P;;
(2)
IH-lGI =0.
It is clear the roots are invariant, for
(3)
0= IKHK' -lKGK'1
= IK(H-lG)K'1 = IKI·IH-lGI·IK'I. On the other hand, these are the only invariants, for given G and H there is
328
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
a K such that KGK' = I and
(4)
11 0
12
0 0
0
0
Ip
KHK' =L=
0
where 11 ~ ···Ip are the roots of (2). (See Theorem A.2.2 of the Appendix.) Theorem 8.6.1. Let x" be 1II1 observatio/l ji"Ol/l N(PIZ~(1) + J3';z~2>, I), where Eaz:(l)z~)' = 0 and Eaz:(l)Z:(I), =A I1 . 2 • The only functions of the sufficient statistics and A 11.2 invariant under Ihe transformations x: = x" + rz~), ;::*(1) = Cz:(ll, and x: = Kxa are the roots of (2), where G = NI and H= ~IAI1.?P;.
The likelihood ratio criterion is a function of
(5)
U=
IGI IG +HI =
IKGK'I IKGK' +KHK'I
III I/+LI
which is clearly invariant under the transformations. Intuitively it would appear that good tests should reject the null hypothesis when the roots in some sense are large, for if PI is very different from 0, then will tend to be large and so will H. Some other criteria that have been suggested are (a) Eli' (b) EljO + I), (c) max Ii' and (d) min Ii. In each case we reject the null hypothesis if the criterion exceeds some specified number.
PI
8.6.2. The Lawley-Hotelling Trace Criterion Let K be the matrix such that KGK' =1 [G=K-I(K,)-I, or G- I =K'KJ and so (4) holds. Then the sum of the roots can be written p
(6)
L Ii = tr L = tr KHK'
;=1
= tr HK' K = tr HG -I. This criterion was suggested by Lawley (1938), Bartlett (1939), and Hotelling (1947), (1951). The test procedure is to reject the hypothesis if (6) is greater than a constant depending on p, m, and n.
8.6
329
OTHER CRITERIA FOR TESTING THE LINEAR HYPOTHESIS
The general distribution t of tr HG - I cannot be characterized as easily as that of Up,m,n' In the case of p = 2, Hotelling (1951) obtained an ell:plicit e:<pression for the distribution of tr HG -I = II + 11 , A slightly different form of this distribution is obtained from the density of the two roots II and I, in Chapter 13. It is
P)
Pr{trHG-I.:5.w}=lw/l1+w)(m-l,II-1)
-
{; r [ h m + n -
r(tm)rOn)
1)
1(1 +w) -l(
n- I)
[ 1
1."/(1+0»'
1
1
~(m -I),~(n-' 1) .
v"here I/a, b) is the incomplete beta junction, that is, the integral of f3(y: a, b) fcom 0 to x. Constantine (966) expressed the density of tr HG - I as an infinite series in generalized Laguerre polynomials and as an infinite series in zonal polynomials; these series, however, converge only for tr HG -I < I. Davis (968) showed that the analytic continuation of these series satisfies a system of linear homogeneous differential equations of order p. Davis (l970a, 1970b) used a solution to compute tables as given in Appendix B. Under the null hypothesis, G is distributed as L'~_IZaZ~ Cn = N - q) and H is distributed as Lb~ I Y"Y,:, where the Za and Y" are independent, each with distribution NCO,:I). Since the roots are invariant under the previously specified linear transformation, we can choose K so that K:IK' = I and let G* = KGK' [= L(KZ.,)CKZ,,)') and H* = KHK'. This is equivalent to assuming at the outset that :I = I. Now
(8) This result follows applying the Cweak) law of large numbers to each element of O/n)G,
(9) Theorem 8.6.2. Let fCH) be a junction whose discontinuities foml a set of probability zero when H is distributed as L7~ I Y"Y,: with the Y,. independent, each Then the limiting distribution of f(NHG -I) is the with distribution NCO, distribution of fCH).
n.
tLawlcy (I93H) purported to derive Ihe exact t1istrihution. hut the resull is in error.
l
330
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Proof This is a straightforward application of a general theorem [for example, Theorem 2 of Chernoff (1956)] to the effect that if the cdf of Xn converges to that of X (at every continuity point of the latter) and if g(x) is a function whose discontinuities form a set of probability 0 according to the distribution of X, then the cdf of g(Xn ) converges to that of g(X). In our case XII consists of the components of Hand G, and X consists of the components of H and I. •
Corollary 8.6.1. The limiting distribution of N tr HG -lor n tr HG -I is the with pql degrees of freedom.
"X 2-distribution
This follows from Theorem 8.6.2, because p
( 10)
tr H =
P
ql
r. hi; = r. r. y,;. ;=1
i=1 u=1
Ito (956), (1960) developed asymptotic formulas, and Fujikoshi (1973) extended them. Let wp.m./a) be the a significance point oftr HG- 1 ; that is, PI)
and let X;( a) be the a-significance point of the X 2-distribution with k degrees of freedom. Then , 1 [p+m+l 4 (12) nli p,m.n(a)=Xp-m(a)+2n pm+2 Xpm(a)
+(P-m+l)X}m(a)]+O(n- 2 ).
Ito also gives the term of order n -2. See also Muirhead (1970). Davis (1970a), (1970b) evaluated the accuJJcy of the approximation (12). Ito also fouml (13)
1 [p+m+l
Pr{ntrHG- ' sz}=Gpm (z)-2n
pm+2
+(p-m
Z2
+ l)gpm(Z)] +O(n- 2 ),
where Gk(z)=Pr{x;sz} and gk(Z) = (djdz)Gk(z). Pillai (1956) suggested another approximation to nWp,m.n(a), and Pillai and Samson (1959) gave moments of tr HG- 1 • Pillai and Young (1971) and Krishnaiah and Chang (1972) evaluated the Laplace transform of tr HG -I and showed how to invert
8.6
OTHER CRITERIA FOR TESTING THE LINEAR HYPOTHESIS
331
the transform. Khatri and Pillai (1966) suggest an approximate distribution based on moments. PiIlai and Young (1971) suggest approximate distributions based on the first three moments. Tables of the significance points are given by Grubbs (1954) for p = 2 and by Davis (1970a) for p = 3 and 4, Davis (1970b) for p = 5, and Davis (1980) for p = 6(1)10; approximate significance points have been given by PiIlai (1960). Davis's tables are reproduced in Table B.2. 8.6.3. The Bartiett-Nanda-Pillai Trace Criterion Another criterion, proposed by Bartlett (1939), Nanda (1950), and PiIIai (1955), is
~ li ()-I V= i..J l+l. =trL I+L
(14)
i=l
I
= tr KHK'(KGK' +KHK,)-I = tr HK'[K(G + H)K'j-1 K =trH(G+H)-\
where as before K is such that KGK' = I and (4) holds. In terms of the roots fj = LJ(1 + l), i = 1, ... , p, of (15)
IH-f(H+G)1 =0,
the criterion is r..f= 1 fi. In principle, the cdf, density, and moments under the null hypothesis can be found from the density of the roots (Sec. 13.2.3),
(16)
p
p
i=1
;=1
,
CTI.t;Hm-p-l) TI (1- !;) ,(n-p-I) TI (J; - fj), i<j
where (17)
1Ttp2rp[ ~(m + n) 1
c= rp(~n)rp(~m)rA~p)
for 1 > fl > ... > fp > 0, and 0 otherwise. If m - p and n - p are odd, the density is a polynomial in fl' ... ' fp. Then the density and cdf of the sum of the roots are polynomials. Many authors have written about the moments, Laplace transforms, densities, and cdfs, using various approaches. Nanda (1950) derived the distribution for p = 2,3,4 and m =p + 1. Pillai (1954),(1956),(1960) and Pillai and
332
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Mijares (1959) calculated the first four moments of V and proposed approximating the distribution by a beta distribution based on the first four moments. Pillai and Jayachandran (1970) show how to evaluate the moment generating function as a weighted sum of determinants whose elements are incomplete gamma functions; they derive exact densities for some special cases and use them for a table of significance points. Krishnaiah and Chang (1972) express the distributions as linear combinations of inverse Laplace transforms of the products of certain double integrals and further develop this technique for finding the distribution. Davis (1972b) showed that the distribution satisfies a differential equation and showed the nature of the solution. Khatri and Piliai (I968) obtained the (nonnull) distributions in series forms. The characteristic function (under the null hypothesis) was given by James (1964). Pillai and Jayachandran (1967) found the nonnuil distribution for p = 2 and computed power functions. For an extensive bibliography see Krishnaiah (1978). We now turn to the asymptotic theory. It follows from Theorem 8.6.2 that nV or NV has a limiting X2-distribution with pm degrees of freedom. Let Up • III • II (a) be defined by
(18) Then Davis (1970a),(1970b), Fujikoshi (1973), and Rothenberg (1':.l77) have shown that
(19)
2
1 [
nup.m. n ( a ) = Xpm( a) + 2n -
p+m+l 4 ( ) pm + 2 Xl'm a
Since we can write (for the likelihood ratio test)
(20)
nup.m.n(a) = x;",(a) + 2ln (p -m + l)x;",(a) +0(n- 2 ),
we have the comparison -2 1 p +m + 1 4 (21) nwp. m.n(a)=nu p,m,,,(a)+2n· pm+2 Xpm(a)+O(n ),
(22)
(-2) . 1 p + m + 1 4 () nUp,m,n(a)=nup,m.n(a ) +2n' pm+2 Xpm a +On
8.6
OTHER CRITERIA FOR TESTING THE LINEAR HYPOTHESIS
333
An asymptotic expansion [Muirhead (1970), Fujikoshi (1973)] is
(23)
Pr{nV ~z} = Gpn(Z) + ~;
[em - p -
l)Gpm(Z)
+2(p + I)Gpm + 2 (Z) - (p +m + I)Gpm + 4 (z)j + 0(n- 2 ). Higher-order terms are given by Muirhead and Fujikoshi.
Tables. Pillai (1960) tabulated 1% and 5% significance points of V for p
= 2(1)8 based on fitting Pearson curves (i.e., beta distributions with ad-
justed ranges) to the first four moments. Mijares (1964) extended the tables to p = 50. Table B.3 of some significance points of (n + m)V1m = tr(1/m)H{[l/(n + m)](G + H)}-1 is from Concise Statistical Tables, and was computed on the same basis as Pillai's. "Schuurman, Krishnaiah, and Chattopodhyay (1975) gave exact significance points of V for p = 2(1)5; a more extensive table is in their technical report (ARL 73-0008). A comparison of some values with those of Concise Statistical Tables (Appendix B) shows a maximum difference of 3 in the third rlecimal place. 8.6.4. The Roy Maximum Root Criterion
Any characteristic root of HG- 1 can be used as a test criterion. Roy (t 953) proposed 11' the maximum characteristic root of HG -1, on the basis of his union-intersection principle. The test procedure is to reject the null hypothesis if II is greater than a certain number, or equivalently, if 11 = 11/0 + 11) = R is greater than a number '1'. m. n< (1) which satisfies Pr{ R ~ r p . m .,,( OI)} =
(24)
01.
The density of the roots 11"'" Ip for p ~ m under the null hypothesis is given in (16). The cdf of R = 11' Pr{fl ~f*l. can be obtained from the joint density by integration over the range 0 ~/p ~ ... ~/l ~f*. If m - p and n - p are both odd, the density of 11" .. , Ip is a polynomial; then the cdf of II is a polynomial in 1* and the density of 11 is a polynomial. The only difficulty in carrying out the integration is keeping track of the different terms. Roy [(1945), (1957), Appendix 9] developed a method of integration that results in a cdf that is a linear combination of products of univariate beta densities and 1 eta cdfs. The cdf of 11 for p = 2 is (25)
Pr{fl ~f} = Ir(m - 1, n - 1)
-
f;rl!(m + n - 1)] , r(4m)rOn) p<m-l)(1-/r
33~
TESTING THE GENERAL LINEAR HYPOTHESIS; MAN OVA
This is dcrived in Section 13.5. Roy (1957), Chapter 8, gives the cdfs for p = 3 and 4 also.
By Theorem 8.6.2 the limiting distribution of the largest characteristic root of nHG" ' , NHG- I , nH(H+G)-I, or NH(H+G)-I is the distribution of the largest characteristic root of H having the distribution W(I, m). The dCllsitil"s of thc roots of H are given in Section 13.3. In principle, the marginal density of the largest root can be obtained from the joint density by integration, but in actual fact the integration is more difficult than that for the density of the roots of HG -I or H(B + G)-I. The literature on this subject is too extensive to summarize here. Nanda (1948) obtained the distribution for p = 2, 3, 4, and 5. Pillai (1954), (1956), (1965), (1967) treated the distribution under the null hypothesis. Other results were obtained by Sugiyama and Fukutomi (1966) and Sugiyama (1967). Pillai (1967) derived an appropriate distribution as a linear combination of incomplete beta functions. Davis (1972a) showed that the density of a single ordered root satisfies a differential equation and (1972b) derived a recurrence relation for it. Hayakawa (1967),' Khatri md Pillai (1968), Pillai and Sugivama (1969), and Khatri (1972) treated th,~ noncentral case. See Krishnai;h 11978) for more references. Tables. Tables of the percentage points have been calculated by Nanda (1951) and Foster and Rees (1957) for p = 2, Foster (19~7) for p = 3, Foster (1958) for p = 4, and Pillai (1960) for p = 2(1)6 on the basis of an approximation. [See also Pillai (1956),(1960),(1964),(1965),(1967).] Heck (1960) presented charts of the significance points for p = 2(1)6. Table B.4 of significance points of nll/m is from Concise Statistical Tables, based on the approximation by Pillai (1967).
8.6.5. Comparison of Powers
The four tests that have been given most consideration are those based on Wilks's U, the Lawley-Hotelling W, the Bartiett-Nanda-Pillai V, and Roy's R. To guide in the choice of one of these four, we would like to compare power functions. The first three have been compared by Rothenberg on the basis of the asymptotic expansions of their distributions in the nonnull case. Let v;", ... , v~'1 be the roots of (26) The distribution of (27)
8.6
335
OTHER CRITERIA FOR TESTING THE LINEAR HYPOTHESIS
is the nonceniral X 2-distribution with pm degrees of freedom and noncentrality parameter Ll=l vt. As N ~ 00, the quantity (l/n)G or (1/N)G approaches :I with probability one. If we let N ~ 00 and A 11.2 is unbounded, the noncentrality parameter grows indefinitely and the power approaches 1. It is more informative to consider a sequence of alternatives such that the powers of the different tests are different. Suppose PI = P~ is a sequence of matrices such that as N ~ 00, (P~ - P~)A 11.2(P~ - ~)' approaches a limit and hence vt, ... , VpN approach some limiting values VI' .•. ' vp, respectively. Then the limiting distribution of NtrHG- I , ntrHG- I , NtrH(H+G)-l, and n tr H(H + G)-I is the noncentral X2-distribution with pm degrees of freedom and noncentrality parar .1eter L;= I Vi. Similarly for - N log U and -nlogU. Rothenberg (1977) has shown under the above conditions that
p
+
L Vh p m+6[ xim(a)] i=1
- [
p i~ v? -
p
+ m + 1 (P + 2 i~
pm
/Ii
)2] gin" x,;", (a)] ) + (n1 ) ' I X[
0
336
(30)
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Pr{tr H( H + G) -I
2:
vp •.m • n ( a)}
= 1 - Gpm [x;m( a)
I;~ Vi]
p
+
L
v/gpm +6 [xim(a)]
;~l
where G/xly) is the noncentral X2-distribution with f degrees of freedom and noncentrality parameter y, and g/x) is the (central) x2-density with f degrees of freedom. The leading terms are the noncentral X2-distribution; the power functions of the three tests agree to this order. The power functions of the two trace tests differ from that of the likelihood ratio test by ±gpm+8[ X;m(a)]j(2n) times
(31)
"p
2_p+m+l m +2 P
L.., V,
;~I
(P"
. )2
;,(
L.., V,
;~l
L..,
Vi
__ )2_ p (p-l)(p+2)_2 V pm + 2 V ,
;~l
where Ii = 'L1_1 vjp. This is positive if (32)
~>
v
(p-l)(p+2) pm +2
where a} = 'Lf_l( Vi - 1i)2 Ip is the (population) variance of VI ••. , vp; the left-hand side of (32) is the coefficient of variation. If the Vi'S are relatively variable in the sense that (32) holds, the power of the Lawley-Hotelling trace test is greater than that of the likelihood ratio test, which in turn is greater than that of the Bartlett-Nanda-Pillai trace test (to order lin); if the inequality (32) is reversed, the ordering of power is reversed. The differences between the powers decrease as n increases for fixed VI' ••. , vp' (However, this comparison is not very meaningful, because increasing n decreases p~ -Pi and increases Z'Z.) A number of numerical comparisons have been made. Schatzoff (1966b) and Olson (1974) have used Monte Carlo methods; Mikhail (1965), Pillai and Jayachandran (1967), and Lee (1971a) have used asymptotic expansions of
8.7
TESTS AND CONFIDENCE REGIONS
337
distributions. All of these results agree with Rothenberg's. Among these three procedures, the Bartlett-Nanda-Pillai trace test is to be preferred if the roots are roughly equal in the alternative, and the Lawley-Hotelling trace is more powerful when the roots are substantially unequal. Wilks's likelihood ratio test seems to come in second best; in a sense it is maximin. As noted in Section 8.6.4, the Roy largest root has a limiting distribution which is not a X2-distribution under the null hypothesis and is not a noncentral X 2-distribution under a sequence of alternative hypotheses. Hence the comparison of Rothenberg cannot be extended to this case. In fact, the distributions .n the nonnull case are difficult to evaluate. However, the Monte Carlo results of Schatzoff (1966b) and Olson (1974) are clear-cut. The maximum root test has greatest power if the alternative is one-dimensional, that is, if V2 = .. , = vp = O. On the other hand, if the alternative is not one-dimensional, then the maximum root test is inferior. These test procedures tend to be robust. Under the null hypothesis the limiting distribution of ~1 - ~~ suitably normalized is normal with mean 0 and covariances the same as if X were normal, as long as its distribution satisfies some condition such as bounded fourth-order moments. Then in = (l/N)G converges with probability one. The limiting distribution of each criterion suitably normalized is the same as if X were normal. Olson (1974) studied the robustness under departures from covariance homogeneity as well as departures from normality. His conclusion was that the t'NO trace tests and the likelihood ratio test were rather robust, and the maximum root test least robust. See also Pillai and Hsu (1979). Berndt and Savin (1977) have noted that
(33) (See Problem 8.19.) If the X2 significance point is used, then a larger criterion may lead to rejection while a smaller one may not.
8.7. TESTS OF HYPOTHESES ABOUT MATRICES OF REGRESSION COEFFICIENTS AND CONFIDENCE REGIONS 8.7.1. Testing Hypotheses Suppose we are given a set of vector observations Xl"'" X N with accompanying fixed vectors Zl"'" ZN, where x" is an observation from N(~z".l). We let ~=(~1 ~2) and Z~=(Z~1)',Z~2)'), where ~1 and z~1), have ql (=q -qc) columns. The null hypothesis is
( 1)
338
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
where 13': is a specified matrix. Suppose the desired significance level is a. A test procedure is to compute
U=
(2)
IN~nl INI)
and compare this number with up. q,j a), the a significance point of the Up.q,.n-distribution. For p = 2, ... ,10 and even m, Table 1 in Appendix B can be used. For m = 2, ... , 10 and even p the same table can be used with m replaced by p and p replaced by m. (M as given in the table remains unchanged.) For p and m both odd, interpolation between even values of either p or m will give sufficient accuracy for most purposes. For reasonably large II. the asymptotic theory can be used. An equivalent procedure is to calculate Pr{Up.m.Tl 5 U}; if this is less than a, the null hypothesis is rejected. Alternatively one can use the Lawley-Hotelling trace criterion
(3)
W = tr ( N i w- N in) ( N in) - I
= tr (~In
- ~nAlldPlll - ~n'( Ninf\,
the Pillai trace criterion
(4)
or the Roy maximl.m root criterion R, where R is the maximum root of
These criteria can be referred to the appropriate tables in Appendix B. We outline an approach to computing the. criterion. If we let Ya =xa ~~ z~.l). then Y.. can be considered as an observation from N(ll.z a , I), where .1. = (.1. 1 .1. z ) = (~I - ~7 ~z)' Then the null hypothesis is H:.1. 1 = 0, and
( 6)
( 7) Thus the problem of testing the hypothesis ~\ = ~~ is equivalent to testing the hypothesis .1.\ = 0, where $Ya = .1.z a . Hence let lis suppose the problem is testing the hypothesis ~I = 0. Then
Niw = LXaX~ -
pzwAzzp;w and
339
1:1.7 TESTS AND CONFIDENCE REGIONS
Nin = LXaX~ - J3nAJ3;l. We have discussed in Section 8.2.2 the computation of J3nAJ3~ and hence Ni n. Then J32wA22J3;w can be computed in a similar manner. If the method is laid out as
(8) the first q2 rows ond columns of A* and of A** are the same as the result of applying the forward solution to the left-hand side of (9) and the first q2 rows of C* and C** are the same as the result of applying the forward solution to the right-hand side of (9). Thus 132 wA22J3;w = Ci Ci* " where C*' = (C1' Ci ') and C**' = (Ci*' Ci* '). The method implies a method for computing a determinant. In Section A.5 of the Appendix it is shown that the result of the forward solution is FA =A*. Thus IFI·IAI = IA*I. Since the determinant of a triangular matrix is the product of its diagonal elements, IFI = 1 and IAI = IA* I = nf~1 This result holds for any positive definite matrix in place of A (with a suitable modification of F) and hence can be used to compute INinl and INiwl.
at.
8.7.2. Confidence Regions Based on U We have considered tests of hypotheses PI = P':', where P~ is specified. In the usual way we can deduce from the family of tests a confidence region for PI. From the theory given before, we know that the probability is 1 - a of drawing a sample so that
(10) Thus if we make the confidence-region statement that
PI
satisfies
(11) where (11) is interpreted as an inequality on p) = PI> then the probability is 1 - a of drawing a sample such that the statement is true.
~
PI
Theorem 8.7.1. The region (11) in the P1-space is a confidence region for with confidence coefficient 1 - a.
340
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Usually the set of PI satisfying (] 1) is difficult to visualize. However, the inequality can be used to determine whether trial matrices are included in the region. 8.7.3. Simultaneous Confidence Intervals Based on the Lawley-Hotelling Trace Each test procedure implies a set of confidence regions. The Lawley-Hotelling trace criterion can be used to develop simultaneous confidence intervals for linear combinations of elements of PI' A confidence region with confidence coefficient 1 - a is
To derive the confidence bounds we generalize Lemma 5.3.2. Lemma 8.7.1.
For positive definite mat/ices A and G, ItrcP'y!5,';trA IcP'GcP';trAY'G-1y.
(13)
Proof Let b = tr cP'Y/tr A -I cP'GcP. Then 05, tr A(Y - bGcPA- I),G- I (Y - bGcPA -I) = tr AY'G- I Y - b tr cP'Y - b tr Y'el> + b 2 tr cP'GcPA- 1
(14)
=trAY'G-Iywhich yields (13).
(trcP'y)2 tr A -I cP'GcP '
•
Now (12) and (13) imply that (15)
I
Itr cP't3w - tr cP'PII = tr cP'(t3ln - PI)
I
5, ,; tr Aj"112 cP' NillcP' tr A II .2(t3111 - PI )'( N in) -I (t3I!, _. PI) 5, Vtr A I / 2 cP' NincP VW p.m • n ( a) holds for all p (16)
tr
X
cP'~w -
m matrices cP. We assert that
VNtr Aj"/2cP'incP "';wp.m,n( a) 5, tr cP'P I 5, tr cP't3 lll + V N tr Aj"I\cP'incP VWp,m,n( a)
holds for all cP with confidence 1 - a.
8.7
341
TESTS AND CONFIDENCE REGIONS
The confidence region (12) can he explored hy use of (16) for various cPo If CPik = 1 for some pair (I, K) and 0 for other elements, then (16) gives an interval for f3IK' If CPik = 1 for a pair (I, K), -1 for (I, L), and 0 otherwise. the interval pertains to f3IK - f3/L' the difference of coefficients of two independent variables. If CPik = 1 for a pair (I, K). - 1 for (J. K). and 0 otherwise, one obtains an interval for f3IK - f3JK' the difference of coefficients for two dependent variables. 8.7.4. Simultaneous Confidence Intervals Based on the Roy Maximum Root Criterion
A confidence region with confidence 1 - a based on the maximum root criterior. is
where ch)(C) denotes the largest characteristic root of C. We can derive simultaneous confidence bOlmds from (17). From Lemma 5.3.2, we find for any vectors a and b
(18)
[a'(Plll- PI)br = {[(PllI- PI)'aj'b}" ~ [(PIU- PI)'aj'AII.J(Plu- PI)'aj'b'Ai/2 b
_ a'(Plu - PI)A II .2 (PI1i -
-
a'Ga
PI)'a.
a
'G
a
'b'A- I b 11·2
with probability 1 - a; the second inequality follows from Theorem A.2.4 of the Appendix. Then a set of confidence intervals on all linear camhi nations a'Plb holding with confidence I - a is
The linear combinations are a'Plb=Lr~lr:/"~laif3j/,b". If a l =1. ai=O. i 1, and b l = 1, b h = 0, h 1, the linear combination is simply f3 11 • If a l = 1, a i = 0, i 1, and b l = 1, b 2 = - L b" = 0, h L 2. the linear combination is f311 - f312'
*"
*"
*"
*"
342
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
We can compare these intervals with (16) for = ab', which is of rank 1. The term suhtracted from and added to tr '~lll = a'~l11 b is the square root of
This is greater than the term subtracted and added to a'~lllb in (19) because wp.mja), pertaining to the sum of the roots, is greater than rp,m,n(a), relating to one root. The bounds (16) hold for all p X m matrices <1>, while (19) holds only for matrices ab' of rank 1. Mudholkar (1966) gives a very general method of constructing simultaneous confidence intervals based on symmetric gauge functions. Gabriel (1969) relates confidence bounds to simultaneous test procedures. Wijsman (1979) showed that under certain conditions the confidence sets based on the maximum root are smallest. [See also Wijsman (1980).]
8.8, TESTING EQUALITY OF MEANS OF SEVERAL NORMAL DISTRIBUTIONS WITH COMMON COVARIANCE MATRIX In univariate analysis it is well known that many hypotheses can be put in the form of hypotheses concerning regression· coefficients. The same is true for the corresponding multivariate cases. As an example we consider testing the hypothesis that the means of, say, q normal distributions with a common covariance matrix are equal. Let y~il he an observation from NC ....(i), 'l), a = 1, ... , Ni , i = 1, ... , q. The null hypothesis is
H: ....(1) = ... = ....(q).
(1)
To put the problem in the form considered earlier ir this chapter, let
(2)
x =
with N = Nl
(3)
(x x,'" x . x I_/'.\
N\ + I
... x ) = N
(y(l) y(2) ... y(l) y(l) ... I 2 N\ I
+ .. , +Nq • Let Z= (ZI
Z2
ZN\
0 0
0 0
0 0
I 0
0 0 0
0
0
0
0
0
1
1
1
ZN\+1
()
ZN)
y(q)) N q
8.8
343
TESTING EQUALITY OF MEANS
that is, Zi,,= 1 if NI + ... +N;-I < a5.N1 + ... +N;, and Zi"=O otherwise, for i = 1, ... , q - 1, and Zqa = 1 (all a). Let P = (PI P2), where PI = ( ....(1)
(q), ... , ....(q-I) - ....(q»),
-
....
(4) P2 = ....(q). Then Xex is an observation from N(Pz" , l;), and the null hypothesis is PI = O. Thus we can use the above theory for finding the criterion for. testing the hypothesis. We have
N
N)
0
0
N)
0
N2
0
N2
0
0
Nq_)
Nq_ 1
N)
N2
Nq_)
N
A= E z"z~ =
(5)
,,=1
N
(6)
C=
E x"z~ = (Ey~l) a=l
Ey~2)
...
Ly~q-I)
ex
ex
ex
Ly~i»). I.
ex
(7) Ey~i)y~i)'
=
-
NYY'
i,a
E(y~i)_ji)(y~i)-ji)'.
=
i, a
For Let
in,
Nin = LX"X~ -
we use the formula
1
0
0
(8)
PnAP~ = LX"X~ - CA-IC'.
0 0
0 0
D= 0
0
i
0
-1
-1
-1
1
344
TESTING THE GENERAL LiNEAR HYPOTHESIS; MANOVA
then
(9)
1
0
0
0
1
U
0
0 1
0 0
D- 1 = 1
0 1
Thus
(10) CA-1C' = CD'D- 1 'A- 1D- 1DC' =
CD' ( DAD') - 1 DC'
0 0
o
(11)
Nin =
L,y~i)y~i)'
0
Nq
-I
L,y~l),
a
L,y~q),
a
- L,N;y(i)y(i),
i, a
= L, (y~) -
y(i»)(y~i)
- y(i»)'.
i,a
It will be seen that iw is the estimator of :t when .....(1) = ... = .....(q) and in is the w(iJhted average of the estimators of :to based on the separate samples.
When the null hypothesis is true, INinl / INiwl is distributed as Up,q_I, n' where n = N - q. Therefore, the rejection region ~t the a significance ievel is (12)
8.8
345
TESTING EQUALITY OF MEANS
The left-hand side of (12) is (11) of Section 8.3, and
(13)
NIw - NIn = .'[y~i)y~i),
-
NJY' - ('[y~i)y~i),
- '[Nji(i)ji(i)')
I.cr
I,ll
I
"N(-(i) -y-)(-(i) =L...;y Y -y-), =H, as implied by (4) and (5) of Section 8.4. Here H has the distribution WeI, q - 1). It will be seen that when p = 1, this test reduces to the usual Ftest
'[N( -(i) __ )2 ; Y
(14)
Y
(y~i)_ji(i»)2
._Il_>F
q-I
(a)
q-I.n
.
We give an example of the analysis. The data are taken from Barnard's study of Egyptian skulls (1935). The 4 ( = q) populations are Late Predynastic (i = 1), Sixth to Twelfth (i = 2), Twelfth to Thirteenth (i = 3), and Ptolemaic Dynasties (i = 4). The 4 (= p) measurements (i.e., components of y~;» are maximum breadth, basialveolar length, nasal height, and basibregmatic height. The numbers of observations are NI = 91, N z = 162, N3 = 70, NJ, = 75. The data are Sllmn arized as
133.582418 _ 98.307692 - ( 50.835165 133.000000 (16)
134.265432 96.462963 51.148148 134.882716
134.371429 95.857143 50.100000 133.642857
135.306667 95.040000 52.093333 131.466667
NIn 9661.997470 _ 445.573 301 - ( 1130.623900 2148.584210
1
445.573301 9073.115027 1239.211990 2255.812722
1130.623900 1239.211990 3938.320351 1271.054662
2148.584210 2255.812722 1271.054 fi62 . 8741.508829
214.197666 9559.460890 1131.716372 2381.126 040
1217 .929248 1131.716372 4088.731856 1133.473898
2019 .. 8202. 16] 238l.l26 040 1133.473898 . 93S2.2-l2 7 20
From these data we find
9785.178098 _ 214.197666 - ( 1217.929248 2019.820216
346
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
We shall use the likelihood ratio test. The ratio of determinants is 5
A
(18)
U= \N: n \ = 2.4269054 X 10 =0J214344. \NlwI 2.9544475xWS
Here N = 398, n = 394, P = 4, and q = 4. Thus k = 393. Since n is very large, we may assume - k log U4,3. 394 is distributed as X 2 with 12 degrees of freedom (when the null hypothesis is true). Here - k log U = 77.30. Since the 1% point of the X~2-distribution is 26.2, the hypothesis of p..(1) = p..(2) = p..(~) = t p..(4) is rejected.
8.9. MULTIVARIATE ANALYSIS OF VARIANCE The univariate analysis of variance has a direct generalization for vector variables le.ading to an analysis of vector sums of squares (i.e., sums such as L.X" x~). In fact, in the preceding section this generalization was considered for an analysis of variance problem involving a single classification. As another example consider a two-way layout. Suppose that we are interested in the question whether the column effects are zero. We shall review the analysis for a scalar variable and then show the analysis for a vector variable. Let Y;j, i = 1, ... , r, j = 1, ... , c, be a set of rc random variables. We assume that i=l, ... ,r,
( I)
j=l, ... ,c,
with the restrictions
(2)
[, A;= [, Vj=O' ;=1
j=1
that the variance of Y;j is a Z, and that the Yfj are independently normally distributed. To test that column effects are zero is to test that
(3)
j = 1, ... , c.
This problem can be treated as a problem of regression by the introduction 'The above computations were given by Bartlett (! 947).
8.9
347
MULTIVARIATE ANALYSIS OF VARIANCE
of dummy fixed variates. Let
(4)
ZOO,ij=l, ZkO,ij =
k=i,
1,
=0, ZOk,ij
= 1, =0,
k *j.
Then (1) can be written r
(5)
,tY;j
=
/-LZOO,ij
+ [
c AkZkO,ij
+ [
k= I
The hypothesis is that the coefficients of fixed variates here,
(6)
IJkZOk.ij'
k= I
ZOk.ij
ZUJ.I1
zOO.rc
ZIO. II
ZIO. rc
Z20. II
z20.rc
ZOe II
zOc,rc
are zero. Since the matrix of
is singular (for example, row 00 is the sum of rows 10,20, ... , rO), one must elaborate the wgression theory. When one does, one finds that the test criterion indicated by the regression theory is the usual F-test of analysis of variance. Let
y=.l"y .. rc i...J lJ' i, j
(7)
348
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
and let
a = "(y. _ Yt. _ Y. ). + Y.. )2 i..J 'J i,j
= "i...J y2 _ e "LJ y2 _ r" + rey2•• ' i..J y2 oj I]
E.
i, j
j
(8) b=r[(Y.}-yj =rLy.2j-reY.~. j
Then the F-statistic is given by
F=E...(e-1)(r-1) a e -1 .
(9)
Under the null hypothesis, this has the F-distribution with e - 1 and (r - 1). (e - 1) degrees of freedom. The likelihood ratio criterion for the hypothesis is the re /2 power of
(10)
a 1 a+b= 1+{(e-l)/[(r-1)(e-l)]}F'
Now let us turn to the multivariate anal)sis of variance. We have a set of p-dimensional random ve::tors ~}' i = 1, ... , r, j = 1, ... , e, with expected values (1), where fL, the A.'s, and the v's are vectors, and with covariance matrix I, and they are independently normally distributed. Then the same algebra may be used to reduce this problem to the regression problem. We define y:.,~., y:} by (7) and A = "(Y. L... IJ - YI. - Y.J. + Y,- )(YIJ - YI. - Y.J
+ Y.• )'
i. j
= [Y;jY;j - e [Y;.Y;'. - r [Y. jY: j i, j
(11)
i
B = r" L... (Y.J - Y.. )(Y.J - Y.. )'
= r [Y. jY.'} - reY.. Y:..
+ reY..Y:.,
8.9
349
MULTIVARIATE ANALYSIS OF VARIANCE
Table 8.1
Varieties Location
M
S
V
T
P
Sums
UF
81 81
105 82
120 80
110 87
98 84
514 414
W
147 100
142 116
151 112
192 148
146 108
778 584
M
82 103
77 105
78 117
131 140
90 130
458 . 595
C
120 99
121 62
124 96
141 126
125 76
631 459
GR
99 66
89 50
69 97
89 62
104 80
450 355
D
87 68
77 67
79 67
102 92
96 94
441 338
Sums
616 517
611 482
621 569
765 655
659 572
3272 2795
A statistic analogous to (10) is
IAI IA +BI'
(12)
Vnder the null hypothesis, this has the distribution of U for p, n = (r - 1). (c - 1) and ql = C - 1 given in Section 8.4. In order for A to be nonsingular (with probal?ility 1), we must require p :'> (r - 1Xc - 1). As an example we use data first published by lmmer, Hayes, and Powers (1934), and later used by Fisher (1947a), by Yates and Cochran (1938), and by Tukey (1949). The first component of the observation vector is the barley . yield in a given year; the second component is the same measurement made the following year. Column indices run over the varieties of barley, and row indices over the locations. The data are given in Table 8.1 [e.g., ~: in the upper left-hand comer indicates a yield of 81 in each year of variety M in location UFJ. The numbers along the borders are sums. We consider the square of (147, 100) to be
( 147)(147 100
100)
= (21,609
14,700
14,700) 10,000 .
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
350
Then
"y Y' _ (380,944 '--. ij ij 315,381
( 13)
1·1
315,381 ) 277,625 '
(14)
E(6Y)(6Y)' = (2,157,924 j .I .I 1,844,346
1,844,346) 1,579,583 '
( 15)
EC 5Y )(5Y)' = (1,874,386 i 1,560,145
1,560,145 ) 1,353,727 '
(16)
(30Y )(30Y )' = (10,750,984 .. .. 9,145,240
I.
I.
9,145,240) 7.812,025 .
Then the error sum of squares is A
( 17)
= (3279
802
802) 4017 '
the row sum of squares is ( 18)
5Ec y _ Y)(Y - Y )' = (18,011 j I. .• I... 7,188
7.188) 10,345 '
and the column sum of squares is (19)
2788
B = ( 2550
2550) 2863'
The test criterion is 3279 (20)
8021
1 IAI -,--80_2_4_0_17--,= 0.4107. IA +BI = 161167 33521 3352
6880
This result is to be compared with the significant point for U2• 4• 20' Using the result of Section 8.4, we see that 1 - ".rQ.4i07 V0.4107
. ..!2. = 2.66 4
is to be compared with the significance point of F: I• 3S ' This is significant at the 5% level. Our data show that there are differer ces between varieties.
8.9
:';51
MULTIVARlATEANALYSIS OF VARIANCE
Now let us see that each F-test in the univariate analysis of variance has analogous tests in the multivariate analysis of variance. In the linear hypothesis model for the univariate analysis of variance, one assumes that the random variables Y1 , ••• , YN have expected values that are linear combinations of unknown parameters (21) where the (3's are the parameters and the z's are the known coefficients. The variables {Y,,} are assumed to be normally and independently distributed with common variance a 2. In this model there are a set of linear combinations, say L~_l Yi"Y'" where the Y's are known, such that
(22) is distributed as a2X2 with n degrees of freedom. There is another set of linear combinations, say L" I/>g" Y", where the I/>'s are known, such that ~23)
is distributed as a 2X2 with m degrees of freedom when the null hypothesis is true and as (J"2 times a noncentral X2 when the null hypothesis is not true; and in either case b is distributed independently of a. Then (24)
b n a m
[C"/3Y,,Y/3 n [d"/3Y,,Y/3· m
has the F-distribution with m and n degrees of freedom, respectively, when the null hypothesis is true. The null hypothesis is that certain {3's are zero. In the multivariate analysis of variance, Y1, • •• 'YN are vector variables with p components. The expected value of Y" is given by (21) where ~g is a vector of p parameters. We assume that the {Y,,} are normally and independently distributed with common covariance matrix I. The linear combinations LYi" Y" can be formed for the vectors. Then N
(25)
L
",/3-1
d"/3Y"Y~
352
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
has the distribution W(l:, a). When the null hypothesis is true,
has the distribution W(l:, m), and B is indeper.dent of A. Then
(27)
I[d"IlY.,Y# I
IAI
IA + BI
=
I[daIlYaY~ + [caf.l YaY~ I
has the Up.m.n-distribution. The argument for the distribution of a and b involves showing that @"La'YiaYa=O and @"La¢g"Y,,=O when certain {3's are equal to zero as specified by the null hypothesis (as identities in the unspecified (3's). Clearly this argument holds for the vector case as well. Secondly, one argues, in the univariate case, that there is an orthogonal matrix'" = (!/Inll) such that when the transformation Yf.l = L" !/If3" Z" is made n
a=
[ a.f.l.y.o
(28)
d af3 !/I"y !/If3SZy Z s = [
Z;,
a=l n+m
h=
[ a.{3.y.S
[
cull !/I"y !/I1l~ZyZIi =
Z;.
a=n+1
Because the transformation is orthogonal, the (Z) are independently and normally distributed with common variance (J'2. Since the Za' a= 1, ... ,n, must be linear combinations of La 'Yi" Y" and since Z", a = n + 1, ... , n + m, must be linear combinations of La¢g"Y", they must have means zero (under the null hypothesis). Thus a/(J'2 and b/(J'2 have the stated independent X 2-distributions. In the multivariate case the transformation Yf3 = L" !/If3"Za is used, where Yf.l and Za are vectors. Then n
A=
[ a,f.l,y,o
ZaZ~,
d"f3!/1"y!/lf.loZyZ~= [
a=1
(29)
n+m
B=
[ a,f.l.y,8
caf.l !/I"y !/If.lSZyZ~ =
[
Z"Z~
a=n+l
because it follows from (28) that La. f3 d af3 !/I"y !/If.lo = 1, 'Y = 0 ~ n, and = 0 otherwise, and La. f3 C"1l !/I"y!/lIlS = 1, n+1~'Y=o~n+m, and =0 otherwise. Since lJ1 is orthogonal, the (Z) are independently normally distributed
353
8.10 SOME OPTIMAL PROPERTIES OF TESTS
with covariance matrix
:~.
The same argument shows
tz" = 0,
ex
= 1, ... ,
n + m, under the null hypothesis. Thus A and B are independently distributed according to
WeI, n)
and
weI, m),
respectively.
H.10. SOME OPTIMAL PROPERTIES OF TESTS 8.10.1. Admissibility of Invariant Tests In this chapter we have considered several tests of a linear hypothesis which are invariant with respect to transformations that leave the null hypothesis invariant. We raise the question of which invariant tests are good tests. In particular we ask for aumissible pl'llceuures, that is, prncedurcs that cannot be improved on in the sense of smaller probabilities of Type I and/or Type II error. The competing tests are not necessarily invariant. Clearly, if an invariant test is admissihle in the class of all tests, it is admissible in the class of invariant tests. Testing the general linear hypothesis as treated here is a generalization of testing the hypothesis concerning one mean vector as treated in Chapter 5. The invariant procedures in Chapter R arc generalizations of the T'-test. One way of showing a procedure is admissible is to display a prior distribution on the parameters such that the Bayes procedure is a given test procedure. This approach requires some ingenuity in constructing the prior. but the verification of the property given the prior is straightforward. Prohlems 8.26 and 8.27 show that the Bartlett-Nanda-Pillai trace criterion V and Wilks's likelihood ratio criterion U yield admissible tests. The disadvantage of this approach to admissibility is that one must invent a prior distribution for each procedure; a general theorem does not cover many cases. The other approach to admissibility is to apply Stein's theorem (Theorem 5.6.5), which yields general results. The invariant tests can be stated in terms of the roo's of the determinantal equation
(1 )
IH-A(H+G)I=o,
where H = P]A]]'2P; = W]W; and G = Nl n = W3W3. There is also a matrix (or Wz ) associated with the nuisance parameters 132' For convenience, we define the canonical form in the following notation. Let W] = X (p x m). W2 = Y (p X r), W3 = Z (p X 11), fX= E, fY=H, and fZ = 0: the columns are independently normally distributed with covariance matrix I. The null hypothesis is E = 0, and the alternative hypothesis is E =1= O. The usual tests are given in terms of the (nonzero) roots of
P2
(2)
lxx' -
A(ZZ'
+ XX') I = IXX'
- A( U - l'Y') 1=
o.
354
TESTlN(j THE (jENERAL LINEAR
HYI'UIHto~I~;
MANUVA
where U = XX' + YY' + ZZ'. Expect for roots that are identically zero, the roots of (2) coincide with the nonzero characteristic roots of X'(U - yy,)-l X. Let V=(X,Y,U) and M(V) =X'(U-yy,)-I X .
(3)
The vector of ordered characteristic roots of M(V) is denoted by (4)
where AI:?: ... :?: Am :?: O. Since the inclusion of zero roots (when m > p) causes no trouble in the sequel, we assume that the tests depend on A(M(V». The admissibility of these tests can be stated in terms of the geometric characteristics of the acceptance regions. Let R~
(5)
= {AERmIAI:?: A2:?:
'" :?: Am:?:
O},
R";= {A E RmIA 1 :?: 0, ... , Am:?: o}.
It seems reasonable that if a set of sample roots leads to acceptance of the null hypothesis, then a set of smaller roots would as well (Figure 8.2).
Vi
Definition 8.10.1. A region A c = 1, ... , m, imply v EA.
R~
is monotone if A E A, v
~ Ai' i
Definition 8.10.2.
where
7I"
For A c
R~
the extended region A* is
ranges over all permutations of 0, ... , m).
Figure 8.2. A monotone acceptance region.
E R~
, and
355
8.10 SOME OPTIMAL PROPERTIES OF TESTS
The main result, first proved by Schwartz (1967), is the following theorem: Theorem 8.10.1. If the region A c R'~ is monotone and if the extended region A* IS closed and convex, then A is the acceptance region of an admissible test. Another characterization of admissible tests is given in terms of majorization. Definition 8.10.3. v=(v 1,· .. ,vm )' if
where A.ri) and order.
V[i)'
i
A vector A = (AI' .. ', Am)' weakly majorizes a vector
= 1, ... , m, are the coordinates rea"anged in nonascending
We use the notation A >- w v or v -< w A if A weakly majorizes v. If A, v E R~, then A >- wV is simply
(8)
Al~Vl'
Al+A2~Vl+V2"'"
Al+···+Am~Vl+·"+Vm'
If the last inequality in (7) is replaced by an equality, we say simply that A majorizes v and denote this by A >- v or v -< A. The theory of majorization and the related inequalities are developed in detail in Marshall and Olkin (1979).
v
Definition 8.10.4. A regionA cR~ is monotone in majorization v -< wA imply v EA. (See Figure 8.3.)
if A EA,
E R~,
Theorem 8.10.2. If a region A c R~ is closed, convex, and monotone in majorization, then A is the acceptance region of an admissible test.
"
Figure 8.3. A region monotone in majorization.
356
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOV A
Theorems 8.10.1 and 8.10.2 are equivalent; it will be convenient to prove Theorem 8.10.2 first. Then. an argument about the extreme points of a certain convex set (Lemma 8.10.11) establishes the equivalence of the two theorems. Theorem 5.6.5 (Stein's theorem) will be used because we can write the distribution of (X, Y, Z) in exponential form. Let V = XX' + YY' + ZZ' = (u ij ) and I-I = (a ij ). For a general matrix C=(CI""'C k ), let vec(C)= (e'I"'" cD'. The density of (X, Y, Z) can be written as
(9) f( X, Y, Z) = K(S,H, '1) exp{tr S'I-I X + trH'I-1 Y - ~tr I-IV}
= K( S, H, I) exp{ w'(I)Y(I) + W'(2)Y(2) + w'(3)Y(3)}, where K(X, H, I) is a constant, 00(2) = vec(I-IH), 00(3) =
-!( a
II,
2a 12, .•• , 2a Ip, a 22 , ... , a PP )',
(10) Y(I) = vec( X) ,
Y(2)
= vec(Y),
If we denote the mapping (X, Y, Z) -+ Y = (Y(I)' Y(2)' Y(3»' by g, Y = g(X, Y, Z), then the measure of a set A in the space of Y is meA) = JL(g-~(A», where JL is the ordinary Lebesgue measure on RP(m+r+n). We note that (X, Y, V) is a sufficient statistic and so is Y = (Y(I)' Y(2)' Y(3l. Because a test that is
admissible with respect to the class of tests based on a sufficient statistic is admissible in the whole class of tests, we consider only tests based on a sufficient statistic. Then the acceptance regions of these tests are subsets in the space of y. The density of Y given by the right-hand side of (9) is of the form of the exponential family, and therefore we ~an apply Stein's theorem. Furthermore, since the transformation (X, Y, U) -+ Y is linear, we prove the convexity of an acceptance region of (X, Y, U). The acceptance region of an invariant test is given in terms of A(M(V» = (''\'1' ... , Am)'. Therefore, in order to prove the admissibility of these tests we have to check that the inverse image of A, namely, A = (VI A(M(V» E A), satisfies the conditions of Stein's theorem, namely, is convex. Suppose V; = (Xi' Xi' U) E A, i = 1,2, that is, A[M(V;») EA. By the convexity of A, pA[M(VI ») + qA[M(V2») EA for 0 5;p = 1- q 5; 1. To show pVI + qV2 EA, that is, A[M(pVI + qV2») EA, we use the property of monotonicity of majorization of A and the following theorem.
8.10
357
SOME OPTIMAL PROPERTIES OF TESTS
>.(Il _ >.( M( 1'1» >.(2) _ >'(M(1'2»
l~~IIIIIIIIIIIi:i·~~~~~_ A(2)
A(M(pl'l
Al
+ qI'2»
Figure 8.4. Theorem 8.10.3.
Theorem 8.10.3.
i.
The proof of Theorem 8.lD.3 (Figure 8.4) follows from the pair of majorizations
(12)
A[M(pVI +qV2 )] >-wA[pM(VI) +qM(Vz )] >- wPA[ M( VI)] + qA [M(Vz )]·
The second majorization in (12) is a special case of the following lemma. Lemma 8.10.1.
For A and B symmetric,
A(A +B) >- wA(A) + A(B).
(13)
Proof By Corollary A.4.2 of the Appendix, k
(14)
LA;(A+8)= max trR'(A+B)R ;=1
R'R=I, ~
max tr R'AR
+
R'R=I,
k
max tr R' BR R'R=I,
k
= L A;(A) + L Ai(B) ;=1
;=1
k
=
L ;=1
{A;CA) + Ai(8)},
k
=
1. .... p.
•
358
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Let A > B mean A - B is positive definite and A ~ B mean A - B is positive semidefinite. The first majorization in (12) follows from several lemmas. Lemma 8.10.2
l1 S)
pUI +qU2
-
(pYI +qYz)(pYI +qY2)' ~P(UI
- YIY;) +q(U2 - Yzyn·
Proof The left-hand side minus the right-hand side is (16)
Pl'IY{ + qYzYz - p2YIY{ - q Z y 2 yZ - pq(YIYz + Y2 Y;) = p( 1 - p)YIY{ + q(l- q)YzYz - pq(Y1YZ + Y2Y;) ~
o.
B > 0, then A -I
~
= pq( YI - Y2 ) (Y1 - Y2)' Lemma 8.10.3.
If A
~
Proof See Problem 8.31. Lemma 8.10.4.
B- 1 •
•
If A> 0, thell f(x, A) =x'A-1x is convex in (x, A).
Proof See Problem 5.17. Lemma 8.10.5.
•
•
If AI > 0, A 2 > 0, then
Proof From Lemma 8.10.4 we have for all Y
(18)
py' H'IA"\IBI Y + qy' B~A21 Bzy - y'(pBI + qB 2 )'(pA I +- qA z ) -1(pBI + qB 2)y = p( Bly)'A"\I( Bly) + q(B2y)'A;;I(B2Y) -(pRIY +qB2y)'(pAI ~O.
+ qAzrl(pBly + qB 2y)
•
Thus the matrix of the quadratic form in Y is positive semidefinite.
•
The relation as in (17) is sometimes called matrix convexity. [See Marshall and Olkin (1979).]
8.10
359
SOME OPTIMAL PROPERTIES OF TESTS
Lemma 8.10.6.
(19) where VI=(X"YI,U,), V2 =(X 2 ,Y2,U2), UI-YIY;>O, U2-Y2Y~>0, OS;p =l-qs;1. Proof Lemmas 8.10.2 and 8.10.3 show that
(20)
[pU I +qU2 - (pYI +qY2)(pYI +qY2 S;
)'r 1
[p(U, - YIY;) +q(U2 - Y2Y2)]-'.
This implies
(21)
M(pVI + qV2 ) S;
(PXl + qX2)'[P(UI
-
YlY;)
+ q( U2 -
l Y2YDr (pX l
+ qX2).
Then Lemma 8.10.5 implies that the right-hand side of (21) is less than or equal to
• Lemma 8.10.7.
If As; B, then A(A) -<
w
A(B).
Proof From Corollary A.4.2 of the Appendix, k
(23)
k
L,A;(A)= max trR'ARs; max trR'BR= L,A;(B), i= I
R'R=lk
i='
R'R=lk
k=l, ... ,p.
•
From Lemma 8.10.7 we obtain the first majorization in (12) and hence Theorem 8.10.3, which in turn implies the convexity of A. Thus the acceptance region satisfies condition 0) of Stein's theorem. Lemma 8.10.8.
For the acceptance region A of Theorem 8.10.1 or Theorem
8.10.2, condition (ij) of Stein's theorem is satisfied. Proof Let w correspond to (<1>, '1',0); then (24)
w'Y =
W'(l)Y(I)
+ w(2)Y(2) + w(3)Y(3)
= tr X + tr 'I' Y I
I
t tr 0 U,
360
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
is symmetric. Suppose that {ylw'Y > c} is disjoint from A = (Vi A(M(V)) EA}. We want to show that in this case 0 is positive semidefi-
where 0
nite. If this were not true,.then
o (25)
-I
o
where D is nonsingular and -lis not vacuous. Let X=(l/y)X o, Y= (l/y)Yo,
U~ (D')-' (~
(26)
0 yI 0
~)D-"
and V = (X, Y, U), where X o, Yo are fixed matrices and y is a positive number. Then
(27)
1 <{)IXo w'y = -tr y
1 W'Y + ztr 1 + -tr o y
(-I0 0
o yI
o
for sufficiently large y. On the other hand, (28)
A(M(V))
=
A{X'(U - IT/) -I X}
~ >H(D')'(:
o yI
o
..... 0 as y ..... 00. Therefore, V E A for sufficiently large y. This is a contradiction. Hence 0 is positive semidefinite. Now let WI correspond to (<{)I' 0, I), where <{)I '" O. Then 1+ A0 is positive definite and «) 1 + A«) '" 0 for sufficiently large A. Hence WI + Aw E n - no for sufficiently large A. • The preceding proof was suggested by Charles Stein. By Theorem 5.6.5, Theorem 8.10.3 and Lemma 8.10.8 now imply Theorem 8.lD.2.
To obtain Theorem 8.lD.1 from Theorem 8.10.2, we use the following lemmas. Lemma 8.10.9. A c R':;, is convex and monotone in majorization if and only if A is monotone and A* is convex.
8.10
361
SOME OPTIMAL PROPERTIES OF TESTS
(0, A , ) · + - - - -....
e_ extreme points
etA)
Figure 8.S
Proof Necessity. If A is monotone in majorization, then it is obviously monotone. A* is convex (see Problem 8.35). Sufficiency. For X E R":: let
(29)
qX)
=
{xlxER~',x>-wX}'
D( X) = {xix
E
R":: ,x >-
,..x}.
It will be proved in Lemma 8.10.10, Lemma 8.10.11, and its corollary lhal monotonicity of A and convexity of A* implies C( X) c A*. Then D( A) = C(X)nR"::cA*nR"::=A. Now suppose l1ER'~ and l1-<wA. Then liE D(X) CA. This shows that A is monotone in majorization. Furthermore. if A* is convex, then A = R":: nA* is convex. (See Figure 8.5,) • Lemma 8.10.10. Let C be compact and conve'(, and let D be conve'(. If the extreme points of C are contained in D, then C cD.
Proof Obvious. Lemma 8.10.11.
•
Every extreme point of C(X) is of the foml
(30) where 71" is a pemlUtation of (1, ... , m) and = 0 for some k.
01
= ... = Ok = 1, 0k+ 1 = '" = 8m
Proof C(x) is convex. (See Problem 8.34.) Now note that C( A) is permutation-s)mmetric, that is. if (xl ..... Xm)'EC(A). then (x~III ..... X"(m/E C(x) for any permutation 71". Therefore, for any permutation rr. rr(C( A)) =
362
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
{(X,,-(Il •...• x"(,,,/Ix E C(A)} coincides with C(A). This implies that if (xl ..... x",)' is an extreme point of C(A), then (X"'(I)""'x,,(,,»)' is also an extreme point. In particular, (X[IJ, •.. , x[mJ) E R'~ is an extreme point. Conversely, if (Xl • . . . ' x,,,) E R'~ is an extreme point of C(A), then (x,,(I),"" x"(" / is an extreme point. We see that once we enumerate the extreme points of C( A) in R"::, the rest of the extreme points can be obtained by permutation. Suppose x E R":: . An extreme point, being the intersection of m hyperplanes. has to satisfy 111 or more of the following 2m equations:
(.31 )
E",:.r m
=[).
F,,, : XI + ... +x",
=
Al + ... + Am'
Suppose that k is the first index such that Ek holds. Then x E Rn~ implies 0= x k ~ X,+ I ~ ... ~ xm ~ O. Therefore, E k , ... , Em hold. The remaining k - 1 = 111 - (111 - k + 1) or more equations are among the F's. We order them as F;" ... ,F;; where i l -: ... I. Since x k = ... = xm = 0, (.32)
F;,:
XI
+ ...
+X k _ 1
= Al + ... +A k- I + ... +A;,.
But XI + ... +X k - l ~ Al + ... +A k- l , and we have Ak + ... +A;, = O. Therefore. 0 = Ak + .. , + A" ~ Ak ~ ... ~ Am ~ O. In this case Fk _ I , ••• , Fm reduce to the same equation XI + ... +X k _ 1 = Al + ... +Ak - l . It follows that x satisfies k - 2 more equations, which have tc' be F l , ••• , F k -- 2 • We have shown that in either case Ek, ... ,E""FI, ... ,E'_1 hold and this gives the point ,B = (Ai''''' Ak - l.' 0, ... ,0), which is in R'~ n CO.. ). Therefore, f3 is an extreme point. •
Corollary 8.10.1.
C( A) c A*.
Proof If A is monotone, then A* is monotone in the sense that if A=(A1.· ..• A",)'EA*, v=(vl, ... ,v",Y. v;~A;, i=l, ... ,I1l, then vEA*.
(See Problem 8..35.) Now the extreme points of C( A) given by (30) are in A* because of permutation symmetry and monotonicity of A*. Hence, by Lemma 8.10.10. C(A)cA*. •
8.10
363
SOME OPTIMAL PROPERTIES OF TESTS
Proof of Theorem B.I0.l. Immediate .from Theorem 8.10.2 and Lemma
8.10.9.
•
Application of the theory of Schur-convex functions yields several corollaries to Theorem 8.10.2 Let g be continuous, nondecreasing, and convex in [0, 1).
Corollary 8.10.2. Let
m
(33)
f(A.) =fP'l'"'' Am)
= [g(A;). i~l
Then a test with the acceptance region A = {A.lf(A.)::; c} is admissible. Proof Being a sum of convex functions f is convex, and hence A is convex. A is closed because f is continuous. We want to show that if f(x)::;c and y -< wX (x,y ERn;), then f(y) ::;c. Let .tk = I:~_lX;, Yk= I:1~lY;' Then y -< wX if and only if Xk '2Yk' k = 1, ... , m. Let f(x) = hU l , ... , xm) = g(il)+E7'~2g(X;-X;-I)' It suffices to show that h(xl, ... ,x m ) is increasing in each xj • For i ::; m - 1 the convexity of g implies that (34)
h( XI, ... , x; + e, ... , xm) - h( xi"'" x;, ... , xm) = g( XI
+ e) - g( x;) - {g( X;+ I)
-
g( x i +I
-
e)} '2 O.
For i = m the monotonicity of g implies
Setting gO.) = -logO - A), g(A) = A/(1 - A), g( A) = A, respectively, shows that Wilks' likelihood ratio test, the Lawley-Hotelling trace test, and the Bartlett-Nanda-Pillai test are admissible. Admissibility of Roy's maximum root test A: A[ ::; c follows directly from Theorem 8.10.1 or Theorem 8.10.2. On the contrary, the minimum root test, A,::; c, where t = min(m,p), does not satisfy the convexity condition. The following theorem shows that this test is actually inadmissible. Theorem 8.10.4. A necessary condition for an invariant test to be admissible is that the extended region in the space of fA;, ... , fA; is convex and monotone.
We shall only sketch the proof of this theorem [following Schwartz (1967)]. Let ,fA; = d j , i = 1, ... , t, and let the density of d l , ... , d , be f(dl v), where v = (vI"'" v,)' is defined in Section 8.6.5 and f(dl v) is given in Chapter 13.
364
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
The ratio f(dlv)lf(dIO) can be extended symmetrically to the unit cube (0 :s;; d; :s;; 1, i = 1, ... , t). The extended ratio is then a convex function and is strictly increasing in each d;. A proper Bayes procedure has an acceptance region (36)
f(dlv) ~C, j (dfOfdI1(v)
where n( v) is a finite measure on the space of v's. Then the symmetric extension of the set of d satisfying (36) is convex and monotone [as shown by Birnbaum (1955)]. The closure (in the weak * topology) of the set of Bayes procedures forms an essentially complete class [Wald (1950)]. In this case the limit of the convex monotone acceptance regions is convex and monotone. The exposition of admissibility here was developed by Anderson and Takemuta (1982). 8.10.2. Unbiasedness of Tests and Monotonicity of Power Functions
A test T is called unbiased if the power achieves its minimum at the null hypothesis. When there is a natural parametrization and a notion of distance in the parameter space, the power function is monotone if the power increases as the distance between the alternative hypothesis and the null hypothesis increases. Note that monotonicity implies unbiasedness. In this section we shall show that the power functions of many of the invariant tests of the general linear hypothesis are monotone in the invariants of the parameters, namely, the roots; these can be considered as measures of distance. To introduce the approach, we consider the acceptance interval (-a, a) fer testing the null hypothesis /L = 0 against the alternative /L 0 on the basis of an observation from N( /L, a 2). In Figure 8.6 the probabilities of acceptanc(; are represented by the shaded regions for three values of /L. It is clear that the probability of acceptance Qecreases monotonically (or equivalently the power increases monotonically) as /L moves away from zero. In fact, this property depends only on the density function being unimodal and symmetric.
*"
~A -a 0 p.
0
Figure 8.6. Three probabilities of acceptance.
-0
0 a
p.
365
8.10 SOME OPTIMAL PROPERTIES OF TESTS
Figure 8.7. Acceptance regions.
In higher dimensions we generalize the interval by a symmetric convex set, and we ask that the density function be symmetric and unimodal in the sense that every contour of constant density surrounds a convex set. In Figure 8.7 we illustnte that in this case the probability of acceptance decreases monotonically. The following theorem is due to Anderson (1955b). Theorem 8.10.5. Let E be a convex set in n-space, symmetric about the f(x) - f( -x), (ij) (xif(x) ~ u} = origin. Let f(x) ~ 0 be a function such that Ku is convex for every u (0 < u < (0), and (iii) fd(x) dx < x. Then
en
jf(x+ky)dx~ jf(x+y)dx
(37)
E
E
for 05 k 51. The proof of Theorem 8.10.5 is based on the following lemma. Lemma 8.10.12.
(38)
Let E, F be convex and symmetric about the origin. Then
V{ ( E + ky) n F}
~
V{ ( E + y) n F} ,
where 0 5 k 5 1 and V denotes the n-dimensional volume. Proof Consider the set aCE + y) + (1- a)(E - y) = aE + (1 - a)E + (2a - l)y which consists of points a(x + y) + (1 - a )(z - y) with x,:, E E. Let ao = (k + 1)/2, so that 2ao -1 = k. Then by convexity of E we have
(39)
ao(E+y) + C1- ao)(E-y) cE +ky.
Hence by convexity of F
ao[(E+y) nFj +(l-ao){CE-y) nFj c(E+ky) nF
]66
TESTING THE GENERAL LINEAR HYPOTIlESIS; MANOVA
and
( 40)
V( a o [( E + )') n F] + ( 1 - a o) [( E - )') n F]}
~ V{ ( E
+ ky) n F} .
Now by the Brunn-Minkowski inequality [e.g., Bonnesen and Fenchel (1948), Section 48], we have
(41)
Vi/n{ao[(E+y)nF] +(1-a o)[(E-y)nF]} ~
aoVi/n{( E + y) n F} + (1 - ao)Vi/n{(E - y) nF}
= aoVi/n{ (E + y) n F} + (1 - ao )Vi/n{ ( -E + y) n ( - F)} =vi/n{(E+y)nF}. The last equality follows from the symmetry of E and F.
•
Proof of Theorem 8.10.5. Let
H(u) = V{(E+Icy) nKul,
(42)
H*(u) = ViCE +y) nKu}·
(43) Then (44)
ff(x+y)dx= f E
f(x)dx E+y
= {Of 1(0 s. u S.}(X)}( u) dxdu o E+y
= {'H*(u)du. o
Similarly, (45)
ff(x + Icy) dx = E
t'H(u) duo 0
By Lemma 8.10.12, H(u) ~ H*(u). Hence Theorem 8.lD.5 follows from (44) and (45). •
8.10
367
SOME OPTIMAL PROPERTIES OF TESTS
We start with the canonical form given in Section 8.10.1. We further simplify the problem as follo·vs. Let t = min(m, p), and let VI"'" V, (VI;:: V2 >- ... ;:: v,) be the nonzero characteristic roots of S'.I -I S, where S = rff X. Lemma 8.10.13.
There exist matrices B (p xp) and F (m
X
m) such that
BSF' = (Dt,O),
(46)
p~m,
p>m, where Dv = diag( VI"'" v.).
Proof We prove this for the case p ~ m and vp > O. Other cases can be proved similarly. By Theorem A.2.2 of the Appendix there is a matrix B such that B.IB' =1,
(47) Let
(48) Then
(49) Let F' = (F;,
F~)
be a full m
X
m orthogonal matrix. Then
(50) and
(51) BSF' =BS(F;,F~) =BS(S'B'D;t,F2) = (DLo).
•
Now let (52)
U=BXF',
V=BZ.
Then the columns of U, V are independently normally distributed with covariance matrix 1 and means when p ~ m
(53)
rffU= (Dt,O), rffV=o.
368
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Invariant tests are given in'terms of characteristic roots i l ,.·., it ([I ~ ... ~ it) of V'(W')-IV. Note that for the admissibility we used the characteristic roots of Ai of V'(VV' + W,)-I V rather than ii = A;/O - AJ Here it is more natural to use ii' which corresponds to the parameter value IIi' The following theorem is given by Das Gupta, Anderson, and Mudholkar (1964). Theorem 8.10.6. If the acceptance region of an invariant test is convex in the space of each column vector of U for each set ofJixed vulues of V and of the other column vectors of v, then the power of the test increases monotonically in each II;. Proof Since UU' is unchanged when any column vector of U is multiplied by -1, the acceptance region is symmetr:c about the origin in each of the column vectors of V. Now the density of V= (tlij), V= (vij) is (54)
f( V, V)
=
(21Tf
t(n+m)p exp [ -
i{trw ,E (~" - F,)' ,tE u,;}], +
+
J*' Applying Theorem 8.10.5 to (54), we see that the power increases monotonically in each ~. • Since the section of a convex set is convex, we have the following corollary. Corollary 8.10.3.
If the acceptance region A of an invariant test is convex in
V for each fixed V, then the power of the test increases monotonically in each v;.
From this we see that Roy's maximum root test A: 11 ~ K and the Lawley-Hotelling trace test A: tr V'(W,)-l V ~ K have power functions that are monotonically increasing in each V;. To see that the acceptance region of the likelihood ratio test
A:
(55)
0(1 +IJ ~K
;=1
satisfies the condition of Theorem 8.10.6 let
(56)
(W')
-I
= T'T,
T:pXp
V* = (uL ... ,u~) = TV.
369
8.10 SOME OPTIMAL PROPERTIES OF TESTS
Then t
(57)
0(1 +1;) =IU'(W,)-IU+II =IU*'u* +/1 i~l
=\U*U*' +/1=luiui' +BI = (ui'B-Iui + 1)\B\ =(u'IT'B-1TuI +1)\B\, where B = uiui' + ... +u~,u~,' + I. Sin<.:c T' BIT is positive definite. (SS) is convex in Ul' Therefore, thc likelihood ratio test has a power function which is monotone increasing in each Vi' The Bartiett-Nanda-Pillai trace test
L i=l
:1 I.
t
A: tr U'(UU' + W') -I U =
(58)
1
::;K
'
has an acceptance region that is an ellipsoid if K < 1 and is convex in each column U j of U provided K ::; L (See Problem 8.36.) For K> 1 (58) may not be convex in each colt mn of U. The reader can work out an example for
p=2. Eaton and Perlman (1974) have shown that if an invariant test is convex in U and W = W', then the power at (vlJ, ••• , v,n) is greater than at (1'1 ..... I) if <{V;, ... ,F,)-<w
,R).
Vvi
~
v.;
(a)
(b)
( c)
Figure 8.8. Contours of power functions.
370
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
and similarly the Eaton-Perlman result does not exclude (b). The last result guarantees that the contour looks like (c) for Roy's maximum root test and the Lawley-Hotelling trace test. These results relate to the fact that these two tests are more likely to detect alternative hypotheses where few v;'s are far from zero. In contrast with this, the likelihood ratio test and the Bartlett-Nanda-Pillai trace test are sensitive to the overall departure from the null hypothesis. It might be noted that the convexity in .r,; -space cannot be translated into the convexity in v-space. By using the noncentral density of l;'s which depends on the parameter values VI"'" V" Perlman and Olkin (1980) showed that any invariant test with monotone acceptance region (in the space of roots) is unbiased. Note that this result covers all the standard tests considered earlier.
8.11. ELLIPTICALLY CONTOURED DISTRIBUTIONS 8.11.1. Observations Elliptically Contoured
The regression model of Section 8.2 can be written
( 1)
x" = ~z" + e",
a=l, ... ,N,
where en is an unobserved disturbance with ge" = 0 and ge"e~ = I. We assume that e" has a density IAI- ~g(e'A-le); then I=(gR 2 /p)A, where R'=e~A"leu' In general the exact distribution of B=l:~_IX"Z:rA-1 and NI = I:~'_l(X" - Bz"Xx" - Bz,,)' is difficult to obtain and cannot be expressed concisely. However, the expected value of B is p, and the covariance matrix of vec B is I ® A -I with A = I:~~ I Z" z~. We can develop a large-' sample distribution for B and Nt. Theorem 8.11.1. Suppose (1/N)A -> Ao, z~ z" < cOllStant, a = 1,2, ... , and either the e",'s are independent identically distributed or the e,,'s are independeflf with (),"le~e,,12+& < constant for some £> O. Then B.4 ~ and !Nvec(B~) has a limiting normal distribution with mean 0 and covariance matrix I®A(;I.
Theorem 8.11.1 appears in Anderson (1971) as Theorem 5.5.13. There are many alternatives to its assumptions in the literature. Under its assumptions in!" I. This result permits a large-sample theory for the criteria for testing null hypotheses about ~. Consider testing the null hypothesis
(2)
8.11
371
ELLIPTICALLY CONTOURED DISTRIBUTIONS
where p* is completely specified. In Section 8.3 a more general hypothesis was considered for P partitioned as P = (PI' P2)' However, as shown in that section by the transformation (4), the hypothesis PI = I3'r can be r~duced to a hypothesis of the form (1) above. Let N
(3)
G=
L
(xa-Bza)(xa-Bza)'=Nin,
a=1
H= (B - P)A(B - P)'.
(4)
Lemma 8.11.1. Under the conditions of Theorem 8.11.1 the limiting distribution of H is weI, q ). Proof Write H as
(5) Then the lemma follows from Theorem 8.11.1 and (4) of Section 8.4.
•
We can express the likelihood ratio criterion in the form -2Iog'\ = -N logU=N loglI + G-IHI
(6)
=N IOgl1 + Theorem 8.11.2. hypothesis is tme,
-I
Under the conditions of Theorem 8.11.1, when the null
(7)
00,
~(~G) HI.
-2Iog'\
d ->
2
Xpq'
Proof We use the fact that N logll +N-ICI = tr C + OpCN- I ) when N-> since II +xCI = 1 +x tr C + O(x 2 ) (Theorem A.4.8).
We have
(8)
tr(~GrIH=N
t E
i,j=1
gii(big-f3;g)agh(bjh-f3jh)
g,h=1
= [vec(B' - P')]'( ~G-l ®A) vec(B' - P') because (1/ N)G !...I, (l/N)A !Nvec(B' - P') is N(.I ®AOI).
->
:!.. X;q
Ao, and the limiting distribution of •
372
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Theorem 8.11.2 agrees with the first term of the asymptotic expansion of -210g A given by Theorem 8.5.2 for sampling from a normal distribution. The test and confidence" procedures discussed in Sections 8.3 and 8.4 can be applied using this X 2-distribution. The criterion U = A2/N can be written as U = nf_1 Vj, where II; is defined in (8) of Section 8.4. The term II; has the form of U; that is, it is the ratio of the sum of squares of residuals of Xi" regressed on XI,,"'" Xi-I, ", Z" to the sum regressed on XI,,"'" Xi-I, ,,' It follows that under the null hypothesis VI"'" ~ are asymptotically independent and -N log II; ~ Thus - N log U = - Nr:.r= I log II; ~ X;q. This argument justifies the step-down procedure asymptotically. Section 8.6 gave several other criteria for the general linear hypothesis: the Lawley-Hotelling trace tr HG- 1 , the Bartiett-Nanda-Pillai trace tr H(G +H)-I, and the Roy maximum root of HG- 1 or H(G +H)-I. The limiting distributions of N tr HG- 1 and N tr H(G + H)-I are again X;q. The limiting distribution of the maximum characteristic root of NHG- I or NH(G + JI)-I is the distribution of the maximum characteristic root of H having the distributions W(I, q) (Lemma 8.11.1). Significance points for these test criteria are available in Appendix B.
xi.
8.11.2. Elliptically Contoured Matrix Distributions
In Section 8.3.2 the p X N matrix of observations on the dependent variable was defined as X = (XI"'" x N ), and the q X N matrix of observations on the independent variables as Z = (Zl"'" ZN); the two matrices are related by tff X = I3Z. Note that in this chapter the matrices of observations have N columns instead of N rows. Let E = (el> ... ,eN) be a p X N random matrix with density IAI- N / 2 g(P- I EE'(F')-I], where A=FF'. Define X by
(9)
X= I3Z +E.
In these terms the least squares estimator of
13 is
(10) where C =XZ' = I:~=IX"Z~ and A = ZZ' = I:~=IZaZ~. Note that the density of E is invariant with respect to multiplication on the right by N X N orthogonal matrices; that is, E' is left spherical. Then E' has the stochastic repres~ntation
(11)
E'! UTF',
8.11
373
ELLIPTICALLY CONTOURED DISTRIBUTIONS
where V has the uniform distribution on V'V = Ip, T is the lower triangular matrix with nonnegative diagonal elements satisfying EE' = TT', and F is a lower triangular matrix with nonnegative diagonal elements satisfying FF' =.1:. We can write
(12)
B-I3=EZ'A- 1 :!.FT'U'Z'A- 1,
:!. FT'V'(Z'A- 1Z)UTF'.
(13)
H = (B -13)A(B -13)' = EZ'A- 1ZE'
(14)
G = (X-I3Z)(X-I3Z)' -H=EE'-H =E(IN -Z'A- 1Z)E' =FT'U'(IN - Z'A- 1Z)VTF'.
It was shown in Section 8.6 that the likelihood ratio criterion for H: 13 = o. the Lawley-Hotelling trace criterion, the Bartlett-Nanda-Pillai trace criterion, and the Roy maximum root test are invariant with respect to linear transformations x --> Kx. Then Corollary 4.5.5 implies the following theorem. Theorem 8.11.3. Under the null hypothesis 13 = O. the distribution of each invariant criterion when the distribution of E' is left spherical is the same liS rh,. distribution under normality. Thus the tests and confidence regions described in Section 8.7 are valid for left-spherical distributions E'. The matrices Z'A- 1Z and I N -Z'A- 1Z are idempotent of ranks q and N - q. There is an orthogonal matrix ON such that
(15)
OZ'A-1Z0' =
.
[Iq 0] 0
0'
The transformation V= O'U is uniformly distributed on V'V= II" and
(16)
O]VK' o '
I
]VIC
o_
N
,
q
where K = FT'. The trace criterion tr HG -1, for example, is
(17) The distribution of any invariant criterion depends only on U (or V), not on T.
374
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
Since G + H = FT'TF', it is independent of U, A selection of a line:u transformation of X can be made on the basis of G + H, Let D be a p X r matrix of rank r that may depend on G + H, Define x: = D' x a ' Then "'/·x~ = CD'ph", and the hypothesis P = 0 implies D'P = O. Let X* = (xf, ... ,x~)=D'X, Po=D'P, E{J=D'E, Ho=D'HD, GD=D'GD. Then E;) = E' D!!" UTF'D'. The invariant test criteria for Po = 0 are those for 13 = 0 and have the same distributions under the null hypothesis as for the normal distribution with p replaced by r.
PROBLEMS 8.1. (Sec. 8.2.2)
Consider the following sample (for N = 8):
Weight of grain Weight of straw Amount of fertilizer
17 19 11
40 53 24
9 10 5
IS 29
6
12
5
9
13
19
30
12
7
27 14
11
18
Let Z20 = 1, and let Zlo be the amount of fertilizer on the ath plot. Estimate for this sample. Test the hypothesis ~I = 0 at ('Ie 0.01 significance level.
P
8.2. (Sec. 8.2) Show that Theorem 3.2.1 is a special case of Theorem 8.2.1. [Hint: Let q = 1, Zo = 1, ~ = JL.) 8.3. (Sec. 8.2)
Prove Theorem 8.2.3.
8.~. (Sec. 8.2)
Show that ~ minimizes the generalized variance
r. (X"'''~Z'')(X''-~Z,,)'I·
I (1'=1
8.5. (Sec. 8.3) In the following data [Woltz, Reid, and Colwell (1948), used by R. L. Anderson and Bancroft (1952») the variables are Xl' rate of cigarette bum; xc' the percentage of nicotine; ZI' the percentage of nitrogen; z2' of chlorine; Z), of potassium; Z4' of phosphorus; Z5, of calcium; and z6' of magnesium; and Z7 = 1; and N = 25: 53.92
N
a~l xa =
N
(42.20)
L
54.03 '
zo=
a=l
62.02 56.00 12.25 , 89.79 24.10 25
N
..
..,
LI (x" -x)(x" -x)
"
=
(0.6690 0.4527
0.4527 ) 6.5921 '
PROBLEMS
375
N
L
(za-z)(za-Z)'
a=1
,
1.8311 -0.3589 -0.0125 -0.0244 1.6379 0.5057 0
-0.3589 8.8102 -0.3469 0.0352 0.7920 0.2173 0
N
L
(za-z)(xa-i)'=
a=l
-0.0125 -0.3469 1.5818 -0.0415 -1.4278 -0.4753 0 0.2501 -1.5136 0.5007 -0.0421 -0.1914 -0.1586 0
-0.0244 0.0352 -0.0415 0.0258 0.0043 0.0154 0
1.6379 0.7920 -1.4278 0.0043 3.7248 0.9120 0
0.5057 0.2173 -0.4753 0.0154 0.9120 0.3828 0
0 0 0 0 0 0 0
2.6691 -2.0617 -0.9503 -0.0187 3.4020 1.1663 0
(a) Estimate the regression of XI and X z on ZI' zs, z6' and Z7' (b) Estimate the regression on all seven variables. (c) Test the hypothesis that tlJ.e regression on zz, Z3, and Z4 is O. 8.6. (Sec. 8.3) Let q = 2, ZI" = w" (scalar), zz" = 1. Show that the U-statistic for testing the hypothesis PI = 0 is a monotonic function of a T 2-statistic, and give the TZ-statistic in a simple form. (See Problem 5.1.) 8.7. (Sec. 8.3) Let
Zq" =
1, let qz = 1, and let
i,j=l,···,ql=q-1.
Prove that
8.8. (Sec. 8.3) Let ql
=
qz· How do you test the hypothesis PI
=
P2?
8.9. (Sec. 8.3) Prove
fa
t'IH
=~x A-IZ(Z»'[~(Z(l)-A A-lz(Z»(z(l)-A A-I.(2»'J-1 i.J a (z(I)-A a 12 22 i...J a 12 22 a a 12 22 "'0' 'l'
a
a
376
TESTING THE
GEN~RAL
LINEAR HYPOTHESIS; MANOVA
8.10. (Sec. 8.4)
By comparing Theorem 8.2.2 and Problem 8.9, prove Lemma 8.4.1.
8.11.
Prove Lemma 8.4.1 by showing that the density of
(Sec. 8.4)
KI exp[ - ~tr
PIll
and
P2w
is
I-I(PIll- P'f)A lI ,2 (PIll -If:)'j ·Kz exp[ - ~tr I -1(P2w - Pz)A Z2 (P2W - Pz)'].
Show that the cdf of U3• 3• n is
8.12. (Sec. 8.4)
3)
I
Iu ( '2 n - 1,'2 +
.{
2utn-l
f(n+2)r[~(n+l)] I r=
r(n -1)r('2n -1)v1T ~
n(n-l)
2u-l-n + -n- Iog
ut
+ --[arcsin(2u -I) -.!.1T] n-l 2
(1+!1=U) 2u-l-n-1(1-U)1} {U + 3(n + 1) .
[Hint: Use Theorem 8.4.4. The region {O:$; ZI :$; 1,0 :$; Z2 :$; 1, ZfZ2 :$; u} is the union of fO :$;ZI:$; 1,0 :$;Z2:$; u} and fO :$;Zl:$; u/z 2 , u :$;Z2:$; I}.] 8.13. (Sec. 8.4)
Find Pr{U4.3.n ~ u}.
8.14. (Sec. 8.4)
Find PrfU4 • 4 • n ~ u}.
8.15. (Sec. 8.4) For p :$; m find ooEU h from the density of G and H. [Hint: Use the fact that the density of K+L:~lViVi' is W(I,s+t} if the density of K is WeI,s) and V], ... ,V; are independently distributed as NCO, I).] 8.16. (Sec. 8.4) (a) Show that when p is even, the characteristic function of Y= log Up • m • n , say
8.17. (Sec. 8.5)
Usc the asymptotic expansion of the distribution to compute Prf-k
logU3•3• n :$;M*} for
(a) n = 8,
M* =
(b) n = 8, (c) n = 16, (d) n = 16,
M*
14.7,
= 21.7,
M* = M* =
14.7, 21.7.
(Either compute to the third decimal place or use the expansion to the k- 4 term.)
377
PROBLEMS
8.18. (Sec. 8.5) In casep = 3, ql = 4, and n = N - q = 20, find the 50'7<: significance point for k log U (a) using - 210g A as X 2 and (b) using - k log U as X 2. Using more terms of this expansion, evaluate the exact significance levels for your answers to (a) and (b).
8.19. (Sec. 8.6.5) Prove for Ii ~ 0, i = 1, ... , p, P
L
I·
1 ~ I :;:; log
i= 1
'
n (I + Ii):;:; L P
1=
P
I
Ii'
;= I
Comment: The inequalities imply an ordering of the value, of the Bartlett-Nanda-Pillai trace, the negative logarithm of the likelihood ratio criterion, and the Lawley-Hotelling trace. 8.20. (Sec. 8.6) The multivariate beta density. Let Hand G be independently distributed according to WO:, m) and WC~:, n), respectively. Let C be a matrL\ such that CC' = H + G, and let
Show that the density of L is
for Land [- L positive definite, and 0 otherwise. 8.21. (Sec. 8.9) Let Yij (a p-component vector) be distributed according to N(ll-i!, l:)' where ooEYij = Il-ij = Il- + Ai + Vj + "Yij' L:iAi = 0 = L:jV j = L:i"Yij = '"5.. j "Yij; the "Yij are the interactions. If m observations are made on each Yij (say Y,)I"'" Yi,m)' how do you test the hypothesis Ai = 0, i = 1, ... , r~ How do you test the hypothesis "Yij= 0, I = 1, ... ,r, j= 1, ... ,c? 8.22. (Sec. 8.9) The Latin square. Let Yij' i, j = 1, ... , r, be distributed according to N(ll-ij' I), where ooEY,'j = Il-ij ='1 + Ai + Vj + Il-k and k = j - i + 1 (mod r) with L:Aj = L:v j = L:ll-k = O.
(a) Give the univariate analysis of variance table for main effects and error (including sum~; of squares, numbers of degrees of freedom. and mean squares). (b) Give the table for the vector case. (e) Indicate in the vector case how to test the hypothesis Ai = 0, i = l. .... T. 8.23. (Sec. 8.9)
Let
XI
be the yield of a process and
.1'2
a quality measure. Let
± 10° (temperature relative to average) :, = ±O.75 lrelatiw measure of flow of one agent), and Z4 = ± 1.50 (relative measure of flow of another
ZI = 1, Zz =
agent). [See Anderson (1955a) for details.] Three ohservations wcre made on x:
378
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
and x 2 for each possible triplet of values of
~ = (58.529 98.675
Sl
= 3.090,
S2
-0.3829 0.1558
Z2'
z), and
-5.050 4.144
Z4'
The estimate of Pis
2.308) . -0.700 '
= 1.619, and r = -0.6632 can be used to compute S or
i.
(a) Formulate an analysis of variance model for this situation. (b) Find a confidence region for the effects of temperature (i.e., f3 12 , f32Z)' (el Test the hypothesis that the two agents have no effect on the yield and quantity. 8.24. (Sec. 8.6) Interpret the transformations referred to in Theorem 8.6.1 in the and z~l). original terms; that is, H: PI =
P'r
8.25. (Sec. 8.6) Find the cdf of tr HG roots given in Chapter 13.] 8.26. (Sec.
I
for p
=
2. [Hint: Use the distribution of the
8.10.1)
Bartlett-Nanda-Pillai V-test as a Bayes procedure. Let matrix no be defined by [f I' l:] = [0, (J + CC') _. I], where the p X m matrix C has a density proportional to II + CC'I - h n+ m), and fl = ('YI"'" 'Ym); let n l be defined by [fl.l:]=[(J+CC')-IC,(J+CC,)-I] where C has a density proportional to II + CC'I- hn+m)e!!rc'u+cCY'c.
"'1."'2"'" "'m+n be independently normally distributed with covariance ! and means xE",; = 'Y;, i = 1, ... , m, ooE",; = 0, i = m + 1, ... , m + n. Let
(a) Show that the measures are finite for n ~p by showing tr C'(J + CC,)-IC < m and verifying that the integral of II + CC'I- 4(n+m) is finite. [Hint: Let C = (Cl>"" cm), D j = I + r,i~ I ciC; = EjE;, Cj = E j _ 1 d j , j = 1, ... , m (Eo = J). Show IDjl = IDj_11 (1 + d;d j ) and hence IDml = [I7'_ I O + i;d). Then refeI to Problem 5.15.] (b) Show that the inequality (26) of Section 5.6 is equivalent to
Hence the Bartlett-Nanda-Pillai V-test is Bayes and thus admissible. 8.27. (Sec. 8.10.1) Likelihood ratio test as a Bayes procedure. Let "'I'"'' wm+n be independently normally distributed with covariance matrix I and means coEw; = 'Y;, i = 1, ... , m, ocE",; = 0, i = m + 1, ... , m + n, with n ~ m + p. Let [Io be defined by [fl,l:]=[O,(J+CC')-I], where the pXm matrix C has a denr.ity proportional to 11+ CC'I - y
379
PROBLEMS
where the m column~ of D are conditionally independently normally distributed with means 0 and covariance matrix [I - C'(J + CC')-IC]-I, and C has (marginal) density proportional to
(a) Sllow the measures are finite. [Hint: See Problem 8.26.] (b) Show that the inequality (26) of Section 5.6 is equivalent to
Hence the likelihood ratio test is Bayes and thus admissible. 8.28. (Sec. 8.10.1) Admissibility of the likelihood ratio test. Show that tile acceptance region IZZ' 1/ IZZ' + XX' I ~ c satisfies the conditions of Theorem 8.10.1. [Hint: The acceptance region can be written n~_lm; > c, where m; = 1- A;, i = 1, ... ,t.]
8.29. (Sec. 8.10.1) Admissibility of the Lawley-Hotelling test. Show that the accep· tance region tr XX'(ZZ,)-I 5; c satisfies the conditions of Theorem 8.10.1. 8.30. (Sec. 8.10.1) Admissibility of the Bartlett-Nanda-Pillai trace test. Show that the acceptance region tr X'(ZZ' +XX,)-IX 5; c satisfies the conditions of Theorem 8.10.1. 8.31. (Sec. 8.10.1) Show that if A and B are positive definite and A - B is positive semidefmite, then B- 1 - A -I is positive semidefinite. 8.32. (Sec. 8.10.1) Show that the boundary of A has m-measure O. [Hint: Show that (closure of A) CA U C, where C = {VI U - IT' is singular}.] 8.33. (Sec. 8.10.1) Show that if A cR~ is convex and monotone in majorization, then A* is convex. [Hint: Show
where
8.34. (Sec. 8.lD.1) Show that CO') is convex. [Hint: Follow the solution of Problem 8.33 to show (px + qy) -< w>' if x -< w>' and y -< w>..] 8.35. (Sec. 8.10.1) Show that if A is monotone, then A* is monotone. [Hint: Use the fact th at X(k] = .
max {min(x;" ... ,x;,)}.j
11.··· ,lk
380
TESTING THE GENERAL LINEAR HYPOTHESIS; MANOVA
8.36. (Sec. 8.10.2) Monotonicity of the power function of the Bartlett-Nanda-Pillai trace test. Show that tr'(uu' +B)(uu' +B+ W)-l 5,K
is convex in u for fixed positive semidefinite B and positive definite B 05, K 5, 1. [Hint: Verify
+ W if
(UU'+B+W)-I =(B+W)'I-
1 _I (B+W)-IUU'(B+W)-I. l+u'(B+W) u
The resulting quadratic form in u involves the matrix (tr A)l- A for A = (B + W)- tB(B + W)- 1; show that this matrix is positive semidefinite by diagonalizing A.] 8.37. (Sec. 8.8) Let x~,), a = 1, ... , N., be observations from N(IL(·), I), v = 1, ... , q. What criterion may be used to test the hypothesis that m
IL(·) =
L
'Y"Ch.
+ IL,
h~1
where Ch. arc given numbers and 'Y., IL arc unknown vectors? [Note: This hypothesis (that the means lie on an m-dimensional hyperplane with ratios of distances known) can be put in the form of the general linear hypothesis.] 8.38. (Sec. 8.2) Let x" be an observation from N(pz" , I), a = 1, ... , N. Suppose there is a known fIxed vector 'Y such that P'Y = O. How do you estimate P? 8.39. (Sec. 8.8) What is the largest group of transformations on y~), a = 1, ... , N;, i = 1, ... , q, that leaves (1) invariant? Prove the test (12) is invariant under this group.
CHAPTER 9
Testing Independence of Sets of Variates
9.1. INTRODUCTION In this section we divide a set of p variates with a joint normal distribution into q subsets and ask whether the q s.ubsets are mutually independent; this is equivalent to testing the hypothesis that each variable in one subset is uncorrelated with each variable in the others. We find the likelihood ratio criterion for this hypothesis, the moments of the criterion under the null hypothesis, some particular distributions, and an asymptotic expansion of the distribution. The likelihood ratio criterion is invariant under "linear transformations within sets; another such criterion is developed. Alternative test procedures are step-down procedures, which are not invariant, but are flexible. In the case of two sets, independence of the two sets is equivalent to the regression of one on the other being 0; the criteria for Chapter 8 are available. Some optimal properties of the likelihood ratio test are treated.
9.2. THE LIKELIHOOD RATIO CRITERION FOR TESTING INDEPENDENCE OF SETS OF VARIATES Let the p-component vector X be distributed according to N(J1, :l). We partition X into q subvectors with PI' P2' .. " Pq components, respectively:
An Introduction to Multivariate Statistical Analysis, Third Edition. By T. W. Andersoll ISBN 0-471-36091-0 Copyright © 2003 John Wiley & Sons, Inc.
381
382
TESTING INDEPENDENCE OF SETS OF VARIATES
that is, X(l)
X(2)
X=
(1)
X(q)
The vector of means J.l and the covariance matrix I are partitioned similarly,
J.l(I) J.l(2) (2)
J.l= J.l,q)
I=
(3)
III
112
Ilq
121
122
1 2q
Iql
Iq2
Iqq
The null hypothesis we wish to test is that the subvectors X(I), ..• , X(q) are mutually independently distriruted, that is, that the density of X factors into the densities of X(I), ••. , X(q). It is q
H:n(xlJ.l,I) = Iln(x(i)IJ.l(i), Iii).
( 4)
j~l
If
Xll), ... , Xlq)
are independent subvectors,
(5) (See Section 2.4.) Conversely, if (5) holds, th·~n (4) is true. Thus the null hypothesis is equivalently H: I ij = 0, i *- j. Thi~ can be stated alternatively as the hypothesis that I is of the form
(6)
III
0
0
0
122
0
0
0
Iqq
Io=
Given a sample xI' ... 'x", of N observations on X, the likelihood ratio
9.2
LIKELIHOOD RATIO CRITERION FOR INDEPENDENCE OF SETS
383
criterion is
,\ = max~,Io L("., .I o) max~,I L(".,.I) ,
(7) where
(8)
*
and L("., .Io) is L(".,.I) with .Iii =0, i j, and where the maximum is taken with respect to all vectors ". and positive definite .I and .Io (i.e., .Ii)' As derived in Section 5.2, Equation (6),
(9) where
(10)
i!l=~A=~
r.
(X,,-i)(x,,-i)'.
a~l
Under the null hypothesis, q
(11)
L("., .I o) = TILi(".(i), .I;;), ;=1
where
(12) Clearly q
(13)
m1l?'L(".,.I o) = TI max L;(".(i),.I ii )
p.,:I o
;=1
p.(l),:I
jj
e-~PN, JN (21T) 21 nf~ III;;) iN I
where
(14)
'"
384
TESTING INDEPENDENCE OF SETS OF VARIATES
If we partition A and 'In as we have l:,
(15)
A=
All
A\2
A lq
AZI
A zz
A Zq
Aql
Aq2
Aqq
'In =
'Ill
'I \2
'I lq
'I ZI
'Izz
'IZq
'Iql
'I qZ
'Iqq
we see that 'I iiw = 'I;; = (1/N)A;;. The likelihood ratio criterion is
(16)
A=max .... l:oL( .... ,l:o)= l'Inl~N max .... l:L( .... ,l:) [1{=II'Iiil~N
IAI~N [1{=IIA ii l!N'
The critical region of the likelihood ratio test is
(17)
A 5 A( e),
where ;l(e) is a number such that the probahility of (17) is e with l: = l:o. (It remains to show that such a number c~n be found.) Let (18)
V=
IAI [1{=IIA;;1 .
Then A = V~N is a monotonic increasing function of V. The critical region (17) can be equivalently written as (19)
V5 V(e).
Theorem 9.2.1. Let Xl"'" x N be a sample of N observations drawn from N( .... , l:), where X a , .... , and l: are partitioned into PI"'" Pq rows (and columns in the case of l:) as indicated in (1), (2), and (3). The likelihood ratio criterion that the q sets of components are mutually independent is given by (16), where A is defined by (10) and partitioned according to (15). The likelihood ratio test is given by (17) and equivalently by (19), where V is defined by (18) and A( e) or V( e) is chosen to obtain the significance level e. Since r;j = a;j/ ..ja;;a jj , we have p
(20)
IAI = IRI na;j> ;=1
9.2
LIKELIHOOD RATIO CRITERION FOR INDEPENDENCE OF SETS
385
where
(21)
R
= (r;j) =
RIl
R12
R lq
R21
R22
R 2q
Rql
Rq2
Rqq
and PI + ... +Pi
(22)
IA"I = IR;;I j=Pl+"
n
+Pi-I +1
aff ·
Thus
(23) That is, V can be expressed entirely in terms of sample correlation coefficients. We can interpret the criterion V in terms of generalized variance. Each set (X;I,,,,,X;N) can be considered as a vector in N-space; the let (X il x; .... , X;N - x) = Z;. say, is the projection on the plane orthogonal to the equiangular line. The determinant IA I is the p-dimensional volume squared of the parallelotope with Zl"'" zp as principal edges. The determinant IA;;I is the pi-dimensional volume squared of the parallelotope having as principal edges the ith set of vectors. If each set of vectors is orthogonal to each other set (i.e., R;f = 0, i "" j), then the volume squared IAI is the product of the: volumes squared IA;;I. For example, if p = 2, PI = P2 = 1, this statement is that the area of a parallelogram is the product of the lengths of the sides if the sides are at right angles. If the sets are almost orthogonal; then IA I is almost DIA;;I, and V is almost 1. The criterion has an invariance property. Let C; be an arbitrary nonsingular matrix of order p; and let
(24)
CI
0
0
0
C2
0
0
0
Cq
C=
Let Cx a + d = x:. Then the criterion for independence in terms of x~ is identical to the criterion in terms of Xa' Let A* = I:a(x~ - x* )(x: - x*)' be
386
TESTING INDEPENDENCE OF SETS OF VARIATES
partitionl!d into submatrices A* I}
(25)
=
Ai> Then
"(x*(j) -- i*Ii))(x*(j) -- i*{j»)' i....J u u
= C" (xli) - r(i»)(x(j) ~, l..J n· n
- i(j»)' e
}
and A'" = CAe'. Thus
V* =
(26)
Ji.:.L = nlA;~1
ICAC'I nlCjAjjC;1
ICI·IAI·IC'I = _I_A1_ = V OICjl·IAj;I·IC;1 nlA;;I.
for 1CI = 01 C;I. Thus the test is invariant with respect to linear transformations within each set. Narain (1950) showed that the test based on V is strictly unbiased; that is, the probability of rejecting the null hypothesis is greater than the significance level if the hypothesis is not true. [See also Daly (1940).]
9.3. THE DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION WHEN THE NULL HYPOTHESIS IS TRUE 9.3.1. Characterization of the Distribution We shall show that under the null hypothesis the distribution of the criterion V is the disuibution of a product of independent variables, each of which has the distribution of a criterion U for the linear hypothesis (Section 804). Let
(1)
V;=
All
A \.i. I
A;_I.I
A;_I.;_I
A;l
Ai.i-l
All
Al.i-l
A;-l.l
A;-l.i-l
Ali
I A;-l.i! Ajj
·IA;;I
i = 2, ... ,q.
9.3
1
DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION
387
Then V = V2V3 '" Vq • Note that V; is the N /2th root of the likelihood ratio criterion for testing the null hypothesis
(2)
H;:lil =O, ... ,li.i-l =0,
that is, that X(i) is independent of (X(I)" H is the intersection of these hypotheses.
... , XU-I) ')'.
The null hypothesis
Theorem 9.3.1. When Hi is truL, V; has the distribution of Up;.p;.n_p;, where n=N-1 andp;=Pl + ... +P;_I, i=2, ... ,q. Proof The matrix A has the distribution of E:~IZ"Z~, where ZI,,,,,Zn are independently distributed according to N(O, I) and Z" is partitioned as (Z~l)', ... , z~q) ')'. Then c()nditional on Z~I) = z~l, ... , Z~ -I) = z~ -I), a = 1, ... , n, the subvectors Z\i), ... , Z~i) are independently distributed, Z~) having a normal distribution with mean
(3)
and covariance matrix
(4)
where
(5)
When the null hypotheris is not assumed, the estimator of !=I; is (5) with ljk replaced by A jk , and the estimator of (4) is (4) with ljk replaced by (l/n)A jk and !=I; replaced by its estimator. Under Hi: !=Ii = and the covariance matrix (4) is liP which is estimated by (l/n)A jj • The N /2th root of the likelihood
°
388
TESTING INDEPENDENCE OF SETS OF VARIATES
ratio criterion for Hi is
Ai-I,i-I
(6)
All
A \,i-I
Ali
Ai_I,1
Ai-I,i-I
Ail
Ai,i-I
Ai-I,i Au
All
A1,i-1
Ai_I,1
Ai-I,i-l
'!Aii!
which is V;. This is the U-statistic for Pi dimensions, Pi components of the conditioning vector, and n - Pi degrees of freedom in the estimator of the • covariance matrix. Theorem 9.3.2. The distribution of V under the null hypothesis is the distribution of V2 V3 .,. ~, where V2 , • •• ,Vq are independently distributed with V; having the distribution of Upi,Pi,n-Pi' where Pi = PI + ... +Pi-I'
Proof From the proof of Theorem 9.3.1, we see that the distribution of V; is that of Upi,Pi,n-Pi not depending on the conditioning Z~k), k = 1, ... , i - 1, a = 1, ... , n. Hence the distribution of V; does not depend on V 2 ,···, V;-I'
•
n'=2 nr'!' I
Theorem 9.3.3. Under the null hypothesis Vis distributed as Xii' where the Xi/s are independent and Xii has the density f3[x! !(n - Pi + 1 -
j),wJ Proof This theorem follows from Theorems 9.3.2 and 8.4.1.
•
9.3.2. Moments Theorem 9.3.4. criterion is
h
(7) Iff V
When the null hypothesis is true, the hth moment of the
=.a {Pi}] r[r [ q
(n-Pi+1-j)+h]r[!Cn+1-j)]} (n - Pi + 1 - j)] r [!C n + 1 - j) + h] .
9.3
DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION
389
Proof Because V2 , ••• , Vq are independent, (8) Theorem 9.3.2 implies J:'v/ = rfU):,P""_ii," Then the theorem follows by substituting from Theorem 8.4.3. • If the Pi are even, say Pi
2ri, i> 1. then by using the duplication formula WI.! can rcauce the hth moment of V to =
Ha + ~)I'(a + 1) = fiirc2a + ])2- 2 ,. for the gamma function
(9)
(%v"=n{n ;=2
=
k=l
f(n+l- p i -2k+2h)r(1l+1-2k)} r(n+l-Pi-2k)r(n+I-2k+2h)
aD q
{
ri
S--I(1l
+ I-Pi - 2k,pi)
'J~IX"+I-P,-2k+2"-I(I_x)P,-1 d1:}. Thus V is distributed as n{_2{nk'~ 1 Y;l}, where the Y;k are independent. and Y;k has density (3Cy; 11 + 1 - Pi - 2k, .0,). In general, the duplication formula for the gamma function can be used to reduce the moments as indicated in Section 8.4. 9.3.3. Some Special Distributions If q = 2, then V is distributed as Up"PI.n-PI' Special cases have been treated in Section 8.4, and references to the literature given. The distribution for PI = P2 = P3 = 1 is given in Problem 9.2, and for PI = P2 = PJ = 2 in Prohlem 9.3. Wilks (1935) gave the distributions for PI = P2 = I, for P3 = P - 2,t for Pl == 1, P2 = P3 = 2, for PI = 1, P2 = 2, P3 = 3, for PI = 1, P2 = 2, P_, = 4, and f'Jr PI = P2 = 2, P3 = 3. Consul (1967a) treated the case PI = 2, P2 = 3, P_' even. Wald and Brookner (1941) gave a method for deriving the distribution if not more than one Pi is odd. It can be seen that the same result can be obtained by integration of products of beta functions after using the duplication formula to reduce the moments. Mathai and Saxena (1973) gave the exact distribution for the general case. Mathai and Katiyar (1979) gave exact significancl.! points for p = 3( I) 10 and 11 = 3(1)20 for significance levels of 5% and 1% Cof - k log V of Section 9.4).
tIn Wilks's form lla
nt(N- 2 -i)) should
he n~(11
-
2-
ill.
390
TESTING INDEPENDENCE OF SETS OF VARIATES
9.4. A.l\I ASYMPTOTIC EXPANSION OF THE DISTRIBUTION OF THE LIKELIHOOD RATIO CRITERION The hth moment of A = V~N is
/AII=K
(1)
nf_lr{HN(l+h)-i]}
n7_1{r1j::'lr{HN(1 +h) -j]}}'
where K is chosen so that rffAD with
= 1. This is of the form of (1) of Section 8.5
b =p,
a =p, (2)
j=Pl+"'+Pi-l+1,,,,,Pl+"'+Pi'
i=l, ... ,q.
Then f= ~[p(p + 1) - Ep/Pi + 1)] = ~(p2 - Epn, 13k = 8 j = ~(1- p)N. In order to make the second term in the expansion vanish we take p a5 p
(3)
+ 9( pZ - Epn 6N(pZ - Ep;)
= 1 _ 2( p3 - Epf)
Let 3 k = pN = N - -2 -
(4)
Then
("2
p.l -
Ep?
( Z "2)' 3 p - '-Pi
= 1z/k2, where [as shown by Box (1949)] (p3 _ "Lp;)2 72(pZ - "Lpn'
(5)
We obtain from Section 8.5 the following expansion: (6)
Pr{ -k log V::; v}
= Pr{ X/
::; v}
+ ;~ [Pr{X/+4 ::; v}
-- Pr{x/::; v}] + O(k- 3 ).
391
9.5 OTHER CRITERIA
Table 9.1 p
f
v
'Y2
N
4 5 6
6
12.592 18.307 24.996
11
10
24 15 '8
15 15 15 16
15
235
4B
k
'Y2/ k2
Second Teon
71
0.0033 0.0142 00393 0.0331
-0.0007 -v.0021 -0.0043 -0.0036
'6 69
'6 67
'6 73
'6
If q = 2, we obtain further terms in the expansion by using the results of Section 8.5. If Pi = 1, we have f=' !p(p - 1),
k=N- 2p+ll 6
(7) 'Y2 =
'Y3
'
P(i8~ 1) (2p2 -
2p -13),
= p(:;O 1) (p - 2)(2p -1)(p + 1);
other terms are given by Box (1949). If Pi = 2 (p = 2q) f= 2q(q -1),
(8)
k=N _ 4q ~ 13, 'Y2 =
q(q7~ 1) (8q2 -
8q -7).
Table 9.1 gives an indication of the order of approximation of (6) for Pi = 1. In each case v is chosen so that the first term is 0.95. If q = 2, the approximate distributions given in Sections 8.5.3 and 8.5.4 are available. [See also Nagao (1973c).j
9.5. OTHER CRITERIA In case q = 2, the criteria considered in Section 8.6 can be used with G + H replaced by All and H replaced by A12A2ilA~l' or G +H replaced by A22 and H replaced by A 21 A l/A 12'
392
TESTING INDEPENDENCE OF SETS OF VARIATES
The null hypothesis of independence is that l: - l:o = 0, where l:o is defined in (6) of Section 9.2.. An appropriate test procedure will reject the null hypothesis if the elements of A - Ao are large compared to the elements of the diagonal blocks of Au (where An is composed of diagonal blocks Ajj and off-diagonal blocks of 0). Let the nonsingular matrix B jj be such that BjjAjjB;j = I, that is, A ~ 1 = B;jBjj' and let Bo be the matrLx with B;; as the ith diagonal block and O's as off-diagonal blocks. Then BoAoB~ = 1 and ()
(1)
Bo(A -Ao)Bo =
III I
A 1211~2
1111 A l'IIl;1'I
B 22 A 2I B'11
0
B22A2qB~q
BqqAqlB'l1
BqqAq2B~2
0
This matrix is invariant with respect to transformations (24) of Section 9.2 operating on A. A different choice of B jj amounts to mUltiplying (1) on the left by Qo and on the right by QQ' where Qo is a matrix with orthogonal diagonal blocks and off-diagonal blocks of O's. A test procedure should reject the null hypothesis if some measure of the numerical values of the elements of (1) is too large. The likelihood ratio criterion is the N /2 power of IBo(A -Ao)B~ +/1 = IBoABol. / Another measure, suggested by Nagao (1973a), is
q
=~
L
tr AijAjjIAjjA~I.
;,}~I
;*}
For q = 2 this measure is the average of the Bartlett-Nanda-Pillai trace criterion with G + H replaced by A JJ and H replaced by A 12 A 221A 21 and the same criterion with G + H replaced by A 22 and H replaced by A 21 A 11 lA 12 • This criterion multiplied by n or N has a limiting x2-distribution with number of degrees of freedom f = ~(p2 - r,'l~ I pl), which is the same number as for -N log V. Nagao obtained an asymptotic expansion of the distribution:
(3)
PrUntr(AAol-/)2~x} = Pr{
x/ ~x}
393
9.6 STEP-DOWN PROCEDURES
9.6. STEp·DOWN PROCEDURES 9.6.1. Step·down by Blocks It was shown in Section 9.3 that the N 12th root of the likelihood ratio criterion, namely V, is the product of q - 1 of these criteria, that is, V2 , ••• , Vq • Th~ ith subcriterion V; provides a likelihood ratio test of the hypothesis Hi [(2) of Section 9.3] that the ith subvector is independent of the preceding i - 1 subvectors. Under the null hypothesis H [= '!~ 2 Hi]' these q - 1 criteria are independent (Theorem 9.3.2). A step-down testing procedure is to accept the null hypothesis if
n
(1)
i= 2 .... ,q,
and reject the null hypothesis if V; < vi(.s) for any i. Here viCe) is the number such that the probability of (1) when Hi is true is 1 - e,. The significance level of the procedure is e satisfying q
(2)
1- e =
n (1 -
ei)'
i~2
The sub tests can be done sequentially, say, in the order 2, ... , q. As soon as a subtest calls for rejection, the procedure is terminated; if no subtest leads to rejection, H is accepted. The ordering of the subvectors is at the discretion of the investigator as well as the ordering of the tests. Suppose, for example, that measurements on an individual are grouped into physiological measurements, measurements of intelligence, and measurements of emotional characteristics. One could test that intelligence is independent of physiology and then that emotions are independent of physiology and intelligence, or the order of these could be reversed. Alternatively, one could test that intelligence is independent of emotions and then that physiology is independent of these two aspects, or the order reversed. There is a third pair of procedures.
394
TESTING INDEPENDENCE OF SETS OF VARIATES
Other' criteria for the linear hypothesis discussed in Section 8.6 can be used to test the component hypotheses H2 , ••• , Hq in a similar fashion. When H, is true, the criterion is distributed independently of X~l), .. . , X~i -1) , ex = 1. ... , N, and hence independently of the criteria for H 2 ,.··, Hi-I' 9.6.2. Step-down by Components
Iri Section 8.4.5 we discussed a componentwise step-down procedure for testing that a submatrix of regression coefficients was a specified matrix. We adapt this procedure to test the null hypothe~is Hi cast in the form -I
(3)
Hi: (l:il
l:i2
...
l:11
l:12
l:1,i-1
l:21
l:22
l:2,i-1
l:i-l.l
l:i-I,1
l:i-I,i-I
l:i,i-I)
=0,
where 0 is of order Pi X Pi' The matrix in (3) consists of the coefficients of the regression of X(i) on (X(I)', ... ,X(i-I),)'. For i = 2, we test in sequence whether the r,~gression of Xp , + I on X(l) = (XI' ... , Xp,)' is 0, whether the regression of X p , +2 on X(I) is 0 in the regression of X p , +2 on X(I) and X p , + I' ••• , a 1d whether the regression of Xp,+p, on X(l) is 0 in the regression of X p'+P2 on X(l), X p ,+1> ... ,Xp ,+P2- 1' These hypotheses are equivalently that the first, second, ... , and P2th rows of the matrix in (3) for i = 2 are O-vectors. Let A;}) be the k X k matrix in the upper left-hand corner of A ii , let AW consist of the upper k rows of A ij , and let A~}) consist of the first k columns of Aji' k = 1, ... , Pi' Then the criterion for testing that the first row of (3) is 0 is (4)
All
AI,i_1
Ai-I,I
Ai-I,i-I
A(l) il
A(I)
i,;-1
All
A1,i-1
A i _ I .1
Ai-l. i - I
A(I) I,
A(l)
I-l,i
A(I)
" AI)
9.6
395
STEP-DOWN PROCEDURES
For k> 1, the criterion for testing that the kth row of the matrix in (3) is 0 is [see (8) in Section 8.4]
(5)
Ai_I,1
Ai-I,i-I
A(k) i-I,i
A (k) il
A(k) i,i-l
A(k)
AI,i_1
1.(k-l)
II
.. Ii
Ai_I,1
Ai-I,i-I
A(k-I) i-Itl
A (k-I) il
A(k-I) i,i-l
A(k-I)
IA\~-I)I
. IA\7)1
,
II
k=2"",Pi'
i=2, ... ,q.
Under the null hypothesis the criterion has the beta density {3[x;i(n - Pi + 1 - j),w;l. For given i, the criteria Xii"'" X iPi are independent (Theorem 8.4.1). The sets for different i are independent by the argument in Section 9.6.1.
A step-down procedure consists of a sequence of tests based on X 2i " ' " X 2P2 ' X 31 , ••• , X qp ,,' A particular component test leads to rejection if
(6) The significance level is
8,
where q
(7)
1-
8
=
Pi
n n (1 -
i~2 j~l
8i
J·
396
TESTING INDEPENDENCE OF SETS OF VARIATES
The s'~quence of subvectors and the sequence of components within each subvector is at the discretion of the investigator. The criterion Vi for testing Hi is Vi = Of!.! Xik , and criterion for the null hypothesis H is q
q
V= nVi=
(8)
i~2
Pi
n nX
i~2 k~l
ik •
These are the random variables described in Theorem 9.3.3.
9.7. AN EXAMPLE We take the following example from an industrial time study [Abruzzi (1950)]. The purpose of the study was to investigate the length of time taken by various operators in a garment factory to do several elements of a pressing operation. The entire pressing operation was divided into the following six elements: 1. Pick up and position garment.
2. 3. 4. 5. 6.
Press and repress short dart. Reposition garment on ironing board. Press three-q~rters of length of long dart. Press balance of long dart. Hang garment on rack.
In this case xa is the vector of measurements on individual a. The component x ia is the time taken to do the ith element of the operation. N is·76. The data (in seconds) are summarized in the sample mean vector and covariance matrix:
( 1)
(2)
9.47 25.56 13.25 i= 31.44 27.29 8.80
s=
2.57 0.85 1.56 1.79 1.33 0.42
0.85 37.00 3.34 13.47 7.59 0.52
1.56 3.34 8.44 5.77 2.00 0.50
1.79 13.47 5.77 34.01 10.50 1.77
1.33 7.59 2.00 10.50 23.01 3.43
0.42 0.52 0.50 1.77 3.43 4.59
397
9.8 THE CASE OF TWO SETS OF VARIATES
The sample standard deviations are (1.604,6.041,2.903,5.832,4.798,2.141). The sample correlation matrix is
(3)
R=
1.000 0.088 0.334 0.191 0.173 0.123
0.334 0.186 1.000 0.343 0.144 0.080
0.088 1.000 0.186 0.384 0.262 0.040
0.191 0.384 0.343 1.000 0.375 0.142
0.173 0.262 0.144 0.375 1.000 0.334
0.123 0.040 0.080 0.142 0.334 1.000
The investigators are interested in testing the hypothesis that the six variates are mutually independent. It often happens in time studies that a new operation is proposed in which the elements are combined in a different way; the new operation may use some of the elements several times and some elements may be omitted. If the times for the different elements in the operation for which data are available are independent, it may reasonably be assumed that they wilJ be independent in a new operation. Then the distribution of time for the new operation can be estimated by using the means 2nd variances of the individual items. In this problem the cr.terion V is V= IRI = (jA72. Since the sample size is large we can use asymptotic theory: k = 4~3, f = 15, and - k log V = 54.1. Since the significance point for the X 2-distribution with 15 degrees of freedom is 30.6 at the 0.01 significance level, we find the result significant. We reject the hypothesis of independence; we cannot consider the times of the elements independent. 9.S. THE CASE OF TWO SETS OF VARIATES In the case of two sets of variates (q = 2), the random vector X, the observation vector X a , the mean vector ~, and the covariance matrix I are partitioned as follows:
X=
(1) ~=
(X(I) )
(X(I) )
X(2)
,
(~(l)
)
~(2)
,
x: 1= (III
Xu
=
2)
12\
,
112 ). 122
The nulJ hypothesis of independence specifies that of the form
(2)
112
= 0, that is. that I is
398
TESTING INDEPENDENCE OF SETS OF VARIATES
The test criterion is (3) It was shown in Section 9.3 that when the null hypothesis is true, this criterion is distributed as Up, . P ,. N-I-p,' the criterion for testing a hypothesis about regression coefficients (Chapter 8). We now wish to study further the relationship between testing the hypothesis of independence of two sets and testing the hypothesis that regression of one set on the other is zero. The conditional distribution of X~I) given X~2) = X~2) is N[~(I) + P(X~2) ~(2», 1 11 .2] = N[P(x~) - i(2» + v, 1 11 . 2], where P = 1 12 1 221, 1 11 '2 = 111 - 112 1221 1 21 , and v = ~(1) + p(i(2) - ~(2». Let X: =X~I), = [(X~2) i (2 1], p* = (P v), and 1* = 1 11 . 2 , Then the conditional distribution of X: is N(P* z!, 1*). This is exactly the distribution studied in Chapter 8. The null hypothesis that 112 = 0 is equivalent to the null hypothesis p = O. Considering X~2) fixed, we know from Chapter 8 that the criterion (based on the likelihood ratio criterion) for testing this hypothesis is
z:'
»'
U=
(4)
IE( x: - P~IZ!)( x: - P~IZ:)' I
IL( x! - ~wZ:(2»)( x: - ~wz:(2))' I'
where
(5)
~w=v=i*=i(1),
Pr1 =
(~!l
~n)
= (A 12 A -I 22
X-(I»).
The matrix in the denominator of U is N
(6)
"' - ' (x{l) - i(11)(i(ll - i(I»), = A 1\' u: u a=\
9.8
399
THE CASE OF TWO SETS OF VARIATES
The matrix in the numerator is N
(7)
L
[x~) -x(1) -AI2A2i(x~) -x(2»)] [x~) -x(1) -A12A221(X~) -X(2»)],
a=l
Therefore,
(8)
IAI
which is exactly V. Now let us see why it is tha: when the null hypothesis ls true the distribution of U = V does not depend on whether the X~2) are held flX,;\d. It was shown in Chapter 8 that when the null hypothesis is true the distribudon of U depends only on p, ql' and N - q2, not on za' Thus the conditional distribution of V given X~2) = X~1) does not depend on X~2); the joint dIstribution of V and X~2) is the product of the distribution of V and the distribution of X~2), and the marginal distribution of V is this conditional distribution. This shows that the distribution of V (under the null hypothesis) does not depend on whether the X~2) are fIXed or have any distribution (normal or not). We can extend this result to show that if q > 2, the distribution of V under the null hypothesis of independence does not depend on the distribution of one set of variates, say X~l). We have V = V2 ... ~, where Vi is defined in (1) of Section 9.3. When the null hypothesis is true, ~ is distributed independently of X~l), ... , x~q -I) by the previous result. In turn we argue that ~ is distributed independently of X~I), ... , x~j -I). Thus V2 ••• Vq is distributed independently of X~I). Theorem 9.8.1. Under the null hypothesis of independence, the distribution of V is that given earlier in this chapter if q - 1 sets are jointly normally distributed, even though one set is not normally distributed. In the case of two sets of variates, we may be interested in a measure of association between the two sets which is a generalization of the correlation coefficient. The square of the correlation between two scalars XI and X 2 l:an be considered as the ratio of the variance of the regression of XI on X 2 to the variance of Xl; this is Y( (3X2)/Y(XI ) = {32a22/al1 = (al~/a22)/al1 = pf2' A corresponding measure for vectors X(I) and X(2) is the ratio of the generalized variance of the regression of X(lJ on X(2J to the generalized
400
TESTING INDEPENDENCE OF SETS OF VARIATES
variance of X(1), namely,
(9)
I $PX(2)(PX(2»), I 11111
If PI = P2' the measure is
(10) In a sense this measure shows how well X(I) can be predicted from X(2). In the case of two scalar variables XI and X 2 the coefficient of alienation is al2/ a 12, where al~2 = ~(XI - {3X2)2 is the variance of XI about it~ regression on X 2 when $ XI = $ X 2 = 0 and (~'(XIIX2) = (3X 2 • In the case of two vectors X(I) and X(2), the regression matrix is P = 1121221, and the generali;:ed variance of X(I) about its regression on X(2) is
(11)
Since the generalized variance of X(I) is coefficient of alienation is
(12)
1111 - I12Izl12J1 11111
I ~ x (1)X(I), I = I1 11 1, the
vector
III
The sample equivalent of (12) is simply V. A measure of association is 1 minus the coefficient of alienation. Either of these two measures of association can be modified to take account of the number of components. In the first case, one can take the Pith root of (9); in the second case, one can subtract the Pith root of the coefficient of alienation from 1. Another measure of association is
(13)
n
tr ~[px(2)(px(2»)'1( ~X(1)X(\)yl
tr I12 I II 21Iil
P
P
Thi.~ mea&ure of association ranges between 0 and 1.
If X(1) can be predicted exactly from X(2) for PI :5,P2 (i.e., 1\1.2 = 0), then this measure is 1. If no linear comtination of X(I) can be predicted exactly, this measure is O.
9.9
ADMISSIBILITY OF THE LIKELIHOOD RATIO TEST
401
'9.9. ADMISSIBILITY OF THE LIKELIHOOD RATIO TEST The admissibility of the likelihood ratio test in the case of the 0-1 loss function can be proved by showing that it is the Bayes procedure with respect to an appropriate a priori distribution of the parameters. (See Section 5.6.) Theorem 9.9.1. The likelihood ratio test of the hypothesis that I is of the form (6) of Section 9.2 i.l· Hayr.l· alld admissihle if N > P + 1.
Proof We shall show that the likelihood ratio test is equivalent to rejection of the hypothesis when
(1 )
~c,
jf(xIO)IlIl(dO)
where x represents thc sample, 0 represents the parameters (~ and I), f(xIO) is the density, and IT I and no are proportional to probability measures of 0 under the alternative and null hypotheses, respectively. Specifically, the left -hand side is to be proportional to the square root of n (," I IA;;I /
IAI. To define IT!> let
(2)
~= (1+ W,)-lW.
where the p-component random vector V has the density proportional to (1 v' v)- tn, n = N - 1, and the conditional distribution of Y given V = v is N[O,(1+v'v)/Nj. Note that the integral of (1+v'v)-t ll is finite if n>p (Problem 5.15). The numerator of (1) is then
+
402
TESTING INDEPENDENCE OF SETS OF VARIATES
The exponent in the integrand of (3) is - 2 times V
Ct~
1
N
N
Lx~(I+~'v')xa-2yv'
(4)
2
LXa+Nlv'(I+vV,)-lV+ 1/V'V Ct~
1 N
L
=
a~l
= tr A
N
X~Xa
+ v'
L
XaX~V - 2yv'Hi + Ny2
a~l
+ v'Av + Hi'x + N(y -
X'V)2,
where A = L~~ lXaX~ - trxi'. We have used v'(J + VV,)-l v + (1 + V'V)-l = 1. [from U+W'y-l =1-(1+v'v)-t vv']. Using \/+vv'\ =l+v'v (Corollary A.3.1), we write (3) as
f ... f
const I' - ~tr A - ~Ni'i
(5)
x
-x
To define
no
x
e - -l"'A,' dv = const\A \ - t e- -ltr A- rNi •i
•
-:::xl
let l: have the form of (6) of Section 9.2. Let
i = 1, ... ,q, where the p(component random vector V(i) has density proportional to 11 + tl(i)' v ti » - ~n, and the conditional distribution of l'j given V(i) = v(i) is N[O,(1 +vU)'vU»/N], and let (Vt,Yt), ... ,(Vq,Yq) be mutually independent. Then the denominator of (1) is q
(7)
n const\A
;=1
ii \-
= const(
t exp[ -
D
\A ii \ -
Htr Aii + Hi(i)'X(i»)] ~) exp[ - ~(tr A + Hi'x)].
The left-hand side of (1) is then proportional to the square root of n{~t\Ail\/\A\.
•
This proof has been adapted from that of Klefer and Schwartz (1965).
9.10. MONOTONICITY OF POWER FUNCTIONS OF TESTS OF INDEPENDENCE OF SETS Let Z" = [Z~\)', Z~2)']', a = 1, ... , n, be distributed according to (1)
9.10
MONOTICITY OF POWER FUNCTIONS
403
We want to test H: l:12 = O. We suppose PI 5,P2 without loss of generality. Let PI' ... ' PPI (PI? ... ~ pp) be the (population) canonical correlation coefficients. (The pl's are the characteristic roots of l:1/ l: 12 l:zll:21' Chapter 12.) Let R = diag( PI' ... , pp) and a = [R, 0] (PI x P2). Lemma 9.10.1.
There exist matrices BI (PI XPI)' B2 (P2 XP2) such that
(2) Proof Let m = P2' B = B I, F' = l:t2 B;, S = l:12l:Z2 t in Lemma 8.10.13. Then F' F = B 2l: 22 B'z = [P2' B Il: 12 B z = BISF = a. •
(This lemma is also contained in Section 12.2.) Let x,,=BIZ~I), y,,=B2Z~2), a=1, ... ,n, and X=(xl, ... ,x.), Y= (YI' ... 'Y.). Then (x~,Y;)', a= 1, ... ,n, are independently distributed according to
(3) The hypothesis H: l: 12 = 0 is equivalent to H: a = 0 (i.e., all the canonical correlation coefficients PI' ... ' PPI are zero). Now given Y, the vectors x"' a = 1, ... , n, are conditionally independently distributed according to N(ay", [ - aa') = N(ay", [ - R2). Then x! = UPI - R2)- tx" is distributed according to N(My", [p) where
M= (D,O), (4)
D
= diag( 81 , •.• , 8p, )'
8j = pj(1- pnt,
i = 1, ... ,PI.
Note that 8/ is a characteristic root of l: 12 l:zll: 21 l:1112' where l:n.2 = l:n -l: 12 l: Z21l:21· Invariant tests depend only on the (sample) canonical correlation coeffiwhere cients rj =
..;c;,
(5) Let Sh =X*y,(yy,)-l lX *"
(6) Se =X* X*' - Sh =X* [[- Y'(YY') -I Y]X*'.
404
TESTING INDEPENDENCE OF SETS OF VARIATES
Then
(7) Now given Y, the problem reduces to the MANOYA problem and we can apply Theorem 8.10.6 as follows. There is an orthogonal transformation (Section 8.3.3) that carries X* to (U,v) such that Sh = UU', Se = W', U=(ul, ... ,u",). V is PIX(n-P2)' u j has the distrihution N(ojEj,T), i = 1, ... , PI (E I being the ith column of I), and N(O, I), i = PI + 1, ... , P2, and the columns of V are independently clistributed according to N(O, Then cl, ... ,cp , are the characteristic roots of UU'(W,)-I, and their distribution depends on the characteristic roots of MYY'M', say, Tt, ... ,Tp~. Now from Theorem 8.10.6, we obtain the following lemma.
n.
Lemma 9.10.2. If the acceptance region of an invariant test is convex in each column of U, given V and the other columns of U, then the conditional po~)er given Y increases in each characteristic root T/ of MYY' M'. Lemma 9.10.3.
If A ~ B, then A;(A) ~ ),-;CB).
Proof By the minimax property of the characteristic roots [see, e.g., Courant and Hilbert (1953)],
(8)
x'Ax
x'Bx
Aj(A) = max min - , - ~ max min - , - = A;(B), Sj
XES;
X X
Sj
xr:S j
where Sj ranges over i-dimensional subspaces.
X X
•
T/
Now Lemma 9.10.3 applied to MYY'M' shows that for every j, is an increasing function of OJ = pj(l- pn t and hence of Pj' Since the marginal distribution of Y does not depend on the p;'s, by taking the unconditional power we obtain the following theorem. Theorem 9.10.1. An invariant test for which the acceptance region is convex in each column of U for each set of fixed V and other columns of U has a power function that is monotonically increasing in each Pj' 9.11. ELLIPTICALLY CONTOURED DISTRIBUTIONS 9.11.1. Observations Elliptically Contoured Let x I' ... , X N be N observations on a random vector X with density
(1)
I AI- tg [ (x - v)' A -I (x - v) 1,
9.11
405
ELLIPTICALLY CONTOURED DISTRIBUTIONS
where $R 4
(2)
1
X=N
N
Ex",
,,=1
s= N~l
N
E (x,,-i)(x,,-i)'. ,,=1
Then
(3)
fNvee( S - 'I)
~ N [0, (K + 1) (II" + Kpp)( l: ® ~) + K vee ~ (Vl~C I)'].
where 1 + K = P $R 4 /[(p + 2)( rfR 2)2J. The likelihood ratio criterion for testing the null hypothesis l:ij = 0, i *- j, is the N 12th power of U = ni=2 Vi' where V; is the U-criterion for testing the null hypothesis 'Iii = 0, ... , 'Ii-I. i = 0 and is given by (1) and (6) of Section 9.3. The form of V; is that of the likelihood ratio criterion U of Chapter 8 with X replaced by X(i), P by Pi given by (5) of Section 9.3, Z by
(4)
and 'I by 'I ii under the null hypothesis Pi = O. The subvector X(i - 1) is uncorrelated with 'X(i), but not independent of X(i) unless (X(i-n,. X(i)'), is normal. Let
(5)
(6)
,4(i,i-l) = (Ail>"" A i. i- I ) =,4(i-I.,),
with similar definitions of l;(i-l), l;(i.i-l), S(i-l) , and S(i·'-I). We write
V; = IG;I I IG; + H,I, where (7) - . . - . = (N -l)S(,,'-l)(S(,-I»
(8)
-I -
. S(,-I.,),
G; =Aj/ -H, = (N -l)S;;- Hi'
Theorem 9.11.1. When X has the density (1) and the null hypothesis is true. the limiting distribution of Hi is W[(1 + K ):':'i" Pi J. where Pi = P I + ... +fI, _I
406
TESTING INDEPENDENCE OF SETS OF VAIUATES
alld PI is the number of components of x(j). Proof Since I(i·i-I) = 0, we have rf,"S(i,i-l) = 0 and
(9) if j, I 5,Pi and k, m > Pi or if j, I> Pi' and k, m 5,Pi' and $SjkStm = 0 otherwise (Theorem 3.6.1). We can write
Since S(i-I) ->p l;(i-') and {iivecS(I·,-I) has a limiting normal distribution, Theorem 9.10.1 follows by (2) of Section 8.4. • Theorem 9.11.2. hypothesis is true
Under tlze conditions of Theorem 9.11.1 when the null
- N log V;
(11 )
Ii
->
2
(1 + K) XPiP ;'
Proof We can write V; = II +N-1(*G)-IHil and use N logll +N-1CI = tr C + O/N- 1 ) and
(12)
tr(~GirIHi=N . .P~'
~
gijSigSg"Sjh
/'J~p,+1 g,"~1
Because
Xli)
is uncorrelated with X(i -I) when the null hypothesis
Hi:i(i.i-I) =0, V; is asymptotically independent of V 2 , ... ,V;-I' When the null hypotheses H 2 , ••• , Hi are true, V; is asymptotically independent of V~,
... , ~i _I' It follows from Theorem 9.10.2 that q
(13)
- N log V = - N
d
L log V; -> xl, i~2
where f= 'L.f-2PiPi = Hp(p + 1) - 'L.'!-lP;CPi + 1)]. The likelihood ratio test of Section 9.2 can be carried out on an asymptotic basis.
9.11
407
ELLIPTICALLY CONTOURED DISTRIBUTIONS
Let Ao = diag(A n , ... , Aqq). Then 2
(14)
ttr(AAi)l
-I) =t
g
L
tr AijAjjlAjiA;; 1
i,j~l
N)
has the xl-distribution when :I = diag(:I ll , ... , :I qq ). The step-down procedure of Section 9.6.1 is also justified on an asymptotic basis. 9.11.2. Elliptically Contoured Matrix Distributions
Let Y (p X N) have the density g(tr YY'). The matrix Y is vector-spherical; that is, vec Y is spherical and has the stochastic representation vec Y = R vec UpXN ' where R2 = (vec Y)' vec Y = tr YY' and vec UpXN has the uniform distribution on the unit sphere (vec UpXN )' vec UpXN = 1. (We use the notation ~}XN to distinguish f'om U uniform on the space UU' = 1/,). Let X= VE'N + CY,
(15)
where A = CC' and C is lower triangular. Then X has the density
(16)
IAI" N/ 2g[trC 1(X-vE'N)(X' -ENV')(C')r
l
= IAI-N/2g[tr (X' - E"V') A-I (X - VE'N )].
*
*
Consider the null hypothesis :I ij = 0, i j, or alternatively A ij =, 0, i j, or alternatively, Rij = 0, i j. Then C = diag(C Ip " " C qq ). Let M=I v -(1/N)E N E'N; since M2=M, M is an idempot.:nt matrix with N - 1 characteristic roots 1 and one root O. Then A =XMX' and Aii = X(i)MX(i)'. The likelihood function is
*
(17)
I AI -n /2 K{ tr A - I [A + N( i-v) ( i-v) ,] }.
The matrix A and the vector i are sufficient statistics, and the likelihood ratio criterion for the hypothesis H is (IAI/nf~IIAiil)N/2, the same as for normality. See Anderson and Fang (1990b). Theorem 9.11.3. that
(18)
Let f(X) be a uector-ualuedfunction of X (p X N) such
408
TESTING INDEPENDENCE OF SETS OF VARIATES
./ for all v and
f(KX) =f(X)
(19)
for all K = diag(K n , ... , Kqq). Then the distribution of f(X), where X has the arbitrary density (16), is the same as the distribution of f(X), where X has the normal density (16). Proof The proof is similar to the proof of Theorem 4.5.4.
•
It follows from Theorem 9.11.3 that V has the same distribution under the null hypothesis H when X has the density (16) and for X normally distributed since V is invariant under the transformation X -> KX. Similarly, Vi and the criterion (14) are invariant, and hence have the distribution under normality.
PROBLEMS 9.1. (Sec. 9.3) Prove
by integration of Vhw(AI~o,n). Hint: Show
K(~o,n)2h) cffV h = K(~ ~o,n
+
f ... fn . IA;; I- w(A, ~ ~o' n + 2h ._1 q
h
)
dA,
where K(I,n) is defined by w(AII,n)=K(I,n)IAlhn-p-I)e-ilrrIA. Use Theorem 7.3.5 to show
n [K(IK(I;;,n) ,n+2h)f f ( )] ... w AjjlIii,n dA;; . q
. h K(Io,n) o\"V = K(I ,n+2h) o
9.2. (Sec. 9.3) Prove that if PI
jj
;-1
=
P2 = P3
=
1 [Wilks (1935)]
[Hint: Use Theorem 9.3.3 and Pr{V ~ v} = 1 - Pr{u ~ V}.]
409
PROBLEMS
9.3. (Sec. 9.3) Prove that if PI = P2 = PJ = 2 [Wilks (1935)]
Pr{V:;v} =IJu (n-5,4)
+ B- 1 (n -'\ 4)v~(n-5){ nj6 - ~(n - l)fV - ~(n - 4)u
- ~(n - 2)u log u - ten - 3)u
3 2 /
10g u}.
[Hint: Use (9).] 9.4. (Sec. 9.3)
Derive some of the distributions obtained by Wilks {l935) and referred to at the end of Section 9.3.3. [Hint: In addition to the results for Problems 9.2 and 9.3, use those of Section 9.3.2.] For the case P, = 2, express k and "Y2' Compute the second term at" (6) when u is chosen so that the first term is 0.95 for P = 4 and 6 and N = 15.
9.5. (Sec. 9.4)
9.6. (Sec.9.5)
Prove that if BAR' = CAe' = I for A positive definite and Band C nons in gular then B = QC where Q is orthogonal.
Prove N times (2) has a limiting X2-distribution with freedom under the null hypothesis.
9.7. (Sec. 9.5)
f
degrees of
9.S. (Sec. 9.8) Give the sample vector coefficient of alienation and the vector correlation coefficient. 9.9. (Sec. 9.8)
If y is the sample vector coefficient of alienation and z the square of the vector correlation coefficient, find .f y" z" when ~ I~ = O.
9.10. (Sec. 9.9)
Prove
f ... f 00
co
1
._00
-co
(1 + r.f~IVn'
I
-r
du 1 ... duP <:x:
if P < n. [Hint: Let Yj = wj ';l + "[,r~j+ 1 y?, j = 1, ... , P - I, il. turn.]
= arithmetic speed, x 2 = arithmetic power, x) = intellectual interest. = soeal interest, x5 = activity interest. Kelley (1928) observed the following
9.11. Let XI x4
correlations between batteries of tests identified as above, based on 109 pupils: 1.0000 0.4249 -0.0552 - 0.0031 0.1927
0.4249 1.0000 -0.0416 0.0495 0.0687
- 0.0552 -0.0416 1.0000 0.7474 0.1691
-0.0031 0.0495 0.7474 1.0000 0.2653
0.1927 0.0687 0.1691 0.2653 1.0000
410
TESTING INDEPENDENCE OF SETS OF VARIATES
Let Xl\)' =(x l .x2) and x (2 ), = (X3,x 4,X S)' Test the hypothesis that independent of x (2 ) at the 1% significance level.
x(1)
is
9.12. Can}" out the same exercise on the data in Problem 3.42. 9.13. Another set of time·study data [Abruzzi (1950)] is summarized by the correlation matrix based on 188 observations:
1.00 -0.27 0.06 0.07 0.02 Test the hypothesis that
-0.27 1.00 -0.01 -0.02 -0.02 aij
=
0, i
0.06 -0.01 1.00 -0.07 -0.04
om -om -0.02
1.00 -0.10
0.02 -0.02 -0.04 -0.10 1.00
*" j, at the 5% significance level.
CHAPTER 10
Testing Hypotheses of Equality of Covariance Matrices and Equality of Mean Vectors and Covariance Matrices
10.1. INTRODUCTION
In this chapter we study the problems of testing hypotheses of equality of covariance matrices and equality of both covariance matrices and mean vectors. In each case (except one) the problem and tests considered are multivariate generalizations of a univariate problem and test. Many of the tests are likelihood ratio tests or modifications of likelihood ratio tests. Invariance considerations lead to other test procedures. First, we consider equality of covariance matrices and equality of covariance matrices and mean vectors of several populations without specifying the common covariance matrix or the common covariance matrix and mean vector. Th(' multivariate analysis of variance with random factors is considered in this context. Later we treat the equality of a covariance matrix to a given matrix and also simultaneous equality of a covariance matrix to a given matrix and equality of a mean vector to a given vector. One other hypothesis considered, the equality of a covariance matrix to a given matrix except for a proportionality factor, has only a trivial corresponding univariate hypothesis. In each case the class of tests for a class of hypotheses leads to a confidence region. Families of simultaneous confidence intervals for covariances and for ratios of covariances are given.
An Introduction to Multivariate Stati.ltical Analysis, Third Edition. By T. W. Anderson ISBN 0-471-36091-0 Copyright © 2003 John Wiley & Sons, Inc.
411
412
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
~ The application of the tests for elliptically contoured distributions is ...rtreated in Section 10.11. .
10.2. CRITERIA FOR TESTING EQUALITY OF SEVERAL COVARIANCE MATRICES In this section we study several normal distributions and consider using a set of samples, one from each population, to test the hypothesis that the covariance matrices of these populations are equal. Let x~), a = 1, ... , N g , g = 1, ... , q, be an observation from the gth population N(IJ.(g), l:g). We wish to test the hypothesis
(1 )
Ng
Ag =
(2)
E
(x~g) - x(g»)( x~g) - x(g))',
g= 1, ... ,q,
,,~l
First we shall obtain the likelihood ratio criterion. The likelihood function is
(3)
L=
n q
g~l
I
g
1
l"N (2 1T ) 2PNI~ '''g''
exp [ - -1 "N (x(g) - n(g»)'l:-c1(x(g) -1J.(g») ] .
2
L..
,,~l
"
,...
8
"
The space n is the parameter space in which each I- g is positive definite and lJ.(g) any vector. The space w is the parameter space in which II = l:2 = ... = l:q (positive definite) and lJ.(g) is any vector. The maximum likelihood are given by estimators of lJ.(g) and l:g in
n
g= 1, ... ,q.
(4)
The maximum likelihood estimators of lJ.(g) in ware given by (4), fJ.t;) =i(g), since the maximizing values of lJ.(g) are the same regardless of l:g. The function to be maximized with respect to l:1 = ... = l:q = l:, say, is
(5)
11
2PNI~ltN (21)T ..
_1."q
exp [
2 g~l L..
N
g ]
"(x(g) -x(g»)'l:-I(X(g) -x(g») . L..
,,~l
"
"
413
10.2 CRITERIA FOR EQUALITY OF COVARIANCE MATRICES
By Lemma 3.2.2, the maximizing value of I- is
(6) and the maximum of the likelihood function is
(7) The likelihood ratio criterion for testing (1) is
(8) The critical region is
(9) where AI(e) is defined so that (9) holds with probability E: when (0 is true. Bartlett (1937a) has suggested modifying Aj in the univariate case hy replacing sample numbers by the numbers of degrees of freedom of the A •. Except for a numerical constant, the statistic he proposes is .
(10) where ng = Ng - 1 and n = L~~j ng = N - q. The numerator is proportional to a power of a weighted geometric mean of the sample generalized variances, and the denominator is proportional to a power of the determinant of a weighted arithmetic mean of the sample covariance matrices. In the scalar case (p = 1) of two samples the criterion (10) is (nl)~nl(n2)~11~ F~'I,
(11)
(njF+ n2)\("'+"') '
si
where sf and are the usual unbiased estimators of population variances) and
(12)
s"
F =~, . s'2
(J"j"
and (J"~ (the two
414
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
Thus the critical region
(13) is based on the F-statistic with n l and n z degrees of freedom, and the inequality (13) implies a particular method of choosing FI(e) and F 2 (t:) for the critical region (14)
Brown (1939) and Scheffe (1942) have shown that (14) yields an unbiased test. Bartlett gave a more intuitive argument for the use of VI in place of AI' He argues that if Np say, is small, A 1 is given too much weight in AI' and other effects may be missed. Perlman (1980) has shown that the test based on VI is unbiased. If one assumes (15) where z~) consists of kg components, and if one estimates the matrix defining
I3g ,
N.
A,.... = "i..J (x(g) - A z(g))(x u - I-"g A z(g»), cr t-'g a a ,
( 16)
,,~I
one uses (10) with fig = Ng - kg. The statistical problem (parameter space .n and null hypothesis w) is invariant with respect to changes of location within populations and a common linear transformation
g=l, ... ,q,
( 17)
where C is nonsingular. Each matrix Ag is invariant under change of location, and the modified criterion (10) is invariant:
(18)
V*= I
Ilqg-I IA*lh g IA*I ~n
Ili_IICAgC'lh ICAe'li n
Similarly, the likelihood ratio criterion (8) is invariant.
415
10.3 CRITERIA FOR TESTING THAT DISTRIBUTIONS ARE IDENTICAL
An alternative invariant test procedure [Nagao (l973a)] is based on the criterion q
(19)
2
q
~ 2:ngtr(SgS-I-I) =~ 2: n g tr(Sg-S)S-I(Sg-S)S-I, g=1
g=l
where Sg = (l/ng)Ag and S = (l/n)A. (See Section 7.8.)
10.3. CRITERIA FOR TESTING THAT SEVERAL NORMAL DISTRIBUTIONS ARE IDENTICAL In Section 8.8 we considered testing the equality of mean vectors when we assumed the covariance matrices were the same; that is, we tested
(1 ) The test of the assumption i.1 H2 was considered in Section 10.2. Now let us consider the hypothesis that both means and covariances are the same; this is a combination of HI and H 2 • We test
(2) As in Section 10.2, let x~g), a
= 1.... , N g, be an observation from
N(fJ.(g), Ig),
g= 1, ... ,q. Then D is the unrestricted parameter space of (fJ.(g), I g }, g= 1, ... , q, where Ig is positive definite, and w* consists of the space restricted by (2). The likelihood function is given by (3) of Section 10.2. The hypothesis HI of Section 10.2 is that the parameter point falls in w; the hypothesis H2 of Section 8.8 is that the parameter point falls in w* given it falls in w ~ w* ; and the hypothesis H here i~; that the parameter point falls in w* given that it is in D. We use the following lemma: Lemma 10.3.1. Let y be an observation vector on a random vector with density fez, 6), where 6 is a parameter vector in a space D. Let Ha be the hypothesis 6 E Da cD, let Hb be the hypothesis 6 E Db' C D a, given 6 E D a, and let Hab be the hypothesis 6 E Db' given 6 E D. If Aa, the likelihood ratio criterion for testing Ha, Ab for H b, and Aab for Hab are uniquely defined for the observation vector y, then
(3)
416
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
Proof. The lemma follows from the definitions: max OEfl f(y,O)
(4)
Aa= max "f( y, 0)' OEfl
(5)
Ab =
(6)
Ab= a
max OEflb f(y,O) max OEfl " f(y,O) ' max o E fl. f(y, 0) max OEfl f(y,O)
•
Thus the likelihood ratio criterion for the hypothesis H is the product of the likelihood ratio criteria for HI and H 2 ,
(7) where q
(8)
B=
Ng
L L g~1
(x~g)-x)(x~,g)-x)'
a~1
q
=A +
L
Ng(x(g) -x)(i(g) -i)'.
g~1
The critical region is defined by
(9)
A:::; A( e),
where A(e) is chosen so that the probability of (9) under H is e. Let
(10)
v 2
= IAI ~n = An / N . IBI~"
2'
this is equivalent to A2 for testing Hz. which is A of (12) of Section 8.8. We might consider
(11)
V= VV = I 2
n
IA I tn, ,g IBl,n
q_ g-l
However. Perlman (1980) has shown that the likelihood ratio test is unbiased.
10.4
417
DISTRIBUTIONS OF THE CRITERIA
10.4. DISTRIBUTIONS OF THE CRITERIA 10.4.1. Characterization of the Distributions First let us consider VI given by (10) of Section 10.2. If
(1)
g=2, ... ,q,
then
(2)
Theorem 10.4.1.
V12 , VI ;, ... , V 1q defined by (1) are independellT when
l:l = ... = l:q and ng "2:.p, g = 1, ... , q. The theorem is a consequence of the following lemma: Lemma 110.4.1. If A and B are independently distributed according to W(l:, m) and W(l:, n), respectively, n "2:. p, m "2:. p, and C is such that C(A + B)C' = I, then A + Band CAC' are independently distributed; A + B has the Wishart distribution with m + n degrees of freedom, and CAe' has the multivariate beta distribution with nand m degrees of freedom. Proof of Lemma. The density of D = A + Band E = CAC' is found by replacing A and B in their joint density by C- lEe' -1 and D - C 1EC' - I = C-I(J - E)C'-l, respectively, and multiplying by the Jacobian, which is mod\C\-(p+l) = \D\ ~(p+l), to obtain
for D, E, and I - E positive definite.
•
418
=
TESTING HYPOTHESES OF EQUAL TY OF COVARIANCE MATRICES
Proof of Theorem. If we let AI + ... +Ag=Dg and CgCA J + ... = I, g = 2, ... , q, then
+Ag_J)C~
E g, where CgD8C~
C-IE C-llt(",+···+n'-I)lc-I(/_E
V =I g g g Ib'
g le-·le,111(1I1+···+lIg) g g
)C-1I in ,
gg
g=2, ... ,q, and E 2 , ... , Eq are independent by Lemma 10.4.1.
•
We shall now find a characterization of the distribution of V lg • A statistic V 1g is of the form
IBlblCI IB + Cl b +c C
(5)
•
Let B; and C; be the upper left-hand square submatrices of Band C, n::spectivdy, of order i. Define b(i) and c(;) by B;_l
(6)
i = 2, ... ,p.
B;= ( b' (I)
Then (5) is (Bo = Co = I,
btl)
= C(l) = 0)
( 7)
IBlblCI IB + Cl b + c = C
P
IB;lbIC;lc
lJ IB;_llbIC;_llc
P
=0 i= I
IB;_1 +C;_llb+c IB; + C;l b + C
b~.; -1 cii ';-1
(b;;_;_1
+ Cjj_;_1 )b+c
419
10.4 DISTRIBUTIONS OF THE CRITERIA
where b jj . j _ 1 = b jj - b(i)Bi=-\ b(i) and c jj ' j - 1 = c jj - c(j)Cj-_\ c(j). The second term for i = 1 is defined as 1. Now we want to argue that the ratios on the right-hand side of (7) are statistically independent when Band C are independently distributed accorrling to W(l:, m) and W(l:, n), respectively. It follows from Theorem 4.3.3 that for B j _ 1 fixed b(i) and bjj.j-] are independently distributed according to N(~(i)' (J'jj.j_IBj-=-\) and (J'jj.j_l X 2 with m - (i -1) degrees of freedom, respectively. Lemma 10.4.1 implies that the first term (which is a function of bjj.j_I/Cjj.j_l) is independent of b j ;';_1 + Cjj . j _ l . We apply the following lemma: Lemma 10.4.2.
For B j _ 1 and Cj _ 1 positive definite
I I I I -- (B-j_11 b(j) - C-j_Ic(j) )'(Bj - l + Cj-I )-I(Bj-I b(j)
-
1 C-j-Ic(i)' )
Proof Use of (B- 1 + C l ) - I = [CI(B + C)B-I]-I = B(B + shows the left-hand side of (8) is (omitting i and i - 1)
C)-IC
(9) b'B-I(B- 1
+ C l f\B- I + C- 1 )b + c'(B- I + C l )(B- 1 + C- I f 1 c l c
-(b+c)'B-I(B- 1 +C I )-IC 1 (b+c) =b'B-I(B- 1
+ c - I f l B-Ib + c'C-I(B- 1 + C- I ) -IC-I C
I -b'B-I(B- 1 +C)
-I
C
1
C-c'C I (B- I +CI)B-Ib,
which is the right-hand side of (8).
•
The denominator of the ith second term in (7) is the numerator plus (8). The conditional distribution of Bj-_Il b(j) - Cj-_\ c(i) is normal with mean Bj--\~(i) - Cj--\'Y(i) and covariance matrix (J'jj.j_I(B j-_\ + C j-_\). The covariance matrix is (J'jj.j -I times the inverse of the second matrix on the right-hand side of (8). Thus (8) is distributed as (J'jj.j_1 X2 with i - I degrees of freedom, independent of Bj _ l , Cj _ 1 , bjj . j _ l , and c jj . j _ l • Then
(10)
bh';_IC~.j_1
( b jj ' j _ l
+coo. )b+c = ( boo. ll"f-] ll'l-l
is distributed as X/O -
b jj ' j _ 1
xy, where
+ Cjj.i-l
)b(
C jj ' j _ l b jj. j _ 1
)C
+ Cjj' j _ 1
Xj has the {3[ ~(m - i + 1), ~(n - i + 1)]
420
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
distribution, i = 1, ... , p. Also (11)
]b+C
bjj'i_I+Cjj.i_1 [
hii .i- I
+C jj . i - I +
(8)
,
i=2, ... ,p,
is distributed as Yi b + c , where Yi has the {3[ ~(m + n) - i + 1, ~(i - 1)] distribution. Then (5) is distributed as nf_Ixt(1-xynf=2l'ib+c, and the factors are mutually independent. Theorem 10.4.2.
where the X's and Y's are independent, X ig has the (3[~(nl + ... +n g_ 1 ~(ng - i + 1)] distribution, and Yi g has the (3[~(nl + ... +ng) - i + distribution.
-
i
+ 1),
qu -1)]
Proof The factors V I2 , ... , V lq are independent by Theorem 10.4.1. Each term V1g is decomposed according to (7), and the factors are independent.
•
The factors of VI can be interpreted as test criteria for sub hypotheses. The term depending on X i2 is the criterion for testing the hypothesis that aYl_1 = (J'i\~l-I' and the term depending on Yi2 is the criterion for testing = given (J'Yl_1 = (J'i~21_1' and I i -1.1 = I i -I,2' The terms depen9ing on X ig and Yi g similarly furnish criteria for testing 1.1 = Ig given II = ... =
am am
I
g - 1•
Now consider the likelihood ratio criterion A given by (7) of Section 10.3 for testing the hypothesis f.t(I) = ... = f.t(q) anc' II = ... = I q • It is equivalent to the criterion
(13)
The two factors of (13) are independent because the first factor is independent of Al + ... +Aq (by Lemma 10.4.1 and the proof of Theorem 10.4.1) and of i(l>, ... , i(q).
421
10.4 DISTRIBUTIONS OF THE CRITERIA
Theorem 10.4.3
wheretheX's, Y's, andZ'sareindependent, X ig has the .B[~(nl + ... +11 .. __ 1 i + 1)¥n g - i + 1)] distribution, Y;g has the .BWn l + '" +n g) - i + 1, ~(i - 1)] distribution, and Zi has the .B[~(n + 1 - iH(q - 1)] distn·bution. Proof The characterization of the first factor in (13) corresponds to that of VI with the expone~ts of X ig and 1 - X ig modified by replacing ng by N g . The second term in Up~~-l. "' and its characterization fo\1ows from Theorem
8.4.1.
•
10.4.2. Moments of the Distributions We now find the moments of VI and of W. Since 0 ::; VI ::; 1 and 0 ::; W::; 1, the moments determine the distributions uniquely. The hth moment of VI we find from the characterization of the distribution in Theorem lOA.2:
(15) q
tffV h = n {nP tffX!(II'+'--+II,-ll h(l-X 1
g~2
.
,~l
rg
)~II-,hnP tffY!(II;+---+II,lI'\j '.~
II;.
1~2
r[ing (1 + h) - iU -1)]r[ Hnl + ... +ng) - i + 1] r[ Hng - i + 1)] r[ Hn l + ... +ng)(1 + h) - i -'- 1]
.fr i=2
r [Hn 1+ ...
r[Hnl
+ n g) (1 + h) - i + 1] r [-t( n I + '" + 11 g - i + 1)] } + ... +ng) -i + 1]r[-hnl + ", +ng)(1 +h) - ~(i -1)]
422
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
The hth moment of W can be found from its representation in Theorem 10.4.3. We have (16)
n n r[.!.(n q
g-~
{
p
2
I
+ ...
+11 g-I + 1- i)
rlH1l1 + '"
;~I
rl Hll
g
2
+ 1 - i + Ngh)]
1
+ '" +Ng-I )]
+11.,_1 + l-i)]r[!{n g + 1-i)]
r[ Hnl + ... +ng) -
r[!(n l + ... +ng) + !h(NI
-np r[.!.(n
+ .!.h(N 2 I
i+
1]
+ '" +Ng) + 1- i]
+"'+n g )+.!.h(N +"'+N)+l-i] 2 I g
;=2
r[!(n l +"'+n g)+l-i]
p r[Hn+1-i+hN)]r[hN-i)]
=
n{n
;=1
g=1
)l r[Hn + 1-i)]1'[!(N+hN-i)] r[!(Ng+hNg-i}]} r[!(N-i)] r[HNg-i)] r[~(N+hN-i)]
1'pOn)
Ii r p[!(ll g+hNg)]
rp(!n+!hN}g=1
rgOllg)
We summarize in the following theorem:
Theorem 10.4.4. Let VI be the criterion defined by (10) of Section ]0.2 jor testing the hypothesis that HI : I I = ". = I", where Ag is ng times the sample covariance matrix and n g + 1 is the size of the sample from the gth population; let W be the criterion defined by (13) for testing the hypothesis H: t-tl = '" = t-t q and HI' where B = A + LgN/i(g) - i)(i(g) -r)'. The hth moment of VI when H I is true is given by (15). The h th moment 0, . W, the criten'on for testing H, is given by (16). This theorem was first proved by Wilks (1932). See Problem 10.5 for an alternative approach.
10.4
423
DISTRIBUTIONS OF THE CRITERIA
If p is even, say p = 2r, we can use the duplication formula for the gamma function irca + ~)I'(a + 1) = {iTf(2a + 1)2-2,,]. Then
{f q f(n g +hn g +I-2 j )1 f(n+I-2j) } J]}] f(ng+I-2j) f(n+hn+I-2j) r
h
(17)
rlVl =
and
n {[nq r
(18)
rlW h =
j=1
g=1
f(n g +hNg +I-2 j f(n g +I-2j)
)1
f(N-2j) } f(N+hN-2j)'
In principle the distributions of the factors can be integrated to obtain the distributions of VI and W. In Section 10.6 we consider VI when p = 2, q = 2 (the case of p = 1, q = 2 being a function of an F-statistic). In other cases, the integrals become unmanageable. To find probabilities we use the asymptotic expansion given in the next section. Box (1949) has given some other approximate distributions. 10.4.3. Step-down Tests The characterizations of the distributions of thc criteria in terms of independent factors suggests testing the hypotheses HI and H by testing component hypotheses sequentially. First, we consider testing HI: I I = I2 for q = 2. Let ()_
(19) X(i1 -
(
X(!?) (,-I) ) g) ,
xi
g) ) fJ.(i-I)
(g) _
fJ.(i) -
f..Llg)
(
,
i=2, ... ,p,
The conditional distribution of
Xi
g
)
g=I,2.
Xlll l ) = x~fll) is
given
(20) where UiW-I = Uj~g) - 0W'I~~\ oW. It is assumed that the components of X have been numbered in descending order of importance. At the ith step the component hypothesis ui~~l-I = Ui~~l-I is tested at significance level Bj by means of an F-test based on sil)i-tlsif)j-I; SI and 8 2 are partitioned like I(I) and I (2). If that hypothesis is accepted, then the hypothesis = (or II~llam=Ii~llam) is tested at significance level 0i on the assumption that Ii~1 = Ii~1 (a hypothesis prwiously accepted). The criterion is
am am
(21)
(
8(1)-1 S(I) ,-I (,)
S(2)-1 S(2»)'(S(I)--1 ,-I
(,)
,-I
(i -
+ S(2)-1 )-1 (S(I)-I S(I) ,-I· ,-I (,) I)Sii i-1 o
S(2)-1 S(2») ,-I
(,)
424
TESTING HYPOTHESES OF EQUALITY OF COY ARlANCEMATRICES
where (n l + n 2 - 2i + 2)SU"i_1 = (n l - i + l)Sg'>i_1 + (n2 - i + l)S}r.~_I. Under the null hypothesis (21) has the F-distribution with i-I and n l + n2 - 2i + 2 degrees of freedom. If this hypothesis is accepted, the (i +l)st ~tep is taken. The overall hypothesis "II ="I 2 is accepted if the 2p - 1 component hypo:heses are accepted. (At the first step, CTW is vacuous.) The overall significance level is p
(22)
p
TI (1 - eJ TI (1 -
1-
i=1
i=2
0;).
If any component null hypothesis is rejected, the overall hypothesis is rejected. If q > 2, the null hypotheses HI:"II = ... ="I q is broken down into a sequence of hypotheses [l/(g - 1)]("I1 + ... + "Ig_l) = "Ig and tested sequentially. Each such matrix hypothesis is tested as "II ="I 2 with S2 replaced by Sg and SI replaced by [l/(n l + ... +ng_I)](A I + ... +A g_ I ). In the case of the hypothesis H, consider first q = 2, "II = "I 2, and fL(1) = fL(2). One can test "II = "I 2 • The steps for testing fL(l) = fL(2) consist of t-tests for /-LI I) = /-L12) based on the conditional distribution of XiI) and XP) given l ) and I ). Alternatively one can test in sequence the equality of the conditional distributions of XP) and XP) given l ) and XU:I). For q> 2, the hypothesis "II = ... ="I q can be tested, and then fLI = ... =fLq. Alternatively, one can test [l/(g-l)]("II+···+"I g_I)="I g and [l/(g - l)](fL(l) + ... + fL(g-I») = fL(g}.
xg:
xg:
xg:
10.5. ASYMPTOTIC EXPANSIONS OF THE DISTRIBUTIONS OF THE CRITERIA Again we make use of Theorem 8.5.1 to obtain asymptotic expansions of the distributions of VI and of A. We assume that ng = kgn, where ~~_Ikg = 1. The asymptotic expansion is in terms of n increasing with k l , ... , kq fixed. (We could assume only lim ng/n = kg > 0.) The hth moment of
n tpn
(1) is
(2)
A*1 = V 1 .
n qg=lngtpn
q
= V1 . g
n ,pn g
TI (-) n
g=1
g
I
[
="
q
1
g=1
g
TI (-) k
kg
]tpn
V1
10.5
425
ASYMPTOTIC EXPANSIONS OF DISTRIBUTIONS OF CRITERIA
This is of the form of (1) of Section 8.6 with
(3)
b=p,
Yj=~n,
7Jj=~(1-j),
a=pq,
xk=~ng,
k=(g-l)p+1.. .. ,gp. k=i,p+i, ... ,(q-l)p+i,
j
= 1, ... ,p,
g
= 1. ....
q.
i= I, .... p.
Then
f= -
(4)
2[ L gk - L 7Jj - Ha -
= - [q
b) 1
i~ (1 - i) - f~ (1 - j) -
(qp - p) ]
=-[-q~p(p-l)+~p(p-l)-(q-l)pl
= Sf
Hq -
1) p( P + 1),
= ~(1- p)n, j = l, ... ,p, and 13k
= ~(1 - p)l1~
= ~(1- p)k~l1, k =
j)p
(g.-
+ 1, ... ,gp. In order to make the (5)
p
sec(~md
term in the expansion vanish, we take
=l-(t~-~) g=Il1g
11
p
2p"+3p-l 6(p+l)(q-l)'
Then
p(p+l) (p-l)(p+2) ( [
(6)
w2 =
g"f" q
1 l1i
1 112
----....:"-------=-----:,.---------~
48p2
Thus
(7)
Pr{ - 2p log Ai :=; z}
=Pr{xJ:=;z}+w 2 [Pr{xl+4:=;z}-i'r{xl:=;z}] +0(11-").
(8)
'J
) -6(q-l)(1-pf .
as
426
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
This is the form (1) of Section 8.5 with q
b=p,
j=1, ... ,p,
Yj= !N=! LNg, g~1
(9)
a=pq,
k=(g-1)p+1, ... ,gp,
g= 1, ... ,q,
k=i,p+i, ... ,(q-1)p+i, i=1, ... ,p. The basic number of degrees of freedom is f = ~p(p + 3Xq - 1). We use (11) of Section 8.5 with 13k = (1 - ph k and sf = (1- p)Yj' To make W 1 = 0, we take
( 10)
1 1)
11
_ (q 2p2 -I- 9p + p-1- gL:INg -N 6(q-1)(p+3)'
Then
The asymptotic expansion of the distribution of - 2 p log ,\ is (12) Pr{ -2p log A :=;z}
=Pr{x/ :=;z} +w2 [Pr{xl+4 :=;z} -Pr{xl:=;z}] +O(n- 3 ). Box (949) considered the case of Ai in considerable detail. In addition to this expansion he considered the use of (13) of Section 8.6. He also gave an F-approximation. As an example, we use one given by E. S. Pearson and Wilks (1933). The measurements are made on tensile strength (XI) and hardness (X 2 ) of aluminum die castings. There are 12 obselVations in each of five samples. The obselVed sums of squares and cross-products in the five samples are
( 13)
A = ( 78.948 I 214.18
214.18 ) 1247.18 '
A = (223.695 2 657.62
657.62 ) 2519.31 '
A = ( 57.448 3 190.63
190.63 ) 1241.78 '
A = (187.618 4 375.91
375.91 ) 1473.44 '
A
259.18 ) 1171.73 '
= ( 5
88.456 259.18
427
10.6 THE CASE OF TWO POPULATIONS
and the sum of these is (14)
LA
= ( 636.165 I
1697.52
1697.52) 7653.44 .
The -log.\i is 5.399. To use the asymptotic expansion we find p = 152/165 = 0.9212 and W2 = 0.0022. Since W2 is small, we can consider -2p log Ai as X 2 with 12 degrees of freedom. Our obselVed criterion, therefore, is clearly not significant. Table B.5 [due to Korin (1969)] gives 5% significance points for - 2 log Ai for N1 = ... = Nq for various q, small values of Ng , and p = 2(1)6. The limiting distribution of the criterion (19) of Section 10.1 is also An asymptotic expansion of the distribution was given by Nagao (1973b) to terms of order l/n involving x2-distlibutions with t, t";' 2, t + 4, and t + 6 degrees of freedom.
xl-
10.6. THE CASE OF TWO POPULATIONS 10.6.1. Invariant Tests When q = 2, the null hypothesis H1 is "I 1 = "I 2 . It is invariant with respect to transformations
(1) where C is nonsingular. The maximal invariant of the parameters under the transformation of locations (C = I) is the pair of covariance matrices ~1' "I 2, and the maximal invariant of the sufficient statistics x(1), S1' x(2), S2 is the pair of matrices S1' S2 (or equivalently AI' A 2 ). The transformation (1) induces the transformations "Ii = C"I 1C', "Ii = C"I 2C', Si = CS 1C', and Si = CS 2 C'. The roots A1 ~ A2 ~ ... ~ Ap of
(2) are invariant under these transformations since
Moreover, the roots are the only invariants because there exists a nonsingular matrix C such that
(4) where A is the diagonal matrix with Ai as the ith diagonal element, i = 1, ... ,p. (See Th~orem A.2.2 of the Appendix.) Similarly, the maximal
428
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
invariants of SI and S2 are the roots II
~
l2
~
.. ,
~
lp of
(5) Theorem 10.6.1. The maximal invariant of the parameters of N(IL(l), '1\) and N(IL(2), '1 2) under the tramformation (1) is the set of roots AI ~ '" ~ Ap of (2). The maximal invariant of the sufficient statistics i(l), SI' i(2), S2 is the set of roots II ~ ... ~ lp of (5).
Any invariant test criterion can be expressed in terms of the roots ll,. .. ,lp' The criterion VI is n{pnlnyn, times
(6)
ISII ~"IIS21 ~II,
I niSI + n~S~1 ~n
where L is the diagonal matrix with lj as the ith diagonal element. The null hypothesis is rejected if the smaller roots are too small or if the larger roots are too large, or both. The null hypothesis is that Al = ... = Ap = 1. Any useful invariant test of the null hypothesis has a rejection region in the space of ll"'" lp that inc\ude~ the points that in some sense are far from II = ... = II' = 1. The power of an invariant test depends on the parameters through the roots AI' ... , Ap. The criterion (19) of Section 10.2 is (with nS = niSI + n 2 S 2)
(7)
~nltr[(SI-S)S-lr + ~n2tr[(S2-S)S-I]2
= ~nl tr [C( SI - S)C'( CSC') _1]2
+ ~n2 tr [C(S2 - S)C'(CsC')-lr =
:1 L+ :~ 1) }(:1 L+ :2 1)-112
~1l1 tr [{L - (
This criterion is a measure of how close II"'" II' arc to 1; the hypothesis is reje.cted if the measure is too large. Under the null hypothesis, (7) has the X 2-distribution with f = 1p( P + 1) degrees of freedom as n l --> 00, n 2 --> 00,
429
10.6 THE CASE OF TWO POPULA nONS
and n 11n 2 approaches a positive constant. Nagao (1973b) gives an asymptotic expansion of this distribution to terms of order lin. Roy (1953) suggested a test based on the largest and smallest roots, II and I p' The procedure is to reject the null hypothesis if II > k I or if I p < k p' where k! and kp are chosen so that the probability of rejection when A = I is the desired significance level. Roy (1957) proposed determining kl and kp so that the test is locally unbiased, that is, that the power functions have a relative minimum at A = I. Since it is hard to determine kl and kp on this basi&, other proposals have been made. The Ii: nit k I can he determined so that Pr{/l>ktlHI} is one-half the significance level, or Pr{/p
(8)
tS'V h = renl + hnl - 1)r(n2 + hn2 - 1)r(n - 1) I r(nl -1)f(n2 - l)r(n +hn -1)
where XI and X 2 are independently distributed according to ,B(xin I - 1. n2 - 1) and ,B(xi n l + n 2 - 2, 1), respectively. Then Pr{V1 :=; v} can be found by integration. (See Problems 10.8 and 10.9.) Anderson (1965a) has shown that a confidence interval for a''Ilala''I2a for all a with confidence coefficient e is given by Upl U, II I L), where Pr{(n 2 - p + l)L:=; n 2 F",.",_p+l} Pr{(n l - p + l)F",_p+l, n,:=; np} = 1 - e. 10.6.2. Components of Variance In Section 8.8 we considered what is equivalent to the one-way analysis of variance with fixed effects. We can write the model in the balanced case (N! = N2 = ". = Nq ) as
(9)
x~g)
=
IL(g)
+ U.!8) a=I, ... ,M,
g=l, ... ,q.
430
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
where rCU(g) = 0 and rCU(g)U(g)' ='1, Vg = fL(g) - fL, and fL = (1/q)E~=lfL(g) (L%,~ I v, = 0). The null hypothesis of no effect is VI = .. , = Vq = O. Let i l " ) = O/M)E~~ I x~g) and i = (l/q)L~=1 i(g). The analysis of variance table is Degrees of Freedom
Sum of Squares
Source q
Effect
H
L
=M
(i(g) - i)(i(g) - i)'
q-l
g=1 '{
Error
G=
M
L L
(x~g) - i(g»)(x~g) - i(g))'
q(M -1)
g= 1 a= 1 q
Total
M
L L
(x~g) - i)(x~~) - i)'
qM-l
g= 1 a=1
Invariant tests of the null hypothesis of no effect are b~lsed on the roots of \H - mG\ = 0 or of ISh -ISe\ = 0, where Sh = [lj(q -1)]H and S" = [1/q(M - O]G. The null hypothesis is rejected if one or more of the roots is too large. The error matrix G has the distribution WO., q(M -1)). The effects matrix H has the distribution W('1, q - 1) when the null hypothesis is true and has the noncentral Wishart distribution when the null hypothesis is not true; its expected value is q
(10)
rCH
= (q
- 1)'1
+M
L (fLlX ) -
fL)(fL(g) - fL)'
g=1 q
= (q -l)~ +M
.
L VgV~. g=1
The MANDY A model with random effects is a= 1, ... , M,
(11 )
g= 1, ... ,q,
where V, has the distribution MO,0). Then x~g) has the distribution +. 0). The null hypothesis of no effect is
MfL, "I
( 12)
0=0.
In this model G again has the distribution W('1, q(M - 1)). Since i(g) = fL + V, + i7(g) has the distribution N(fL, (l/M)'1 + 0), H has the distributior + M0, q - 1). The null hypothesis (12) is equivalent to the equality 0
wet
10.7 TESTING HYPOTHESIS OF PROPORTIONALITY; SPHERICITY TEST
431
the covariance matrices in these two Wishart distributions; that is, :I = "I + Ma. The matrices G and H correspond to Al and A2 in Section 1O.6.l. However, here the alternative to the null hypothesis is that ("I + M a) -"I is positive semidefinite, rather than "I 1 "* "I 2 • The null hypothesis is to be rejected if H is too large relative to G. Any of the criteria presented in Section 10.2 can be used to test the null hypothesis here, and its distribution under the null hypothesis is the same as given there. The likelihood ratio criterion for testing a = 0 must take into account the fact that a is positive semidefinite; that is, the maximum likelihood estimators of "I and "I + Ma under n must be such that the estimator of a is positive semidefinite. Let II > I~ > ... > II' be the roots of
(13) (Note {l/[q(M - l)]}G and (l/q)H maximize the likelihood without regard to a being positive definite.) Let Ij = Ii if Ii> 1, and let Ii = 1 if I j ::;; 1. Then the likelihood ratio criterion for testing the hypothesis a = 0 against the alternative a positive semidefinite and a"* 0 is
(14)
where k is the number of roots of (13) greater than 1. [See Anderson (1946b), (l984a), (1989a), Morris and Olkin (1964), and Klotz and Putter (1969).]
10.7. TESTING THE HYPOTHESIS THAT A COVARIANCE MATRIX IS PROPORTIONAL TO A GIVEN MATRIX; THE SPHERICITY TEST 10.7.1. The Hypothesis In many statistical analyses that are considered univariate, the assumption is made that a set of random variables are independent and have a common variance. In this section we consider a test of these assumptions based on repeated sets of observations. More precisely, we use a sample of p-eomponent vectors XI"'" X N from N(fL,"I) to test the hypothesis H:"I = (721, where (72 is not specified. The hypothesis can be given an algebraic interpretation in terms of the characteristic roots of "I, that is, the roots of
(1)
1"I-eM =0.
432
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
The hypothesis is true if and only if all the roots of (1) are equal. t Another way of putting it is that the arithmetic mean of roots rPI' ... ' rPp is equal to the geometric mean, that is,
nf-lrPl/ p 1:.1-1 rPJp
(2)
=1'1II/P =1 tr tip .
The lengths squared of the principal axes of the ellipsoids of constant density are proportional to the roots rPi (see Chapter 11); the hypothesis specifies that these are equal, that is, that the ellipsoids are spheres. The hypothesis H is equivalent to the more general form '1}1 = u 2 '1}1o, with 'I}1o specified, having observation vectors YI' ... ' YN from N( v, '1}1). Let C be a matrix such that (3) and let p.* = Cv, '1* = C'I}1C', x: = Cy". Then xi, ... ,x~ are observations from N( p.* , "I *), and the hypothesis is transformed into H:'1 * = u 2I. 10.7.2. The Criterion In the canonical form the hypothesis H is a combination of the hypothesis HI :'1 is diagonal or the components of X are independent and H 2 : the diagonal elements of "I are equal given that "I is diagonal or tht. variances of the components of X are equal given that the components are independent. Thus by Lemma 10.3.1 the likelihood ratio criterion A for H is the product of the criterion AI for HI and A2 for H 2 • From Section 9.2 we see that the criterion for HI is (4) where N
(5)
A=
E
,,-1
(x" -i)(x" -i)' =
(aiJ
and r ij = a i / Vajjajj. We use the results of Section 10.2 to obtain A2 by considering the ith component of x" as the a th observation from the ith population. (p here is q in Section 10.2; N here is Ng there; pN here is N tThis follows from the fact that 1: = O'
is a diagonal matrix with roots as diagonal
10.7 TESTING HYPOTHESIS OF PROPORTIONALITY; SPHERICITY TEST
433
there.) Thus
(6)
Thus the criterion for H is
(7) It will be obseIVed that A resembles (2). If II' ... ' I p are the roots of
IS -lii = 0,
(8)
where S = (l/n)A, the criterion is a power of the ratio of the geometric mean to the arithmetic mean,
_ (nl,IIP )Y>N --
(9)
A-
'f,ljp
Now let us go back to the hypothesis 'IT = (T 2 '11o, given obseIVation vectors YI, ... ,YN frem N(v, 'IT). In the transformed variables {x~} tht' criterion is IA*I tN(tr A* /p)- tpN, where N
(10)
E
A* =
(x: - i*)( x: - i* ) ,
a=1 N
=C
E
(Ya - Y)(Ya - y)'C'
a=1
=CBC',
where N
(11)
B=
E (Ya-Y)(Ya-Y)'· a=1
434
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
From (3) we have '110 = C-I(C')-I = (C'C)-I. Thus
(12)
tr A* = tr CBC' = tr BC'C
The results can be summarized.
Theorem 10.7.1. Given a set of p-component observation vectors Y1>···' J'N from N( v, '11), the likelihood ratio criterion for testing the hypothesis H: '11 = U 2 'I'll' where '110 is specified and u 2 is not specified, is
( 13)
Mauchly (1940) gave this criterion and its moments under the null hypothesis. The maximum likelihood estimator of u 2 under the null hypothesis is tr BWll I l(pN), which is tr A/(pN) in canonical form; an unbiased estimator is tr B '110 I l[p(N - 1)] or tr A/[ peN - 1)] in canonical form [Hotelling (1951)]. Then tr B WU- I I u 2 has the X2-distribution with peN - 1) degrees of freedom.
10.7.3. The Distribution and Moments of the Criterion The distribution of the likelihood ratio criterion under the null hypothesis can be characterized by the facts that A = Al A2 and Al and A2 are independent and by the characterizations of Al and A2. As was obseIVed in Section 7.6. when l: is diagonal the correlation coefficients {r ij } are distributed independently of the variances {aiJ(N - l)}. Since Al depends only on {ri) and A2 depends only on {all}' they are independently distributed when the nlill hypothesis is true. Let W = A2/ N, WI = AV N, W Z = AY N. From Theorem 9.3.3. we see that WI is distributed as ni'~z Xi' where X 2 , ••• , Xp are i + 1), 1)], where n = independent and Xi has the density (3 [xl N _. l. From Theorem 10.4.2 with W z = PI'VI2/1I, we find that W2 is distributed as ppnr~2 Yji- I (J - Yj), where Y2, .. ·, Yp are independent and 1] has the density (3Cy\ tn(j - 1Hn). Then W is distributed as WIWZ ' where WI and W 2 are independent.
tCn -
tCi -
10.7 TESTING HYPOTHESIS OF PROPORTIONALITY; SPHERICITY TEST
435
The moments of W can be found from this characterization or from Theorems 9.3.4 and 10.4.4. We have (14)
(15) It follows that
( 16)
¥,
For p = 2 we have (17)
h h r(n) tC'W =4 r(n+2h)
Dr[!Cn+1-i)+h] r[t(n+l-i)] 2
_ r(n)r(n-l+2h) _ n-1 - r(n+2h)r(n-1) - n-1+2h = (n -1) {zn-2+2h dz, o by use of the duplication formula for the gamma function. Thus W is distributed as Z2, where Z has 'he density (n _l) z n-2, and W has the density ~(n - l)w t(n - 3). The cdf is (18)
Pr{W.::;; w} =F(w) = wt
This result can also be found from the joint distribution of 11 ,/2 , the roots of (8). The density for p = 3, 4, and 6 has been obtained by Consul (l967b). See also Pillai and Nagarsenkar (1971). 10.7.4. Asymptotic Expansion of the Distribution
From (16) we see that the rth moment of
(19)
wtn = Z, say,
is
436
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
This is of the form of (1), Section 8.5, with
(20)
a =p, b = 1,
k= 1, ... ,p, YI = tnp,
711
= O.
Thus the expansion of Section 8.5 is vard with f = w(p + 1) - 1. To make the second term in the expansion zero we take p so 1
(21)
- p=
2p2+p+2 6pn
Then (22)
(p + 2)(p -1)(p - 2)(2 p 3 + 6p2 + 3p + 2) 288p 2n 2p2
Thus the cdf of W is found from (23)
Pr{-2plogZ~z}
=Pr{-nplogW~z}
/
= Pr{x/ ~z} + w 2 (Pr{xl+4 ~z} - Pr{xl ~z}) +O(n- 3 ). Factors c(n, p, e) have been tabulated in Table B.6 such that (24)
Pr{ -np log W ~ c(n,p, e) X1p(n+1l-'1 (e)} = e.
Nagarsenkar and Pillai (1973a) have tables for W. 10.7.5. Invariant Tests
The null hypothesis H: 'I = ( j 2 I is invariant with respect to transformations X* = cQX + v, where c is a scalar and Q is an orthogonal matrix. The invariant of the sufficient statistic under shift of location is A, the invariants of A under orthogonal transformations are the characteristic roots 11>"" Ip, and the invariants of the roots under scale transformations are functions that are homogeneous of degree 0, such as the ratios of roots, say Id/2, ... ,lp_I/lp' Invariant tests are based on such hnctions; the likelihood ratio criterion is such a function.
10.7
TESTING HYPOTHESIS OF PROPORTIONALITY; SPHERICITY TEST
437
Nagao (1973a) proposed the criterion (25)
-ntr I
(
2
tf 1 S ) -P- ( S- tr 1 S ) -P S- P tr S p tr S
=!n tr (tiS S _1)2 = ~n(1,tr S2 - p] - (tr Sr = !n[~2
~ 12 _ ] ("(I 1.)2.'::' I p '-,= I /-1
=
where
i = Ef= Il;/P.
~n Lf=I(!; -
I
12
i)2
The left-hand side of (25) is based on the loss function
LqCI., G) of St:ction 7.8; the right-hand side shows it is proportional to the square of the coefficient of variation of the characteristic roots of the sample covariance matrix S. Another criterion is II/Ip. Percentage points have been given by Krishnaiah and Schuurmann (1974).
10.7.6. Confidence Regions Given observations Y I' •.. , YN from N( v, '1'), we can test 'I' = U 2 \II 0 for any specified \('0' From this family of tests we can set up a confidence region for lV. If any matrix is in the confidence region, all multiples of it are. This kind of confidence region is of interest if all components of y" are measured in the same unit, but the investigator wants a region iIllkpcndcnt of this common unit. The confidence region of confidence 1 - e: consists of all matrices '1'* satisfying
(26) where A(e:) is the e: significance level for the criterion. Consider the case of p = 2. If the common unit of measurement is irrelevant, the investigator is interested in T= 1/111 /1/122 and p = 1/1121..J 1/1 11 1/122 . In this case
(27)
-P#II1/122) 1/111
'1'-1 =
- pr:;). T
438
TESTING HYPOTHESES OF EQUALITY OF COYARIANCEMATRICES
The region in terms of
T
and p is
(28)
Hickman
(l~53)
has given an example of such a confidence region.
10.8. TESTING THE HYPOTHESIS THAT A COVARIANCE MATRIX IS EQUAL TO A GIVEN MATRIX 10.8.1. The Criteria If l" is distributed according to N( v, 'IT), we wish to test HI that '\{I = '\{IQ, where 'ITo is a given positive definite matrix. By the argument of the preceding section we see that this is equivalent to testing the hypothesis HI:"I = J, where "I is the covariance matrix of a vector X distributed according to N(IJ.,"I). Given a sample xI' ... 'X N, the likelihood ratio criterion is
max ... L(IJ., 1)
( I)
AI = max .... :!: L(~) IJ., .. ,
where the likelihood function is
Results in Chapter 3 show that
(3)
"N ( X,,-X-)'( x,,-x-)] (2 7T ) -tPN exp [1 -2L...,,~1 Al =
,
N
'
(2-rr)-W l(l/N)AI-,Ne-tpN
where
(4) Sugiura and Nagao (1968) have shown that the likelihood ratio test is biased, but the modified likelihood ratio test based on (5)
10.8 TESTING THAT A COVARIANCE MATRIX IS EQUAL TO A GIVEN MATRIX
439
where S = (l/n)A, is unbiased. Note that
(6)
2 - nlog Ai = tr S ~ 10giSI - p =L/(I, S),
where L/(I, S) is the loss function for estimating 1 by S defined in (2) of Section 7.8. In terms of the characteristic roots of S the criterion (6) is a constant plus p
(7)
p
E(-logDlj-p= E (Ij-Iogl i -I); j=1
;=1
;=1
for each i the minimum of (7) is at I; = 1. Using thp- algebra of the preceding section, we see that given y" ... , YN as observation vectors of p components from N( v, 'IT), the modified likelihood ratio criterion for testing the hypothesis HI: 'IT = 'ITo, where 'ITo is sp.ecified, is
(8) where N
(9)
B=
E (Y,,-Y)(Y,,-Y)'. a o ·1
10.8.2. The Distribution and Moments of the Modified Likelihood Ratio Criterion The null hypothesis HI: I = 1 is the intersection of the null hypothesis of Section 10.7, H: I = (721, and the null hypothesis (72 = 1 given 1=(721. The likelihood ratio criterion for HI given by (3) is the product of (7) of Section 10.7 and tpN
(10)
tr A )' (pN
A + 4nN e _ llr 2 ".
'
which is the likelihood ratio criterion for testing the hypothesis (72 = 1 given 1= (721. The modified criterion Ai is the product of IAI in /(tr A/p)tpn and
(11) these two factors are independent (Lemma 1004.1). The characterization of the distribution of the modified criterion can be obtained from Section
440
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
10.7.3. The quantity tr A has the x2-distribution with np degrees of freedom under the null hypothesis. Instead of obtaining the moments and characteristic function of Ai [defined by (5)] from the preceding characterization, we shall find them by use of the fact that A has the distribution W(:t, n). We shall calculate
Since
(13)
l:t -1 + hll ~("+"h)IAll(,,+nh-p- 1 ) e-l1r(l:-1 +hl)A 2lJ!("+"h)I~}[ tn(l
+ h)]
2lpnhl:tllnhrp[tn(1 +h)j
I1+ h:tll(Hnhlrpctn)
the hth moment of
Ai
is
(14)
Then the characteristic function of - 210g A* is
(15)
,ce-2illog A1 = ,cAr- 2i, 2e)-i pnl
=(
n
l:tl-inl P r[t(n+l-j)-intj 1/-2it:tlln-in,)] r[Hn+l-j)j
10,8
TESTING THAT A COVARIANCE MATRIX IS EQUAL TO A GIVEN MATRIX
When the null hypothesis is true, I
=
441
I, and
(16) 2 ,ce-2iIIOgAT=(~)
-ipnl
n
p
(1_2it)-~p(n-2inl)n j~l
r[ 2,I ( ~+ 1 -)')
. ] -mt ,
rhCn+l-j)]
This characteristic function is the product of p terms such as
( 17)
cfJ( t) = ( 2e ) -inl (l _ 2it) - ,(,," j"rI r[ ~(n + 1 - j) ~ inl] , }
r[Hn+l--])]
n
Thus - 2 log A~ is distributed as the sum of p independent variates, the characteristic function of the jth being (17), Using Stirling's approximation for the gamma function, we have
+ 1 - )') - mt , ]~(n-jl-inl e -[l(n+l-J')-inr][l( ' 2 n e-[1(n+l-nl[±(n - j + 1)]1(n-jl - 1 2' - ~ (1 -( - II) , 1-
(
il (j -- I)
\ ~,(" ,I
~(I1-j+l)(1-2il»)
i)
2j - 1 ) -1111 n(I-2it)
---,--'---:c--:-
x/
As n --> 00, cfJ/t) --> (1 - 2it)- ii, which is the characteristic function of (X 2 with j degrees of freedom), Thus - 2 log Ai is asymptotically distributed as Lr~l xl, which is X 2 with Lr~lj = -b(p + 1) degrees of freedom, The
distribution of Ai can be further expanded [Korin (1968), Davis (1971)] as
(19)
Pr{-2plogAj~z}
=Pr{x!~z}+ p;~2(pr{xl+4~z}-pr{xl~z})+O(W3), where
(20) (21)
2p2 + 3p - 1 p=l- 6N(p+l) , _ p(2p4 + 6 p 3 + p2 - 12Jl - 13) 288( P + 1)
'Y2 -
442
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
Nagarsenker and Pillai (1973b) found exact distributions and tabulated 5% and 1% significant points, as did Davis and Field (1971), for p = 2(1)10 and n = 6(1)30(5)50,60,120. Table B.7 [due to Korin (1968)] gives some 5% and 1% significance points of - 210g Ai for small values of nand p = 2(1)10.
10.8.3. Invariant Tests The null hypothesis H: 1. = I is invariant with respect to transformations X* = QX + v, where Q is an orthogonal matrix. The invariants of the sufficient statistics are the characteristic roots II,"" Ip of S, and the invariants of the parameters are the characteristic roots of 1.. Invariant tests are based on the roots of S; the modified likelihood ratio criterion is one of them. Nagao 0973a) suggested the criterion p
(22)
tn tr(S _1)2 = tn L
(li - 1)2.
;=1
Under the null hypothesis this criterion has a limiting X2-distribution with ~p( p + 1) degrees of freedom. Roy (957), Section 6.4, proposed a test based on the largest and smalles1 characteristic roots 1\ and Ip: Reject the null hypothesis if (23) where
(24) and e is the significance level. Clemm, Krishnaiah, and Waikar (1973) giv, tables of u = 1/1. See also Schuurman and Waikar (1973).
10.8.4. Confidence Bounds for Quadratic Forms The test procedure based on the smallest and largest characteristic roots ca be inverted to give confidence bounds on qudaratic forms in 1.. Suppose n has the distribution WO:, n). Let C be a nonsingular matrix such thl 1. = C' C. Then nS* = nC' - \ SC- I has the distribution W(I, n). Since l; : a' S* a/a' a < Ii for all a, where I; and Ii are the smallest and large characteristic roots of S* (Sections 11.2 and A.2), (25)
Pr { I 5,
a'S*a li'il
} 5, u 't/ a*"O = 1 - e,
where (26)
Pr{/5, I; 5, Ii 5,
u} = 1 - e.
10.8 TESTING THAT A COVARIANCE MATRIX IS EOUAI. TO A GIVEN MATRIX
Let a = Cb. Then a'a (25) is (27)
=
b'C'Cb = b'1b and a'S*a
1 - e = Pr { l:o:;
b'Sb l?fJi
:0:;
'V b
u
=
443
b'C'S*Cb = b'Sb. Thus
*" 0 }
= pr{ b'Sb < b'~b < b'Sb u-"-l
Vb}
v.
Given an observed S, one can assert (28)
'Vb
with confidence 1 - e. If b has 1 in the ith position and D's elsewhere, (28) is Sjjlu :0:; (J"ii :0:; siill. If b has 1 in the ith position, - 1 in the jth position, i j, and D's elsewhere, then (28) is
*"
(29) Manipulation of tnesc inequalities yields (30)
Sij Sii + Sjj ( 1 1) T - --2- 7 - Ii
:0:; (J"ij :0:;
Sij Si; + Sjj ( 1 1) Ii + -2-- 7 - Ii '
i *"j.
We can obtain simultaneously confidence intervals on all elements of I. From (27) we can obtain
e
1 b'Sb l-e=Pr { Ii b'b
(31)
:0:;
n S it ::;;
b'1b
1 b'Sb
:O:;7i'lJ:O:;7 b'b
b'1b 1 . a'Sa Pr { - mlll-,-:O:; -b'b u
= pr{ ~lp
a
:0:;
aa
Ap
:0:;
AI
:0:;
:0:;
'Vb
}
1 a'Sa -l max-,a aa
il1}'
l1
where and lp are the large~t and smallest characteristic roots of Sand 1..1 and Ap are the largest and smallest characteristic roots of I. Then
st (32) is a confidence interval for all characteristic roots of I with confidence at least 1 - e. In Section 11.6 we give tighter bounds on 1..(1) with exact confidence.
444
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
10.9. TESTING THE HYPOTHESIS THAT A MEAN VECTOR AND A COVARIANCE MATRIX ARE EQUAL TO A r.IVEN VECTOR AND MATRIX In Chapter 3 we pointed out that if '11 is known, (y - VO)''I101(y - vo) is suitable for testing
(1)
given
'11 == '11 0 •
Now let us combine HI of Section 10.8 and H 2 , and test
(2)
H:v=vo,
on the basis of a sample Yl' ... ' YN from N( v, '11). Let
(3) where
(4)
C'I1oC' = I.
Then x!> ... ,xN constitutes a sample from N(j..L,:I), and the hypothesis is
(5)
:I = 1.
The likelihood ratio criterion for H2 : j..L = 0, given :I = I, is
(6) The likelihood ratio criterion for H is (by Lemma 10.3.1)
(7)
The likelihood ratio test (rejecting H if A is less than a suitable constant) is unbiased [Srivastava and Khatri (1979), Theorem 10.4.5]. The two factors Al and A2 are independent because Al is a function of A and A2 is a function of i, and A and i are independent. Since
(8)
445
10.9 TESTING MEAN VECfOR AND COVARIANCE MATRIX
the hth moment of A is
under the null hypothesis. Then
(10)
-210g A = -210g Al - 210g A2
has asymptotically the X 2-distrihution with f = p( p + 1)/2 + P degrees of freedom. In fact, an asymptotic expansion of the distribution [Davis (1971)] of -2p log A is
(11)
Pr{ - 2p log A~ z}
= p;{x/ ~z} + ;22 (Pr{X/+4 pN
~z} - Pr{x/ ~z}) + O(N- 3 ),
wh,~re
+ 9p - 11 6N(p + 3) ,
= 1 _ 2p2
(12)
P
(13)
'Y2
=
p(2p4 + 18 p 3 + 49p2 + 36p 288(p - 3)
13)
Nagarsenker and Pillai (1974) used the moments to derive exact distrihutions and tabulated the 5% and 1% significance points for p = 2(1)6 and N = 4(1)20(2)40(5)100. r>;ow let us return to the observations YI"'" YN' Then
a
=trA+NX'i
= tr( B'I1ol) + NO -
VO
)''I10 1 (ji - vo)
and
(15) Theorem 10.9.1. Given the p-component observation vectors YI' ... , y", from N( v, '11), the likelihood ratio en'ten'on for testing the h.lpothesis H: v = vl"
446
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
(16 ) When the null hypothesis is trut', - 2log A is asymptotically distributed as X 2 with iP(p + 1) + P degrees of freedom.
10.10. ADMISSIBILITY OF TESTS
We shall consider some Bayes solutions to the problem of testing the hypothesis
"I[= ... ="I q
(1 )
as in Section 10.2. Under the alternative hypott. esis, let g = 1, ... ,q,
where the p X 'g matrix Cg has density proportional to \1 + Cb C~I- tn" ng = Ng - 1, the 'g-component random vector ylg) has the conditional normal distribution with mean 0 and covariance matrix (l/Ng)[Ir, - C~(Ip + CgC~)-ICg]-1 given Cg, and (C1,y(l»), ... ,(Cq,y(q») are independently distributed. As we shall see, we need to choose suitable integers '1' ... ' 'q. Note that the integral of 11 + CgC~I-1/1, is finite if ng?.p + 'g. Then the numerator of the Bayes ratio is q
(3)
canst
x
x
-x
-x
n f··· f
g~
\
11 + CgC~1 tN,
IN,
·exp {
-"2
ar:\ [x~gl -
l '
(1 + CgC~r Cgylglj
.(1 + CgC~)[ x~g) - (1 + Cgc~r[Cgy(glj} ·II + CgC~I-!n'II-
C~(1 + CKC~rlCg It
. exp{ - !Ngy(gl' [1 - C~(1 + CgC~
r1qgljy
447
10.10 ADMISSIBILITY OF TESTS
= const
n exp {I -"2 E x~g), x~g) f ... f g
q
N
g=1
a=1
}
00
00
_00_00
Under the null hypothesis let
(4) where the p X r matrix C has density proportional to II + CC'I- tn, n = I:~a 1 n g , the r-component vecto~ y(g) has the conditional normal distribution with mean 0 and covariance matrix (l/Ng)[/, - C'Up + CC,)-IC]-I given C, and y(l), ... , y
.(1 + CC')[ x~) - (I + CC') -I C/
·1/ + CC'I- hl / _C'(I + CC') -Iclt
f )]}
448
TESTING HYPOTHESES OF EQUALITY OF COY ARIANCE MATRICES
The Bayes test procedure is to reject the hypothesis if
(6)
=,.
For invariance we want Li~I'g The binding constraint on the choice of '1>"" rq is 'g:-:;; ng - p, g = 1, ... ,q. It is possible in some special cases to choose rl, ... ,rq so that (r 1 , ••• ,rq) is proportional to (NI , ••• , Nq) and hence yield the likelihood ratio test or proportional to (n l , .•. , nq) and hence yield the modified likelihood ratio test, but since r l , ... , rq have to be integers, it may not be possible to choose them in either such way. Next we consider an extension of this approach that involves the choice of numbers tl"'" t q , and t as well as rl, ... ,rq , and r. Suppose 2(p-1)
(7)
10.11
449
ELLIPTICALLY CONTOURED DISTRIBUTIONS
For invariance we want I = L:i~ I 18 , If t l , ••• , Iq are taken so rg + tg = kNg and p - 1 < kNg < Ng - p, g = 1, ... , q, for some k, then (7) is the likelihood ratio test; if rg + tg = kng and p - 1 < kng < ng + 1 - p, g = 1, ... , q, for some k, then (7) is the modified test [i.e., (p -l)/min g Ng < k < 1 - P /ming N,) Theorem 10.10.1.' If 2p < Ng + 1, g = 1, ... , q, then the likelihood ratio test and the modified likelihood ratio test of the null hypothesis (1) are admissible. Now consider the hypothesis
(8) The alternative hypothesis has been treated before. For the null hypothesis let
(9) where the p X r matrix C has the density proportional to 11 + CC' 1- ~(.\' - II and the r-component vector y has the conditional normal distribution with mean 0 and covariance matrix (1/N)[I-C'(I+CC,)-ICr l given C. Then the Bayes procedure is to reject the null hypothesis (8) if
(10) If 2p < Ng + 1, g = 1, ... , q, the prior distribution can be modified as before to obtain the likelihood ratio test and modified likelihood ratio test.
Theorem 10.10.2. If 2p
10.11. ELLIPTICALLY CONTOURED DISTRIBUTIONS 10.11.1. Observations Elliptically Contoured Let x~gl,
a
= 1, ... , Ng , be Ng observations on
X(g)
having the density
(1)
A;
where cS'[(X - v(g»)' I(X_ v(g»)F = t!'R~ < 00, g = 1, ...• q. Note that the same function gO is used for the density in all q populations. Define N, A g'
450
TESTING HYPOTHESES OF EQUALITY OF COY ARIANCE MATRICES
g = 1, .... q, and A by 0) of Section 10.2. Let Sg = O/ng)A g, where ng = Ng - 1, and S = O/n)A, where n = I:~=lng. Since the likelihood ratio criterion Al is invariant under the transformation X lg ) = eXlg) + vl.!:J, under the null hypothesis we can take :II = ... =:I q = I and vii) = ... = v lg ) = O. Then
(2)
- 2log Al
=
-
[g~l N
g
logl ignl - N logl iwl]
=-
[gtl Nglogll + (ign -I) 1- N logII + (iw -I) I]
= -
t~l Ng[tr(ign-l) - !tr(ign -I)2 + Op(Ng- )1 3
-N[tr(i w -I)-!tr(i w -I)2 +O/N- 3
=
~
)J}
q
E Ng[vec(ign-l)]' vec(ign-l) g=l
By Theorem 3.6.2
and n.Sg = Ngign' g = 1, ... , q, are independent. Let N g= kgN, g = 1, ... , q, -> 00. In terms of this asymptotic theory the limiting distribution of vec(SI - I), ... , vec(Sq - I) is the same as the distribution of ytil, .... ylq) of Section 8.8, with :I of Section 8.8 replaced by (K + 1)(11" + Kpl') + K vec lp (vec lp)'. When :I = T, the variance of the limiting distribution of ,fi{(s},g) -1) is
1::% = I k ~ = 1, and let N
3K
+ 2;
the covariance of the limiting distribution of
-
1) and
sir, i '* j, is K + 1; the set W(s\f' I).... , IN (s;;~,> - 1) is indepcndent of the set (si!», i '* j; and the sir, i <j,
";N~ tsJr
- 1), i '* j, is
{N; (Siig ) -
K;
the variance of
arc mutually uncorrelated (as in Section 7.9.1).
10.11
ELLIPTICALLY CONTOURED DISTRIBUTIONS
(4)
-210g A=!
451
q
L
NlYg - Y)'{Y g - Y)
g=1
q
=
td L
Ng{Yg - Y){Yg -
y),
g=1
Let Q be a q X q orthogonal matrix with last column Define
(J NI / N , ... , WN)'.
(5) Then
Wq
= {Ny and q
L
(6)
q-I
NgYgY~ -Njji =
g=1
L
WgW~.
g=1
In these terms q-I
-210g Al =!
(7)
L
W~Wg
+ O/N- 3 ),
g=1
and WI' ... ' wq_ 1 are asymptotically independent, Wg having the covariance matrix of {Ny g; that is, (K+1)(Jp2+K pp )+Kveclp (veclp). Then W~Wg= Ei,i= l(wfl»)2 = Ef= I( wfl»)2 + 2Ei < / W0g»)2. The covariance matrix of wW, ... ,w~~) is 2(K+1)lp+KEE, where £=(1, ... ,1)'. The characteristic roots of this matrix are 2( K + 1) of multiplicity p - 1 and a single root of 2(K+1)+PK. Thus Ef=I(Whg»)2 has the distribution of 2(K+l)X;_I+ g [2(K + 1) + p.d X12. The distribution of 2E;
(8)
-210g AI
When sampling from (1) and the null hypothesis is !me,
:!.. (K + 1) X(~-I)(P-I)(P+2)/2 + [( K + 1) + pK/2] xi-I.
When K=O, -210gAI !!. X(~-I)p(p+l)/2 is in agreement with (12) of Section 10.5. The validity of the distributions derived in Section 10.4 depend on the observations being normally distributed; Theorem 10.11.1 shows that even the asymptotic theory depends on nonnormality.
452
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
The likelihood criteria for testing the null hypothesis (2) of Section 10.3 is the product AI A2 or VI V2. Lemma 10.4.1 states that under normality VI and V 2 (or equivalently AI and A2) are independent. In the elliptically contoured case we want to show that log VI and log V2 are asymptotically independent. Lemma 10.11.1.
10.2 with l:1 = independent.
l:2
Let Al =nlS I and A2 = n2S2 be defined by (2) of Section =/. Then AI(A I +A 2)-1 and Al +A2 are asymptoticaily
(9)
Then
(11)
By application of Lemma 10.11.1 in succession to AI and Al +A 2, to Al +A2 and Al +A2 +A 3, etc., we establish that AIA-I,A2A-I, ... ,AqA-1 are independent of A =AI + ... +Aq.1t follows that VI and V2 are asymptotically independent. Theorem 10.11.2.
(12)
When
l:1
= ... = l:g and ....(1) = ... = ....(g),
-210g AIA2 = -210g Al - 2 log A2 d
->
2 2 + Xp(q-I)· 2 (K + 1) X(q-l)(p-l)(p+2)/2 + [ (K + 1) + pK/21Xq-I
10.11
453
ELLIPTICALLY CONTOURED DISTRIBUTIONS
The hypothesis of sphericity is that I = u 21 (or A = An The criteron is AIA2' where
(13)
The first factor is the criteron for independence of the components of X, and the second is that the variances of the components are equal. For the first we set q = P and Pi = 1 in Theorem 9.10, and for the second we set q = P and p = 1. Thus .
10.11.2. Elliptically Contoured Matrix Distributions Consider the density
q
(15)
n \A
g=1
g
\-Ng/2g[trEA-l(X(8l_V,E' g s NIf, )(X(8l_ V,~,E'.)'] N. c
g=1
In this density (A g , xg ), g = 1, ... , q, is a sufficient set of statistics, and the likelihood ratio criterion is (8) of Section 10.2, the same as for normality [Anderson and Fang (1990b)j. Theorem 10.11.3. Let .r(X) (X(1), ... , X(q») (p X N) such that
be a
vector-valued function
of X
=
for every (v(l), ... , v(q») and
( 17)
f( CX(l), ... , CX(q»)
=
f( X(I), ... , Xli!»
for every nonsinguiar C. Then the distribution of f(X) where X has the arbitral!' density (15) with Al = ... = A q is the same as the distribution off( X) where X has the normal density (15).
454
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
The proof of Theorem 10.11.3 is similar to the proof of Theorem 4.5.4. The theorem implies that the distribution of the criterion VI of (10) of Section 10.2 when the density of X is (15) with A I = .. , = A q is the Same as for normality. Hence the distributions and their asymptotic expansions are those discussed in Sections 10.4 and 10.5. Corollary 10.11.1. that
Let f(X) be a vector-valued function of X (p X N) such
(18)
f(X+VE',y) =f(X)
for every v and (17) holds. Then the distribution of f(X), where X has the arbitrary' density (15) with A I = '" = A q and v(l) = '" = v(q), is the same as the distribution of f(X), where X has the normal density (15).
If follows that the distribution of the criterion A of (7) or V of (11) of Section 10.3 is the same for the density (15) as for X being normally distributed. Let X (p x N) have the dellsity
(19)
I AI -N /2 g [tr A -I (X - v E',y )( X -
V E'N )'] .
Then the likelihood ratio criterion for testing the null hypothesis A = A.I for some A> 0 is (7) of Section 10.7, and its distribution under the null hypothesis is the same as for X being normally distributed. For more detail see Anderson and Fang (1990b) and Fang and Zhang (1990). PROBLEMS 10.1. (Sec. 10.2) Sums of squares and cross-products of deviations from the means of four measurements are given below (from Table 3.4). The populations are Iris versicolor (I), Iris serosa (2), and Iris virginica (3); each sample consists of 50 observations: A = 1
A = 2
(13.0552 4.1740 8.9620 2.7332
( 6.0882
4.8616 0.8014 0.5062
( 19.812' A, =
4.5944 14.8612 2.4056
4.1740 4.H2S0 4.0500 2.0190 4.8616 7.0408 0.5732 0.4556 4.5944 5.0962 3.4976 2.3338
8.9620 4.0500 10.8200 3.5820 0.8014 0.5732 1.4778 0.2974 14.8612 3.4976 14.9248 2.3924
2n32)
2.0\90 3.5820 ' 1.9162
0%') 0.4556 . 0.2974 ' 0.5442
2.4056) 2.3338 2.3924 . 3.6962
,~
455
PROBLEMS
(a) Test the hypothesis :II =:I2 at the 5% significance level. (b) Test the hypothesis :II = :I2 = :I3 at the 5% significance level.
10.2. (Sec. 10.2) (a) Let y(g), g = 1, ... , q, be a set of random vectors each with p components. Suppose tCy(g) =
0,
Let C be an orthogonal matrix of order q such that each element of the last row is
Define q
Z(g)
=
L
g= 1, ... ,q.
cg"y(h),
h~l
Show that g=I, ... ,q-l,
if and only if
(b) Let x~g), a = 1, ... , N, be a random sample from N(IL(g), :I g), g = 1, ... , q. Use the result from (a) to construct a test of the hypothesis
based on a test of independence of Z(q) and the set Z(l), •.• , Z(q-I). Find the exact distribution of the criterion for the case p = 2.
10.3. (Sec. 10.2) Unbiasedness of the modified likelihood ratio test of a} = a{ Show that (14) is unbiased. [Hint: Let G =n I F/n 2 , r=a}/u}, and c i
=
const
f
C2
I I r,nlG,nl-1 (1
+ rG) !(n , -
I
+n ) ,
dG
CI
= const[C'H4 nl - I (1 + H) - 4(n l +n,) dH.
"'I Show that the derivative of the above with respect to r is positive for 0 < r < 1, o for r = 1, and negative for r > 1.]
456
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
xl,
10.4. (Sec. 10.2) Prove that the limiting distribution of (19) is where f= tp(p + lXq - 1). [Hint: Let :I = I. Show that the limiting distribution of (19) is the limiting distribution of
where
S(g)
=
(s}'»),
S
=
(s,), and the -{r/;(s}t> - 8ij)' i ::;,j, are independent in
the limiting distribution, the limiting distribution of
-{r/; (sf?) -
1) is N(0,2),
and the limiting distribution of -{r/;sf,>, i <j, is NCO, 1).] 10.5. (Sec. 10.4) Prove (15) by integration of Wishart densities. [Hint: .cvt = .cni"IAgll-n.IAI- t" can be written as the integral of a constant times IAI- ,nnz~ ,w(ARI:I, ng + hn g). Integration over LZ=I Ag = A gives a constant times w(AI:I, n).] 10.6. (Sec. 10.4) Prove (16) by integration of Wishart and normal den~ities. [Hint: L~_, Nii(K) - i)(i(K) - i)' is distributed as Ll: IYf Yf' Use the hint of Prob· lem 10.5.] 10.7. (Sec. 10.6) Let x\"), ... , x}J') be observations from N( ....(·). I.), let A.= Dx~·) -i(v»)(x~:) -i(·»)'.
I' =
1.2. and
(a) Prove that the likelihood ratio test for H::II =:I2 is equivalent to rejecting H if
(b) Let dr, d~, .. . , d; be the roots of l:I j
-
A:I 21 = 0, and let
D=
Show that T is distributed as IB,I·IB2 1/IB, +B212. where B, is dis· tributed according to WeD". N - 1) and B2 is distributed according to 2 W(J,N- 1). Show that T is distributed as IDC,DI'IC2 1/IDCI D+C21 , where C, is distributed according to W(l, N - 1). 10.8. (Sec. 10.6) For p
=
2 show
Pr{V, ::;,u} =IQ(n,-1,n 2 -1)
+ B-' (nj - 1, n 2 - l)u("I-t n,-2)/n fb x- 2n ,/n(1_ XI) -nl/n dx l a
457
PROBLEMS
where a < b are the two roots of x~l(1 - Xl)'" = u::; n~IIl~' /n". [Hint: This foJlows from integrating the density defined by (8).] 10.9.
(Sec. 10.6)
For p = 2 and
III
= n~ = m, say, show
=2Ia(m-1,m-l)+2B-l(m-l,m-l)ul-ll/mllog
1+
Iwhere a =
W- h - 4u
l
/
m
11- 4[;1(". fI"=- 4 1 0• , m
].
10.10. (Sec. 10.7) Find the distribution of W for p = 2 under the null hypothesis Cal directly from the distribution of A and (b) from the distribution of the characteristic roots (Chapter 13). 10.11. (Sec. 10.7) Let x[o"" X N be a sample from N(fJ., I} What is the likelihood ratio criterion for testing the hypothesis fJ. = kfJ.o, I = kZI II • where I-'-n and III are specified and k is unspecified? 10.12. (Sec. 10.7) Let xli), ... , x~~ be a sample from NCfJ.(1), I [). and xF' ..... x~' be a sample from N(fJ.(2), I 2 ). What is the likelihood ratio criterion for testing the hypothesis that I I = k 2);.2' where k is unspecified? What is the likelihood ratio criterion for testing the hypothesis that 1-'-(1) = kl-'-(~l and I I = k~I~. where k is unspecified? 10.13. (Sec. 10.7) Let xa of p components, a = 1. ... ,N, be observations from N(fJ., I). We define the following hypotheses:
H 2 :fJ.=0, In each case k 2 is unspecified, but Io is specified. Find the likelihood ratil) criterion A2 for testing Hz. Give thc asymptotic distrihution of -:! log'\~ under Hz. Obtain the exact distribution of a suitahle monotonic function of A, under H2 • 10.14. (Sec. 10.7) Find the likelihood ratio criterion A for testing H of Problem 10.13 (given XI"'" x N ). What is the asymptotic distribution of - ~ log ,\ umkr
Ii? 10.15. (Sec. 10.7) Show that A = AIA2' where A is defined in Problem 10.14. A~ j, defined in Problem 10.13, and Al is the likelihood ratio criterion for HI in Problem 10.13. Are Al and Az independently distributed under H? Prove your answer.
458
TESTING HYPOTHESES OF EQUALITY OF COVARIANCE MATRICES
10.16. (Sec. 10.7) Verify that trB'l'ol has the X2-di~;tribution with p(N-1) degrees of freedom. 10.17. tSec. 10.7.0 Admissibility of sphericilY lest. Prov, that the likelihood ratio test of sphericity is admissible. [Hint: Under the null hypothesis let :l = [1/(1 + 1)')]1. and let 7) have the density (1 + 7)2)- ~"P(7)2)P- ~.l 10.18. (Sec. 10.10.1)
Show that fer r?p
r .,. r It. x,.ril~lll _-x
-x i=1
+
t. Xi.ril-~n ndr < j
i=l
00
1=1
if 2p - 1 < I + r + p < n + 1. [Hinl: IAI /11 +AI $; 1 if A is positive semidefinite. Also, IL;~I.rj.ril has the distribution of X}X,2_ 1 ••• X/-p+l if XI""'X, are independently distributed according to N(O, n.] 10.19. (Sec. 1O.IlU)
Show
where C is p x r. [Hint: CC' has the dbtribution W(A proportional to e - ~Ir C".4C.] 10.20. (Sec. 10.10.1)
I,
r) if C has a density
Using Problem 10.18, complete the proof of Theorem 10.10.1.
CHAPTER 11
Principal Components
11.1. INTRODUCTION Principal components are linear combinations of random or statistical variables which have special properties in terms of variances. For example, the first principal component is the normalized linear combination (the sum of squares of the coefficients being one) with maximum variance. In effect, transforming the original v,~ctor variable to the vector of principal components amounts to a rotation of coordinate axes to a new coordinate system that has inherent statistical proper~ies. This choosing of a coordinate system is to be contrasted with the many problems treated previously where the coordinate system is irrelevant. The principal components turn out to be the characteristic vectors of ~he covariance matrix. Thus the study of principal components can be con~idered as putting into statistical terms the usual developments of characteristic roots and vectors (for positive semidefinite matrices). From the point of view of statistical theory, the set of principal components yields a convenient set of coordinates, and the accompanying variances of the components characterize their statistical properties. In statistical practice, the method of principal components is used to find the linear combinations with large variance. In many exploratory studies the number of variables under consideration is too large to handle. Since it is the deviations in these studies that are of interest, a way of reducing the number of variables to be treated is to discard the linear combinations which have small variances and study only those with large variances. For example, a physical anthropologist may make dozens of measurements of lengths and breadths of
An Introduction to Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0-471-36091-0 Copyright © 2003 John Wiley & Sons, Inc.
459
460
PRINCIPAL COMPONENTS
each of a number of individuals, such measurements as ear length, ear breadth, facial length, facial breadth, and so forth. He may be interested in describing and analyzing how individuals differ in these kinds of physiological characteristics. Eventually he will want to explain these differences, but first he wants to know what measurements or combinations of measurements show considerable variation; that is, which should have further study. The principal components give a new set of linearly combined measurements. It may be that most of the variation from individual to individual resides in three linear combinations; then the anthropologist can direct his study to these three quantities; the other linear combinations vary so little from one person to the next that study of them will tell little of individual variation. Hotelling (1933), who developed many of these ideas, gave a rather thorough discussion. In Section 11.2 we define principal components in the population to have the prorerties described above; they define all orthogonal transfonnation to a diagonal covariance matrix. The maximum likelihood estimators have similar properties in the sample (Section 11.3). A brief discussion of computation is gIVen in Section 11.4, and a numerical example is carried out in Section 11.5. Asymptotic distributions of the coefficients of .the sample principal components and the sample variances are derived and applied to obtain large-sample tests and confidence interval~ for individual parameters (Section 11.6); exact confidence bounds are found for the characteristic roots of a covariance matrix. In Section 11.7 we consider other tests of hypotheses about these roots.
11.2. DEFINITION OF PRINCIPAL COMPONENTS IN THE POPULATION Suppose the random vector X of p components has the covariance matrix :I. Since we shall be interested only in variances and covariances in this chapter, we shall assume that the mean vector is O. Moreover, in developing the ideas and algebra here, the actual distribution of X is irrelevant except for the covariance matrix; however, if X is nonnally distributed, more meaning ca~ be given to the principal components. In the following treatment we shall not m.e the usual theory of characteristic roots and vectors; as a matter of fact, that theory will be derived implicitly. The treatment will include the cases where :I is singular (Le., positive semidefinite) and where :I has multiple roots. Let Il be a p-component column vector such that Il'Il = 1. The variance of Il'X is
(1)
11.2
461
DEFINITION OF PRINCIPAL COMPONENTS IN THE POPULATION
To determine the normalized linear combination \3' X with mc.ximum variance, we must find a vector Il satisfying \3'1l = 1 which maximizes (n Let
(2)
>
= \3' 'I Il - A( \3'1l- 1) =
1: f3i
£Tjj
I,}
f3j
-
A( E 13/ -
1),
,
where A is a Lagrange multiplier. The vector of partial derivatives (clef> / is
ilf3j)
(3) Theorem A.4.3 of the Appendix). Since \3''I1l and everywhere in a region containing \3'1l = 1, a vector must satisfy the expression (3) set equal to 0; that is ~by
(4)
(~-
AI)1l = o.
In order to get a solution of (4) with in other words, A must satisfy
(5)
\3'1l have derivatives Il maximizing \3'~1l
\3'1l = 1 we must have 'I - AI singular;
I'I _. AIl = o.
The function I'I - All is a polynomial in Aof degree p. Therefore (5) has p roots; let these be Al ~ A2 ~ ... ~ Ap. [Il' complex conjugate in (6) proves A rea!.] If we multiply (4) on the left by \3', we obtain
(6) This shows that if Il satisfies (4) (and Il'Il = 1), then the variance of \3' X [given by (1)] is A. Thus for the maximum variance we should use in (4) the largest root AI' Let Il(ll be a normalized solution of ('I - AII)Il = O. Then UI == 1l(1), X is a normalized linear combination with maximum variance. [If 'I - All is of rank p - 1, then there is only one solution to ('I - AJ)1l = 0 and \3'1l = 1.] Now let us find a normalized combination \3' X that has maximum variance of all linear combinations uncorrelated with U I • Lack of correlation means
(7) since :rp(l) = AIIl(!). Thus jl'X is orthogonal to U in both thc statistkal sense (of lack of correlation) and the geometric sense (of the inner product of the vectors p and p(1) being zero). (That is, AIIl'P(I) = 0 only if \3'Il(l) = 0 when Al *- 0, and Al *- 0 if 'I *- 0; the case of 'I = 0 is trivial and is nOI trL'ated.)
462
PRINCIPAL COMPONENTS
We now want to maximize ( 8) where A and vI are Lagrange multiplers. The vector of partial derivatives is (9) and we set this equal to O. From (9) we obtain by multiplying on the left by 1l(1) ,
by (7). Therefore, VI = 0 and Il must satisfy (4), and therefore A must satisfy (5). Let \2) be the maximum of AI"'" Ap such that there is a vector Il satisfying (I - \~) nil = 0, Wil = 1, and (7); call this vector 1l(2) and the corresponding linear combination V 2 = 1l(2) , X. (It will be shown eventually that A(2) = A1 · We define \1) = AI') This procedure is continued; at the (r + 1)st step, we want to find a vectorIl such that Il' X has maximum variance of all normalized linear combinations which are uncorrelated with VI"'" V" that is, such that (11 )
i
= 1, ... , r.
We want to maximize r
(12)
4>'+1
=
E vjIl'IIl(l),
WIIl- A(Il'Il- 1) - 2
i=\
where A and derivatives is
VI"'"
v, are Lagrange multipliers. The vector of partial
a~
(13)
r
~ all = 2Ia.. - 2Aa.. - 2 "Ia(i) t... v, .. , j=1
and we set this equal to O. Multiplying (13) on the left by Il(i)', we obtain
(14) If \J) *- 0, this gives - 2 vj\j) = 0 and Vj = O. If A(j) = 0, then I p(j) = \j)p(j) = 0 and the jth term in the sum in (13) vanishes Thus Il must satisfy (4), and therefore A must satisfy (5).
11.2 DEFINITION OF PRINCIPAL COMPONENTS IN THE POPULATION
463
Let \,+ I) be the maximum of AI"'" Ap such that there is a vector Il satisfying (I - A('+I)1)1l = 0, Il'Il' = 1, and (11); caU this vector 1l(,+I), and the corresponding linear combination u,+ I = Il('+ I), X. If \'+ I) = 0 and AU) = 0, j *- r + 1, then 1l(j)'"I Il('+ I) = 0 does not imply 1l(j)'Il('+ I) = O. However, Il('+ I) can be replaced by a linear combination .of Il('+ I) and the pU)'s with A(j)'S being 0, so that the new Il('+ I) is orthogonal to aU Il(j), j = 1, ... , r. This procedure is carried on until at the (m + 1)st stage one cannot find a vector Il satisfying Il'Il = 1, (4), and (11). Either m =p or m
(15)
\1)
0
0
0
A(2)
0
0
0
A(p)
A=
The equations "I1l(') = \,)Il(') can be written in matrix form as
(16)
"IP =
PA,
and the equations 1l{')'Il(') = 1 and P(')'Il{S) = 0, r *- s, can be written as
(17)
P'P=I.
From (16) and (17) we obtain
(18)
P'"IP=A.
464
PRINCIPAL COMPONENTS
From the fact that
(19)
II -
All = 11=1' I ·1 I
-
All '11=11
= II=I'II=I- AI=I'I=II = IA - All
= n(AU) - A) we see that the roots of (19) are the diagonal clements of A; that is, A(1) = AI' A(2) = A2,··., A(p) = Ap. We have proved the following theorem: Theorem 11.2.1. Let the p-component random vector X have C X = 0 and CXX' = I. Then there exists an orthogonal linear transformation
(20) such that the covariance matrix of U is CUU' = A and
(21)
A=
o o
where Al ~ A2 ~ ... ~ Ap ~ 0 are the roots of (5). The rth column of 1=1, per), satisfies (I - ArI)p(r) = O. The rth component of U, Vr = per), X, has maximum variance of all normalized linear combinations uncorrelated with VI' ...• Vr_ 1 • The vector U is defined as the vector of principal components of X. It will be observed that we have proved Theorem A.2.1 of Appendix A for B positive semidefinite, and indeed, the proof holds for any symmetric B. It might be noted that once the transformation to VI"'" Vp has been made, it is obvious that VI is the normalized linear combination with maximum variance, for if V* = LC;U;, where LC; = 1 (V* also. being a normalized linear corr.bination of the X's), then Var(U*) = Lc;A; = AI + Ll=2 c;( A; - AI) (since c~ = 1 - Lfcf), which is clearly maximum for c; = 0, i = 2, ... , p. Similarly, V 2 is the normalized linear combination uncorrelated with VI which has maximum variance (V* = LC;U; being uncorreIated with VI implying c I = 0); in turn the maximal properties of V3 , ••. , Vp are verified. Some other consequences can be derived. Corollary 11.2.1. Suppose Ar+1 = .. , = Ar+m = I' (i.e., v is a root of multiplicity in); then I - vI is of rank p - m. Furthermore 1=1* = (p(r+ I) p(r+m) is uniquely determined except for multiplication on the right by an orthogonal matrix.
465
11.2 DEFINITION OF PRINCIPAL COMPONENTS IN THE POPULATION
Proof. From the derivation of the theorem we have (l: - vJ)pli) = 0, i = r + 1, ... , r + m; that is, p(r+ I), ... , b(r+m) are m linearly independent solutions of (l: - vI)p = o. To show that there cannot be another linearly
independent solution, take I:f~lx;p(i',. where the x; are scalars. If it is a solution, we have vI:x;W;) = l:(f-x;P(i») = I:x;l:p(i) = I:x;A;p(i). Since I'X; = A;x;, we must have x; = 0 unless i = r + 1, ... , r + m. Thus the rank is p - m. If p* is one set of solutions to (l: - vl)P = 0, then any other set of solutions are linear combinations of the others, that is, are j:l*A for A nonsingular. However, the orthogonality conditions j:l* 'p* = I applied to the linear combinations give 1= (P*A)'(j:l*A) =A'P* 'P*A =A'A, and thus A must be orthogonal. • Theorem 11.2;2. An orthogonal transformation V = cx of a random vector X leaves invariant the generalized variance and the sum of the variances of the components. Proof Let tS'X= 0 and tS'XX' generalized variance of V is
(22)
=
l:. Then tS'V= 0 and cG'W'
=
Cl:C'. The
I C l: C 'I = I CI . Il:l . I C' I = Il: I . I CC' I = Il:l •
which is the generalized variance of X. The sum of the variances of the components of V is
(23)
L tS'v? = tr( Cl:C') = tr( l:C'C) = tr( l:l) = tr l: = L cG' X?
•
Corollary 11.2.2. The generalized variance of rhe vector of principal components is the generalized variance of the original vector, and the sum of the variances of the principal components is the sum of the variances of the original variates. Another approach to the above theory can be based on the surfaces of constant density of the normal distribution with mean vector 0 and covariance matrix l: (nonsingular). The density is
(24) and surfaces of constant density are ellipsoids
(25) A principal axis of this ellipsoid is defined as the line from - y to y, where y is a pOint on the ellipsoid where the squared distance x' x has a stationary
-'66
PRINCIPAL COMPONENTS
point. Using the method of Lagrange multipliers, we determine the stationary points by considering
(26) where A is a Lagrange multiplier. We differen~iate '" with respect to the components of x, and the derivatives set equal to 0 are (27)
a", ax
= 2x - 2Al:- 1x = 0
'
or
(28) Multiplication by I gives
(29)
l:x= Ax.
This equation is the same as (4) and the same algebra can be developed. Thus the vectors 13(11, ... , p(P) give the principal axis of the ellipsoid. The transformation u = j:l'x is a rotation of the coordinate axes so that the new axes are in the direction of the principal axes of the ellipsoid. In the new coordinates the ellipsoid is (30)
2M.
Thus the length of the ith principal axis is A third approach to the same results is in terms of planes of closest fit [Pearson (1901)]. Consider a plane through the origin, at' x = 0, where at' at = 1. The distance of a point x from this plane is at' x. Let us find the coefficients of a plane such that the expected distance squared of a random point X from the plane is a minimum, where ex = 0 and exX' = l:. Thus we wish to minimize .c'(at'X)2=.c'at'xx.'at=a'l:a, subject to the restriction a' 0: = 1. Comparison with the first approach immediately shows that the solution is at = p(P). Analysis into principal components is most suitable when all the components of X are measured in the same units. If they are not measured in the same units, ~he rationale of maximizing p'l:P relative to P'P is questionable; in fact, the analysis will depend on the various units of measurement. Suppose ~ is a diagonal matrix, and let Y = ~ X. For .example, one component of X may be measured in inches .and the corresponding component of Y may be measured in feet; another component of X may be in pounds and the
11.3
467
MAXIMUM LIKELIHOOD ESTIMATORS
corresponding one of Y in ounces. The covariance matrix of Y is CYY' = Ca XX' a = 4 I a = '1", say. Then analysis of Y into principal components involves maximizing C( 'Y ' y)2 = "'I' 'I" 'Y relative to 'Y' 'Y and leads to the equation 0 = ('I" - vI)'Y = (al:a- vI)'Y, where v must satisfy 1'1"- vII = O. Multiplication on the left by a-I gives
(31) Let a'Y = Ot; that is, 'Y'Y = 'Y ' a x = Ot' X. Then (31) results from maximizing C(Ot' X)2 = Ot'IOt relative to Ot' a- 2 Ot. This last quadratic form is a weighted sum of squares, the weights being the diagonal elements of a- 2 • It might be noted that if a- 2 is taken to be the matrix
a- 2 =
(32)
(Tn
0
0
(T22
0 0
0
0
(Tpp
then 'I" is the matrix of correlations.
11.3. MAXIMUM LIKELIHOOD ESTIMATORS OF THE PRINCIPAL COMPONENTS AND THEIR VARIANCES A primary problem of statistical inference in principal component analysis is to estimate the vectors 13(1), ... , p(p) and the scalars AI, ... ' Ap. We apply the algebra of the preceding section to an estimate of the covariance matrix. Theorem 11.3.1. Let XI, ••• ,XN be N (>p) observations from N(Jl.,I), where l: is a matrix with p different characteristic roots. Then a set of maximum likelihood estimators of AI' ... ' Ap and pm, ... , p(p) defined in Theorem 11.2.1 consists of the roots k 1 > ... > k p of
li-kIl =0
(1)
and a set of corresponding vectors b(l), .•. , b(p) satisfying
(i -
(2)
kJ)b(i) = 0,
(3) where
i
is the maximum likelihood estimate of l:.
468
PRINCIPAL COMPONENTS
Proof When the roots of II - All = 0 are different, each vector fl(i) is uniquely defined except that fl!i) can be replaced by -fl(i). If we require that the first nonzero component of fl(i) be positive, then fl(i) is uniquely defined, and p., A, I=l is a single-valued function of p., I. By Corollary 3.2.1, the set of maximum likelihood estimates of p., A, I=l is the same function of iL, i. This function is defined by (1), (2), and (3) with the corresponding restriction that the first nonzero component of b(i) must be positive. [It can be shown that if III 0, the probability is 1 that the roots of (1) are different, hecause the conditions on i for the roots to have multiplicities higher than 1 determine a region in the space of i of dimensionality less than !P(p + 1); see Okamoto (1973).] From (18) of Section 11.2 we see that
*"
(4) and by the same algebra
(5) Replacing b(i) by -b(i) clearly does not change '[,kib(i)b(i)'. Since the likelihood function depends only on i (see Section 3.2), the maximum of the likelihood function is attained by taking any set of solutions of (2) and (3) .
•
It is possible to assume explicitly arbitrary multiplicities of roots of I. If these mUltiplicities are not all unity, the maximum likelihood estimates are not defined as in Theorem 11.3.1. [See Anderson (t 963a).l As an example suppose that we aSSume that the equation II - All = 0 has one root of multiplicity p. Let this root be A\. Then by Corollary 11.2.1, 'Z - All is of rank 0; that is, I - All = 0 or I = All. If X is distributed according to N(p., I) = Nlp., A\ J), the components of X are independently distributed with variance AI' Thus the maximum likelihood esdmator of AI is
(6) and i = All, and ~ can be any orthogonal matrix. It might be pointed out that in Section 10.7 we considered a test of the hypothesis that I = All (with AI unspecified), that is, the hypothesis is that I has one characteristic root of multiplicity p. In most applications of principal component ana lysis it can be assumed that the roots of I are different. It might also he pOinted out that in some uses of this method the algebra is applied to the matrix of correlation
11.4
MAXIMUM LIKELIHOOD ESTIMATES OF THE PRINCIPLE COMPONENTS
469
coefficients rather than to the covariance matrix. In general this leads to dilferent roots and vectors.
11.4. COMPUTATION OF THE MAXIMUM LIKELIHOOD ESTIMATES OF THE PRINCIPAL COMPONENTS There are several ways of computing the characteristic roots and characteristic vectors (principal components) of a matrix l: or i. We shall indicate some of them. One method for small p involves expanding the determinantal equation
(1)
0= II - All
2.nd solving the resulting pth-degree equation in A (e.g., by Newton's method or the secant method) for the roots Al > A2 > ... > Ap. Then I - AJ is of rank p - 1, and a solution of (I - Aj I)I3(i) = 0 can be obtained by taking f3t) as the cofactor of the element in the first (or any other fixed) column and jth row of I - AJ. The second method iterates using the equation for a characteristic root and the corresponding characteristic vector
(2)
Ix= Ax,
where we have written the equation for the population. Let not orthogonal to the first characteristic vector, and define
(3)
be any vector
x(O)
i
=
0, 1.2, ....
It can be shown (Problem 11.12) that
(4) The rate of convergence depends on the ratio A2/ AI; the closer this ratio is to 1, the slower the convergence. To find the second root and vector define
(5) Then
(6)
Izl3(i) = II3(i) - AI I3(1)I3(1)'I3(j)
= II3(i) = AjW i )
470
PRINCIPAL CO MPONENTS
if i"* 1. and (7) Thus '\2 is the largest root of 12 and 13(2) is the corresponding vector. The iteration process is now applied to 12 to find A2 and 13(2). Defining 13 = 12 - '\213(2'13(2)', we can find A3 and 13(3), and so forth. There are several ways in which the labor of the iteration procedure may be reduced. One is to raise I to a power before proceeding with the iteration. Thus one can use I 2, defining
(8)
i =0, 1,2, ....
This procedure will give twice as rapid convergence as the use of (3). Using I· = 12 I 2 will lead to convergence four times as rapid, and so on. It should be noted that since 12 is symmetric, there are only pep + 1)/2 elements to be found. Efficient computation, however, uses other methods. One method is the QR or QL algorithm. Let Io = I. Define recursively the orthogonal Qj and lower triangular L, by I,=Q;L; and I;+l =L;Q; (=Q;IiQ), i= 1,2, .... (The Gram-Schmidt orthogonalization is a way of finding Qi and Lj; the QR method replaces a lower triangular matrix L by an upper triangular matrix R.) If the characteristic roots of I are distinct, lim,_ coli + 1 = A*, where A* is the diagonal matrix with the roots usually ordered in ascending order. The characteristic vectors are the columns of Iim;_cc Q;Q;-l ... Q'l (which is com-· puted recursively). A more efficient algorithm (for the symmetric I) uses a sequence of Householder transformations to carry I to tridiagonal form. A Householder m£loir is H = I - 2 a a' where a' a = 1. Such a matrix is orthogonal and symmetric. A Householder transformation of the symmetric matrix I is HIH. It is symmetric and has the same characteristic roots as I; its characteristic vectors are H times those of I. A tridiagonal matrix is one with all entries 0 except on the main diagonal, the first superdiagonal, and the first subdiagonal. A sequence of p- 2 Householder transformations carries the symmetric I to tridiagonal fOlm. (The first one inserts O's into the last p - 2 entries of the first column and row of HIH, etc. See Problem 11.13.)
471
11.5 AN EXAMPLE
The QL method is applied to the tridiagonal form. At the ith step let the tridiagonal matrix be TJi); let lj(i) be a block-diagonal matrix (Givens matrix)
p'O
(9)
J
~
o
0 cos OJ
-sin OJ
0
sin OJ
cos OJ
0
0
o
[:
where cos OJ is the jth and j + 1st diagonal element; and let 1j(i) = P~~j1j~)I' j = 1, ... , p - 1. Here OJ is chosen so that the element in position j, j + 1 in 1j is O. Then p(i) = pfi)p~i) ••. P~~ 1 is orthogonal and p(i)TJi) = R(i) is lower triangular. Then TJi + 1) = R(i) p(i)' (= p(i)TJi)p(i)') is symmetric and tridiagonal. It converges to A* (if the roots are all different). For more details see Chapters 11/2 and 11/3 of Wilkinson and Reinsch (1971), Chapt.!r 5 of Wilkinson (1965), and Chapters ), 7, and 8 of Golub and Van Loan (1989). A sequence of one-sided Householder transformation (H 1:) can carry 1: to R (upper triangular), thus effecting the QR decomposition.
11.5. AN EXAMPLE In Table 3.4 we presented three samples of observations on varieties of iris [Fisher (1936)]; as an example of principal component analysis we use one of those samples, namely Iris versicolor. There are 50 observations (N = 50, n = N - 1 = 49). Each observation consists of four measurements on a plant: Xl is sepal length, X 2 is sepal width, X3 is petal length, and X 4 is petal width. The observed sums of squares and cross products of deviations from means are
(1)
13.0552 4.1740 A= a~l (x,,-x)(x,,-x) = 8.9620 ( . 2.7332 50
_
_ ,
4.1740 4.8250 4.0500 2.0190
8.9620 4.0500 10.8200 3.5820
2.7332) 2.0190 3.5820 ' 1.9162
and an estimate of I is
(2)
S = ...!.A = 49
0.266433 0.085184 ( 0.182899 0.055780
0.085184 0.098469 0.082653 0.041204
0.182899 0.082653 0.220816 0.073102
0.0557801 0.041204 0.073102 . 0.039106
472
PRINCIPAL COMPONENTS
We use the iterative procedure to find the first principal component, by computing in ·turn z(j) = SZ(j-l). As an· initial approximation, we use z(O)' = (1,0, 1,0). It is not necessary to normalize the vector at each iteration; but to compare successive vectors, we compute zij) /z}i -I) = ,;(j), each of which is an approximation to iI' the largest root of S. After seven iterations, ,P) agree to within two units in the fifth decimal place (fifth significant figure). This vector is normalized, and S is applied to the normalized vector. The ratios, ,;(8), agree to within two units in the sixth place; the value of i l is (nearly accurate to the sixth place) i l = 0.487875. The normalized eighth iterated vector is our estimate of p(1), namely,
b(l) =
(3)
0.6867244] 0.3053463 ( 0.6236628 . 0.2149837
This vector agrees with the normalized seventh iterate to about one unit in the sixth place. It should be pointed out that i l and b(l) have to be calculated more accurately than i2 and b(2), and so forth. The trace of S is 0.624824, which is the sum of the roots. Thus i l is more than three times the sum of the other roots. We next compute
(4)
S2 =S -i1b(1)b(I),
0.0363559 _ -0.0171179 - ( -0.0260502 -0.0162472
-0.0171179 0.0529813 -0.0102546 0.0091777
-fl.0260502 -0.0102546 0.0310544 0.0076890
- 0.016 2472] 0.0091777 0.0076890 ' 0.0165574
and iterate z(j) = S2 Z(j-I), using z(O), = (0, 1,0,0). (In the actual computation S2 was multiplied by 10 and the first row and column were mUltiplied by -1.) In this case the iteration does not proceed as rapidly; as will be seen, the ratio of i2 to i3 is approximately 1.32. On the last iteration, the ratios agree to within four units in the fifth significant figure. We obtain i2 = 0.072 382 8 and
(5)
-0.669033] 0.567484 0.343309 . 0.335307
The third principal component is found from S3 = S2 -i2 b(2;b(2)', and the fourth from S4 = S3 -i3 b(3)b(3),.
11.6
473
STATISTICAL INFERENCE
The results may be summarized as follows:
(7)
B=
0.6867 0.3053 ( 0.6237 0.2150
-0.6690 0.5675 0.3433 0.3353
-0.2651 -0.7296 0.6272 0.0637
0.1023 -0.2289 -0.3160 0.9150
The sum of the four roots is r.;~ Ii; = 0.6249, compared with the trace of the sample covariance matrix, tr S = 0.624 824. The first accounts for 78CfC of the total variance in the four measurements; the last accounts for a little more than 1%. In fact, the variance of 0.7xI + 0.3.1.'2 + 0.6.1.') + 0.2x~ (an approximation to the first principal component) is 0.478, which is almost 77CfC of the total variance. If one is interested in studying the variations in conditions that lead to variations of (XI' x 2 , x 3 , x 4 ), one can look for variations in conditions that lead to variations of 0.7xI + 0.3X2 + 0.6x, + 0.2x~. It is not very important if the other variations in (XI' x 2 , x 3 , x 4 ) are neglected in exploratory investigations.
11.6. STATISTICAL INFERENCE 11.6.1. Asymptotic Distributioli1s In Section 13.3 we shall derive the exact distribution of the sample characteristic roots and vectors when the population covariance matrix is I or proportional to I, that is, in the case of all population roots equal. The exact distribution of roots and vectors when the population roots are nor all equal involves a multiply infinite series of zonal polynomials; that development is beyond the scope of this hook. [See Muirhead (19HZ).] We derive the asymptotic distrihution of the 'roots and vectors when the population roots are all different (Theorem 13.5.1) and also when one root is multiple (Theorem 13.5.2). Since it can usually be assumed that the population roots are different unless there is information to the contrary, we summarize here Theorem 13.5.1. As earlier, let the characteristic roots of l: be Al > '" > AI' and the corresponding characteristic vectors be 13(1), ... , WP). normalized so W')'W il = 1 and satisfying (31i ~ 0, i = 1, ... , p. Let the roots and vectors of 51 be 11 >", > Ip and b(1), ... ,b(P) normalized so b(i)'b(i) = 1 and satisfying b li ~ O. i = 1, ... , p. Let d i = Iii (Ii - A) and g(i) = {/; (b(i) - 13(i)). i = 1, .... p. Then in the limiting normal distribution the sets d I' ... , d P and gil>, ... , g' P J are independent and d l , ••• , d p are mutually independent. The element d, has
PRINCIPAL COMPONENTS
the limiting distribution N(O, V .. limiting distribution are
n. The covariances of
g(I), ... , g(p)
in the
(1)
(2) See Theorem 13.5.1. In making inferences about a single ordered root, one treats Ii as approximately normal with mean A; and variance 2),}/n. Since Ii is a consistent estimate of Ai' the limiting distribution of
In I, -
A; n fil;
(3)
is N(O, 1). A two-tailed test of the hypothesis A; = Ai has the (asymptotic) acceptance region
rn 1- 1..
0
(4)
-z(s) ~
V 2-'-0-' ~z(s), A;
where the value of the N(O,1) distribution beyond z(s) is !s. The interval (4) can be inverted to give a confidence interval for A; with confidence 1 - s: [.
(5)
[.
--==,'=--- < A· <
1 + ';2/nz(s) -
=='=---
,- 1- \f2/nz(s)'
Note that the confidence coefficient should be taken large enough so ';2/n z(d < 1. Alternatively, one can use the fact that the limiting distribution of In (Jog I; - log A) is N(O, 2) by Theorem 4.2.3. Inference about components of a vector 13(;) can be based on treating b(l) as being approximately normal with mean 13(;) and (singular) covariance matrix lin times (D. 11.6.2. Confidence Region for a Characteristic Vector We use the asymptotic distribution of the sample characteristic vectors to obtain a large-sample confidence region for the ith characteristic vector of I (Anderson (1963a)]. The covariance matrix (1) can be written
( 6)
475
11.6 STATISTICAL INFERENCE
where Ai is the p Xp diagonal matrix with 0 as the ith diagonal element and ,; AiAj /(A i - Aj) as the jth diagonal element, j i; Ai is the (p - 1) X (p-1) diagonal matrix obtained from Ai by deleting the ith row and column; and is the p X (p - 1) matrix formed by deleting the ith column from j:l. Then h(i) = Ai -I '.[,1 (b(i) - 13(i» has a limiting normal distribution with mean 0 and covariance matrix
*"
P7
P7
(7) ~nd
(8) has a limiting X2-distribution with p - 1 degrees of freedom. The matrix of the quadradc form in .[,1 (b(i) - 13(i» is
because j:lA-Ij:l' = 1:-1, j:lj:l' =1, and j:lAj:l' = 1:. Then (8) is (10)
n(b(i) -13(i»),[ Ai1:- 1 - 2I + (1/ A;) 1: ] (b(i) -13(i»)
= nb(i), [Ai1: -I
-
2I + (1/ A;) 1: ]b(i)
= n[ Aib(i)'rlb(i) + (1/ Ai )b(i)'1:b(i) - 2], because 13(i)' is a characteristic vector of 1: with root Ai' and of 1: -I with root 1/ Ai. On the left-hand side of (10) we can replace l: and Ai by the consistent estimators S and Ii to obtain (11)
n(b(i) -13(i»)'[l iS-1 - 2I + (1/li)S](b(i) -13(i»)
= n[lil3(i)'S-II3(i) + (1/l i )I3(i)'SI3(i) -
2],
which has a limiting X 2-distribution with p - 1 degrees of freedom. A confidence region for the ith characteristic vector of l: with confidence 1 - e consists of the intersection of l3(i)'I3(i) = 1 and the set of l3(i) such that the right-hand side of (11) is less than xi-I(e), where Pr{ xi-I> xi-lee)} = e. Note that the matrix of the quadratic form (9) is positive semidefinite.
476
PRINCIPAL COMPONENTS
This approach also provides a test of the null hypothesis that the ith characteristic vector is a specified 13~) (13~)'13~) = 1). The hypothesis is rejected if the right-hand side of (11) with I3 U ) replaced by 13\!) exceeds xi-I(e). Mallows (1961) suggested a test of whether some characteristic vector of I is 130' Let 130 be p X (p - 1) matrix such that 130130 = O. if the null hypothesis is true, WaX and 13;)X are independent (because 130 is a nonsingular transform of the set of other characteristic vectors). The test is based on the multiple correlation between WuX amI I3:JX, In principle, the test procedure can be inverted to obtain a confidence region. The usefulness of these procedures is limited by the fact that the hypothesized vector is not attached to a characteristic root; the interpretation depends on the root (e.g., largest versus smallest). Tyler (981), (1983b) has generalized the confidence region (11) to indud'! the vectors in a linear subspace. He has also studied casing the restrictions of a normally distributed parent population.
11.6.3. Exact Confidence Limits on the Characteristic Roots We now consider a confidence interval for the entire set of characteristic roots of I, namely, AI :?= ... :?= \' [Anderson (1965a)]. We use the facts that I3 U) 'II3(i) = Aj l3(i)'I3U) = 1, i = 1, p, and 13(1)'II3(P) = 0 = 13(1)'I3(p). Then 13(1)'X and l3(p)'X are uncorrelated and have variances AI and Ap ' respectively. Hence nl3(1)' SI3(!) / A] and nl3(p), SI3(p) / Ap are independently distributed as X 2 with n degrees of freedom. Let Z and u be two numbers such that
(12) Then
(13)
~
b'Sb b'Sb} Pr { min - - ~ Ap ' AI ~ max -Zb'b~1
U
b'b~l
~.
11.6
477
STATISTICAL INFERENCE
Theorem 11.6.1. A confidence interval for the characteristic roots of I with confidence at least 1 - e is
(14) where I and u satisfy (12).
A tighter inequality can lead to a better lower bound. The matrix H = nil"'" nIl' hecause p is orthogonal. We use the following lemma.
nf:l'Sp has characteristic roots Lemma 11.6.1.
For any positive definite matrix H i = 1, ... p,
(15)
0
where H- 1 = (h;j) and ch/H) and chl(H) are the minimum and maximum characteristic roots of H, respective/yo Proof From Theorem A.2.4 in the Appendix we have ch/H) ~ hi; ~ chl(H) and
(16)
i
=1
0 ... ,
p.
Since ch/H) = Ijch l (H- 1 ) and ch/H) = Ijch p (H- 1 ), the lemma follows .
•
The argument for Theorem 5.2.2 shows that Ij(A p is distributed as X 2 with n - p + 1 degrees of freedom, and Theorem 4.3.3 shows that h pp is independent of h 11 • Let [' and u' be two numbers such that h PP )
(17) Then (18)
1 - e= Pr{n[' ~ x}}Pr{ Xn2_p+1 ~ nu'}.
478
PRINCIPAL COMPONENTS
Theorem 11.6.2. A confidence interval for the characteristic roots of I with confidence at least 1 - e is
(19)
I I ..L<, u' - I\p -<'1\1 -<...!. I' ,
"'here I' and u' satisfy (17). Anderson (1 965 a, 1965b) showed that the above confidence bounds are optimal within the class of b<Junds (20) where f and g are homogeneous of degree 1 and are monotonically non decreasing in each argument for fixed values of the others. If (20) holds with probability at least 1 - e, then a pair of numbers u' and l' can be found to satisfy (17) and
(21) The homogeneity condition means that the confidence bounds are multiplied by C C if the obselVed vectors are multiplied by c (which is a kind of scale invariance), The monotonicity conditions imply that an increase in the size of S results in an increase in the limits for I (which is a kind of consistency). The confidence bounds given in (31) of Section 10.8 for the roots of I based on the distribution of the roots of S when I = I are grcater.
11.7. TESTING HYPOTHESES ABOUT THE CHARACTERISTIC ROOTS OF A COVARIANCE MATRIX 11.7.1. Testing a Hypothesis about the Sum of the Smallest Characteristic Roots An investigator may raise the question whether the, last p - m principal components may be ignored, that is, whether the first In principal components furnish a good approximation to X, He may want to do this if the sl1m of the variances of the last principal components is less than some specified amount, say y, Consider the null hypothesis (1)
11.7
479
CHARACTERISTIC ROOTS OF A COVARIANCE MATRIX
where 'Y is specified, against tile alternative that the sum is less than 'Y. If the characteristic roots of I are different, it follows from Theorem 13.5.1 that
rn(i~t+l/i- i~t+l Ai)
(2)
Ar
has a limiting normal distribution with mean 0 and variance 2r.f~m + I The variance can be consistently estim«ted by 2r.f~ m + I !J. Then a rejection region with (large-sample) significance level s is
(3)
p
i~EI-I Ii < 'Y-
V2r.f~m+l/;
rn
z(2s),
where z(2e) is the upper 3ignificance point of the standard normal distribution for significance level s. The (large-sample) probability of rejection is s if equality holds in (1) and is less than s if inequality holds. The investigator may alternatively want an upper confidence interval for r.f~m + I Ai with at least approximate confidence level 1 - s. It is
(4)
p
V2r.f~m+ II;
rn
p
i=~ I Ai ~ i=E-1 Ii +
z(2s).
If the right-hand side is sufficiently small (in particular less than 'Y), the investigator has confidence that the sum of the variances of the smallest p - m principal components is so small they can be neglected. Anderson (1963a) gave this analysis also in the case that Am + I = ... = Ap.
11.7.2. Testing a Hypothesis about the Sum of the Smallest Characteristic Roots Relative t9 the Sum of All the Roots The investigator may want to ignore the last p - m principal components if their sum is small relative to the sum of all the roots (which is the trace of the covariance matrix). Consider the null hypothesis
(5)
H : f( A.) =
A
+ ... +A P + ... +Ap
III + I
Al
> 8
,
where 8 is specified, against the alternative that t(A.) < 8. We use the fact that af(A.)
---ar;- = (6)
at( A.) _
---ar;- -
Am+! -
+ ...
+Ap
(AI + ... +Ap)2' Al + ... + Am
(AI + ... +AS'
i = l,oo.,m,
i=m+l,oo.,p.
480
PJ'.INCIP AL COMPONENTS
Then the asymptotic variance of f(l) is
when equality holds in (5), by Theorem 4.2.3. The null hypothesis H is rejected if m(f(l) - 8] is less than the appropriate significance point of the standard normal distribution times the square root of (7) with A's replaced by I's and tr I by tr S. Alternatively one can construct a large-sample confidence region for f( A.). A confidence region of approximate confid~nce 1 - e is [z =z(2e)]
(8)
'[,f~'''+1 A,
<
'[,;=1 A;
-
'[,;~m+ll; + [2('[,;=m+ll;)2'[,~~111 + 2('[,~=1Ij)2'[,f=m+llfP '[,;=llj
z
m('[,;=1 Ij)2
If the right-hand side is sufficiently small, the investigator may be willing to let the first principal components represent the entire vector of measurements.
11.7.3. Testing Equality of the Smallest Roots Suppose the observed X is given by V + U + IL, where V and U are unobservable random vectors with means 0 and IL is an unobservable vector of constants. If CUU' = (I 2/, then U can be interpreted as composed of errors of measurement: uncorrelated components with equal variances. (It is assumed that all components of X are in the same units.) Then V can be interpreted as made up of the systematic parts and is supposed to lie in an m-dimensional space. Then CW' =
nf=m+llj ( )p-m P"' p-m . ('[,f'. III + II;)
11.7
481
CHARACTERISTIC ROOTS OF A COVARIANCE MATRIX
It is also the likelihood ratio criterion, but we shall not derive it. [See Anderson (1963a).] Let In (Ii - Am+ I) = d i , i = m + 1, ... , p. The logarithm of (9) multiplied by - n is asymptotically equivalent under the null hypothesis
to p
(10)
-n log = -n
n Ii + n i=m+1
LP
i=m+1
=n{-
"ff-m+l /,. p - m) log--==--m p
U
+n-~d)
+n(p-m)log ,=m+1
m+I,
IOg(I+~)+(p-m)IOg(l+ i=m+1 Am+ln-
f.
=n {
log(A
(
LP [d. --'-1 Am+1n'
i=m+1
+ ( p -m )[
1 -[ = -Z 2Am+1
LP
;=m+1
d
(A 111+1 + 11~d) . , p-m
~)}
d E;=III+l , (p-m)Am+lll-
2
-,,-,-'- + "-
]
_A;;'+ln
(Ef=m+ldi)z
Ef=m+ldi
+ .. _J} ,,-. Z 2 ( P - m) A", + I n 2( p - m) "-m + III
dJ -
1 (P ----:::m L d ) 2] + 0/1)_ p i=m+1 i
It is shown in Section 13.5.2 that the limiting distribution of d m + I , . , . , dp is the same as the distribution of the roots of a symmetric matrix Un = (U,.,.), i, j = m + 1, ... , p, whose functionally independent elements are independe'nt and normal with mean 0; all off-diagonal element u ij , i <j, has variance A~+l' and a diagonal element U ii has variance 2A~+I' See Theorem 13.5.2. Then (0) has the limiting distribution of
(11)
~82
PRINCIPAL COMPONENTS
Thus L; "' jll~/ A;" + 1 is asymptotically X 2 with ~(p - mXp - m - 1) degrees of freedom: ~[L!:''''+llIij - (L{:'",+llIiY/(P - m)1!A~+1 is asymptotically X2 with p - m - 1 degrees of freedom. Then (0) has a limiting X 2-distribution p -/11 + 2)(p - m - 1) degrees of freedom. The hypothesis is rejected with if the left-hand side of (0) is greater than the upper-tailed significance point of the ,\ 2-distribution. If the hypothesis is not rejected, the investigator may consider the last p - m principal components to be composed entirely of error. When the units of measurement are not all the same, the three hypotheses considered in Section 11.7 have questionable meaning. Corresponding hypotheses for the correlation matrix also have doubtful· interpretation. Moreover, the last criterion does not have (usually) a X 2-distribution. More discussion is given by Anderson (1963a). The criterion (9) corresponds to the sphericity criterion of Section 10./, and the number of degrees of freedom of the corresponding X 2-distribution is ±Cp - m)(p - m + 1) - 1.
±(
11.8. ELLIPTICALLY CONTOURED DISTRIBUTIONS 11.8.1. Observations Elliptically Contoured Let x I"
.. , X ,\
be N observations on a random vector X with density
( 1) where'" is a positive definite matrix, R 2 =(x-v)''IT- 1(x-v), and J'R~<x. Define K=p0R~/[(cX'R2)2(p+2)1-1. Then
"*
(2)
483
PROBLEMS
The covariance of gi and gj is
(3)
For inference about a single ordered root Aj the limiting standard normal distribution of IN(li - A)/(J2(2 + 3K) I) can be used. For inference about a single vector the right-hand side of (11) in Section 11.6.2 can be used with S replaced by (1 + K)S and S-I by S-I/(1 + K). It is shown in Section 13.7.1 that the limiting distribution of the logarithm of the likelihood ratio criterion for testing the equality of the q = p - m smallest roots is the distribution of (1 + K) X;(Q-ll/2-1' 11.8.2. Elliptically Contoured Matrix Distributions Suppose the density of X =
(Xi>""
x N ) is
11J1I-N/2g[tr(X - E'NV)'qr-I(X- E'NV)]
= Iqrl-N/2 g [tr Aqr-I + N(x- V)'qr-I(X- v)], where A = (X - E'Ni)(X - E'NX)' = nS and n = N - 1. Thus X and A are a sufficient set of statistics. Now consider A = YY' having the density g(tr A). Let A = BLB', where L is diagonal with diagonal elements II> .. , > Ip and B is orthogonal with Pi! ~ O. Then Land B are independent; the roots II"'" Ip have the density (18) of Section 13.7, and the matrix B has the conditional Haar invariant distribution.
PROBLEMS 11.1. (Sec. 11.2) Prove that the characteristic vectors of
(~ ~)
are
1/1i) ( ljli ) ( 1/1i and -1/1i' corresponding to roots 1 + I' and 1 - p. 11.2. (Sec. 11.2) Verify that the proof of Theorem 11.2.1 yields a proof of Theorem A.2.1 of the Appendix for any real symmetric matrix.
484
PRINCIPAL COMPONENTS
11.3. (Sec. 11.2) Let z = y +x, where rly = rlx = 0, rlyy' =~, rlu' = u 2[, rlyx' = O. The p components of y can be called systematic parts, and the components of x errors.
(a) Find the linear combination "I' z of unit variance that has minimum error variance (i.e., "I' x has minimum variance). (b) Suppose 4>ii + u 2 = 1, i = 1, ... ,p. Find the linear function "I I z of unit variance that maximizes the sum of squares of the correlations between Zj and 'Y'z, i = 1, ... ,p. (c) Relate these results to principal components. 11.4. (Sec. 11.2) Let I = «I> + U 2[, where «I> is positive semidefinite of rank m. Prove that each characteristic vector of «I> is a vector of I and each root of I is a root of «I> plus u 2.
11.5. (Sec. 11.2) Let the characteristic roots of
~ be Al ~ A2 ~ ... ~ Ap ~ O.
(a) What is the form of I if Al = A2 = .. ' = Ap > O? What is the shape of an ellipsoid of constant density? (b) What is the form of I if Al > A2 = .. , = Ap > O? What is ttoe shape of an ellipsoid of constant density? (c) What is the form of I if Al = ... = Ap_ 1 > Ap > O? What is the shape of the ellipsoid of constant density?
11.6. (Sec. 11.2)
lntraclass correlation. Let
where E = (1, ... , 1)'. Show that for p > 0, the largest characteristic root is u 2 [1 + (p -1)p] and the corresponding characteristic vector is E. Show that if E' x = 0, then x is a characteristic vector corresponding to the root u 2(1 - p). Show that the root u 2(1 - p) has multiplicity p - 1.
11.7. (Sec. 11.3) In the example of Section 9.6, consider the three pressing operations (X2' x 4 , x 5 ). Find the first principal component of this estimated covariance matrix. [Hint: Start with the vector (1, 1, 1) and iterate.) 11.S. (Sec. 11.3) Prove directly the sample analog of Theorem 11.2.1, where LXa = 0, LXaX~ = A. 11.9. (Sec. 11.3) Let II and Ip be the largest and smallest characteristic roots of S, respectively. Prove rill ~ Al and rllp 5; Ap' 11.10. (Sec. 11.3) Let U I = jl(l)'X be the first population principal component with variance 'Y(UI ) = AI' and let VI =b(l)'X be the first sample principal component with (sample) variance II (based on S). Let S* be the covariance matrix of a second (independent) sample. Show rlb(l)'S*b(l) 5; AI'
485
PROBLEMS
11.11. (Sec. 11.3) Suppose that U;j> 0 for every i, j [l: = (u;j)]' Show that (a) the coefficients of the first principal component are all of the same sign. and (b) the coefficients of eat h other principal component cannot be all of the same sign. 11.12. (Sec. 11.4)
Prove (4) when Al >
(a) Show l:' (b) Show
=
,1,2'
pA;p'.
where I; = fl)_osj and Sj = (e) Show
II "';X;OX(O.
where Ell has 1 in the upper left-hand position and O's elsewhere. (d) Show limi~JI;AP2 = I/(jl(ll'x(IIl)2. (e) Conclude the proof. 11.13. (Sec. 11.4)
Let
q;,I] l:,,'
K
=[10 H' 0]
where H = Ip -I - 2 a a' and a has p - 1 components. Show that a can be chosen so that in
HU(I)
has all 0 components except the first.
n.14. (Sec. 11.6) Show that log I; -
fl
z( e) < log A; < log I; +
fl
z( e)
is a confidence interval for log A; with approximate confidence 1 - e. 11.15. (Sec. 11.6)
Prove that
1/'
< II if I'
=
I and p > 2.
11.16. (Sec. 11.6) Prove that 1/ < 1/* if I = /* and p > 2, where I'" and and 1/ of Section 10.8.4.
1/'"
are the I
486
PRINCIPAL COMPONENTS
11.17. The lengths. widths. and heights (in millimeters) of 24 male painted turtles [Jolicoeur and Mosimann (1960)] are given below. Find the (sample) principd components and their variances. Case No. 1 2 3 4 5 6 7 8 9 10 11
12
Length
Width
Height
93 94 96 101 102 103 104 106 107 112 113 114
74 78 80 84 85 81 83 83 82 89 88 86
37 35 35 39 38 37 39 39 38 40 40 40
Case No. 13 14 15 16 17 18 19 20 21 22 23
24
Length
Width
Height
116 117 117 119 120 120 121 125 127 128 131 135
90 90 91 93 89 93 95 93 96 95 95 106
43 41 41 41 40
44 42 45 45 46 46 47
CHAPTER 12
Canonical Correlations and Canonical Variables
12.1. INTRODUCTION
In this section we consider two sets of variates with a joint distribution, and we analyze the correlations between the variables of one set and those of the other set. We find a new coordinate system in the space of each set of variates in such a way that the new coordinates display unambiguously the system of correlation. More preci3ely, we find linear combinations of variables in the sets that have maximum correlation; these linear combinations are the first coordinates in the new systems. Then a second linear combination in each set is sought such that the correlation between these is the maximum of correlations between such linear combinations as are uncorrelated with the first linear combinations. The procedure is continued until the two new coordinate systems are completely specified. The statistical method outlined is of particular usefulness in exploratory studies. The investigator may have two large sets of variates and may want to study the interrelations. If the two sets are very large, he may want to consider only a few linear combinations of each set. Then he will want to study those l.near combinations most highly correlated. For example, one set of variables may be measurements of physical characteristics, such as various lengths and breadths of skulls; the other variables may be measurements of mental characteristics, such as scores on intelligence tests. If the investigator is interested in relating these, he may find that the interrelation is almost
An Introduction to Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0·471-36091-0 Copyright © 2003 John Wiley & Solis, Inc.
487
488
CANONICAL CORRELAnONS AND CANONICAL VARIABLES
completely described by the correlation between the first few canonical variates. The basic theory was developed by Hotelling (1935), (1936). In Section 12.2 the canonical correlations and variates in the popUlation are defined; they imply a linear transformation to canonical form. Maximum likelihood estimators are sample analogs. Tests of independence and of the rank of a correlation matrix are developed on the basis of asymptotic theory in Section 12.4. Another formulation of canonical correlations and variates is made in the case of one set being random and the other set consisting of nonstochastic variables; the expected values of the random variables are linear combinations of the nonstochastic variables (Section 12.6). This is the model of Section 8.2. One set of canonical variables consists of linear combinations of the random variables and the other set consists of the nonstochastic variables; the effect of thc rcgrcssion of a member of the first sct on a mcmber of the second is maximized. Linear functional relationships are studied in this framework. Simultaneous equations models are studied in Section 12.7. Estimation of a ~ingle equation in this model is formally identical to estimation of a single linear functional relationship. The limited-information maximum likelihood estimator and the two-stage least squares estimator are developed.
12.2. CANONICAL CORRELATIONS AND VARIATES IN THE POPULATION
Suppose the random vector X of P components has the covaria.lce matrix I (which is assumed to be positive definite). Since we are only interested in variances and covariances in this chapter, we shall assume ex = 0 when treating the population. In developing the concepts and algebra we do not need to assume that X is normally distributed, though this latter assumption will be made to develop sampling theory. We partition X into two subvectors of PI and pz components, respectively,
(1 )
X=
X(1)) ( X(Z) •
For convenience we shall assume PI ~J z. The covariance matrix is partitioned similarly into Pl and P2 rows and columns,
(2)
12.2 CORRELATIONS AND VARIATES IN THE POPULATION
489
In the previous chapter we developed a rotation of coordinate axes to a new system in which the variance properties were clearly exhibited. Here we shaH develop a transformation of the first PI coordinate axes and a trallsformation of the last P2 coordinate axes to a new (PI + p)-system that wiH exhibit clearly the intercorrelations between X(l) and X(2). Consider an arbitrary linear combination, V = a' X(l), of the components of X(l), and an arbitrary linear function, V = 'Y ' X(2), of the components of XO). We first ask for the linear functions that have maximum correlation. Since the correlation of a multiple of V and a multiple of V is the same as the correlation of V and V, we can make an arbitrary normalization of a and 'Y. We therefore require a and 'Y to be such that V and V have unit variance, that is,
(J) (4) We note that ,cV = ,ca' X(I) = a',c X(I) = 0 and similarly ,c V = O. Then the correlation between V and V is
(5) Thus the algebraic problem is to find a and 'Y to maximize (5) subject to (3) and (4). Let
where A and /L are Lagrange multipliers. We differentiate '" with respect to the elements of a and 'Y. The vectors of derivatives set equal to zero are
(7) (8) Multiplication of (7) on the left by a' and (8) on the left by 'Y' gives
(9) (10) Since a '1 11 a = 1 and 'Y '1 22 ,)' = 1, this shows that A = /L = a 'l; 12 'Y. Thus (7)
490
CA'\IONICAL CORRELATIONS AND CANONICAL VARIABLES
and (8) can be written as (11)
-Al:lla
(12)
+ l:12'Y =0,
l:21 a - Al: 22 'Y
= 0,
since :I'\2 = :I 21' In one matrix equation this is
(13) In order that there be a nontrivial solution [which is necessary for a solution satisfying (3) and (4)], the matrix on the left must be singular; that is,
(14) The determinant on the left is a polynomial of degree p. To demonstrate this, consider a Laplace expansion by minors of the first PI columns. One term is I - A:I III ·1 - A:I221 = ( - A)P, +p 'I:I III ·1 :I 22 I. The other terms in the expansion are of lower degree in A because one or more rows of each minor in the first PI columns does not contain A. Since :I is positive definite, I:I III ·1l: 22 1 0 (Corollary Al.3 of the Appendix). This shows that (14) is a polynomial equation of degree P and has P roots, say Al ~ A2 ~ ... ~ Ap. [a' and 'Y' complex conjugate in (9) and (0) prove A reaL] From (9) we see that A = a':I 12 'Y is the correlation between V = a' Xl!) and V = 'Y' X (2) when a and 'Y satisfy (13) for some villue of A. Since we want the maximum correlation, we take A= AI' Let a solution to (13) for A = Al be a(l), 'Y(I), and let VI = a(I)' X(I) and VI = 'Y(l)' X(2) Then VI and VI are normalized linear combinations of X(l) and X(2l, respectively, with maximum correlation. We now consider finding a second linear combination of X(I), say L' = a' X(l), and a second linear combination of X(2), say V = 'Y' X(2), such that of all linear combinations uncorrelated with VI and VI these have maximum correlation. This procedure is continued. At the rth step we have obtained linear combinations VI = a(I), xOl, VI = 'YO)' X(2), .•• , ~ = a(r), X(l), V, = 'Y(r), X(2) with corresponding correlations [roots of (4)] A(I) = AI, A(2), ... , A(r). We ask for a linear combination of XO), V = a' X(l), and a linear combination of X(2l, V = 'Y' X (2) , that among all linear combinations uncorrelated with VI' VI"'" Vr , v" have maximum correlation. The condition that V be uncorrelated with U; is
"*
(15)
491
12.2 CORRELATIONS AND VARIATES IN THE POPULATION
Then (16) The condition that V be uncorrelated with V; is (17)
By the same argument we have (18) We now maximize tCUr+I~+I' choosing a and "{ to satisfy (3), (4), (15), and (17) for i = 1,2, ... , r. Consider
r
+
L
r
via '111 a(i) +
L
0i"{
'I 22 ,,{(i),
°
where A, 11-, VI"'" Vr' 1" " , Or are Lagrange multipliers. The vectors of partial derivatives of I/Ir+1 with respect to the elements of a and "{ are set equal to zero, giving (20)
(21)
Multiplication of (20) on the left by
aU)'
and (21) on the left by
,,{(il'
gives
(22)
(23)
Thus (20) and (21) are simply (11) and (12) or alternatively (13). We therefore take the largest A;, say, A(r+ I), such that there is a solution to (13) satisfying (1), (4), (15), and (17) for i= 1, ... ,r. Let this solution be a(r+I), ,,{(r+ lJ , and let u,,+1 = a(r+I)'x(1) and ~+I = ,,{(r+I)'x(2). This procedure is continued step by step as long as successive solutions can be found which satisfy the conditions, namely, (13) for some Ai' (3), (4), (15), and (17). Let m be the number of steps for which this can be done. Now
492
CANONICAL CORRELATIONS AND CANONICo.L VARIABLES
we shall show that m = PI' (5. P2)' Let A = (a(I) ... and
a(m»,
1'1 = ('Y(ll '"
'Y(ml),
o o
(24)
o
o
The conditions (3) and (15) can be summarized as
(25)
A'I11A=I.
Since 111 is of rank PI and I is of rankm, we have m 5.PI' Now let us show that m < PI leads to a contradiction by showing that in this case there is another vector satisfying the conditions. Since A'Ill is m XPI' there exists a PI X(PI-m) matrix E (of rank Pl-m) such that A'IllE=O. Similarly there is a P2 X (P2 - m) matrix F (of rank P2 - m) such that fjI22F <-= O. We also have fjI 2I E = AA'I11E = 0 and A'I12F= AfjI 22 F·= O. Since E is of rank PI - m, E' I 11 E is nons in gular (if m < PI)' and similarly F' I 22 F is nonsingular. Thus there is at least one root of -
(26) 1
vE'IllE
E'I12F 1 _ vF'I22F = 0,
F'I2IE
"*
because IE'IIlEI'IF'I22FI O. From the preceding algebra we see that there exist vectors a and b such that (27)
E'I I2 Fb
= vE'IIl Ea ,
(28)
F'I2IEa
= vF'I 22 Fb.
Let Ea = g and Fb = h. We now want to show that v, g, and h form a new solution A(m+l),a(m+I),'Y(m+I). Let Ij/I l2 h=k. Since A'Iuk= A'Il2 Fb = 0, k is orthogonal to the rows of A'IIl and therefore is a linen combination of the columns of E, say Ec. Thus the equation II2h = Ink can be written
(29) Multiplication by E' on the left gives (30) Since E'I 11 E is nonsingular, comparison of (27) and (30) shows that c = va,
11.2
CORRELATIONS Al'ID VARIATES IN THE POPULATION
493
and therefore k = vg. Thus
(31)
In a similar fashion we show that (32) Therefore v=A(m+I),g=a(m+ll,h='Y(m+l) is another solution. But this is contrary to the assumption that A(m), a(m), 'Y(m) was the last possible solution. Thus m =PI. The conditions on the A's, a's and 'Y's can be summarized as
(31)
A'IlIA=I,
(34)
A'II2rl = A,
(35)
r;I22rl =1.
Let
r 2 = ('Y(P, + I) ••• 'Y(P,»
be a P2
(P2 - PI) matrix satisfying
X
(36)
r;I22rl = 0,
(37)
r;I22r2=1.
Any r 2 can be multiplied on the right by an arbitrary (P2 - PI) X (pz - PI) orthogonal matrix. This matrix can be formed one column at a time: 'Y(P, + 11 is a vector orthogonal to Inri and normalized so 'Y(P,+I)'I n 'Y(P,+1l = 1; 'Y(p,+2) is a vector orthogonal to In(r l 'Y(p,+I» and normalized so ,y(p,+2)'I n 'Y(p,+2) = 1; and so forth. Let r = (r l r 2); this square matrix is nonsingular since r I I 22 r = I. Consider the determinant
A' (38)
o o -AI
=
A
I o
-~
J
= ( _ A)P'-p,\ - AI A
A\
-AI
=(-A)P'-P'I-AII·I-Al-A(-AI)-lAI
= ( _ A)P,-P'I A2/_ A21 = (-A)P,-P'I1(A 2 - AU)2).
494
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
Except for a constant factor the above polynomial is
(39)
Thus the roots of (14) are the roots of (38) set equal to zero, namely, A = ± A(i), i = 1, ... , PI' and A = 0 (of multiplicity pz - PI)' Thus (AI"'" Ap) = (AI"'" ApI' 0, ... ,0, - API''''' - AI)' The set {A(i)Z}, i = 1, ... , PI' is the set (An, i = 1, ... , PI' To show that the set (A(i)}, i = 1, ... , PI' is the set {A), i = 1, ... , PI' we only need to show that A(i) is nonnegative (and therefore i!> one of the A;, i = 1, ... , PI)' We observe that (40) (41 )
thus, if A(r), aU>, -y(r) is a solution, so is - A(r), - a(r), -y(r). If A(r) were negative, then - A(r) would be nonnegative and - A(r) ~ A(r). But since A(r) wasta be maximum, we must have A(r);;:: - A(r) and therefore A(r);;:: O. Since the set (A(i)} is the same as (A,-l, i = 1, ... , PI> we must have A(i) = A,.. Let
( 42)
U=
=A'X(I),
( 43)
( 44)
The components of U are one set of canonical variates, and the components
12.2
CORRELATIONS AND VARIATES IN THE POPULATION
495
of V= (V(I), V(Z)')' are the other set. We have
(45)
c( ~(I)l (V'
A' v(1)'
V(2)') =
V(2)
(
~
LJ
where
(46)
Al 0
0
0
Az
0
0
0
ApI
A=
Definition 12.2.1. Let X = (X(I), X(2),y, where X(I) has PI components and X(Z) has pz (= p - PI ~ PI) components. Ihe rth pair of canonical variates is the pair of linear combinations Vr = a(r)' X(1) and = ,,/(r), X(2), each of unit variance and uncorrelated with the first r - 1 pairs of canonical variates and having maximum correlation. The correlation is the rth canonical correlation.
v,.
Theorem 12.2.1. Let X = (X(I)' X(Z) ,), be a random vector with covariance matrix I. The rth canonical correlation between X(1) and X (2) is the rth largest root of (14). The coefficients of a(r)' (1) and ,,/(r), X(Z) defining the rth pair of canonical variates satisfy (13) for A = Ar and (3) and (4).
x
We can now verify (without differentiation) that VI,VI have maximum correlation. The linear combmations a'V = (a'A')X(I) and b'V= (bT')X(Z) are normalized by a'a = 1 and b'b = 1. Since A and rare nonsingular, any vector a can be written as Aa and any vector "/ can be written as rb, and hence any linear combinations a' X(I) and ,,/' X(Z) can be written as a' V and b'V. The correlation between them is P,
(47)
a'(A
O)b= LA;ajb j. ;=1
496
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
with respect to b is fOl hj = c j ' since Lcjhj is the cosine of the angle between the vector band (cp ... ,cp"O, ... ,O). Then (47) is
and this is maximized by taking a j = 0, i = 2, ... , PI. Thus the maximized linear combinations are UI and VI. In verifying that U2 and V2 form the second pair of canonical variates we note that lack of correlation between UI and a linear combination a'U means 0= ,cUla'U = ,cUI La;ll; = a l and lack of correlation between VI and b'V means 0= hi. The algebra used above gives the desired result with sums starting with i = 2. We can derive a single matrix equation for a or -y. If we multiply (11) by A and (12) by Ii21, we have ( 48)
AI 12 -y=A 2I
( 49)
Ii} 1 21 a = A-y.
ll a,
Substitution from (49) into (48) gives
(50) or
(51) The quantities
Ai, ... , A;,
satisfy
(52) and a(1), ••• , a(p,) satisfy (51) for A2 = Ai, ... , A;" respectively. The similar equations for -y(l), ... , -y(p,) occur when >t2 = Ai, ... , A;, are substituted with
(53) Th~orem 12.2.2. The canonical correlations are invariant with respect to transformations X(ih = CjX(i), where C j is nonsingular, i = 1,2, and any function of I that is invariant is a function of the canonical correlatio '1s.
Proof Equation (14) is transformed to
CII\2C~ I = ICI
-AC 2 1
22
C2
0
12.2
CORRELATIONS AND VARIATES IN THE POPULATION
497
and hence the roots arc unehangcu. Conversely, let !("i.1I,"i.12,"i.22) be a vector-valued function of I such that !(CI1IIC;, CI1I~C~' C~1~~C;) = /(111,112,122) for allnonsingular C I and C,. If C; =A and C,=I", then (54) is (38), which depends only on the canonical correlations. Then ! = /{I, (A, 0),
n.
•
We can make another interpretation of these developments in terms of prediction. Consider two random variables U and V with means 0 and variances u} and uu2 and correlation p. Consider approximating U hy a multiple of V, say bV; then the mean squared error of approximation is
(55)
= a} (I -- p ~) + (ba;, - pa;,) ~. This is minimized by taking b = a;, pi a; .. We can consider bV as a linear prediction of U from V; then a;,2(l - p2) is the mean squared error of prediction. The ratio of the mean squared error of prediction to the variance of U is uu2(l- p2)/a;,2 = 1 - p2; the complement is a measure of the relative effect of V on U or the relative effectiveness of V in predicting U. Thus the greater p2 or Ipi is, the mOre effective is V in predicting U. Now consider the random vector X partitioned according to (1), and consider using a linear combination V = 'Y' X(2) to predict a linear combination U = a ' X(I). Then V predicts U best if the correlation between U and V is a maximum. Thus we can say that a(I)'X(1) is the linear combination of X(1) that can be predicted best, and ,),(I)'X(2) is the best predictor [Hotelling (1935)]. The mean squared effect of V on U can be measured as
(56) and the relative mean squared effect can be measured by the ratio tf(bV)2/
498
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
would mean treating X'I> wi'h mean ,cX", = 0y~3); the elements of the covariance matrix would be the partial covariances of the first p elements of Y. The interpretation of canonical variates may be facilitated by considering the correlations between the canonical variates and the components of the original vectors [e.g., Darlington, Weinberg, and Wahlberg (1973)]. The covariance between the jth canonical variate ~. and Xi is
(57)
,cUjXi=,c
Since the variance of
~
PI
1',
k~l
k~l
L a~ilXkXi= L akj)uki'
is 1, the correlation bl tween
Corr( ~, Xi) =
(58)
~.
and Xi is
r.Cl.l a~jhki
r;;..
YUii
An advantage of this measure is that it does not depend on the units of measurement of Xj' However, it is not a scalar multiple of the weight of Xi in ~. (namely. a)j). A special case is III = I, 122 = I. Then (59)
A'A=I,
r'f=I,
From these we obtain (fiO)
where A and r are orthogonal and A is diagonal. This relationship is known as the singular value decomposition of 1 12 , The elements of A are the square roots of the characteristic roots of 112 1\2' and the columns of A are characteristic vectors. The diagonal elements of A are square roots of the (possibly nonzero) roots of 1'12 I 12, and the columns of r are the characteristic vectors.
12.3. ESTIMATION OF CANONICAL CORRELATIONS AND VARIATES 12.3.1. Estimation Let X l " ' " x" be N observations from N(tJ., I). Let xa be partitioned into t\vo suhvectors of PI and P2 components, respectively, ( I)
a=
1, ... ,N.
12.3
499
ESTIMATION OF CANONICAL CORRELATIONS AND VARIATES
The maximum likelihood estimator of I [partitioned as in (2) of Section 12.21 is
(2)
i=(~ll ~IZ]=~t(Xa-i)(Xa-i)/ \121
122
a=1
(E( x~) -i(I»)( X~I) -i(1))' E( X~I) -i(1))( x~) -i(Z»)/]. N E( x~) - i(2») ( X~I) - i(I») / E( x~Z) - i(Z») ( X~2) - i(2») / 1
=
The maximum likelihood estimators of the canonical correlations A and the canonical variates defined by A and r involve applying the algebra of the previous section to i. The matrices A, A, and r l are.uniquely defined if we assume the canonical corr~lations different and that the first nonzero element of each column of A is positive. The indeterminacy in r z allows mUltiplication on the right by a (pz - PI) X (pz - PI) orthogonal matrix; this indeterminacy can be removed by various types of requirements, for example, that the submatrix formed by the lower pz - PI rows be upper or lower triangular with positive diagonal elements. Application of Corollary 3.2.1 then shows that the maximum likelihood estimators of AI' ... ' Ap are the roots of
(3) and the jth columns of A and
f\
satisfy
(4)
(5)
r
2
satisfies
(6) (7) When the other restrictions on defined.
r2
are made,
A, f,
and
A are
uniquely
500
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
Theorem 12.3.1. LetxJ"",x N be Nobseruationsfrom N(v., I). Let I be partitioned into PJ and pz (PI :;;P2) rows and columns as in (2) in Section 12.2, and let X(Ji be similarly partitioned as in (1). The maximum likelihood estimators of the canonical correlations are the roots of (3), where Iij are defined by (2). Th:: maximum likelihood estimators of the coefficients of the jth canonical comprments satisfy (4) and (5), j = 1, ... , PJ; the remaining components satisfy (6) and (7).
In the population the canonical correlations and canonical variates were found in terms of maximizing correlations of linear combinations of two sets of variates. The entire argument can be carried out in terms of the sample. Thus &(J),x~) and .y(J)'xf) have maximum sample correlation between any linear combinations of x~) and xfl, and this correlation is II' Similarly, &(2)IX~) and .y(2)Ixfl have the second ma;"imum sample correlation, and so forth. It may also be observed that we could define the sample canonical variates and correlations in terms of S, the unbiased estimator of I. Then a(j) = N - l)jN &(j), c(j) = N - l)jN y.(j), and lj satisfy
v(
V(
(8)
SJ2 c(j) = IjSJl a(j),
(9)
SZJ a(j) = IjSzzc U ),
(10) We shall call the linear combinations aU) X~I) and c(j) xf) the sample canonical variates. We can also derive the sample canonical variates from the sample correlation matrix, I
(11)
R=
Uij
(
~...[&;;
)
= ( -Sij- ) =(r.) = (RJl ";SjjSjj
IJ
Let
(12)
/s-::
0
0
0
.;s;;
0
0
0
,;s;;;:
SJ=
R ZJ
12.3
501
ESTIMATION OF CANONICAL CORRELATIONS AND VARIATES
o o
o o
(13)
o
o
Then we can write (8) through (10) as
(14)
RdS2C(j») =ljRll(Sla(j»),
(15)
R 21 (Sla(j») = IjR22(S2c(f)),
(16)
(Sla(j»)'R ll (Sla U») = 1,
(S 2c Cil)'Rzz(S 2C Cil) = 1.
We can give these developments a geometric interpretation. The rows of the matrix (XI"'" x N ) can be interpreted as p vectors in an N-dimensional space, and the rows of (x 1 - i, ... , X N - i) are the P vectors projected on the (N - 1)-dimensional subspace orthogonal to the equiangular line. Denote these as xi, ... , x;. Any vector u' with components a '(X\I) - i(1l, ... , x~l i(l» = a1xi + ... + ap\x;\ is in the PI-space spanned by xi, ... , x;\' and a ve.ctor v* with componlnts 'Y'(x\2)-i(2), ... ,x\~)-i(1)=YIX;'TI + ... +'Yp,x; is in the P2-space spanned by X;\+I""'X;. The cosine of the angle between these two vectors is the correlation between u = a 'x~l) and va = 'Y 'xl!), a = 1, ... , N..Finding a and 'Y to maximize the correlation is equ;valent to finding the veetors in the PI-space and the pz-space such that the angle between them is least (i.e., has the greatest cosine). This gives the first canonical variates, and the first canonical correlation is the cosine of the angle. Similarly, the second canonical variates correspond to vectors orthogonal to the first canonical variates and with the angle minimized. Q
12.3.2. Computation We shall discuss briefly computation in terms of the population quantities. Equations (50), (51), or (52) of Section 12.2 can be used. The computation of !,12IZ21121 can be accomplished by solving 121 = 122 F for 1221121 and then multiplying by 1\2' If PI is sufficiently small, the determinant 11121221121 - vIlli can be expanded into a polynomial in v, and the polynomial equation may be solved for v. The solutions are then inserted . into (51) to arrive at the vectors a. In many cases PI is too large for this procedure to be efficient. Then one can use an iterative procedure
(17)
502
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
starting with an initial approximation a(O); the vector a(i + 1) may be normalized by
(18)
a(i
+ l)/Il1a(i + 1) = 1.
The AZ(i + 1) converges to Ai and aCi + 1) converges to a(1) (if AI> A2). This can be demonstrated in a fashion similar to that used for principal components, using (19) from (45) of Section 12.2. See Problem 12.9. The right-hand side of (19) is l:f.! 1a(i)A;ci(i)', where ci(i)' is the ith row of A-1. From the fact that A/InA = I, we find that A/In = A-I and thus a(i)/I ll = ci(i)/. Now PI
(20)
I
111
I12IzttZI - Aia(l)ci(l)' =
E a(i)A;ci(i), i=2
=A
o o o A~
o
o o
0
The maximum characteristic root of this matrix is A~. If we now use this matrix for iteration, we will obtain A~ and a(2). The procedure is continued to find as many A; and a(i) as desired. Given Aj and a(il, we find "I (i) from I Z1 a(i) = AiI 22 "1(i) or equivalently 0/ Ajn:zzlIzl ali) = "I(i). A check on the computations is provided by com- A paring I 1Z "I(i) and AiIll ali). For the sample we perform these calculations with 'I. ij or Sij substituted for Iii" It is often convenient to use Rij in the computation (because , - 1 < r IJ < 1) to obtain S 1 a(j) and S2 c(j)· from these a(j) and c(j) can be '1i,' ' .'1. computed. Modern computational procedures are available for canonical correlations 1 and variates similar to those sketched for principal components. Let A
(21)
Z1 =
(22)
Z2 = (x(Z) 1
(X(I) .:.... X-(I) 1 "."
x(l) -
N
x-(l)) ,
- x-(Z) , ••• , x(2) N - X-(2») •
'
J4 {!
503
12.4 STATISTICAL INFERENCE
The QR decomposition of the transpose of these matrices (Section 11.4) is Z; = Q;R;, where Q;Q; = Ipi and R; is upper triangular. Then S;j = Z;Zj = RiQ:·QjRj , i, j = 1,2, and Sjj = R;R;, i = 1,2. The canonical correlations are the singular values of Q 1Q2 and the square roots of the characteri:;tic roots of (QIQ2 XQ'IQ2)' (by Theorem 12.2.2). Then the singular value decomposition of Q;Q2 is peL O)T, where P and T are orthogonal and L is diagonal. To effect the decomposition Householder transformations are applied to the left and right of Q1Q2 to obtain an upper bidiagonal matrix, that is, a matrix with entries on the main diagonal and first superdiagonal. Givens matrices are used to reduce this matrix to a matrix that is diagonal to the degree of approximation required. For more detail see Kennedy and Gentle (1980), Section 7.2 and 12.2, Chambers (1977), Bjorck and Golub (1973), Golub and Luk (1976), and Golub and Van Loan (1989).
12.4. STATISTICAL INFERENCE 12.4.1. Tests of Independence and of Rank In Chapter 9 we considered testing the null hypothesis that X(l) and X(2) are independent, which is equivalent to the null hypothesis that l:12 = 0. Since A'I I2 r = (A 0), it is seen that the hypothesis is equivalent to A = 0, that is, PI = ... = PPI = O. The likelihood ratio criterion for testing this null hypothesis is the N /2 power of
(1)
I
°
A
°
A
I ,,--0.,-::;-O...,-::;-_/....!. =
1/1·1/1
IA~
AI A
I
A
= II - A21 =
n (1- rn, PI
;=1
where 'I = II ~ ... ~ rpi = Ipi ~ 0 are the PI possibly nonzero sample canonical correlations. Under the null hypothesis, the limiting distribution of Bartlett's modification of - 2 times the logarithm of the likelihood ratio criterion, namely, PI
(2)
- [N -
Hp + 3)] L log(l-r?), ;=1
504
CANONICAL CORRELA nONS AND CANONICAL VARIABLES
is x 2 with PI P2 degrees of freedom. (See Section 9.4.) Note that it is approximately 1',
(3)
N
L r;2 = Ntr A ilIA 12AZIIA21' i~1
which is N times Nagao's criterion [(2) of Section 9.5]. If I 12 0, an interesting question is how many population canonical correlations are different from 0; that is, how many canonical variates are needed to explain the correlations between X(l) and X(2)? The number of nonzero canonical correlations is equal to the rank of I 12. The likelihood ratio criterion for testing the null hypothesis Hk : Pk + I = ... = PI', = 0, that is,
*"
that the rank of II2 is not greater tha:n k, is (1974)]. Under the null hypothesis
nr';k+ I(1 -
r?)tN [Fujikosri
Pi
(4)
- [N -!(p + 3)] L 10g(1-rn ;~k
+I
has approximately the X2-distribution with (PI - k)(P2 - k) degrees of freedom. [Glynn and Muirhead (1978) suggest multiplying the sum in (4) by N - k - t(p + 3) + I:}~ I(1/r?); see also Lawley (1959).] To determine the numbers of nonzero and zero population canonical correlations one can test that all the roots are 0; if that hypothesis is rejected, test that the PI - 1 smallest roots are 0; etc. Of course, these procedures are not statistically independent, even asymptotically. Alternatively, one could use a sequence of tests in the opposite direction: Test PI', = 0, then PI', -1 = PI', = 0, and so on, until a hypothesis is rejected or until I12 = 0 is accepted. Yet another procedure (which can only be carried out for small PI) is to test Pp , = 0, then Pp , _ I = 0, and so forth. In this procedure one would use rj to test the hypothesis Pj = O. The relevant asymptotic distribution will be discussed in Section 12.4.2. 12.4.2. Distributions of Canonical Correlati£'ns The density of the canonical correlations is given in Section 13.4 for the case that I12 = 0, that is, all the population correlations are O. The density when some population correlations are different from 0 has been given by Constantine (1963) in terms of a hypergeometric function of two matrix arguments. The large-sample theory is more manageable. Suppose the first k canonical correlations are positive, less than 1, and Jifferent, and suppose that
12.5
505
AN EXAMPLE
PI - k correlations are O. Let
i
=
I .. ... k.
(5) Zi
= Nrl,
i=k+l, .... Pl·
Then in the limiting distribution Z 1" .. , Zk and the set :k, 1" '" ::'" arc mutually independent, Zi has the limiting distribution MO, 1), i = 1, ... , k. anrl the density of the limiting distribution of Zk + 1"'" zp, is
(6)
2t(p,-k)(p,-klrPI-k [!( P I - k)] r PI -k [!( p 2 - k)] 2 2 PI
.n i~k+l
PI
Zl(P'-P,-I)
n
(Zi
-z)).
i.j~k+l
,. ..::..1
This is the density (11) of Section 13.3 of the characteristic roots of a (PI - k)-order matrix with distribution W(Jp,_k' P2 - k). Note that the normalizing factor for the squared correlations corresponding to nonzero population correlations is IN, while the factor corresponding to zero population correlation is N. See Chapter 13. In large samples we treat r? as N[ Pi2 ,(1/N)4p?(1- p?)~] or ri as N[ Pi,(1/N)(1- p?)2] (by Theorem 4.2.3) to obtain tests of Pi or confidence intervals for Pi' Lawley (1959) has shown that the transformation Zi = tanh -I (r) [see Section 4.2.3] does not stabilize the variance and has a significant bias in estimating ~i = tanh - 1 ( Pi)'
12.5. AN EXAMPLE In this section we consider a simple illustrative example. Rao [(1 Q52), p. 245] gives some measurements on the first and second adult sons in a sample of 25 families. (These have been used in Problem 3.1 and Problem 4.41.) Let x lo be the head length of the first son in the IX th family, x 2" be the head breadth of the first son, X3a be the head length of the second son, and x." be the head breadth of the second son. We shall investigate the relations between the measurements for the first son and for the second. Thus X~I \' = (x I,,' X 2" )
506
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
and x~;)' = (x 3". (l)
X-l,,)'
The data can be summarized as t
i' = (185.72.l5l.l2, l83.84, l49.24),
S=
95.2933 52.8683 ( 69.6617 46.lll7
52.8683 54.3600 51.3117 35.0533
69.6617 51.3ll7 100.8067 56.5400
46.1117 35.0533 56.5400 45.0233
The matrix of correlations is 1.0000
R=
(2)
0.7346: 0.7108
0.7040
.9..:~3_4~ __ ~ :.O.9Q~-i-~!i~~~ __~.?~~~ ( 0.7108 0.7040
0.6932: 1.0000 0.7086: 0.8392
0.8392 1.0000
All of the correlations are about 0.7 except for the correlation between the two measurements on second sons. In particular, RI2 is nearly of rank one, and hence the second canonical will be near zero. WG compute _)
(3)
R22 R21
R R-IR
(4)
12
22
21
=
(0.405769 0.363480
0.333205 ) 0.428976 '
=
(0.544311 0.538841
0.53R841 ) 0.534950 .
The determinantal equation is 0= 10.544311- 1.0000v 0.538841 - 0.73461'
(5)
=
0.460363v 2
-
0.538841- 0.7346vl 0.534950 -1.0000v
0.287596v+ n.000830.
The roots are 0.621816 and 0.002900; thus II = 0.788553 and 12 = 0.053852. Corresponding to these roots are the vectors
S
(6)
atl) 1
=
0.552166] ( 0.521548 '
S a (2 ) = I
(
_
1.366501] 1.378467 '
where (7)
t
Rao':-,.
cnmpulations arc in error: his last "uiffcrcncc" is incorrect.
507
12.5 AN EXAMPLE
We apply (1/l)Ri} R21 to SI a(i) to obtain ( 8)
S 2
C (1)
= (0.504511) 0.538242 '
S
C (2)
= (
2
1.767281) -1.757288 '
where
o 1 ..;s;;
= (
(9)
10.0402 0
0) 6.7099'
We check these computations by calculating
(10)
r;1 R11-I R 12 ( S2 C(I)) -_
(0.552157) 0.521560 '
l.R-I ( (2)) _ l2 11 R12 S2 C -
(
1.365151) -1.376741 .
The first vector in (10) corresponds closely to the first vector in (6); in fact, it is a slight improvement, for the computation is equivalent to an iteration on Sla(l). The second vector in (10) does not correspond as closely to the second vector in (6). One reason is that l2 is correct to only four or five significant and thus the components of S2C(2) can be correct to figures (as is 112 = only as many significant figures; secondly, the fact that S2C(2) corresponds to the smaller root means that the iteration decreases the accuracy instead of increasing it. Our final results are
In
(2) 0.054,
(1)
Ii = 0.789,
(11)
(i)
= (0.0566 )
(
0.1400) -0.1870 '
(i)
= ( 0.0502 )
(
0.1760) -0.2619 .
a
C
0.0707 '
0.0802 '
The larger of the two canoni::al correlations, 0.789, is larger than any of the individual correlations of a variable of the first set with a variable of the other. The second canonical correlation is very near zero. This means that to study the relation between two head dimensions of first sons and second sons we can confine our attention to the first canonical variates; the second canonical variates are correlated only slightly. The first canonical variate in each set is approximately proportional to the sum of the two measurements divided by their respective standard deviations; the second canonical variate in each set is approximately proportional to the difference of the two standardized measurements.
508
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
12.6. LINEARLY RELATED EXPECfED VALUES 12.6.1. Canonical Analysis of Regression Matrices In this section we develop canonical correlations and variates for one stochastic vector and one nonstochastic vector. The expected value of the stochastic vector is a linear function of the nonstochastic vector (Chapter 8). We find new coordinate systems so that the expected value of each coordinate of the stochastic vector depends on only one coordinate of the nonstochastic vector; the coordinates of the stochastic vector are uncorrelated in the stc·chastic sense, and the coordinates of the nonstochastic vector are uncorrelated in the sample. The coordinates are ordered according to the eff;ct sum of squares relative to the variance. The algebra is similar to that developed in Section 12.2. If X has the normal distribution N(fJ" l;) with X, fJ" and I partitioned as in (1) and (2) of Section 12.2 and fJ, = (fJ,(I)" fJ,(2),)" the conditional distribution of X(I) given X(2) is normal with mean
(1 ) and covariance matrix
(2) Since we consider a set of random vectors xII>, ... , X~J> with expected values depending on X\2), ••• , x~J> (nonstochastic), we em write the conditional expected value of X,V) as T + P(x~) _i(2», where T = fJ,ll) + P(i(2) - fJ,(2» can be considered as a parameter vector. This is the model of Section 8.2 with a slight change of notation. The model of this section is
(3)
4>= 1, ... ,N,
where X\2), ••• , x~) are a set of nonstochastic vectors (q X 1) and N- I L~= I x~). The covariance matrix is
(4)
tI> tI> tI>
=,.,
i(2)
=
•
Consider a linear combination of X~l>, say Uti> = ex' X~I). Tht:n Uti> has variance ex'''' ex and expected value
(5)
509
12.6 LINEARLY RELATED EXPECTED VALUES
The mean expected value is (1/N)L,~=I0'V.. = a'T, and the mean sum of squares due to X(2) is N
fz L
(6)
N
(0'V>-a'T)2=fz
>=1
L
a'P(x;)-i l2 »)(x;)-i l2 »)'p'a
>=1
= a'pS22P'a.
We can ask for the linear combination that maximizes the mean sum of squares relative to its variance; that is, the linear combination of depentknt variables on which the independent variables have greatest effect. We want to maximize (6) subject to a' W a = 1. That leads to the vector equation
(7) for
K
satisfying
I PS22P' -
(8)
K
wi =
o.
Multiplication of (7) on the left by a' shows that a 'PS22P' a = K for a and K satisfying a' Wa = 1 and (7); to obtain the maximum we take the largest root of (8), say K I • Denote this vector by a(1), and the corresponding random variable by VI .. = a(1)' X~I). The expected value of this first canonical variable is 0'V1 .. =a(I)'[p(x~)-i(2»)+Tl. Let a(l)'p=k",(l)', where k is determined so 1=
(9)
fz
E(",(1)'x~) - ~ E",(1),x~;»)2 ~=1
>=1 N
=
fz L
",(l),(x~) -i(2»)(X~) _i (2 »)'",lll
>=1
= ",(l)'S22·y(1).
-r;;.
-r;;
Then k = Let VI .. = ",(l),(x~) - i(2»). Then ,cVI .. = V~I) + all)'T. Next let us obtain a linear combination V> = a 'X~I) that has maximum effect sum of squares among all linear combinations with variance 1 and uncorrelated with VI .. ' that is, 0 = ,cCU.. - ,cV.. )(Vld> - ,cUI .. )' = a'Wa(l). As in Section 12.2, we can set up this maximization problem with Lagrange multipliers and find that a satisfies (7) for some K satisfying (8) and a 'Wa = 1. The process is continued in a manner similar to that in Section 12.2. We summarize the results. The jth canonical random variable is ~d> = a(j)'X~I). where a lil satisfies (7) for K = Kj and aU)'Wa U ) = 1; KI ~ K2 ~ .. , ~ Kpi are the roots of (8).
510
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
We shall assume the rank of P is PI ~P2' (Then K p , > 0.) ~ has the largest effect sum of squares of linear combinations that have unit variance and are C)AI I t d WI'th UId>"'" Uj-I.d>· Let 'Y (j) -- (1 I I VKj uncorreae ... a (j) , I l j- - a (/)1 'f,an d Vjd> = 'Y U ) I(X~) - i(~»). Then
(10)
(11 )
(12)
i"* j.
(13)
4>=l, ... ,N,
(14)
(15)
nI L N
(
I',/> -
1 N ) N1 LtV)( I'" I',/> - N L v~ = I. I
~=
<1>= I
I
1)= I
The random canonical variates are uncorrelated and have variance 1. The expected value of each random canonical variate is a multiple of the corre sponding nonstochastic canonical variate plus a constant. The nonstochastic canonical variates have sample variance 1 and are uncorrelated in the sample. If PI> P2' the maximum rank of Pis P2 and Kp,+1 = ... = K p , = O. In that case we define A I =(a(1), ... ,a(P'» and A 2 =(a(p,+I) ... ,a(p,», where a (1), •.. , alp,) (corresponding to positive K'S) are defined as before and a(p:+l\ ... , alp,) are any vectors satisfying a(j)lwa(j) = 1 and a(j)lwa(i) = 0, i"* j. Then
CUJi) =
8iV~)
+ Vi' i = 1, ... , P2' and CUJiI =
Vi'
i=
P2
+
1, .. ',Pl'
In either case if the rank of P is r ~ min(Pl' P2)' there are r roots of (8) that are non~ero and hence C UJi) = 8iV~) + Vi for i = 1, ... , r.
511
12.6 LINEARLY RELATED EXPECTED VALUES
12.6.2. Estimation Let xP), ... , x~) be a set of observations on XP), ... , X~) with the probability structure developed in Section 12.6.1, and let X~2), ••• , x~) be the set of corresponding independent variates. Then we can estimate 'T, ~, and "IJ1 by N
T=
(16)
~ L x~) =i(1), >=1
N
(18)
= SI2 S 22 1,
13 =A12A221
(17)
W= ~ L [x~) -
i(l) -
13 (x~) -
i(2»)][ x~) -
i(l) -
13 (x~) -
i(2»)
r
>=1
= ~(All
-AI2A22IA21)
= SlI
- SI2 S;IS 21 ,
where the A's and S's are defined as before. (It is convenient to divide by n = N - 1 instead of by N; the latter would yield maximum likelihood estimators.) The sample analogs of (7) and (8) are
(19)
(20)
0= 1 pS22p' -
kWI
= IS I2 S 22 1S21 -k(SIl -SI2S22IS21)1.
The roots k 1 ;:: '" ;:: k 1', of (20) estimate the roots K 1 ;:: .,. ;:: KI" of (8), and the corresponding solutions a(1), ... , a(p,) of (19), normalized by a(i)'Wa(i) = 1, estimate 0:(1), .•• , o:(p,). Then c(J) = 0/ vk)p'a(J) estimates 'Y(J), and nj = a(J)'i(l) estimates llj' The sample canonical variates are a(j),x~) and c(j)'(x~) - i(2», j = 1, ... , PI' ¢ = 1, ... , N. If PI> P2' then PI - P2 more aUl's can be defined satisfying a(J)''Va(J) = 1 and a(J)''Va(i) = 0, i j.
*"
12.6.3. Relations Between Canonical Variates In Section 12.3, the roots II
(21)
~
...
~ I pt
were defined to satisfy
512
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
Since (20) can be written (22)
we see that I? = kJ{1 + k i ) and k; = If /(1 in Section 12.3 satisfies (23)
m, i = 1, ... , Pl. The vector a(i)
0= (S12Sils21 -lfSI1)a(i) = (S12 S;IS2l - 1 = 1
:;k;
SII )a(i)
~k; [S I2 S;IS21 -k;(SI1 -S12 SZ2 IS21)]a(i),
which is equivalent to (19) for k = k;. Comparison of the normalizations a(i)'Slla(i)=1 and a(i)'(SII-SI~S;IS~I)a(i)=1 shows that ii(i)= (1/ -Ina(i). Then c(J) = (1/ {k;)S;IS 2Ia(j) = c(il. . We see that canonical variable analysis can be applied when the two vectors are jointly random and when one vector is random and the other is nonstochastic. The canonical variables defined by the two approacpes are the same except for norma:ization. The measure of relationship between corresponding canonical variables can be the (canonical) correlation or it can be the ratio of "explained" to "unexplained" variance.
V1
12.6.4. Testing Rank
The number of roots K j that are different from 0 is the rank of the regression matrix p. It is the number of linear combinations of the regression variables that are needed to express the expected values of X~l). We can ask whether the rank is k (1 ~k~PI if PI ~P2) against the alternative that the rank is greater than k. The hypothesis is
(24) The likelihood ratio criterion [Anderson (1951b)] is a power of PI
(25)
PI
n (l+k;)-I= i-k+! n (1+/:). i-k+!
Note that this is the same criterion as for the case of both vectors stochastic (Section 12.4). Then PI
(26)
-[N-t(p+3)]
L i-k+l
log(l-/:)
12.6
513
LINEARLY RELATED EXPEC'TED VALUES
has approximately the X 2 -distribution with (p I - k XPZ - k) degrees of freedom. The determination of the rank as any number between 0 and PI can be done as in Section 12.4. 12.6.5. Linear Functional Relationships
The study of Section 12.6 can be carried out in other terms. For example, the balanced one-way analysis of variance can be set up as
a=l, ... ,m,
(27)
j=l, ... ,I,
a= l. ... ,m,
(28)
where 0 is q X PI of rank q « PI)' This is a special case of the model of Section 12.6.1 with c/J = 1, ... , N, replaced by the pair of indices (a, j), X~l) = Yaj , 'T = ..... , and P(x~) - x(2») = Va by use of dummy variables as in Section 8.8. The rank of (v I' ... , v m ) is that of p, namely, r = PI - q. Thc!re are q roots of (8) equal to 0 with m
(29)
pS22P'=1
L
vav~.
a~l
The model (27) can be interpreted as repeated observations on Va + ..... with error. The component equations of (28) are the linear functional relationships. Let Ya = (1/z)I:~_1 Yaj and Y = O/m)I:';_1 Ya· The sum of squares for effect is m
(30)
H =I
L
(Ya - Y)(Ya - Y)' = n~S22~'
a~l
with m - 1 degrees of freedom, and the sum of squares for e"or is m
(31)
G=
I
L L
(Yaj-Ya)(Yaj-Y a)' =nW
a~l j~l
with m(l- 1) degrees of freedom. The case PI < P2 corresponds to PI < [. Then a maximum likelihood estimator of 0 is
(32)
514
CANON ICAL CORRELATlONS AND CANONICAL VARIABLE'>
and the maximum likelihood estimators of va are
a= 1, ... ,n.
(33)
The estimator (32) can be multiplied by any nonsingular q X q matrix on the left to obtain another. For a fuller discussion, see Anderson (1984a) and Kendall and Stuart (1973).
12.7. REDUCED RANK REGRESSION Reduced rank regression involves estimating the regression matrix P in # x( 1) IX(2) = px(2) by a matrix of preassigned rank k. In the limited-information maximum likelihood method of estimating an equation that is part of a system of simultaneous equations (Section 12.8), the regres~ioll matrix is assumed to be of rank one less than the order of the matrix. Anderson (1951a) derived the maximum likelihood estimator of P when the model is
P
XaO ) = 'T +
(1 )
A(XI2) ~
a
_i (2 )) + Z0 '
a=I, ... ,N,
xW
the rank of P is specified to be k (5, PI)' the vectors X\2), ... , are nonstochastic, and Za is normally distributed. On the basis of a sample Xl"'" x N , define i by (2) of Section 12.3 and A, A, and t by (3), (4), and (5). Partition A=diag(A l,A 2 ), A=(A 1,A 2 ), and t=Ct l ,t2 ), where AI' A[, and tl have k columns. Let <1>1 = Al(Jk - A;)-~. Definition 12.7.1 (Reduced Rank Regression) The reduced rank regressicn estimator in (1) is (2) A
A
I
A
A
A
whereB=Il212Z and Ii i =I l1 -BI 22 B'.
The maximum likelihood estimator of P of rank k is the same for X(1) and X!2l normally distributed because the density of X=(X I 1)',X(2),), factors as
Reduced rank regression has been applied in many disciplines, induding econometrics, time series analysis, and signal processing. See, for example, Johansen (1995) for use of reduced rank regression in estimation of cointegration in economic time series, Tsay and Tiao (1985) and Ahn and Reinsel (1988) for applications in stationary processes, and Stoica and Viberg (1996)
12.8
515
SIMULTANEOUS EQUATIONS MODELS
for utilization in signal processing. In general the estimated reonced rank regression is a better estimator in a regression model than the unrestricted estimator. In Section 13.7 the asymptotic distribution of the reduced rank regression estimator is obtained under the assumptions that are sufficient for the asymptotic normality of the least squares estimator B = i 12 iii. The asymptotic distribution of Bk has been obtained by Ryan, Hubert, Carter, Sprague, and Parrott (1992), Schmidli (1996), Stoica and Viberg (1996), and Reinsel and Velu (1998) by use of the expected Fisher information on the assumption that Z", is normally distributed. Izenman (1975) suggested the term reduced rank regression.
12.8. SIMULTANEOUS EQUATIONS MODELS 12.8.1. The Model Inference for structural equation models in econometrics is related to canonical correlations. The general model is
(1)
By,
+ fz, = Up
t
= 1, ... ,T,
where B is G X G and r is G X K. Here Y, is composed of G jointly dependent variables (endogenous), Z, is composed of K predetermined variables (emgenous and lagged dependent) which are treated as "independent" variables, and U j consists of G unobservable random variables with
(2)
,cU, = 0,
,cUtU; = I.
We requir<.; B to be nonsingular. This model wa~ initiated by Haavelmo (1944) and was developed by Koopmans, Marschak, Hurwicz, Anderson, Rubin, Leipnik, et aI., 1944-1954, at the Cowles Commission for Research in Economics. Each component equation represents the behavior of some group (such as consumers or producers) and has economic meaning. The set of structural equations (1) can be solved for Y, (because B is nonsingular ):
Y, = I1z, + v"
(3) where
(4) with
(5)
,cV, = 0,
516
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
say. The equation (3) is called the reduced form of the model. It is a multivariate regression modeL In principle, it is observable. 12.8.2. Identification by Specified Zeros The structural equation (1) can be multiplied on the left by an arbitrary nonsingular matrix. To determine component equations that are economically meaningful, restrictions must be imposed. For example, in the case of demand and supply the equation describing demand IDly be distinguished by the fact that it includes consumer income and excludes cost of raw materials, which is in the supply equation. The exclusion of the latter amounts to specifying that its coefficient in the demand equation is O. We consider identification of a structural equation by specifying certain coefficients to be O. It is convenient to treat the first equation. Suppose the variables are numbered so that the first G 1 jointly dependent variables are included in the first equation and the remaining G 2 = G - G 1 are not and the first K 1 predetermined variables are included and K 2 = K - Klare excluded. Then we can partition the coefficient matrices as
(6)
(B
n=[P'
o
,,/'
where the vectors p, 0, ,,/, and 0 have G 1, G 2 , K 1 , and K2 components, respectively. The reduced form is partitioned c(,nformally into G 1 and G 2 sets of rows and Kl and K2 sets of columns:
(7) The relation between B,
r,
and II can be expressed as
The upper right-hand corner of (8) yields
(9) To determine need (10)
P (G 1 X
1) uniquely except for a constant of proportionality we
12.8 SIMULTANEOUS EQUATIONS MODELS
517
This implies (11)
Addition of G 2 to (11) gives the order condition
(12) The number of specified D's in an identified equation must be at least equal to 1 less than the number of equations (or jointly dependent variables). It can be shown that when B is nonsingular (10) holds if and only if the rank of the matrix consisting of the columns of (B f) with specified D's in the first row is G - 1. 12.8.3. Estimation of the Reduced Form The model (3) is a typical multivariate regression model. The observations are
(YI),,,.,(Yr). ZI zr
(13)
The usual estimators of II and 0 (Section 8.2) are
(14)
P
=
LT Y,z; (TL Z,Z; )-1 , t~1
(IS)
~
1
0=
T
t~1
T
L (yl-PZI)(yl-PZ,)'. t~1
These are maximum likelihood estimators if the v, are normal. If the z, are exogenous (regardless of normality), then
(16)
$ vec P
= vec II,
C(vecP} =A- 1 ®O,
where T
(17)
A =
L Z,Z; '~1
and vec(d l , ..• , d m ) = (d;, ... , d",) , . If, furthermore, the l', are normaL then P is normal and TO has the Wishart distribution with covariance matrix 0 and T - K degrees of freedom.
518
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
12.8.4. Estimation of the Coefficients of an Equation First. consider the estimation of the vector of coefficients GI-l.Let
/3
when K2 =
(18 )
be partitioned as the equation
n.
Then the probability is 1 that rank(P12) = G I
-
1 and
(19) has a nontrivial solution that is unique except for a constant of proportionality. This is the maximum likelihood estimator when the disturbance terms are normal. If K~ ~ G 1, then the probability is 1 that rank(P12) = G 1 and (19) has only the trivial solution ~ = 0, which is unsatisfactory. To obtain a suitable estimator we find ~ to minimize ~'PI~ in som: sense relative to another function of ~'. Let z, be partitioned into subvectors of Kl and K2 components:
( 20)
(21 )
(22) Let y, and 0 be partitioned into Gland G 2 components:
(23)
y, = (y~l) y;2) ) ,
(24)
0= ( 011 O~I
°l~ ).
0 22
12.8
519
SIMULTANEOUS EQUATIONS MODELS
l'iIow set up the multivariate analysis of variance table for y(l): Sum of Squares
Source T
Y?)Z~I)' A III Z~l)y;I)'
L s.I=1
PI2 A 22'I P ;2 T
Error
L (y~1) -
PI Izjl)
-
P I2 zj2»(y,(I)
-
P II zjl)
-
P I2 Z~2»'
1=1 T
Total
Ly;t)y;l), 1=1
The first term in the table is the (vector) sum of squares of y;l) due to the effect of zP). The second term is due to the effect of Z~2) beyond the effect of z?). The two add to (PAP')w which is the total effect of z" the predetermined variables. We propose to find the vector ~ such that effect of Z~2) and ~'y;l) beyond the effect of zP) is minimized relative to the error sum of squares of ~'y;l). We minimize
(25)
~'( P 12 S 22 .1P;2 )~
(~' Pl2 )S22'I(~' P I2 ),
~'nll~
~'nll~
Tn
= L;= I YIY, - P/P'. This estimator has been called the least variance ratio estimator. Under normality and based only on the 0 restrictions on
where
the coefficients of this single equation, the estimator is maximum likelihood and is known as the limited-information maximum likelihood (LIML) estimator [Anderson and Rubin (1949)]. The algebra of minimizing (25) is to find the smallest root, say II, of (26)
and the corresponding vector satisfying (27) The vector is normalized according to some rule. A frequently used rule is
LO
520
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
set one (nonzero) coefficient equal to 1, say the first, ~] = 1. If we write
~ =(~* ),
(28)
(29)
ll12
= lli2
( n]2 )
~=(~*), ) Pu=(P12 Pi2 '
'
) Ou= (WII ~(]) 0* ' A
(30)
A
W(l)
11
then (27) can be replaced by the linear equation
The first component equation in (27) has been dropped because it is linearly dependent on the other equations [because v is a root of (26)]. 12.8.5. Relation to the Linear Functional Relationship
We now show that the model for the single linear functional relationship (q = 1) is identical to the model for structural equations in the special case that G 2 0 (y?) y,) and z~l) :; 1 (K j 1). Write the two models as
=
=
=
a=l, ... ,n,
(32)
j=I, ... ,k,
where
(33) and t = 1, ... ,T,
(34)
where II = (II] Il 2 ). The correspondence between the models is p
<->
G = G],
(35) (36)
(37)
(a,j)
<->
t,
nk <-> T,
11" <-> 0,
We can write the model for the linear functional relationship with dummy
521
12.8 SIMULTANEOUS EQUATIONS MODELS
variables. Define
o (38)
<--
a th position,
a= 1, ... ,11- 1.
o (39)
Then a= 1, ... ,n,
(40)
where j may be suppressed. Note ( 41)
VII
= - ( V I + ... + VII _ I ) .
The correspondence is S
(42)
UJ
.... ,
(43) n-l<->K~,
( 44)
Let P
=
(PI P 2 ). In terms of the statistics we have the correspondence
( 46)
iJ..=X<->ji,
(47) The effect matrix is n
(48)
H=k
L ,,~1
(x,,-x)(x,,-x)'
<->P2 A 22I P;,
522
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
and the error matrix is n
(49)
G=
k
T
L L ,,~l j~
(xaj-ia)(xaj-ia)' <->TO=
I
L
(Yr-Pzr)(Yr-Pzr)'·
r~l
Then the estimator B of the linear functional relationship for q = 1 is identical to the LlML estimator [Anderson (1951b), (1976), (1984a)]. 12.8.6. Asymptotic Theory as T ..... 00 We shall find the limiting distribution of v'T(~*
- /3*) defined by (28) and
(31) by showing that ~* is asymptotically equivalent to
( 50) This derivation is essentially the same as that given in Anderson and Rubin (1950) except for notation. The estimator defined by (50), known as the two stage least squares (TSLS) estimator, is an approximation to the LIML estimator obtained by dropping the terms VOrl and vW(l) from (31). Let ~* = ~tIML' We assume the conditions for v'T(P - n) having a limiting normal distribution. (See Theorem 8.11.1.) Lemma 12.8.1.
Suppose Cl/T)A ..... AO, a positive definite matrix, as T .....
x. lhen v = O/l/T), where v is the smallest root of (26).
Proof Let
(51)
Pn = v'T (P l2 -
/3' Pl2 S 22.1 P I2 /3
/3' 0 11 /3
/3' U l2 = 0
/3' [ U I2 + (1/v'T)P I2 ] S22-l [ U l2 + (l/v'T )Pl2 ], /3 /3' 0 11 /3
Since
(52)
the lemma follows.
U l2 ). Then because
•
523
12.8 SIMULTANEOUS EQUATIONS MODELS
The import of Lemma 12.8.1 is that the difference between the LIML estimator and the TSLS estima.tor is O/l/T). We have
•
- (ptz S22'1 Pi.; -
= [( Pf2S2HPii)
vn ll )
-I·
(Pi2 S22.1 P'12 - VW(I»)
-I - (Pi2 S 22'I Pi{ •
+ (Pi'2S22'IPi.; - vn ll )
villi rlj Pi2 S 22'IP'17
-I VW(I)
= - v( Pi'2S22 . 1Pin -I il l1 ( Pi'2S22'IPi{ •
+ v( Pi2S22'IPi2 - vOl1)
villi) -I Pi2S22'1 p'J2
-I
00(1)
Consider
(54)
P;2 + Pf{/3* = Pi2/3 =Ai}.1
Note that ~~l5 (y', O)z, = u 1,.
-
T
T
/=1
/=1
L zj2.I)y{I)/3 =Az-l 1 L zj2'I)U I
I'
/3* = -(P;iS22.IPtz)-lpi2S22'IP{2/3 and (/3',O)YI
+
Theorem 12.8.1. Under the conditions of Theorem 8.11.1
p
Proof The theorem follows from (55),
S22.1
-+
Sf2 ,1'
and P 12
-+
fIJ2 .
•
Because of the correspondence between the LIML estimator and the maximum likelihood estimator for the linear functional relationship as outlined in Section 12.7.5, this asymptotic theory can be tr;mslated for the latter.
524
CANONICAL CORRELATIONS AND CANONICAL VARIABLES
Suppose the single linear functional relationship is written as
(57)
0= WVa =
(1
a= 1, ... ,n,
where
a= 1, ... ,n.
(58)
Let n ( - K) be fixed, and let the number of replications k ---> 00 (corresponding to T /K ---> 00 for fixed K). Let a 2 = Wl}I~. Since TI 12 A 22 . 1TI'12 corresponds to kL:~lV:V:', ~* here has the approximate distribution
(59) Although Anderson and Rubin (1950) showed that vat1 and VW(I) could be dropped from (31) defining ~tIML and hence that ~tsLS was asymptotically equivalent to ~tsLS' they did not explicitly propose ~tsLS. [As part of the Cowles Commission program, Chernoff and Divinsky (1953) developed a computational program of ~LIML.J The TSLS estimator was proposed by Basmann (J 957) and Theil (J 96 n. It corresponds in thc linear functional relationship setup to ordinary least squares on the first coordinate. If some other coefficient of ~ were set equal to one, the minimization would be in the direction of that coordinate. Con~idcr the gcncral linear fUllctional relationship when the error covariance matrix is unknown and there are replications. Constrain B to be
(60)
B=(Im
B*).
Partition
(61 ) Then the least squares estimator of B* is
12.8 SIMULTANEOUS EQUATIONS MODELS
For n fixed and k
~ 00
A
and Hi's
I'
~
525
H* and
[See Anderson (1984b).] It was shown by Anderson (1951c) that the q smallest sample roots are of such a probability order that the maximum likelihood estimator is asymptotically equivalent, that is, the limiting distribution of Ik vecOl~L - H*) is the right-hand side of (63). 12.8.7. Other Asymptotic Theory In terms of the linear functional relationship it may be mOle natural to consider 11 ~ 00 and k fixed. When k = 1 and the error covariance matrix is o·2Ip , Gieser (1981) has given the asymptotic theory. For the simultaneous equations modd, the corresponding conditions are that K2 ~ "', T ~ :xl, and K2/T approaches a positive limit. Kunitomo (1980) has given an asymptotic expansion of the distribution in the case of p = 2 and m = q = 1. When n ~ 00, the least squares estimator (i.e., minimizing the sum of squares of the residuals in one fixed direction) is not consistent; the LlML and TSLS estimators are not asymptotically equivalent. 12.8.8. Distributions of Estimators Econometricians have studied intensively the distributions of TSLS and LIML estimator, particularly in the case of two endogenous variables. Exact distributions have been given by Basmann (1961),(1%3), Richardson (968), Sawa (1969), Mariano and Sawa (1972), Phillips (\980), and Anderson and Sawa (1982). These have not been very informative because they are usually given in terms of infinite series the properties of which are unknown or irrelevant. A more useful approach is by approximating the distributions. Asymptotic expansions of distributions have been made by Sargan and Mikhail (1971), Anderson and Sawa (1973), Anderson (1974), Kunitomo (1980), and others. Phillips (1982) studied the Pade approach. See also Anderson (1977). Tables of the distributions of the TSLS and LIML estimators in the case of two endogenous variables have been given by Anderson and Sawa (1977),(1979), and Anderson, Kunitomo, and Sawa (1983a). Anderson, Kunitomo, and Sawa (1983b) graphed densities of the maximum likelihood estimator and the least squares estimator (minimizing in one direction) for the linear functional relationship (Section 12.6) for the case
526 p
= 2,
CANONICAL CORRELATIONS AND CANONICAL VARlABLES
m = q = 1, W = a 2wo and for various values of {3, n, and
( 64)
PROBLEMS 12.1. (Sec. 12.2) Let Zo = Zio = 1, a = 1, ... , n, and ~ = \3. Verify that Relate this result to the discriminant function (Chapter 6). 12.2. (Sec. 12.2)
a(l)
= 1- 1 \3.
Prove that the roots of (14) are real.
12.3. (Sec. 12.2)
U=a'X\Il, V=-y'X(2), CU 2 =1=CV 2, where a and -yare vectorS. Show that choosing a and -y to maximize C UV is equivalent to choosing a and -y to minimize the generalized variance of (U V). lb) Let X' = (X(I)' X(2)' X(3)'), C X = 0,
(e) (d)
le) If)
U=a'X(I), V=-y'X(23), W=\3'X(.1), CU 2 =CV2=CW2=1. Consider finding a, -y, \3 to minimize the generalized variance of (U, V, W). Show that this minimum is invariant with respect to transformations X*(i) = AIXU),IAII '" O. By using such transformations, transform I into the simplest possible form. In the case of X(I) consisting of two components, reduce the problem (of minimizing the generalized variance) to its simplest form. In this case give the derivative equations. Show that the minimum generalized variance is 1 if and only if 112 = 0, 113 = 0, 123 = O. (Note: This extension of the notion of canonical variates does not lend itself to a "nice" explicit treatment.)
12.4. (Sec. 12.2)
Let XO)
=AZ
+ yO),
X(2) = BZ + y(2),
527
PROBLEMS
where y(1), y(2), Z are independent with mean zero and covariance matrices I with appropriate dimensionalities. Let A = (a l , ... , ak)' B = (b l , •.. , bk ), and suppose that A' A, B' B are diagonal with positive diagonal elements. Show that the canonical variables for nonzero canonical correlations are proportional to a~X{l), b;X(2). Obtain the canonical correlation coefficients and appropriate normalizing coefficients for the canonical variables. 12.5. (Sec. 12.2) Let AI;;:>: A2;;:>: ... ; :>: Aq > 0 be the positive roots of (14), where 111 and 122 are q X q nonsingular matrices. (a) What is the rank of I12? (b) Write ni_IA~ as the determinant of a rational function of III' 1 12 , 1 21 , and 1 22 , Justify your answer. (c) If Aq = I, what is the rank of
12.6. (Sec. 12.2) Let 11\ = (1 - g )Ip , + gE p,E~" I zz = (1 - h)Ip2 + he p, E~2' 112 = kEp,E~2' where -1/(PI - 1) AT and aU + 1) -> a(l) if a(O) is such that a'(O)Iua(l) '" O. [Hint: U~C IIII II2Iz21 121 = A A2A - I.J 12.10. (Sec. 12.6)
Prove (9), (10), and (11).
12.11. Let Al ~ A2;;:>: ... ; :>: Aq be the roots of 1 II - AIzl q X q positive definite covariance matrices.
=
0, where II and 12 are
(a) What does Al = Aq = 1 imply about the relationship of II and I2? (b) What does Aq> 1 imply about the relationships of the ellipsoids x'I;-lx = c and x'I2'ix = c? (c) What does Al > 1 and Aq < 1 imply about the relationships of the ellipsoids x'I;-lx=c and x'I2'lx=c? For q = ~ express the criterion (2) of Section 9.5 in terms of canonical correlations.
12.12. (Sec. 12.4)
12.13. Find the canonical correlations for the data in Problem 9.11.
CHAPTER 13
The Distributions of Characteristic Roots and Vectors
13.1. INTRODUCTION In this chapter we find the distribution of the sample principal component vectors and their sample variances when all population variances are 1 (Section 13.3). We also find the distribution of the sample canonical correlations and one set of canonical vectors w'1en the two set!> of original variates are independent. This second distribution will be shown to be equivalent to the distribution of roots and vectors obtained in the next section. The distribution of the roots is particularly of interest because many invariant tests are functions of these roots. For example, invariant tests of the general linear hypothesis (Section 8.6) depend on the sample only through the roots of the deterrninantal equation
(1 ) If the hypothesis is true, the roots have the distribution given in Theorem 13.2.2 or 13.2.3. Thus the significance level of any invariant test of the general linear hypothesis can be obtained from the distribution derived in the next section. If the test criterion is one of the ordered roots (e.g., the largest root), then the desired distribution is a marginal distribution of the joint distribution of roots. The limiting distributions of the roots are obtained under fairly general conditions. These are needed to obtain other limiting distributions, such as the distribution of the criterion for testing that the smallest variances of
An Introduction to Multivariate Statistical Analysis. Third Edition. By T. W. Anderson ISBN 0-471-36091-0 Copyright © 2003 John Wiley & Sons. Inc.
528
13.2 THE CASE OF TWO WISHART MATRICES
529
principal components are equal. Some limiting distributions are obtained for elliptically contoured distributions.
13.2. THE CASE OF TWO WISHART MATRICES 13.2.1. The Transformation Let us consider A* and B* (p xp) distributed independently according to WCI,m) and W(I,n) respectively (m,n ~p). We shall call the roots of
(1)
IA* -IB*1 = 0
the characteristic roots of A* in the metric of B* and the vectors satisfying
(2)
(A* -IB*)x* =0
the characteristic vectors of A* in the metric of B*. In this section we shall consider the distribution of th.::se roots and vectors. Later it will be shown that the squares of canonical correlation coefficients have this distribution if the population canonical correlations are all zero. First we shall transform A* and B* so that the distributions do not involve an arbitrary matrix h. Let C be a matrix such that CIC' = I. Let
(3)
A = CA*C',
B=CB*C'.
Then A and B are independently distributed according to W(I, m) and W(J, n) respectively (Section 7.3.3). Since
IA -IBI = ICA*C' -ICB*C'1 =\C(A* -IB*)C'\ = ICI·IA* -IB* I·IC'I, the roots of (1) are the roots of
(4)
IA -IBI = o.
The corresponding vectors satisfying
(5)
(A-IB)x=O
satisfy
(6)
0= C1(A -IB)x =CI(CA*C' -/CB*C')x = (A* -IB*)C'x.
Thus the vectors x* are the vectors C'x.
530
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECTORS
It will be convenient to consider the routs uf \A - f( A + B) \
(7)
=0
and the vectors y satisfying
( S)
[A - f(A +B)]y = O.
The latter equation can be written
o = (A -
(9)
Since the probability that is
fA -.fB) y = [( 1 - f) A -
f = 1 (i.e.,
that \ -
fB] y.
B\ = 0) is 0, the above equation
( 10) Thus the roots of (4) are related to the roots of (7) by I = flO - f) or = II( I + [), and the vectors satisfying (5) are equal (or proportional) to those satisfying (S·). We now consider finding the distribution of the roots and vectors satisfying 0) and (S). Let the roots be ordered fl > f2 > ... > fp > 0 since the probability of two roots being equal is 0 [Okamoto (1973)]. Let
f
fl 0
0
0
f2
0
0
0
fp
F=
(11)
Suppose the corresponding vector solutions of (S) normalized by y'(A+B)y=1
(12) are
YI"'"
Yr' These vectors must satisfy
(13)
because y;AYj = fjy;(A + B)Yj and y;AYj if (13) holds (fi *- f/ Let the p x p matrix Y be (14)
=
f.y;(A
+ B)Yj' and this can be only
13.2 THE CASE OF TWO WISHART MATRICES
531
Equation (8) can be summarized as
(15)
AY= (A + B)YF,
and (12) and (13) give
(16)
Y'(A + B)Y=I.
From (15) we have
(17)
Y'AY= Y'(A +B)YF=F.
Multiplication of (16) and (17) on the left by (y,)-I and on the right by y- I gives
A +B = (y,)-Iy-I,
(18)
A = (Y') -I FY- I •
Now let y- I = E. Then
(19)
A+B=E'E, A =E'FE, B=E'(I-F)E.
We now consider the joint distribution of E and F. From (19) we see that E and F define A and B uniquely. From (7) and (11) and the ordering fl> '" > fp we see that A and B define F uniquely. Equations (8) for f= fi and (12) define Yi uniquely except for multiplication by -1 (i.e., replacing Yi by - yJ Since YE = I, this means that E is defined uniquely except that rows of E can be mUltiplied by -1. To remove this indeterminacy we require that eil ~ O. (The probability that eil = 0 is 0.) Thus E and F are uniquely defined in terms of A and B. 13.2.2. The Jacobian To find the density of E and F we substitute in the density of A and B according to (19) and multiply by the Jacobian of the transformation. We devote this subsection to finding the Jacobian
(20)
a(A,B)1
l a(E,F) .
Since the transformation from A and B to A and G = A + B has Jacobian unity, we shall find
(21)
a(A,G)I=la(A,B)1 la(E,F) a(E,F)'
532
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECTORS
First we notice that if x" = faCYl' ... ' Yn), a = 1, ... , n, is a one-to-one transformation, the Jacobian is the determinant of the linear transformation
(22) where dx a and dY/3 are only formally differentials (i.e., we write these as a mnemonic device). If fa(Yl' ... ' Yn) is a polynomial, then afa/ aY/3 is the coefficient of Y; in the expansion of fa(Yl + yj, ... , Yn + Y:) [in fact the coefficient in the expansion of f/YI' ... 'Y/3-I'Y/3+Y;'Y/3+I, ... ,Yn)]. The elements of A and G are polynomials in E and F. Thus the c'erivative of an element of A is the coefficient of an element of E* and F* in the expansion of (E + E* )'(F + F* )(E + E*) and the derivative of an element of G is the coefficient of an clement of E* and F* in the expansion of (E + E* )'(E + E*). Thus the Jacobian of the transformation from A, G to E, F is the determinant of the linear transformation
(23)
dA = (dE)'FE
(24)
dG
+ E'(dF)E + E' F(dE),
= (dE)'E+E'(dE).
Since A and G (dA and dG) are symmetric, only the functionally independent component equations above are used. Multiply (23) and (24) on the left by E,-I and on the right by E- l to obtain (25)
E,-l(dA)E- 1 =E'-I(dE)'F+dF+F(dE)E- 1 ,
(26)
E'-l(dG)E- l =E,-I(dE)'
+ (dE)E- 1
It should be kept in mind that (23) and (24) are now considered as a linear transformation without regard to how the equations were obtained. Let
(27)
E,-l(dA)E- l =dA,
(28)
E,-I(dG)E- 1 =dG,
(29)
( dE) E- I =, dW.
Then
(30)
dA = (dW)'F +dF+F(dW),
(31)
dG =dW' +dW.
533
13.2 THE CASE OF TWO WISHART MATRICES
The linear transformation from dE, dF to dA, dG is considered as the linear transformtion from dE, dF to dW, dF with determinant IE-III' = lEI -I' (because each row of dE is transformed by E- I ), followed by the linear transformation from dW, dF to dA, dG, followed by the linear transformation from dA,dG to dA=E'(dA}E,dG=E'(dG)E with determinant IEl p + l . IEI P + 1 (from Section 7.3.3); and the determinant Jf the linear transformation from dE, dF to dA, dG is the product of the determinants of the three component transformations. The transformation (30), (3]) is written in components as da jj = dlj
+ 2/j dw j /
dUi} = J; dWjI
(32) dg jj
+ j; dw,j ,
i
<J,
= 2dwjj ,
dgi} = dWjj
+ dw jj ,
1<',.
The determinant is
da ii
dW jj
I
2F
dW jj (i
<j)
dW jj (i
0
> j)
0
0
2I
0
0
dUij (i
<j)
0
0
M
N
dg'j (i
<j)
0
0
I
I
dgi ,
(33)
dlj
=I~ 2FI·IM 2I I
~1=2PIM-Nl,
where
(34)
11
0
0
0
:, , ,
0
o II , 0 0 :, - - -- -- - - -- - -,-- -- - - - - -- - -- -, _. - -
o ...
0:
12
'"
0 0
0:
M=
I
,I
o
0
, 0
12:,
0
- - - - - - - - - - - - 1 - - - ___ - ______ -I- - - - - -'- - - - • 1 I I _~
o
I ___ _ _________ I______________ ' L ____ ~
0
: I
~
0
0
I
: ,
Ip-I
.
534
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECTORS
and (35)
0
f"
0
0
:
0
I I
I I
o fp 0 0: 0 ------------r-------------r----'---o ... 0 : f3 ... 0 : : 0 I
N=
I
I I I
o
I I I
0:0
I I I
fp::O ~
~
_
------------r-------------r----~---. I I I I ____ I ___ _ __________ IL _____________ L
o
0
: 0
0 : : fp
I
I
I
Then
IM-Ni = n(/;-fj).
(36)
i<j
The determit1ant of the linear transformation (23), (24) is (37)
IEI-PIEIP+IIEIP+12pn (1; - fj) = 2 P IEl p + 2 n i0
Theorem 13.2.1. value of (37).
U - fj).
i0
The Jacobian of the transformation (19) is the absolute
13.2.3. The Joint Distribution of the Matrix E and the Roots
The joint density of A and B is
where
(39) Therefore the joint density of E and F is
(40)
C,IE' FEI ~(m-p-l )IE'(I - F) EI ~(n-p-J) 'e- .\tr E'E2 P IE' EI ~(p+2) n (Ii - fj). i<j
535
13.2 THE CASE OF TWO WISHART MATRICES
Since IE'FEI = !E'I·IFI·IEI = IFI·IE'EI =Of_I.t;IE'EI and IE'(I-F)EI = II-FI·IE'EI = Of_I(1- f)IE'_'!:I, the density of E and F is
( 41) p
i-I
n (1 - .t;)t(n-p-I ni<j (J; --fJ, p
2 P C I IE'EI t(m+n-p le- tlrE'E nfit(m-p-l 1
1
i-I
Clearly, E and F are statistically independent because the density factors into a function of E and a function of F. To determine the marginal densities we have only to find the two normalizing constants (the product of which is 2 P C I ). Let us evaluate (42) where the integration is 0 < ei1 < 00, - 00 < eij < 00, j *- 1. The value of (42) is unchanged if we let - 00 < eil < 00 and multiply by 2- p • Thus (42) is
Except for the constant (2'IT)W', (43) is a definition of the expectation of the t(rn + n - p )th power of IE' EI when the eij have as density the function within brackets. This expected value is the t(rn + n - p )th moment of the generalized variance IE' EI when E' E has the distribution W(I, p). (See Section 7.5.) Thus (43) is I
(44)
(2 'IT ) 2P
,rP [-21(rn +n)] r p (12P )
I
22P(n+m-p)
.
Thus the density of E is
(45)
The density of fi is (41) divided by (45); that is, the density of .t; is
536 for 0 ~fp
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECTORS ~
...
~fl ~
1, where . 1T.jp'f
p
(47)
[Hm+n)]
C 2= f p (I'in ) fp (I'im ) fp (I'iP ).
The density of Ii is obtained from (46) by letting
(48) we have
dfi 1 dl i = (Ii + 1) 2 '
(49)
fi -
~=
I, -Ij (Ii + 1)(lj + 1)' 1
1- fi = Ii + 1 . Thus the density of I, is (50)
Theorem 13.2.2.
If A and B are distributed independently according to
WeI, m) and weI, n) respectively (m ~p, n ~p), the joint density of the roots of IA -IBI = 0 is (50) where C 2 is defined by (47). The joint density of Y can be found from (45) and the fact that the Jacobian is IYj-2 P. (See Theorem A.4.6 of the Appendix.) 13.2.4. The Distribution for A Singular The matrix A above can be represented as A = WI W{, where the columns of WI (p X m) are independently distributed, each according to N(O, I). We now treat the case of m <po If we let B + WIW{ = G = CC' and WI = CU, then the roots of
(51)
0= IA - f(A +B)I = IWIW{ - fGI
= ICUU'C' -fCC'1 = ICI·IUU' -fIpl'ICI
537
13.2 THE CASE OF TWO WISHART MATRICES
are the roots of (52)
We shall show that the nonzero roots fl > .. , > fm (these roots being distinct with probability 1) are the roots of (53) For each root
f *" 0 of (52) there
is a vector x satisfying
(54)
Multiplication by V' on the left gives (55)
0= V'(VV' - f/p)x
= (V'V-f/m)(V'x). Thus V' x is a characteristic vector of VV' and f is the corresponding root. It was shown in Section 8.4 that the density of V = V; is (for Ip - VV' positive definite or 1m - U * V; positive definite)
where p* = m, n* - p* - 1 = n - p - 1, and m* = p. Thus fl"'" f".n must be distributed according to (46) with p replaced by m, m by p, and n by n + m - p, that is, (57)
1Ttm'rm[!(m +n)] rm(~m)rm[Hm + n - p')]=rm-""(-;-tp-'-)
.fI [fl(p-m-l l{1- f;)t(n-p-ll] n(j; -fj)· .=1
'<j
Theorem 13.2.3. If A is ,iistributed as W1W;, where the m columns of WI are independent, each distribUled according to N(O, "I), m :s,p, and B is indepelldently distributed according to W(l;, n), n ~p, then the density of the nonzero roots of \A - f(A + B)\ = 0 is given by (57). These distributions of roots were found independently and at about the same time by Fisher (1939), Girshick (1939), Hsu (1939a), Mood (1951), and Roy (1939). The development of the Jacobian in Section 13.2.2 is due mainly to Hsu [as reported by Deemer and Olkin (1951)].
538
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECTORS
13.3. THE CASE OF ONE NONSINGULAR WISHART MATRIX In this section we shall find the distribution of the roots of
(1 )
IA
-lI\ = 0,
where the matrix A has the distribution W(J, n). It will be observed that the variances of the principal components of a sample of n + 1 from N(p., J) are lin times the roots of (1). We shall find the following theorem useful: Theorem 13.3.1.
If the s/mmetric matrix B has a density of the form
g([I' ... ' Ip), where 11> ... > Ip are the characteristic roots of B, then the density of the roots is
W' gUl'···' Ip)Di < i( Ii -Ii)
7T
(2)
fpOp)
Proof From Theorem A.2.1 of the Appendix: we know that there exists an orthogonal matrix C such that
(3)
B=C'LC,
where
(4)
I)
0
0
0
12
0
0
0
Ip
L=
If the l's are numbered in descending order of magnitude and if cil ~ 0, then (with probability 1) the transformation from B to Land C is unique. Let the matrix C be given the coordinates c 1, ••• ,cp(p-l)/2' and let the Jacobian of the transformation be f(L, C). Then the joint density of Land C is gUt> ... , Ip)f(L, C). To prove the theorem we must show that
(5)
J... Jf(L,C)dc
w' Di
7T 1
r(l)
···dc p (p-1)/2=
p
2P
.
We show this by taking a special case where B = UU' and U (p x m, m ~p) has the density
(6)
,
1
f [-2 (m + n)]
-,mp p 7T
fpOn)
,
II-UU'I;:(n-p-1) .
539
13.3 THE CASE OF ONE NONSINGULAR WISHART MATRIX
rrhen by Lemma 13.3.1, which will be stated below, B has the density
(7)
The joint density of Land C is f(L, C)g* (II"'" Ip)' In the preceding section we proved that the marginal density of L is (50). Thus
(8)
j ... !g*(lI, ... ,lp)f(L,C)dC=g*(lI, ... ,lp)j ... jf(L,C)dC
-I) fp(~~) J g*(lJ, ... ,lp )'
1T~p2n(l.
= This proves (5) and hence the theorem.
•
The statement above (7) is basej on the following lemma:
'f
Lemma 13.3.1. B = YY' is
(9)
If the density of Y (p x rn) is f(YY'), then the density of
iBI t(m-p-I) f( B) 1Twm fpOrn)
The proof of this, like thaI of Theorem 13.3.1, depends on exhibiting a special case; let f(YY') = (21T)- tpm e- ttr YY', then (9) is w(Bi], rn). Now let u~ find the density of the roots of (1). The density of A is
(10)
iAi i(lI-p-l) e- tlrA 2wnrpOn)
nf~ 1 / 1(11 -]>-1) exp( -
tr.f~J ()
2wnfp( tn)
Thus by the theorem we obtain as the density of the roots of A
(11)
540
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECTORS
Theorem 13.3.2. teristic roots (II ~ 12 the density is not O.
~
If A (p x p) has the distribution W(/, n), then the charac· ... ~ lp ~ 0) have the density (ll) over the range where
Corollary 13.3.1. Let VI ~ ••• ~ vp be the sample variances of the sample principal components of a sample of size N = n + 1 from N(p., a 21). Then (n/ a 2 )vj are distributed with density (11). The characteristic vectors of A are uniquely defined (except for multiplication by -1) with probability 1 by (12)
(A-lI)y=O,
y'y = 1,
since the roots are different with probability 1. Let the vectors with
Ylj ~
0 be
(13) Then
(14)
AY=YL.
From Section 11.2 we know that
(15)
Y'Y=I.
Multplication of (14) on the right by y- I = Y' gives A = YLY'.
(16)
Thus Y' = C, defined above. Now let us consider the joint distribution of Land C. The matrix A has the distribution of n
(17)
A=
E X"X~, a=J
where the X" are independently distributed, each according to N(O, I). Let
(I8) where Q is any orthogonal matrix. Then the tributed according to N(O, I) and n
(19)
A* =
X:
E X:X:' =QAQ' a=1
are independently dis-
541
13.3 THE CASE OF ONE NONSINGULAR WISHART MATRIX
is distributed according to W(J, n). The roots of A* are the roots of A; thus A* = C** 'LC**,
(20)
(21)
C** 'C**
define C** if we require
eil*
=I
~ O. Let
(22)
C* =CQ'.
Let
o
o
o (23)
o
J( C*) =
o
o
with eill Ieill = 1 if e~ = O. Thus J(C*) is a diagonal matrix; the ith diagonal element is 1 if eil ~ 0 and is - 1 if eil < O. Thus
(24)
C** =J( C* )C* =J( CQ')CQ'.
The distribution of C** is the same as that of C. We now shall show that this fact defines the distribution of C. Definition 13.3.1. If the random orthogonal matrix E of order p has a distn'bution sueh that EQ' has the same distribution for every orthogonal Q, (hen E is said to have the Haar invariant distribution (or normalized measure). The definition is possible because it has been proved that there is only one distribution with the required invariance property [Halmos (195u)). It has also been shown that this distribution is the only one invariant under multiplication on the left by an orthogonal matrix (i.e., the distribution of QE is the same as that 01 E). From this it follows that the probability is 1/2 P that E is such that e il :? O. This can be seen as follows. Let J I , ••• , J2 P be the 2 P diagonal matrices with elements + 1 and -1. Since the distribution of JiE is the same as that of E, the probability that e il ~ 0 is the same as the probability that the elements in the first column of Ji E are nonnegative. These events for i = 1, ... , 2 P are mutually exclusive and exhaustive (except for elements being 0, which have probability 0), and thus the probability of anyone is 1/2 P •
542
THE DISTRIBUTIONS OF CHARACfERISTICROOTS AND VECfORS
The conditional distribution of E given eil ;;:: 0 is 2 P times the Haar invariant distrihution over this part of the spac~. We shall call it the conditiollal Haar invariam distribution. Lemma 13.3.2. If the orthogonal matrix E has a distribution such that ei! ;;:: 0 and if E** = J(EQ')EQ' has the same distribution for every orthogonal Q, then E has the conditional Haar invariant distribution.
Proof Let the space V of orthogonal matrices be partitioned into the subspaces Vp ... , Vz- so that Ji~ = VI' say, where J l = I and VI is the set for which e jl ;;:: O. Let JLl be the measure in Vj defined by the distribution of E assumed in the lemma. The measure JL(W) of a (measurable) set W in V; is defined as (lj2 P )JL1(JjW), Now we want to show that JL is the Haar invariant measure. Let W be any (measurable) set in Vj' The lemma assumes that 2"JL(W) = JL/W) = Pr{E E W} = Pr{E** E W} = LJLMJWQ' n V;]) = 2P~t(WQ'). If U is any (measurable) set in V, then U = U 7~j(U n Vj). Since JL(Un~,;)=(1j2p)JLMj(UnVj)], by the above this is JL[(UnVj)Q']. Thus JLCU) = JLWQ '). Thus JL is invariant and JLl is the conditional invariant • distribution. From the lemma we see that the matrix C has the conditional Haar invariant distribution. Since the distribution of C conditional on L is tht: same, C and L are independent. Theorem 13.3.3. If C = Y', where Y = (Yp ... , yp) are the normalized characceristic vectors of A with Yli;;:: 0 and where A is distributed according to W(J, n), then C has the conditional Haar invariant distribution and C is distributed independently of the charactelistic roots. From the preceding work we can generalize Theorem 13.3.1. Theorem 13.3.4. g(ll .... . Ip)' where 11
If the symmetric matrix B has a density of the form
> ... > lp are the characteristic roots of B, then the joint
density of the roots is (2) and the matrix of normalized characteristic vectors Y (Ylj;;:: 0) is independently distributed according to the conditional Haar invariant distribution. Proof The density of QBQ', where QQ' =1, is the same as that of B (for the roots are invariant), and therefore the distribution of J(Y'Q')Y'Q' is the • same as that of Y'. Then Theorem 13.3.4 follows from Lemma 13.3.2. We shall give an application of this theorem to the case where B = B' is normally distributed with the (functionally independent) components of B independent with means 0 and variances $bj~ = 1 and $bi~ = ~ (j <j).
13.4
543
CANONICAL CORRELATIONS
Theorem 13.3.5.
Let B = B' have the density
(25) Then the characteristic roots 11 > ... > Ip of B have the density
and the matrix Y of the normalized characteristic vectors (Yli ~ 0) is independently distributed according to the conditional Haar invariant distribution. Proof Since the characteristic roots of B2 are the theorem follows directly. •
If, ... ,I;
and tr B2 =
r.zl,
Corollary 13.3.2. Let nS be distributed according to W(I, n), and define the diagonal matrix Land B by S = C'LC, C'C = I, 11 > ... > Ip, _and c il ~ 0, i = 1, ... , p. Then the density of the limiting distribution of in (L - J) = D diagonal is (26) with l; replaced by d;, and the matrix C is independently distributed according to the conditional Haar measure.
rn
Proof The density of the limiting distribution of (S - J) is (25), and the (S - 1) and the diagonal elements of D are the characteristic roots of • columns of C' are the characteristic vectors.
rn
13.4. CANONICAL CORRELATIONS The sample canonical correlations were shown in Section 12.3 to be the square roots of the roots of
(1) where N
(2)
A;j =
L
(X~1) -
X(i»)( X~j) - XU»),,
a~1
and the distribution of
(3)
X
=
(X(1)) X(2)
i,j=1,2,
544
THE DISTRIBUTIONS OF CHARACfERISTIC ROOTS AND VECfORS
is N(/L, I), where
(4) From Section 3.3 we know that the distribution of
Aij
is the same as that of
n
(5)
Aij
=
I: y;i)y;i)',
i,j = 1,2,
a=J
where n = N - 1 and y = (y(1») y(Z)
(6)
is distributed according to N(O, I). Let us assume that the dimensionality of y(1), say PI' is not greater than the dimensionality of y(2), say Pz' Then there are PI nonzero roots of (1), say
(7) Now we shall find the distribution of
CO when
(8) For the moment assume (YP)} to be fixed. Then A zz is fixed, and
(9) is the matrix of regression coefficients of y(1) on y(Z). From Section 4.3 we know that n
(10)
A 11 . Z =
I:
(y~l) -By~Z»)(y;l) -BYP»)' =A11 -BAzzB'
a=1
and
(11) a~cording to W(I II' n - pz) and W(I 11' Pz), respectively. In terms of Q the equation (1) defining f is
(13 = 0) are independently distributed (12)
IQ - f(A 11 .Z + Q) I= O.
13.5
ASYMPTOTIC DISTRIBUTIONS IN CASE OF ONE WISHART MATRIX
545
The distribution of f;, i = 1, ... , PI' is the distribution of the nonzero roots of (12), and the density is given by (see Section 13.2)
(13)
Since the conditional density (13) does not depend upon y(2), (13) is the unconditional density of the squares of the sample canonical correlation coefficients of the two sets X~I) and X~2), a = 1, ... , N. The density (13) also holds when the X(Z) are actually fixed variate vectors or have any distribution, so long as X(I) and X(Z) are independently distributed and X(I) has a multivariate normal distribution. In the special case when PI = 1, P2 = P - 1, (13) reduces to
(14) which is the density of the square of the sample multiple correlation coefficient between X(I) (p I = 1) and X(2) (P2 = P - 1).
13.5. ASYMPTOTIC DISTRIBUTIONS IN THE CASE OF ONE WISHART MATRIX 13.5.1. All Population Roots Different In Section 13.3 we found the density of the diagonal matrix L and the orthogonal matrix B defined by S = BLB', II ~ ... ~ Ip, and b l ; ~ 0, i = 1, ... , P, when nS is distributed according to W(I, n). In this section we find the asymptotic distribution of Land B when nS is distributed according to wet, n) and the characteristic roots of :t are different. (Corollary 13.3.2 gave the asymptotic distribution when :t = J.) Theorem 13.S.1.
Suppose nS has the distribution W(:t, 11). Define diagonal 13 alld B by
A and L and orthogonal
(1)
S =BLB',
AI> A2 > ... > Ap, II ~ I z ~ ... ~ Ip, 13 1; ~ 0, b l ; ~ 0, i = 1, ... , p. Define in (B - 13) and diagonal J') = in (L - A). Then the limiting distriblltion of
G=
546
THE DISTRIBUTIONS OF CHARACfERISTICROOTS AND VECfORS
D and G is normal with D and G independent, and the diagonal elements of D are independelll. The diagonal element d j has the limiting distribution N(O, 2 AT). The covariance matrix of gj in the limiting distribution of G = (gl' ... , gp) is
( 2)
Pp )'
where ~ = (PI"'" distrihution is
The covariance matrix of gj and gj in the limiting
(3) Proof The matrix
nT=n~'S~
(4)
is distributed according to W(A,n). Let T= YLY',
where Y is orthogonal. In order that (4) determine Y uniquely, we require ~ O. Let /ii (T - A) = V and (Y - I) = W. Then (4) can be written
m
Vjj
A+ _I_V = (I + _1_ w ) (A + _I_D) (I + _1_ w )',
(5)
m
m
m
m
which is equivalent to (6)
V
=
WA + D + A W' +
1
I
vn
1 (WD + WA W' + DW') + - WDW' . n
From 1 = YY' = [I + (11 m)W][1 + (llm)W'], we have
o = W + W' +
(7)
1 m WW' .
We shall proceed heuristically and justify the method later. If we neglect and 1In (6) and (7), we obtain terms of order 1I
m
(8)
V= WA +D+ AW',
(9)
O=W+W'.
When we substitute W' = - W from (9) into (8) and write the result in components, we obtain W ii = 0,
i = 1, ... ,p,
( 10) (11 )
i¥),
i,j=I, ... ,p.
13.5
ASYMPTOTIC DISTRIBUTIONS IN CASE OF ONE WISHART MATRIX
547
(Note wij = - wji .) distribution of U independent with A;Aj , i *" j. Then
From Theorem 3.4.4 we know that in the limiting normal the functionally independent elements are statistically means 0 and variances dY(u ii ) = 21.; and dY(u ij ) = the limiting distribution of D and W is normal, and dl, •.• ,dp,W12,WI3, ••• ,Wp_l.p are independent with means 0 and variances dYCd;) = 21.;, i = 1, ... , p, and dY(wij ) = A;AJCAj - AY, j = i + 1, ... , p, i = 1, ... ,p - 1. Each column of B is ± the corresponding column of J3Y; since Y ~ I, we have J3Y ~ J3, and with arbitrarily high probability each column of B is nearly identical to the corresponding column of J3Y. Then G= (B - J3) has the limiting distribution of J3m CY - I) = J3W. The asymptotic variances and covariances follow. Now we justify the limiting distribution of D and W. The equations T = YLY' and 1= YY' and conditions II> ... > Ip, Yii> 0, i = 1, ... , p, define a 1-1 transformation of T to Y, L except for a set of measure O. The transformation from Y, L to T is continuously differentiable. The inverse is continuously differentiable. in a neighborhood of Y = I and L = A, since the equations (8) and (9) can be solved uniquely. Hence Y, L as a function of T satisfies the conditions of Theorem 4.2.3. •
m
13.5.2. One Root of Higher Multiplicity In Section 11.7.3 we used the asymptotic distribution of the q smallest sample roots when the q smallest population roots are equal. We shall now derive that distribution. Let
(12)
where the diagonal elements of the diagonal matrix A I are different and are larger than 1.* (> 0). Let
(13)
Then T ~ A, which implies L ~ A, Yll ~ I, Yl2 ~ 0, Y21 ~ 0, but Yn does not have a probability limit. Ho'vever, Y22 Yh ~Iq. Let the singular value decomposition of Y22 be ElF, where J is diagonal and E and Fare orthogonal. Define C2 = EF, which is orthogonal. Let U = (I - A) and D= (L - A) be partitioned similarly to T and L. Define Wll = m(YIl -I), Wl2 = myl2 , W21 = my21 , and W22 = m(Y22 = C2 ) = mE(J-
m
m
548
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECTORS
Iq)F. Then (4) can be written
(~I
(14)
= [(
V 12 V 22
21
Ip~q C0) + In1 (WII W 2
.[( ~1 .[
0) + In1 (VII V
A*Iq
21
0)
A*Iq +
1 (DI0
In
)
I2 W )] W22
;J]
W~1 )] (Ip~q C;o ) + In1 (Wll W{2 W 22
=
(~1
0) + In1 [( Dl0
A*Iq
+
(WIIAI W21
A1
J
C2;2 C
(A
21 )] A1WWiz 1 +-M W A*C n' 2
1W{1 A*W12 C;) A*W22 C; + A*C2 12
where the sub matrices of M are sums of products of C2, and 1/ In. The orthogonality of Y Up = IT') implies
AI' A*Iq, D
k,
Wkl ,
where the submatrices of N are sums of products of Wkl • From (14) and (15) we find that
(16) The limiting distribution of (1/ A* )V22 has the density (25) of Section 13.3 with p replaced by q. Then the limiting distribution of D2 and C2 is the distribution of Di and Y2i defined by Vi;. = Y2~ Di Y2~" where (1/ A* )Vi;. has the density (25) of Section 13.3.
13.6 ASYMPTOTIC DISTRIBUTIONS IN CASE OF TWO WISHART MATRICES
549
Theorem 13.5.2. Under the conditions of Theorem 13.5.1 and A = diag(A l' A* lq), the density of the limiting distribution of d p _ q + I " ' " d p is
To justify the preceding derivation we note that D z and Yn are functions of U depending on n that converge to the solution of Uf = Y~; D!~ Y~;'. We can use the following theorem given by Anderson (1963a) and due to Rubin. Theorem 13.5.3. Let F.(u) be the cumulative distribution function of a random matrix Un. Let v" be a matrix-valued function of Un' Vn =!n(u n), and Let Gn(v) be the (induced) distribution of v". Suppose F.(u) ...... F(u) in every continuity point of F(u), ana suppose for every continuity point u of !(u), !n(u n) ...... !(u) when Un ...... u. Let G(v) be the distribution of the random matrix V = !(U), where U has the d.istribution F(u) If the probability of the set of discontinuities of!(u) according to F(u) is 0, then lim Gn ( v) = G ( v)
(18)
n~oo
in every continuity point of G(v).
The details of verifying that U(n) and (19) satisfy the conditions of the theorem have been given by Anderson (1963a).
13.6. ASYMPTOTIC DISTRIBUTIONS IN THE CASE OF TWO WISHART MATRICES 13.6.1. All Population Roots Different In Section 13.2 we studied the distributions of the roots II
(1) ~nd
(2)
\S* -IT* \ = 0
the vectors satisfying (S* - IT* ) x* = 0
2.
12
2. ... 2.
Ip of
550
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECfORS
and x* 'T* x* = 1 when A* = mS* and B* = nT* are distributed independently according to wet, m) and wet, n), respectively. In this section we study the asymptotic distributions of the roots and vectors as n -+ 00 when A* and B* are distributed independently according to W(<, m) and W(I, n), respectively, and min -+ 1/ > O. We shall assume that the roots of (3)
I«I>-AII =0
are distinct. (In Section 13.2 Al
= ... = Ap = 1.)
Theorem 13.6.1. Let mS* and nT* be independently distributed according to W( «1>, m) and W(I, n), respectively. Let AI:> A2 > ... > Ap (> 0) be tite roots of (3), and let A be the diagonal matrix with the roots as diagonal elements in descending order; let 'Y I' ... , 'Y p be the solutions to i = 1, ... ,p,
(4)
'Y'I'Y = 1, and 'Ylj ~ 0, and let r = ('Yl' .. . ,'Yp). Letll ~ ... ~ lp (> 0) be the roots of 0), and let L be the diagonal matrix with the roots as diagonal elements in descending order; let xi, ... ,x; be the solutions to (2) for I = lj, i = 1, ... , p, x* 'T* x* = 1, and xt > 0, and let X* = (xi, . .. , x;). Define Z* = .;n (X* - f) and diagonal D = .;n (L - A). Then the limiting distribution of D and Z* is nOnTUli with means 0 as n -+ 00, m -> 00, and min -> 1/ (> 0). The asymptotic variances and co variances that are not 0 are
(5) (6) (7)
<w€( d j , zi) = Aj'Yj,
(8)
d€(zj,zj)=-
* *
AjAj(l
+ 1/)
1/(\ -
AJ
,
2'Yj'Yj,
i *j.
Proof Let
(9)
S= r's*r,
T= r'T*r.
Then mS and nT are distributed independently according to W( A, m) and W(J, n), respectively (Section 7.3.3). Then 11' ... ' lp are the roots of (10)
Is-lTI =0.
13.6 ASYMPTOTIC DISTRIBUTIONS IN CASE OF TWO WISHART MATRICES
Let
Xl' ..• ' Xp
551
be the solutions to (S -ljT)x = 0,
(11)
i = 1, ... ,p,
and x'Tx= 1, and let X=(xl, ... ,xp ). Then xi = rX i and X* =rx except possibly for multiplication of columns of X (or X*) by -1. If Z = (X - I), then Z* = rz (except possibly for multiplication of columns by -1). We shall now find the limiting distribution of D and Z. Let (S - A) = U and (T - J) = V. Then U and V have independent limiting nonnal distributions with means o. The functionally independent elements of U and V are s,tatistically independent in the limiting distribution. The variances are tCu~; ;;, 2(n/m)A; -> 2).;/7/; tCu;j = (l'/m)A;A j -> A;A/7/, i *" j; tCuli = 2; tCui} = 1, i +j. From the definition of L and X we have SX = TXL, X'TX. = I, and X'SX=L."If we let X-I = G, we obtain
rn
rn
rn
S = G'LG,
(12)
T=G'G.
We require gjj > 0, i = 1, ... , p. Since S .£, A and T'£' I, we have L .£, A and G .£, I. Let (G - J) = H. Then we write (12) as
rn
(13) (14) These can be rewritten
~ (DH+H'D+H'AH) + 1..n H ·I)H,
(15)
U=lJ+AH+H'A+
(16)
V=H+H'+ rnH'H.
yn
1
If we neglect the terms of order 1/ can write
rn and l/n (as in Section 13.5), we
(17)
U=D+AH+H'A,
(18)
V=H+H',
(19)
U-VA=D+AH-HA.
552
THE DISTRIBUTIONS OF CHARACfERISTICROOTS AND VECTORS
The diagonal elements of (18) and the components of (19) are
(20) (21)
i ¥).
(22)
The limiting distribution of Hand D is normal with means O. The pairs (h;j' hj) of off-diagonal elements of H are independent with variances (23)
and covariances (24)
The pairs (d;, h) of diagonal elements of D and H are independent with variances (5),
(25) and covariance (26)
The diagonal elements of D and H are independent of the off-diagonal elements of H. That the limiting distribution of D and H is normal is justified by H.eorem 4.2.3. Sand T are polynomials in Land G, and their derivatives are polynomials and hence continuous. Since the equations (12) with auxiliary conditions can be solved uniquely for Land G, the inverse function is also continuously differentiable at L = A and G = I. By Theorem 4.2.3, D= (L - A) and H = (G - J) have a limiting normal distribution. In turn, X = G -I is continuously differentiable at G = I, and Z = (X - J) = m(G-1 - J) has the limiting distribution of -H. (Expand m{[I + 0/ m)Hj-1 - J}.) Since G!.. I, X!.. I, and Xii> 0, i = 1, ... , P with probability approaching 1. Then Z* = m(X* - r) has the limiting distribution of rz. (Since x!2. I, we have X* = r X!.. r and Xli> 0, i = 1, ... , p, with probability approaching 1.) The asymptotic variances and covariances (6) to • (8) are ohtained from (23) to (26).
in
m
m
13.6
ASYMPTOTIC DISTRIBUTIONS IN CASE OF TWO WISHART MATRICES
553
Anderson (1989b), has derived the limiting distribution of the characteristic roots and vectors of one sample covariance matrix in the metric of another with population roots of arbitrary multiplicities.
13.6.2. One Root of Higher Multiplicity In Section 13.6.1 it was assumed that mS* and nT* were distributed independently according to W( «1>, m) and W(I, n), respectively, and that the roots of I«I> - All = 0 were distinct. In this section we assume that the k larger roots are distinct and greater than the p - k smaller roots, which arc assumed equal. Let the diagonal matrix A of characteristic roots be A = diag(A I , A*lp _ k ), and let r be a matrix satisfying
(27)
«I>r=IrA,
Define Sand T by (9) and diagonal Land G by (12). Then S.!2" A, T.!2" Ip. and L.!2" A. Partition S, T, L, and G as
(28)
p
p
p
where Sl1' TIl' L I, and G l1 are k X k. Then Gil -+Ik• G I2 -> 0, and G 21 -> O. but G 22 does not have a probability limit. Instead G'22G22.!2" I p _ k ' Let the singular value decomposition of G 22 be EJF, where E and F are orthogonal and J is diagonal. Let C 2 = EF. The limiting distribution of U = (S - A) and V = In (T - I) is normal with the covariance structure given above (12) with Ak + I = .. , : \ ' = A* . Define D = m(L - A), HI! = In(G l1 - I), HI2 = In G 12 , H21 =../11 Gel' and H22 = In(G 22 - C 2 ) = InE(J -lp_k)F. Then (13) and (15) are replaced by
m
(29)
AOI (
*
A Ip _ k
o
j
1 + In Dc _
554
THE DISTRIBUTIONS OF CHARACfERIST'CROOTS AND VI:.CrORS
o ] A*lp_k
+
1
In
[DI 0
and (14) and (16) are replaced by
I + -
[
~ H;I
VI1
1 H'
Y, \11
=
I [0
I,
0] + In1 [HII H:I
lp_k
H21]
1 ' C + -WHo H 22 2 n
If we neglect the terms of order 1/ In and 1/11, instead of (1 i) we can write (31)
VII - VII AI [ U,I - V,IA I
(A*I-A I )H I2 C 2 ]
C;D 2 C2
Then Vii = 211;;, i = 1, ... , k; !I;; - A;v;; = d;, i = 1, . .. , k; !Ii; - v;; A; = (A; - Aj)h,j' i *" j, i, j = 1, ... , k; Vn - 1..* V22 = C 2D 2C 2; C 2(UZI - V21 AI) = H,I(A*I-A 1 ); and (U1:-A*VI:)C;=(A*I-AI)H12. The limiting distribution of V 22 - 1..* V22 is normal with mean 0;
•
13.7
555
ASYMPTOTIC DISTRIBUTION IN A REGRESSION MODEL
13.7. ASYMPTOTIC DISTRIBUTION IN A REGRESSION MODEL 13.7.1. Both Sets of Variates Stochastic The sample canonical correlations II"'" Ip2 and vectors a 1"'" aPI' and 'YI, ... ,'Y P2 are defined in Section 12.3. The set 'YI, ... ,'YP2 and 11, ... ,lp2 are defined by
( 1) The asymptotic distribution of these quantities was given by Anderson (1999a) when X = (X(I)" X(Z),), h'lS a normal distribution and also when X(I) is normally distributed with a linear function of nonstochastic X(Z) as expected value. We shall now find the asymptotic distribution wh'!n X has a normal distribution. The model in regression form is
(2)
X(I) =
PX(Z)
+ Z,
where X(Z) and Z are independently normally distributed with expected values $X(Z) = 0 and tffZ = 0 and covariances tffX(Z)X(2) = :I 22 , tffZZ' = :I zz ($X(2)Z' = 0). Then tffX(1) = 0 and tffX(1)X(I), =:Ill = :I zz + P:Izzp' and tff X (1) X(Z)' = P:I 22 • Inference is based on a sample of X of n observations. First we transform to canonical variables U = A' X(I), V = r' X(Z), and W = A' Z. Then (1) is transformed to
(3)
U=0V+ W,
where 0=~'p(r/)-I, tffUU/=:Iuu=I (tW/=:I VV =Ip2 ' (tUV/=:I vv p" = (A,O) = A, tffWW/ = :I ww = Ipi - AZ, and tffVW/ = O. [See (33) to (37) and (45) of Section 12.2.] Let the sample covariance matrices be Suu = A/SllA/, Suv=A/Slzr, and sVV=r'S22L Let the sample vectors constitute H=r-li'=r-I(.yl,"".yp,). Then H satisfies
(4) where A+ = diag(A I , ... , Apl'O, ... ,0); if PI I p2 , Suv -'-+ A. Then A = dlag(A t , ••• , Ap) -> A. Let ~.
(5)
~
~
556
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECTORS
where HII is of (4) are
PI
XPI and Hn is (P2 - PI)
X
(P2 - PI)' The first PI columns
(6) the last P2 - PI columns of (4) are SUVH2 = O. Then Hu ~ IPi' H12 ~ 0, and H21 40, but the probability limit of (4) only implies H22H22.4 Ip2 - p, ' Let the singular value decomposition of H22 be H22 = EJF. Define Suu = m(Suu - Ip), Stv = mCS vv - Ip), Suv = mCSuv - A), Hi = mCH I - I(p), and A* = [mCA - A), 0], where I(p,) = (IPI'O)'. Then expansion of (6) yields
r
(A' + ),z stu)( Ip, + ),z Suu (A + ),z Suv)( l(Pd + kHt) l
(7)
= (I p2 +
fn
stv )( I(p,) +
k
Hi )( A +
k
A*
r.
From (7) we obtain
(8)
StuA1(p,) - A'SuuAI(p,) + A'SuvI(P,) + A'AHi 2 = Stv I(p,)A + HiA2 + 2I(p,)A A* +op(1).
(9) In terms of partitioned matrices (8) is
(10)
The lower submatrix equation [(P2 - PI) XPI] of (10) is
13.7
557
ASYMPTOTIC DISTRIBUTION IN A REGRESSION MODEL
A diagonal element of the upper submatrix equation of (10) is
The right-hand side of (12) is the expansion of the sample correlation coefficient of u;a and Via' See Section 4.2.3. The limiting distribution of A; is N[O,(l- An2]. The (i, j)th component of H~I in (10) is (13) n
(AJ -
Af)h;j =
~ L
vn
(AjV;aUja + A;u;avja - A;\u;auja -
AJ V;aVja) + op(1).
<>=1
The asymptotic covariance of (AJ - Anh;j and
(A7 -
A])hj; is
(1.- AJ)( A7 + A; - 2AT An [ (1- A7)(1- AJ)( A7 + An
(14)
The pair (hij , hJ) is uncorrel~ted with other pairs. Suppose P1=P2' Then r*=rH~I=rH*. Let r=(1'I""'1'P\)' r= (.y I' ... , .yp ). Then .y = l:.f,l Ii'; hij , where hij, i j, is obtained from (13) and hi; from (9). We obtain
*
t
(16)
Anderson (1999a), has also given the asymptotic covariances of eX j and of 'Yj and Note that hi; depends linearly on (U ia , via) and that the pairs (U ia , Via) and (U ja , Vja), i j, are uncorrelated. The covariances (14) do not depend on (U, V) being normal. Now suppose that the rank of r l2 is k
a/.
*
558
THE DISTRIBUTIONS OF CHARACfERISTICROOTS AND VECfORS
The last PI - k columns of (4) are
Hence
13.7.2. One Set of Variates Stochastic and the Other Set Nonstochastic Now consider the case that Xl,) in (2) is nonstochastic, where ':"Zu = 0 and I'ZaZ~ = l:zz. We observe X = XI"'" XII' We assume
(19) and
(20)
~~,
is nonsingular. Then
Sll =
~
t x~l)x~l)'
=
PSZ2P + Szzp' + P S2Z + Szz
~ Pl:22P' + l:zz,
a~1
( 21) Define A, <x, and "I by solutions to
-A(l:zz
(22)
[
+ PS22P')
S22P
PS22 ][<X] -AS 2:! "I
= 0,
'y'S22"1 - 1.
(23)
We shall first assume PI = P2 and AI> ." :> ApI> O. Then (22) and (23) and ai, > 0 define
Let U=A'"X(ll, t'a=r~xa' a=l" .. ,n, W=A'"Z, 0=A'"p(r~)-I=A, H = 1',,-1 Then H and A satisfy (4). Then Svv = T,
t.
(25) (26)
SUI' = 0S vv + Swv = 0 + Swv
P
->
0,
13.7 ASYMPTOTIC DISTRIBUTION IN A REGRESSION MODEL
559
Then (4) can be written (27)
Z
(A+Svw) ( A +ASvw+SwvA+Sww
)-1 (A+Swv)H=HA. ~Z
Note that Svw 1'. 0, Sww 1'.1- AZ, and hence H ~ I, A ~ A. Let stw = [,lsvw, S~w = [,l[Sww - (I - AZ)]. Then (27) leads to
(28)
(1- AZ)StwA + ASh(l- AZ) - AS;:'wA + AZH* = H*Az + 2A A* + op(l).
A diagonal term of (28) gives
Since
.1'( VjaWja / = v?a(1- An,
(30)
@"[ Wj~ - (1 - An]z = 2(1- An
(31)
z
under the assumption that W is normally distributed, the limiting distribution of [,leA j - A) is N[O,(1- An z(1- ~A~)]. Note that this variance is smaller than in the case of X(Z) stochastic. From (28) we find
(32) (A;-A7)h;j=
.~
vn
t
[(1-ADVjawjaAj+Ajwjavja(1-AJ)-A;wjawjaAj]
<>=1
Then
The equation H'SvvH=1 implies H'H=I, leading to H* = -H*' +0/1), that is, h;j = - hj; + 0/1). Now suppose that the rank of P is k < PI = Pz. Then A = diag( A 1,0), where Al =diag(A I , ••• Ak ). Let r=(rl,r z), where r l has k columns and r Z has PI - k columns. Define the partition (5) to be made into k and PI - k rows and columns. The probability limit of (4) implies H11 ~ i k , H12.4 0, HZI .4 0, and H~z H zz .4/. Let the singular value decomposition of H zz be
560
THE DISTRIBUTIONS OF CHARACfERISTICROOTS AND VECfORS
EJF, where J is a diagonal matrix of ~rder PI - k and E and Fare orthogonal matrices of order PI - k. Define C2 = EF. The expansion of (4)
in terms of Stw=vn(Svw-A), SI;,v=vn(SUv-A), S:;'w=vn[Sww(J-A 2)], Hi;=vn(H 11 -J), H't2=vn H I 2' H'h=IiIH 21 , and Hti= vn(H22 - C2) = vnEU - J)F yields (34)
[
AIS:;'j~ (I - A~) + (I - A~)St~AI - AIS;:}~A I St~AI
= [2A1A~ +HiIA~ - A21Hii HiIA"1 . The ith diagonal term of (34) is (29) for i = I, ... , k. The i, jth element of the upper left-hand submatrix is (32) for i*j and i,j= I, ... ,k. Two other submatrix equations of (34) are
(35)
A I H~2
= -SIl,'~C2 + op( I),
(36)
HilAI =St~ +op(l).
The equation I=H'SvvH=H'H yields (37)
HI~ +Hl~ [ C;H'h + Hi2'
The off-diagonal submatrices of (37) agree with (35) and (36). 13.7.3. Reduced Rank Regression Estimator When the rank of J3 is specified to be k «PI)' the maximum likelihood estimator of P is
(38) See Section 12.7. In terms of (3) the reduced-rank regression estimator of 0 is (39) Suppose X(2) is stochastic and 0 = diar,(0 1,0) = diag(AI,O). We define 0Z = vn(0 k - 0), Hi = vn(H j - I(k»' Suv - vn(Suv - A), and Sh = vn(Svv-/). From H;SvvHj =/ we find H~ +H~' = +op(l).
-sW
561
13.7 ASYMPTOTIC DISTRIBUTION IN A REGRESSION MODEL
From (39) and (9) we obtain
(40) _
S*l1 wv
- [ S*21 IVV We can compare 0 k with the maximum likelihood estimator unrestricted by a rank condition 0 = SuvSv~. Then ( 41)
since Svv !.,.J. The effect of the rank restriction is to replace the lower right-hand submatrix of Sl\tv by 0 (the parameter value). Since Sl\tv = (1/ [,l)E:. 1WaV~, we have vec Sl\tv = 0/ Ynn:~.~ 10';, ® W,,). Because Va and W", are independent,
(42)
@"vecSl\tv(vecSIJ,v)'
= @"W'
® @"WW' = I ®
(I - A2) = diag(I - A2 , .•.• [ - .IV).
where A = diag(AI,O) and 1- A2 = diag(I - A21, I). On the other hand
(43)
+0 (1) ·m "'~t [WV(l)' (w,,(l))V(2)') 0
1 vec0* =vec-.
k
I
'"
Q
a
,
P
-}., .t [~:: : (44)
= diag( Jpl
-
where there are k blocks of Jpl
A2 , ... , Ipl -
-
A2 ,
Ik -
A21 .0 .....
Ik -
A21 .O).
A2 and PI - k blocks of diag(Ik - A~, 0).
562
THE DISTRIBUTIONS OF CHARACfERISTlCROOTS AND VECfORS
In the original coordinate system (45)
vec(.8k -p)=vec[CA')-'(0 k -0)r'] = [ r ® ( A') - '] vecC 0 - 0) = [cr"r2) ®IzzCAj,A2)(I-A2)-']vecC0-0).
From (44) and (45) we obtain (46)
"~'vecn(Bk-p)[vec(Bk-p)l' -->
[(r,r;) ®IzzA(Ip-
A2f']
,r j A',I zz ]
+ [r2r; ® IzzA,(Ik - A2 = [r,r; ® I zz ] + =
Ixl ® I zz -
Ifwe define n = Iyxr, We have
=
[r2r~ ® IzzA,(Ik - A~r' A',I zz ]
(r2r; ® I zz A2 A'2Izz)·
I zz A, A ,(J - A~)-' and n = r" then p = nn'.
(47)
n(n'Iz~nr' n' = I zz - IzzA2A'2Izz,
(48)
n(n'Ixxn) -, n' = r,r; = Iil - r2r;.
Thus (46) can be written (49) /' vec B: (vec
.8n'
-->
Iil ® I zz - [Iil- n(n'I xx n) -I n'] ®[ I zz -
n(n'Iz~nr'nl
Theorem 13.7.1. Let (X(1)', X(2)')', a = 1, ... , n, be observations on the random vector x., with mean 0 and covariance matrix I. Let p = I 12I2"z'. Let the columns of f\ satisfy (1) and 'Yjj > O. Suppose that X(I) - px(2) = Z is independent of X (2). Then the limiting distribution of vec.8: = In vec(.8k - P), with Bk = S12 f\ t;, is normal with mean 0 and covariance matrix (46) or (49).
Note that p = nn' = nM'(nM-')' for arbitrary nonsingular M; however, (47) and (48) are invariant with respect to the transforrration n --> nM and n --> nM-', Thus (49) holds for any factorization p = nn',
13.8
563
ELLIPTICALLY CONTOURED DISTRIBUTIONS
The
limiting
distribution
of
Bt
only
depends
on
.;nSZ2S221 =
(A')-IS;t,vsv~r' and hence holds under the same conditions as the asymp-
totic normality of the least s,-/uares estimator B. Now suppose that X~2) = X~2), a = 1, ... , n, is nonstochastic and that (19) holds. The model is (2); in the transformed coordinates [U = A: n X(I), va = r~x~2), W= A:nZ, 0 = A:nJ3(r~)-1 = A] the model is (3). HI = r- I r l satisfies (34) and (37). Again (39) holds. Further, (42) and (43) hold with Va = Va nonstochastic.
Corollary 13.7.1. Let xFl, ... , X~2) be a set of vectors such that (19) holds. Let X~I) = Jh~2) + Za' a = 1, ... , n, where Za is an observation on a random vector Z with ,ffZ = 0 and ,ffZZ' = I zz . Suppose J3 has rank k. Then the limiting distribution of Iii vec(B k - J3) is normal with mean 0 and covariance (46) or (49).
13.8. ELLIPTICALLY CONTOURED DISTRIBUTIONS 13.8.1. Observations Elliptically Contoured Let XI"'" x N be N observations on a random vector X with density
(1) is a positive definite matrix, R2 = (x - V )'lJI- I (x - v), and ,ffR 2 < 00. Define K = P ,ffR4j[(,ffR2)2(p + 2)] - 1. Then ,ffX = v = J.I- and ,ff(X - v Xx - v)' = (,ff R2 jp)lJI = I. Define i and S as the sample mean and covariance matrix. Define the orthogonal matrices J3 and B c..nd the diagonal matrices A and L by
where '"
(2)
I = J3AJ3',
S =BLB',
AI> .,. > Ap, II > '" > lp, f3il ~ 0, bil ~ 0, i = 1, ... , p. As in Section 13.5.1, define T= J3'SJ3 = YLY', where Y= J3'B is orthogonal and Yil ~ O. Then ,ffT= J3'IJ3 = A. The limiting covariances of IN vec(S - I) and IN vec(T - A) are
(3)
lim N,ff vec( S - I )[ vec( S - I ) l' N-+OO
= (K + 1)(Ip2 +Kpp)(I ® I) + K vec I (4)
lim N,ff vec( T - A) [vec( T - A)]' N~OO
.
(vec I)',
564
THE DISTRIBUTIONS OF CHARACfERISTIC ROOTS AND VECfORS
In terms of components ,g t;j = A; O;j and
(5)
(:
Let !Fi(T- A) = U, !Fi(L - A) =D, and !Fi(Y-Ip ) = W. The set ulI, ••• ,u pp are asymptotically independent of the set (U 12 , •.• ,U p _ 1. ) ; the covariances u ij ' i j, are mutually independent with variances (K + 1)A;Aj; the variance of U u = d; converges to (3K + 2)A;; the covariance of U u = d; and u kk = d k , i k, converges to KAjAk' The limiting distribution of wij ' i j, is the limiting distribution of uiJ(\ - A). Thus the W;j, i <j, are asymptotiThese varically mutually independent with ,gWj~ = (K + l)Ai AJ(\ ances and covariances for the normal case hold for K = O.
*
*
*
AY.
Theorem 13.8.1. Define diagonal A and L and orthogonal P and B by (2), Al > ... > Ap, II > ... > Ip, f3;1 ~ 0, b;I ~ 0, i = 1, ... , p. Define G = !Fi (BP) and diagonal D = IN (L - A). Then the limiting distribution of G and D is normal with G and D independent. The variance of d; is (2 + 3KJAf, and the covariance of d j and d k is KA;Ak' The covariance of gj is
(6)
The covariance matrix of gi and gj is
(7)
i
*.i.
Proof The proof is the same as for Theorem 13.5.1 except that (4) is used instead of (4) with K = O. •
In Section 11.7.3 we used the asymptotic distribution of the smallest q sample roots when the smallest q population roots are equal. Let A = diag(A I' A* I q ), where the diagonal elements of (diagonal) A I are different and are larger than A*. As before, let U = IN (T - A), and let U22 be the lower right-hand q X q submatrix of U. Let D2 and Y22 be the lower right-hand q X q submatrices of D and Y. It was shown in Section 13.5.2 that U22 = Y22 D2 Y22 + 0/1).
1 Ii
.8
565
ELLIPTICALLY CONTOURED DISTRIBUTIONS
= ... = \' is
The criterion for testing the null hypothesis Ap _ q+ I
8) ("[,f=p-q + I
IJ q )
q .
n Section 11.7.3 it was shown that - N times the logarithm of (8) has the imiting distribution of
1 [2 = ----;z 2A
Lp
i~p-q+l
,
Uij
+
Lp i=p-q+1
,
uii
-
1 (P L q i~p-q+1
-
!Iii
)-] .
'<J
The term "[" < Jufj has the limiting distribution of (1 + K )A* "Xq"(q -1)/". The limiting distribution of (u p _ q _ l . p_q+ 1' ... ' u pp ) is normal with mean 0 and covariance matrix A*2[2(1 + K)Iq + KE:e']A*2. The limiting distribution of ["[,Ufi - ("[,UiY /q ]A*2 is 2(1 + K )A* 2xi_l. Hence, the limiting distribution of (9) is the distribution of (1 + K)xi(q+I)/1-1. We are also interested in the characteristic roots and vectors of one covariance matrix in the metric of another covariance matrix.
or
Theorem 13.8.2. Let S* be the sample cavan·ance matrix of II slIlIlple si::e M from (1), and let T* be the sample covariance matrix of a sample of size N from (1) with -qr replaced by "I. Let A be the diagonal matrix with '\1 > ... > \' (> 0) as the diagonal elements, where AI> ... ' Ap ar? the roots of \ \ff - Al:\ = O. Let r=(I'I' ... 'I'p) be the matrix with "Ii the solution of (\ff-A,.l:)-y=0. 1" I, I' = 1, and I' Ii ~ O. Let X* = (xi, ... , x;) and diagonal L* callSist of the solutions to
(10)
(S* -IT*)x* = 0,
x* 'T* x* = 1, and xl' ~ O. As M ....... 00, N ....... 00, M / N ....... 7), the limiting distn·bution of z* = IN (X* - r) and diagonal D* = IN (L - A) is normal with the following covariances:
(11) (12)
566
THE DISTRIBUTIONS OF CHARACTERISTIC ROOTS AND VECTORS
( 13)
( I ..f)
(15)
Proof Transform S* and T* to S = f' S* f and T = r 'T* f, and I to A==f'I' and l=f'If, and X* to X=f- 1 X*=G- 1• Let D= {N(L-A), H={N(G-/). U=/N(S-A), and V={N(T-J). He matrices U and V and D and H have limiting normal distributions; they are related by (20), (21), and (22) of Section 13.6. From there and the covariances of the limiting distributions we derive (11) to (16). •
13.8.2. Elliptically Contoured Matrix Distributions Let r (p X N) have the density g(tr YY'). Then A = YY' has the density (Lemma 13.3.1) ( 17)
Let A = BLB', where L is diagonal with diagonal clements II> ... > I" and B is orthogonal with bil ~ O. Since g(tr A) = g(L.f'" II), the density of i l , ... , lp is (Theorem 13.3.4)
( l~)
7T+ P'
g( [f',., IIi)
n i < j(li -I j )
I~Op)
and the matrix B is independently distributed according to the conditional Haar invariant distribution. Su ppose Y* (p X m) and Z* (p X n) have the density w(m+,,)/2 g
[ fr( y* ''11- 1 Y*
+ Z* ''11- 1 Z*) 1
(m,n>p).
Let C bc a matrix such that ewe' = I. Then Y = CY* and Z = CZ* have the density g[tr(YY'+ZZ')]. Let A*=Y*Y*', B*=Z*Z*', A=YY', and B = ZZ'. The roots of IA* - IB* I = 0 are the roots of (A -IBI = O. Let the
567
PROBLEMS
roots of IA - f(A + B)I = 0 be fl> ... > f p' and let F = diag(fl"" ,fp)' Define E (pxp) by A+B=E'E, and A=E'FE, and en ~O, i= 1, ... ,p. Theorem 13.8.3.
The matrices E and F are independent. The
densi~J
of F is
(19) the density of E is
(20) In the development in Section 13.2 the observations Y, Z have the density
and in Section 13.7 g[tr(Y'Y + Z'Z)] = g[tr(A + B)]. The distribution of the roots does not depend on the form of gO; the distribution of E depends only on E' E = A + B. The algebra in Section 13.2 carries over to this more general case.
PROBLEMS 13.1. (Sec. B.2)
Prove Theorem 13.2.1 for p
=
2 hy calculating the Jacobian
directly. 13.2. (Sec. 13.2)
Prove Theorem 13.3.2 for p = 2 directly by representing the orthogonal matrix C in terms of the cosine and sine of an angle.
13.3. (Sec. 13.2) Consider the distribution of the roots of IA -IBI = 0 when A and B are of order two and are distributed according to wet, m) and wet, n),
respectively. (a) Find the distribution of the larger root. (b) Find the distribution of the smaller root. (c) Find the distribution of the sum of the roots. Prove that the Jacobian I a(G, A)I a(E, F)I is fI(li - fj) times a function of E by showing that the Jacobian vanishes for Ij = Ij and that its degree in I; is the same as that of fl(J; - fj).
13.4. (Sec. 13.2)
Give the J-:laar invariant distribution explicitly for the 2 x 2 orthogonal matrix represented in terms of the cosine and sine of an angle.
13.5. (Sec. 13.3)
568
THE DISTRIBUTIONS OF CHARACl ERISTICROOTS AND VESTORS
13.6. (Sec. 13.3) Let A and B be distributed according to W(I, m) and W(I, n) respectively. Let 11 > ... 7> Ip be the roots of IA -IBI = 0 and m 1 > ... > mp be the roots of IA - mIl = O. Find the distribution of the m's from that of the l's by letting n .... 00. 13.7. (Sec. 13.3) Prove Lemma 13.3.1 in as much detail as Theorem 13.3.1. 13.S. Let A be distributed according to W(I, n). In case of p = 2 find the distribution of the characteristic roots of A. [Hint: Transform so that I goes into a diagonal matrix.] 13.9. From the result in Problem 13.6 find the (when the null hypothesis is not true). 13.10. (Sec. 13.3) Show that X (p the density
X
di~tribution
of the sphericity criterion
n) has the density fx(X' X) if and only if T has
where T is the lower triangular matrix with positive diagonal elements such that IT' =X'X. [Srivastava and Khatri (1979)]. [Hint: Compare Lemma 13.3.1 with Corollary 7.2.1.] 13.11. (Sec. 13.5.2) In the case that the covariance matrix is (12) find the limiting distribution of D I , WII , W12 • and W21 • 13.12. (Sec. 13.3)
Prove (6) of Section 12.4.
CHAPTER 14
Factor Analysis
14.1. INTRODUCTION
Factor analysis is based 0<1 a model in which the observed vector is partitioned into an unobserved systematic part and an unobserved e"or part. The components of the error vector are considered as uncorrelated or independent, while the systematic part is taken as a linear combination of a relatively small number of unobserved factor variables. The analysis separates the effects of the factors, which are of basic intcrest, from the error,. From another point of view the analysis gives a description or explanation of the interdependence of a set of variables in terms of the factors without regard to tho~ observed variability. This approach is to be compared with principal component analysis, which describes Or "explains" the variability observed. Factor analysis was developed originally for the analysis of scores on mental tests; however, the methods are useful in a much wider range of situations, for example, analyzing sets of tests of attitudes, sets of physical measure:nents, and sets of economic quantities. When a battery of tests is given to a group of individuals, it is observed that the score of an individual on a given test is more related to his scores on other tests than to the scores of other individuals on the other tests; that is, usually the scores for ~ny particular individual are interrelated to some degree. This interrelation is "explained" oy considering a test score of an individual as made up of a part which is peculiar to this particular test (called error) and a part which is a function of more fundamental quantities called scores of prim my abilities or factor scores. Since they enter several test scores, it is their effect that connects the various
An Introduction to Multivariate Statistical Ana!,".'i.\". Third Edition. By T. \\'. Anderson ISBN 0-471-36091-0 Copyright © 2UllJ John Wiley & Sons. Inc.
569
570
FACfOR ANALYSIS
test scores. Roughly, the idea is that a person who is more intelligent in some respects will do better on many tests than someone who is less intelligent. The model for factor analysis is defined and discussed in Section 14.2. Maximum likelihood estimators of the parameters are derived in the case that the factor scores and errors are normally distributed, and a test that the model fits is developed. The large-sample distribution theory is given for the estimators and test criterion (Section 14.3). Maximum likelihood estimators for fixed factors do not exist, but alternative estimation procedures are suggested (Section 14.4). Some aspects of interpretation are treated in Section 14.5. The maximum likelihood estimators are derived when the factors are normal and identification is effected by specified zero loadings. Finally the estimation of factor scores is considered. Anderson (1984a) discusses the relationship of factor analysis to principal components and linear functional and structural relationships.
14.2. THE MODEL 14.2.1. Definition of the Model
Let the observable vector X be written as
(1)
X=Af+V+J1.,
where X, V, and J1. are column vectors of p components, f is a colU'nn vector of m (5,p) components, and A is a p X m matrix. We assume that V is distributed independently of f and with mean GV = 0 and covariance matrix GVV' = 'IT, which is diagonal. The vector f will be treated alternatively as a random vector and as a vector of parameters that varies from observation to observation. In terms of mental tests each component of X is a score on a test or battery of tests. The corresponding component of J1. is the average score of this test in the population. The components of f are the scores of the mental factors; linear combinations of these enter into the test scores. The coefficients of these linear combinations are the elements of A, and these are called factor loadings. Sometimes the elements of f are called common factors because they are cOmmon to several different tests; in the first presentation of this kind of model [Spearman (1904)] f consisted of one component and was termed the general factor. A component of V is the part of the test score not "explained" by the common factors. This is considered as made up of the error of measurement in the test plus a specific factor, having to do only with this particular test. Since in our model (with one set of observations on each individual) we cannot distinguish between these two
571
14.2 THE MODEL
components of the coordinate of U, we shall simply term the element of U the error of measurement. The specification of a given component of X is similar to that in regression theory (or analysis of variance) in that it is a linear combination of other variables. Here, however, f, which plays the role of the independent variable, is not observed. We can distinguish between two kinds of models. In one we consider the vector f to be a random vector, and in the other we consider f to be a vector of nonrandom quantities that varies from one individual to another. In the second case, :t is more accurate to write Xa = Afa + U + J1.. The nonrandom factor score vector may seem a better description of the systematic part, but it poses problems of inference because the likelihood function may not have a maximum. In principle, the model with random factors is appropriate when different samples consist of different individuals; the nonrandom factor model is suitable when the specific individuals involved and not just the structure are of interest. When f is taken as random, we assume Gf= o. (Otherwise, GX= A Gf + J1., and J1. can be redefined to absorb A Gf.) Let Gff' = <1>. Our analysis will be made in terms of first and second moments. Usually, we shall consider f and U to have normal distributions. If f is not random, then f= fa for the ath individual. Therl we shaH assume usuaHy (1/N)L~_da = 0 and (1/N)L~_daf~ = <1>. There is a fundamental indeterminacy in this model. Let f = Cf* (J* = C-1J) and A* = AC, where C is a nonsingular m X m matrix. Then (1) can be written as
(2)
X=A*f*+U+J1..
When f is random, Gf* f*' = C- I (C- I ), = <1>*; when f is nonrandom, (1/N)L~=d:I:' = <1>*. The model with A and f is equivalent to the model with A* and f*; that is, by observing X we cannot distinguish between these two models. Some of the indeterminacy in the model can be eliminated by requiring that Gff' =1 if f is random, or L~-dJ~ =NI if f is not random; In this case the factors are said to be orthogonal; if is not diagonal, the fectors are said to be. oblique. When we assume = I, then Gf* f* ' = C- I (C- 1), = 1 (I = CC'). The indeterminacy is equivalent to multiplication by an orthogonal matrix; this is called the problem of rotation. Requiring that be diagonal means that the components of f are independently distributed when f is assumed normal. This has an appeal to psychologists because one idea of common mental factors is (by definition) that they are independent or uncorrelated quantities.
572
FACfOR ANALYSIS
A crucial assumption is that the components of U are uncorrelated. Our viewpoint is that the errorS of observation and the specific factors are by definition uncorrelated. That is, the interrelationships of the test scores are caused by the common factors, and that is what we want to investigate. There is another point of view on factor analysis that is fundamentally quite different; that is, that the common factors are supposed to explain or account for as much of the variance of the test scores as possible. To follow this point of view, we should use a different model. A geometric picture helps the intuition. Consider a p-dimensional space. The columns of A can be considered as m vectors in this space. They span some m-dimensional subspace; in fact, they can he considered as coordinate axes in the m-dimensional space, and J can be considered as coordinates of a point in that space referred to this particular axis system. This subspace is called the factor space. Multiplying A on the right hy a matrix corresponds to taking a new set of coordinate axes in the factor space. If the factors are random, the covariance matrix of the observed X is
(3) I
~
cC(X - J-l)(X - J-l)' = cC( AJ + U)( AJ+ U)' = AtPA' + 'It.
If the factors are orthogonal (cCJf' = I), then (3) is
( 4)
I=AA'+'It.
If J and U are normal, all the information about the structure comes from (3) [or (4)] and CCX= J-l.
14.2.2. Identification Given a covariance matrix I and a number m of factors, we cun ask whether there exist a triplet A, tP positive definite, and 'It positive definite and diagonal to satisfy (3); if so, is the triplet unique? Since any triplet can be transformed into an equivalent structure A C, C- I tPC' -I, and 'It, we can put m 2 independent conditions on A and <1> to rule out this indeterminl\cy. The number of components in the ohservahle I. and the number of conditions (for uniqueness) is W(p + 1) + m~; the numbers of parameters in A, tP, and 'It are pm, ~m(m + 1), and p, respectively. If the excess of observed quantities and conditions over number of parameters, namely, H(p - m)2 - p - m], is positive, we can expect a problem of existence but can anticipate uniqueness if a set of parameters does exist. If the excess is negative, we can expect existence but possibly not uniquelless; if the excess is 0, we can hope for both existence and uniqueness (or at least a finite number of solutions). The question of existence of a solution is whether there exists a diagonal
14.2 THE MODEL
573
matrix 'IJ1 with nonnegative diagonal entries such that I - 'IJ1 is poslt\ve semidefinite of rank m. Anderson and Rubin (1956) include most of the known results on this problem. If a solution exists and is unique, the model is said to be i1entified. As noted above, some m 2 conditions have to be put on A and tP to eliminate a transformation A* = AC and <1>* = C-1C,-1. We have referred above to the condition tP = I, which forces a transformation C to be orthogonal. [There are ~m(1Il + 1) component equations in tP = I.] For some purposes. it is convenient to add the restrictions that
(5) iR diagonal. If the diagonal elements of r are ordered and different (YII > 'Y22 > ... > Ymm), A is uniquely determined. Alternative conditions are that the first m rows of· A form a lower triangular matrix. A generalization of this ondition is to require that the first m rows of B A form a lower triangular matrix, where B is given in advance. (This condition is implied by the so-called centroid method.) Simple Structure These are conditions proposed by Thurstone (1947, p. 335) for choosing a matrix out of the class AC that will have particular psychological meaning. If Aja = 0, then the a th factor does not enter into the ith test. The general idea of simple stntcture is that many tests should not depend on all the factors when the factors have real psychological meaning. This suggests that, given a A, one should consider all rotations, that is, all matrices A C where C is orthogonal, and choose the one giving most 0 coefficients. This matrix can be considered as giving the simplest structure and presumably the one with most meaningful psychological interpretation. It should be remembered that the psychologist can construct his or her tests so that they depend on the assumed factors in different ways. The positions of the O's are not chosen in advance, but rotations Care tried until a A is found satisfying these conditions. It is not clear that these conditions effect identification. Reiers~l (1950) modified Thurstone's conditions so that there is only one rotation that satisfies the conditions. thus effecting identification. Zero Elements in Specified Positions Here we consider a set of conditions that requires of the investigator more a priori information. He or she must know that some particular tests do not depend on some specific factol s. In this case, the conditions are that Aja = 0 for specified pairs (i, a); that is, that the a th factor does not affect the ith
574
FACfOR ANALYSIS
test score. Then we do not assume that tf,'ff' = 1. These conditions are similar to some used in econometric models. The coefficients of the ath column are identified except for multiplication by a scale factor if (a) there are at least m - 1 zero elements in that column and if (b) the rank of Na) is /11 - L where Na) is the matrix composed of the rows containing the assigned D's in the ath column with those assigned D's deleted (i.e., the ath column deleted). (See Problem 14.1.) The multiplication of a column by a scale constant can be eliminated by a normalization, such as CPaa = 1 or A.;a = 1 for some i for each a. If CPaa = 1, a = 1, ... , m, then ~ is a correlation matrix. It will be seen that there are m normalizations and a minimum of m(m - 1) zero conditions. This is equal to the number of elements of C. If there are more than m - 1 zero elements specified in one or more columns of A. then there may be more conditions than are required to take out the indeterminacy in A C; in this case thc conditions may restrict A ~ A'. As an example, consider the model
( 6)
V
A21 V A31 V + A32 a
A42 a
+v
a for the scores on five tests, where v and a are measures of verbal and arithmetic ability. The first two tests are specified to depend only on verbal ability while the last two tests depend only on arithmetic ability. The normalizations put verbal ability into the scale of the first test and arithmetic ability into the scale of the fifth test. Koopmans and Reiers~l (1950), Anderson and Rubin (956), and Howe (1955) suggested the use of preassigned D's for identification and developed maximum likelihood estimation under normality for this case. [See also Lawley (958).] Joreskog (1969) called factor analysis under these identification conditions confinnatory factor analysis; with arbitrary conditions or with rotation to simple structure, it has been called exploratory factor analysis.
575
14.2 THE MODEL
Other Conditions A convenient set of conditions is to require the upper square sub matrix of A to be the identity. This assumes that the upper square matrix without this condition is nonsingular. In fact, if A* = (A~', A*2')' is an arbitrary p X m matrix with A~ square and nonsingular, then A = A*A~ -I = (Im' A' 2)' satisfies the condition. (This specification of the leading m X m submatrix of A as 1m is convenient identification condition and does not imply any substantive meaning.)
a
14.2.3. Units of Measurement We have considered factor analysis methods applied to covariance matrices. In many cases the unit of measurelPent of each component of X i., arbitrary. For instance, in psychological tests the unit of scoring has nO intrinsic meaning. Changing the units of measurement means multiplying each component of X by a constant; these constants are not necessarily equal. When a given test score is multiplied by a constant, the factor loadings for the test are multiplied by the same constant and the error variance is multiplied by square of the constant. Suppose DX = X* , where D is a diagonal matrix with positive diagonal elements. Then (1) becomes
(7)
X*=A*J+V*+J-l*,
where J-l* = $ X* = DJ-l, A* = D A, and V* = DV has covariance matrix 'IJI* = D'IJI D. Then
(8)
$( X* - J-l*)( X* - J-l*)' = A* A*' + 'IJI* = I'*,
where '1* = D'1D. Note that if the identification conditions are = I and A' 'IJI- 1A diagonal, then A* satisfies the latter condition. If A is identified by specified O's and the normalization is by cf>aa = 1, a = 1, ... , m (Le., is a correlation matrix), then A* = DA is similarly identified. (If the normalization is Aiu = 1 for specified i for each a, each column of DA has to be renormalized.) A particular diagonal matrix D consists of the reciprocals of the observable standard deviations d ij = 1/ Then l* = DlD is the correlation matrix. We shall see later that the maximum likelihood estimators with identificaLon conditions r diagonal or specified O's transform in the above fashion; that is, the transformation x~=Dxcr' a=l, ... ,N, induces A*=DA and
ru:.
q,* =Dq,D.
576
FACfOR ANALYSIS
14.3. MAXIMUM LIKELIHOOD ESTIMATORS FOR RANDOM ORTHOGONAL FACTORS. 14.3.1. Maximum Likelihood Estimators In this section we find the maximum likelihood estimators of the parameters when the observations are normally distributed, that is, the factor scores and errors are normal [Lawley (1940)]. Then I. = A A , + W. We impose conditions on A and to make them just identified. These do not restrict AA'; it is a positive definite matrix of rank m. For convenience we suppose that = I (i.e., the factors are orthogonal or uncorrelated) and that r = A'W-IA is diagonal. Then the likelihood depends on the mean J1. and I. = A A' + \ft. The maximum likelihood estimators of A and under some other conditions effecting just identificction [e.g., A = (1m' A' 2)'] are transformations of the maximum likelihood estimators of A under the preceding conditions. If XI'"'' x N are a set of N observations on X, the likelihood function for this sample is
(1)
L=(21T)-WNII.I-4-Nexp[-!
a~1 (X a -J1.)'I.-I(Xu -J1.)]'
The maximum likelihood estimator of the mean J1. is ji = Let
x=
(1/N)r.~_1 xu'
N
(2)
E
A =
(xu -i)(x", -i)'.
a-I
Next we shall maximize the logarithm of (1) with J1. replaced by ji; this is t
(3) (This is the logarithm of the concentrated likelihood.) From 1; I. -I = I, we obtain for any parameter e
(4) Then the partial derivative of (3) with regard to "'ii' a diagonal element of \fT, is -N/2 times p
(5)
er
ii
-
E
Ckjerj~ik,
k.j-O
tWe could add the restriction that the off·diagonal elements of A "1,-1 A are 0 with Lagrange multipliers, but then the Lagrange multipliers become 0 when the derivatives are set equal to O. Such restrictions do not affect the maximum.
14.3
577
ESTIMATORS FOR RANDOM ORTHOGONAL FACTORS
where I. -I = (u i }) and (c i ) = C = (l/N)A. In matrix notation, (5) set equal to 0 yields diag I.-I = diag I. -I CI. -I ,
(6)
where diag H indicates the diagonal terms of the matrix H. Equivalently diag I. -I (I. - C)!, -I = diag O. The derivative of (3) with respect to Ak , is - N times p
p
j~1
h,g.}~1
E uk}A}T- E
(7)
UkhChga-S\T'
k=l, ... ,p,
T=I, ... ,I11.
In matrix notation (7) set equal to 0 yields
(8) We have
From this we obtain W-IA(f + I)-I = I.-IA. Multiply (8) by I, and use the above to obtain
A(f +1) = CW-IA,
(10) or
(11) Next we want to show that I,-I_I,-ICI,-1 =I,-I(I,-C)I,'-1 is 'IJ.-l(I. - C)W- 1 when (8) holds. Multiply the latter by I. on the left and on the right to obtain .
(12)
I,'lrl(I. - C) '1'-1 I, = (A A' + '1') '1'-1 ('I' + A A' - C) W- 1( A A' + '1') =W+AA'-C because
(13)
A A'W- 1(W + A A' - C) = A A' + A fA' - A A'W- 1 C = A [(I + f) A' - A'W- 1C]
=0 by virtue of (10). Thus
(14)
578
FACfOR ANALYSIS
Then (6) is equivalent to diag '11 -I (I - C) '11 - 1 = diag o. Since '11 is diagonal. this equation is equivalent to ( 15)
diag( A A' + '11) = diag C.
The estimators .\ and Ware determined by (0), (15), and the requiremept that A' '11 - 1\ is diagonal. We can multiply (11) on the ldt by '11 - ~ to obtain
(16) which shows that the columns of '11- ~A are characteristic vectors of '11- ~(C - '11)'11- ~ = '11- !CW- -l -I and the corresponding diagonal elements of r are the characteristic roots. [In fact, the characteristic vectors of '11- ~cw- ± -I are the characteristic vectors of '11- -lC'W- t because ('11- ~cW- ~ - J)x = yx is equivalent to '11- tcw- -lx = (1 + y)x.] The vectors are normalized by ( '11 - ! A )'( '11 - ~ A ) = A' '11 - 1A = r. The character;stic roots are chosen to maximize the likelihood. To evaluate the maximized likelihood function we calculate ( 17)
tr ci- I = tr ci -I(i - A A ')W- 1 =tr[CW 1-(d;-IA)A'W- 1 ] =tr[CW-1-AA'W- 1 ]
= tr[ (A A' + W)W-I - A A'W- 1 ] =p. The third equality follows from (8) multiplied on the left by i; the fourth equality follows from (15) and the fact that 'P is diagonal. Next we find (18)
Ii I = I W±I·I W-~A A'W·· t
+ IJI w-ll
= I W1.\ A'W--lw-ti~ +Im \ = I WI ·1 P
r + I", 1 m
= n~iin()'j+ ;~
I
j~
I
1).
The second equality is Illll' + I) = Ill'll + I", I for V p X m, which is proved as in (14) of Section XA. From the fact that the characteristic roots of
14.3
579
ESTIMATORS FOR RANDOM ORTHOGONAL FACfORS
'11- t(e - '11)'11- t are the roots 'YI > 'Y2> .,. > 'Yp of 0 = IC- '11 - 'Y'I11 IC-(1 + 'Y)'I1I,
=
(19) [Note that the roots 1 + 'Yi of '11- tCW-l: are positive. The roots 'Yi of 'I1- t(C - W)'I1 ! are not necessarily positive; usually some will be negative.] Then 00
(20)
I ~ 1= I
+ Yj) nf=l(l + Yi)
IClnjEs(l
=
ICI
n j.. s (1 + Yj)'
where S is the set of indices corresponding to the roots in of the maximized likelihood function is (21)
t. The logarithm
-~pNlog21T-~NlogICI-~NElog(1+.yj)-~Np . . j"S
The largest roots Y\ > ... > Ym should be selected for diagonal elements of Then S = {l, ... , m}. The logarithm of the concentrated likelihood (3) is a function of 'I = A A' + W. This matrix is positive definite for every A and every diagonal 'I' that is positive definite; it is also positive definite for some diagonal '11 's that are not positive definite. Hence there is not necessarily a relative maximum for 'I' positive definite. The concentrated likelihood function may increase as one or more diagonal elements of 'I' approaches O. In that case the derivative equations may not be satisfied for 'I' positive definite. The equations for the estinlators (11) and (15) can be written as polynomial equations [multiplying (11) by 1'111], but cannot be solved directly. There are various iterative procedures for finding a maximum of the likelihood function, including steepest descent, Newton - Raphson, scoring (using the information matrix), and Fletcher-Powell. [See Lawley and Maxwell (1971), Appendix II, for a discussion.] Since there may not be a relativ'! maximum in the region for which .pii > 0, i = 1, ... , p, an iterative procedure may define a sequence of values of A and q, that includes ~ii < 0 for some indices i. Such negative values are inadmissible because .pii is interpreted as the variance of an error. One may impose the condition that .pii:2: 0, i = 1, ... ,p. Then the maximum may occur on the boundary (and not all of the derivative equations will be satisfied). For some indices i the estimated variance of the error is 0; that is, some test scores are exactly linear combinations of factor scores. If the identification conditions
t.
580
FACfOR ANALYSIS
= I and A' 'I' -I A diagonal are dropped, we can find a coordinate system for the factors such that the test scores with 0 error variance can be interpreted as (transformed) factor scores. That interpretation does not seem useful. [See Lawley and Maxwell (1971) for further discussion.] An alternative to requiring .p;; to be positive is to require .p;; to be bounded away from O. A possibility is .p;; ~ eu;; for some small e, such as 0.005. Of course, the value of e is arbitrary; increasing e will decrease the value of the maximum if the maximum is not in the interior of the restricted region, and the derivative equations will not all be satisfied. The nature of the concentrated likelihood is such that more than one relative maximum may be possible. Which maximum an iterative procedure approaches will depend on the initial values. Rubin and Thayer (1982) have given an example of three sets of estimates from three different initial estimates using the EM algorithm. The EM (expectation-maximization) algorithm is a possible computational device for maximum likelihood estimation [Dempster, Laird, and Rubin (1977), Rubin and Thayer (1982)]. The idea is to treat the unobservable J's as missing data. Under the assumption that f and V have a joint normal distribution, the sufficient statistics are the means and covariances of the X's and J's. The E-step of the algorithm is to obtain the expectation of the covariances on the basis of trial values of the param".:ters. The M-step is to maximize the likelihood function on the basis of these covariances; this step provides updated values of the parameters. The steps alternate, and the procedure usually converges to the maximum likelihood estimators. (See Problem 14.3.) As noted in Section 14.2, the structure is equivariant and the factor scores are invariant under changes in the units of measurement of the observed variables X --> DX, where D is a diagonal matrix with positive diagonal elements and A is identified by A' '1'-1 A is diagonal. If we let D A = A*, D'\'jI D = '1'*, and DCD = C*, then the logarithm of the likelihood function is a constant plus a constant times
(22)
-log\W* + A*A*'\- trC*(W* + A*A*,)-1
= -loglW + A A'I - tr C(W + A A,)-I - 210g1DI. The maximum likelihood estimators of A* and '1'* are A* = DA and W* = Dq,D, and A* 'q,*-IA* = Aq,-IA is diagonal. Tha: is, the estimated factor loadings and error variances are merely changed by the units of measurement. It is often convenient to use d;; = 1/,;e;, so DCD = (r;) is made up of the sample correlation coefficients. The analysis is independent of the units of measurement. This fact is related to the fact that psychological test scores do not have natural units.
' .. ,
't.
14.3
ESTIMATORS FOR RANDOM ORTHOGONAL FACTORS
581
The fact that the factors do not depend on the location and scale factors is one reason for considering factor analysis as an analysis of interdependence. It is convenient to give some rules of thumb for initial estimates of the' cOlllmunalities, L:;~ 1 A7j = 1 - '''ii' in terms of observed correlations. One rule is to use the Rf.I ..... i- I. i+ I . .... I'. Another is to use max", il/'i"l.
14.3.2. Test of the Hypothesis That the Model Fits We shall derive the likelihood ratio test that the model fits; that is. that for a specified m the covariance matrix can be written as I = 'II + A A' for some diagonal positive definite'll and some p x m matrix A. The likelihood ratio criterion' is
(23)
max .... A.'i' L(p.., 'II + A A') max .... l:L(p..,I)
because the unrestricted maximum likelihood estimator of I is e, tr e( q, + AA)-l=p by (17), and lei/Iii =rrJ~'II+I(1+)yN from (20). The null hypothesis is rejected if (23) is too small. We can usc - 2 times the logarithm of the likelihood ratio criterion: p
E
-N
(24)
log(l
+ Yj)
j=m+ I
and reject the null hypothesis if (24) is too large. If the regularity conditions for q, and A to be asymptotically normally distributed hold, the limiting distribution of (24) under the null hypothesis is X 2 with degrees of freedom p - 111)2 - P - 111], which is the number of elements of I plus the number of identifying restrictions minus the number of parameters in 'II and A. Bartlett (1950) suggested replacing .v by t N - (2p + 11)/6 - 2m/3. See also Amemiya and Anderson (990). From (15) and the fact that YI"'" Yp are the characteristi,; roots of q,- t(e - q, )q,- t we have
H(
(25)
0= tr q,- t( e ~ q, - A k) q,-1 = tr
q,-l(e- q,).q,- ~ -
= tr
q,-l( e - q,) q,-l - trf'
P
=
III
q,- lA A'q,-!
I'
E Yi- E Yi= E ;= I
tr
Yi'
i=m+l
tThis factor is heuristic. If m = O. the factor from Chapter 9 is N - Cp + II lib: Bartlett suggested replacing Nand p by N - m and p - m. respectively.
582
FACfOR ANALYSIS
If I)) < 1 for j = m + 1. ... , p, we can expand (24) using (25) as p
f>
(26)
·-N
E ( Yf - H/ + H/ - .. , ) = ~ N E (Y/ - H/ + .,. ). j~",+
j~m+l
1
Y/.
The criterion is approximately ~Nr.f~m+l The estimators q, and A are found so that C - q, - A A' is small in a statistical sense or, equivalently, so C - q, is approximately of rank m. Then the smallest p - m roots of q,- +lC - q,)q,- + should be near O. The crit~rion measures the deviations of these roots from O. Since Ym + 1 , ••• ,Yp are the nonzero roots of q,.- ;(C - I )q,- ~, we see that p
(27)
2" j =",E+ I yj = ~tr[ q,-l(C- I)q,-ij2 = ttrq,-l(C-i)q,-l(C-i)
=
E (c jj i<j
(TjJ"
~ii~i
because the diagonal elements of C - i are O. In many situations the investigator does not know a value of m to hypothesize. He or she wants to determine the smallest number of factors such that the model is consistent with the data. It is customary to test successive values of m. The investigator starts with a test that the number of factors is a specified mo (possibly 0 Of 1). If that hypothesis is rejected, one proceeds to test that the number is mo + 1. One continues in that fashion until a hypothesis is accepted or until H(p - m)2 - p - m)::; O. In the last event one concludes that no nontrivial factor model fits. Unfortunately, the probabilities of errors under this procedure are unknown, even asymptotically.
14.3.3. Asymptotic Distributions of the Estimators The maximum likelihood estimators A and q, maximize the average concentrated log likelihood functions L*(C, A*, '1'*) given by (3) divided by N for r' ='1'* +A*A*', subject to A*'W*-lA* being diagonal. If C is a consistent estimator of I (the "true" covariance matrix), then L*(C,A*,W*)-> U( 'If + A A', A*, '1'*) uniformly in probability in a neighborhood of A, '11, and L*(W + A A', A*, '1'*) has a unique maximum at 'V* = 'I' and A* A. Because the function is continuous, the A*, W* that maximize U (c. A* , 'I' *) must converge stochastically to A, W. =.0
14.3
ESTIMATORS FOR RANDOM ORTHOGONAL FACfORS
583
Theorem 14.3.1. If A and 'It are identified by A''It-1A being diagonal, if the diagonal elements are different and ordered, and if C .!!.. 'It + A A', then .q,.!!..'ltand A'!!"A. A sufficient condition for C.!!.. 'I is that (I' U')' has a distribution with finite second-order moments. The estimators A and .q, are the solutions to the equations (10), (15), and the requirement that A' 'It -I A is diagonal. These equations are polynomial equations. The derivatives of A and .q, as functions of C are continuous unless they become infinite. Anderson and Rubin (1956) investigated conditions for the derivative to be finite and proved the following theorem: Theorem 14.3.2.
(28)
Let ((lij)=@=W-A(A'W-1A)"IA"
If «(Ii}) is nonsingular, if A and Ware identified by the condition that A' W- l A is diagonal and the diagonal elements are different and ordered, if C .!!.. 'It + A A', and if IN (C - 'I) has a limiting normal distribution, then IN (A - A) and IN (.q, - W) have a limiting normal distribution.
For example, IN(C - 'I) will have a limiting distribution if (J' U')' has a distribution vith finite fourth moments. The covariance matrix of the limiting distribution of IN (A - A) and IN (.q, - W) is too complicated to derive or even present here. Lawley U953) found covariances for IN (A - A) appropriate for W known, and Lawley (1967) extended his work to the case of W estimated. [See also Lawley and Maxwell (1971).] lennrich and Thayer (1973) corrected an error in his work. The covariance of 1N(~ii - !/Ii) and 1N($jj - !/Iii) in the limiting distribution is (29)
i,j= 1,: .. ,p,
where (gij) = «(li})-l. The other covariances are too involved to give here. While the asymptotic covariances are too complicated to give insight into the sampling variability, they can be programmed for computation. In that case the parameters are replaced by their consistent estimators. 14.3.4. Minimum-Distance Methods
An alternative to maximum likelihood is generalized least squares. The estimators are the val1les of Wand A that minimize
(30)
tr(C-'I)H(C-'I)H,
584
FACfORANALYSIS
where I=W+AA' and H=I-I or some consistent estimator of I-I. When H = I-I, the objective function is of the form (31)
[C -
(J" (
'11 , A)]' [cov c] - 1 [ C -
(J"
('11 , A)] ,
where t: represents the elements of C arranged in a vector, (J"( '11, A) is '11 + A A' arranged in a corresponding vector, and cov c is the covariance matrix (jf c under normality [Anderson (1973a)]. J6reskog r.nd Goldberger (1972) use C- I for H and minimize (32) The matrix of derivatives with respect to the elements of A set equal to 0 forms the matrix equation (33) This can be rewritten as (34) Multiplication on the left by I-ICI- 1 yields (8\ which leads to (10). This estimator of A given '11 is the same as the maximum likelihood estimator except for normalization of columns. The equation obtained by setting the derivatives of (32) with respect to '11 equal to 0 is (35)
diag C
1
[('11 + A A') - C]C 1 = diagO.
An ait«rnative is to minimize (36)
!tr {( '11 + A A ') -I [ C -
2
( '11 + A A ') ] }
•
This leads to (8) or (10) and (37) Browne (1974) showed that the generalized least squares estimator of '11 has the same asymptotic distribution as the maximum likelihood estimator. Dahm and Fuller (1981) showed that if cov c in (31) is replaced by a matrix converging to cov c and '11, A, and «II depend on some parameters, then the asymptotic distributions are the same as for maximum likelihood. 14.3.5. Relation to Principal Component Analysis What is the relation of maximum likelihood to the principal component analysis proposed by Hotelling (1933)? As explained in Chapter 11, the vector of sample principal components is the orthogonal transformation B' X, where
14.3 ESTIMATORS FOR RANDOM ORTHOGONAL FACfORS
585
e
normalized by B' B = T.
th; columns of B are the characteristic vectors of Then p
(38)
e = BTB' = L: b;t;b;, ;=)
where T is the diagonal matrix with diagonal clements 1 1 , ••• ,11" the characteristic roots of C. If Im+), ... ,lp are small, e can be approximated by (39)
B)T)B; =
L: b;t;b;, ;~
)
where T1 is the diagonal matrix with diagonal elements approximated by
11" .• ,tm ,
and X is
1/1
(40)
B)BIX=
L: b;(b;X). i=)
Then the sample covariance of the difference between X and the approximation (40).is the sample covariance of
(41) which is B::T2 B'z = I:.[=m+) b;t;b;, and the sum of the variances of the components is I:.[=m + 1 Ii' Here 1'2 is the diagonal matrix with tm+) , .•. , t p as diagonal elements. This analysis is in terms of some common unit of measurement. The first m components "explain" a large proportion of the" variance," tr e. When the units of measurement are not t he same (e.g., when the units are arbitrary), it is customary to standardize each measurement to (sample) variance l. However, then the principal components do not have the interpretation in telms of variance. Another difference between principal component analysis and factor analysis is that the former does not separate the error from the systematic part. This fault is easily remedied, however. Thomson (1934) proposed the following estimation procedure for the factor analysis model. A diagonal matrix '11 is subtracted from C, and the principal component analysis is carried out on C - '11. However, '11 is determined so e - 'II is close to rank m. The equations are
(42) (43)
(44)
(e-'I1)A=AL,
diag( 'II + A A) = diag e, N A = L diagonal.
The last equation is a normalization and takes out the indeterminacy in A. This method allows for the error terms, but still depends on the units of
586
FACTOR ANALYSIS
measurement. The estimators are consistent but not (asymptotically) efficient in the usual factor analysis model. 14.3.6. The Centroid Method Before the availability of high-speed computers, the centroid method was used almost exclusively because of its computational ease. For the sake of history we give a sketch of the method. Let R* be the correlation reduced matrix, that is, the matrix consisting of r;j, i j, and 1 - $/;, where $/t; is a, initial estimate of the error variance in standard deviation units. Thomson's principal components approach is first to find the m characteristic vectors of Ro = R* corresponding to the m largest characteristic roots. As indicated in Chapter 11, one computational method involves starting with an initial estimate of the first vector, say x(°l, calculating X(I) = Rox(O), and iterating. At the rth step x(r) is approximately 'YI x(r-I), where 'Yl is the largest root and x(r)'x(r) - 'Y~xlr-l)'x(r-l). Then YI =x(r)/V'YIX(r-l)'x(r-l) is approximatdy
*"
the first characteristic vector normalized so y'IYl = 'Yl' To oJtain the second vector. apply the same procedure to RI = R* - Y I y;. The centroid method can be considered as a very rough approximation to the principal component approach. With psychological tests the correlation matrix usually consists of positive entries, and the first characteristic vector has all positive components, often of about the same value. The centroid method uses E = 0, ... , 1)' as the initial estimate of the first vector. Then R* E = x(l) is the first iterate and should be an approximation to the first characteristic vector. An approximation to the first characteristic root is E . R* E / E' E. Then Y I = x( I) / ,; E ' R* E is an approximation to the first characteristic vector of R* normalized to have length squared 'Yl' The operations can he carried out on an adding machine or on a desk calculator because R* E amounts to adding across rows and E' R* E is the sum of those row totals. The second characteristic vector is orthogonal to the first. A vector orthogonal to E is E* consisting of p/21'sand p/2 -l's. Then RIE* =x 2 is an approximation to the second characteristic vector, and E*' R] E* /E* 'E* approximates the second characteristic root. These operations involve changing signs of entries of R I and adding. The positions of the -l's in E* are selected to maximize E*' R( E*. The procedure can be continued.
14.4. ESTIMATION FOR FIXED FACTORS Let x"
(1)
=
(x\o,"" xpa)' be an observation on Xu given by
14.5
587
FACfOR INTERPRETATION AND TRANSFORMATION
with fa being a nonstochastic vector (an incidental parameter), a = 1, ... , N, satisfying 'L~~da = O. The likelihoud function is
(2)
L
=
n exp - 1." 2 P
1
[(27T/nf~I"'iil
N/2
i~I
{
N
L."
a~I
(
x iu
_
_ J1-i
~)2
'-i~1 A, ilia \"'m
",..
\ f
.
II
This likelihood function does not have a maximum. To show this fact, let J1-1 = 0, All = 1, A lj = (j *" n, h, =Xlu' Then XI" - J1-1 - 'Lj!:1 Aijfjet = 0, and "'11 does not appear in the exponent but appears only in the constant. As "'11 -> 0, L -> 00. Thus the likelihood does not have a maximum, and the maximum likelihood estimators do not exist [Anderson and Rubin (1956)]. Lawley (1941) set the partial derivatives of the likelihood equal to 0, but Solari (1969) showed that the solution is only a stationary value, not a. maximum. Since maximum lil:elihood estimators do not exist in the case of fixed factors, what estimation methods can be used? One possibility is to use the maximum likelihood method appropriate for random factors. It was stated by Anderson and Rubin (1956) and proved by Fuller, Pantula, and Amemiya (1982) in the case of identification by D's that the asymptotic normal distribution of the maximum likelihood estimators for the random case is the same as for fixed factors. The sample covariance matrix under normality has the nonccntral Wishart distribution [Anderson (1946a)] depending on 'IT, A
°
14.5. FACTOR INTERPRETATION AND TRANSFORMATION 14.5.1. Interpretation The identification restrictions of A 'W- I A diagonal or the first m rows of A being 1m may be convenient for computing the maximum likelihood estimators, but the components of the factor score vector may not have any intrinsic meaning. We saw in Section 14.2 that 0 coefficients may give meaning to a factor by the fact that this factor does not affect certain tests. Similarly, large factor loadings may help in interpreting a factor. The coefficient of verbal ability, for example, should be large on tests that look like they are verbal. In psychology each variable or factor usually has a natural positive direction: more answers right on a test and more of 'the ability represented by the factor. It is usually expected t1.at more ability leads to higher performance; that is, the factor loading should be positive if it is not O. Therefore, roughly
588
FACfORANALYSIS
.(All. A12) '(A21o A22) '(A31o A32)
------------+------------Ail
Figure 14.1. Rows of
A.
speaking, for the sake of interpretation, one may look for factor loadings that are either 0 or positive and large. 14.5.2. Transformations The maximum likelihood estimators on the basis of some arbitrary identification conditions including «I> = I are A and W. We consider transformations (1) If the factors are to be orthogonal, then «1>* = I and P is orthogonal. If the factors are permitted to be oblique, P can be an arbitrary nOllsingular matrix and cP* an arbitrary positive definite matrix. The rows of A can be plotted in an m-dimensional space. Figure 14.1 is a plot of the rows of a 5 x 2 matrix A. The coordinates refer to factors and the points refer to tests. If «1>* is required to be 1m' we are seeking a rotation of coordinate axes in this space. In the example that is graphed, a rotation of 45 0 would put all of the points into the positive quadrant, that is, At ~ O. One of the new coordinates would be large for each of the first three points and small for the other two points, and the other coordinate would be small for the first three and large for the last two. The first factor is representative of what is common to the first three tests, a ..1d the second factor of what is common to the last two tests. If m > 2, a general rotation can be approximated manually by a sequence of tVlo-d;rnensional rotations.
14.5
FACTOR INTERPRETATION' "NO TRANSFORMATION
589
If «1>* is not required to be 1m , the transformation P is simply nonsingular. If the normalization of th,; jth column of A is Ai(j),j = 1, then m
(2)
1=
Xi(j),j =
L
Ai(j).kPkj;
k~l
each column of P satisfies such a constraint. If the normalization is ¢jj = 1, then
(3)
1 = ¢jj
=
L
(pjkf,
k~l
where (pjk)=p-l. Of the various computational procedures that are based on optimizing an objective function, we describe the varimax method proposed by Kaiser (1958) to be carried out on pairs of factors. Horst (1965), Chapter 18, extended the method to be done on all factors simultaneously. A modified criterion is
(4) which is proportional to the sum of the column variances of the squares of the transformed factor loadings. The orthogonal matrix P is selected so as to maximize (4). The procedure tends to maximize the scatter of Ai/ within columns. Since Ai/ ~ 0, there is a tendency to obtain some large loadings and some near O. Kaiser's original criterion was (4) with Aj/ replaced by Ar/ /Eh~l Ait . Lawley and Maxwell (I971) describe other criteria. One of them is a measure of similarity to a predetermined P x m matrix of 1's and D's. 14.5.3. Orthogonal versus Oblique Factors In the case of orthogonal factors the components are uncorrelated in the population or in the sample according to whether the factors are considered random or fixed. The idea of uncorrelated factor scores has appeal. Some psychologists claim that the orthogonality of the factor scores is essential if one is to consider the factor scores more basic than the test scores. Considerable debate has gone on among psychologists concerning this point. On the other side, Thurstone (1947), page vii, says "it seems just as unnecessary to require that mental traits shaH be uncorrelated in the general population as to require that height and weight be uncorrelated in the general population." As we have seen, given a pair of matrices A, «II, equivalent pairs are given by A P, p-l «IIp,-1 for nonsingular P's. The pair may be selected (i.e .. the P
590
FACTOR ANALYSIS
given A. <1» as the one with the most meaningful interpretation in terms of the subject matter of the tests. The idea of simple structure is that with (} factor loadings in certain patterns the component factor scores can be given meaning regardless of the moment matrix. Permitting to be an arbitrary positive definite matrix allows more O's in A. Another consideration in selecting transformations or identification conditions is autonomy, or permanencc, or invariance with regard to certain changes. For example, what happens if a selection of the constituents of a population is made? In case of intelligence tests, suppose a selection is made, such as college admittees out of high school seniors, that can be assumed to involve the primary abilities. One can envisage that the relation between unobselved factor scores / and observed test scores x is unaffected by the selection, that is, that the matrix of factor loadings A is unchanged. The variance of the errors (and specific factors), the diagonal elements of'll, may also be considered as unchanged by the selection because the errors are uncorrelated with the factors (primary abilities). Suppose there is a tme model. A, <1>, 'II, and the investigator applies identification conditions that permit him to discover it. Next, suppose there is a selection that results in a new population of factor scores so that their covariance matrix is <1>*. When the investigator analyzes the new observed covariance matrix'll + A <1>* A " will he find A again? If part of the identification conditions are that the factor moment matrix is I, then he wiil obtain a different factor loading matrix. On the ott er hand, if the identification conditions are entirely on the factor loadings (specified D's and l's), the factor loading matrix from the analysis is the same as before. The same consideration is relevant in comparing two populations. It may be reasonable to consider tt.at 'II I = '11 2, A I = A 2' but I <1>2' To test the hypothesis that <1>1 = <1>2' one wants to use identification conditions that agree with A I = A 2 (rather than A I = A 2C), The condition should be on the factor loadings. What happens if more tests are added (or deleted)? In addition to observing X= A/+ /.I. + U, suppose one observes X* = A*/+ /.1.* + U*, wlll:n: U* is uncorrc\atcd with U. Since the common factors / are unchanged, is unchanged. However, the (arbitrary) condition that A'W-IA is diagonal is changed; use of this type of condition would lead to a rotation oHA' A*').
'*
14.6. ESTIMATION FOR IDENTIFICATION BY SPECIFIED ZEROS We now consider estimation of A, 'II, and when is unrestricted and A is identified by specified O's and 1's. We assume that each column of A has at
14.7
ESTIMATION OF FACTOR SCORES
591
least m + 1 O's in specified positions and that the submatrix consisting of the rows of A containing the O's specified for a given column is of rank m - 1. (See Section 14.2.2.) We further assume that each column of A has 1 in a specified position or, alternatively, that the diagonal element of «I> corresponding to that column is 1. Then the model is identified. The likelihood function is given by (1) of Section 14.3. The derivatives of the likelihood function set equal to 0 are
(1)
diag I-I [c - ('IJI + A«I>A')]I-I = diagO,
(2)
A'I-I[C- ('IJI + A«I>A')]I-1A = 0
for positions in «I> that are not specified, and (3) for positions in A not specified, where (4)
I=W+A«I>A.
These equations cannot be simplified as in Section 14.3.1 because (3) holds only for unspecified positions in A, and hence one cannot multiply by I on the left. [See Howe (1955), Anderson and Rubin (1956), and Lawley (1958).] These equations are not useful for computation. The likelihood function, however, can be maximized nume:·ically. As noted before, a change in units of measurement, X* = DX, results in a corresponding change in the parameters A and 'IJI if identification is by 0 in specified positions of A and normalization is by CPjj = 1, j = 1, ... , m. It is readily verified that the derivative equations (1), (2), (3), and (4) are changed in a corresponding manner. Anderson and Amemiya (1988a) have derived the asymptotic distribution of the estimators under general conditions. Normality of the observations is not required. See also Anderson and Amemiya (1988b).
14.7. ESTIMATION OF FACTOR SCORES It is frequently of interest to estimate the factor scores of the individuals in the group being studied. In the model with nonstochastic factors the factor scores are incidental paraneters that characterize the individuab. As we have seen (Section 14.4), the maximum likelihood estimators of the parameters ('II, A, fL, II' ... ,IN) do not exist. We shan therefore study the estimation of the factor scores on the basis that the structural parameters ('IJI, ,\, fL) are known.
592
FACfORANALYSlS
When fa is considered as an incidental parameter, Xa - fl. is an observation from a distribution with mean A fa and covariance matrix '11. The weighted least squares estimator of fa is (1)
fa = (A ''11- A) 1
-I
= r-1A'W-I(x a
A''IT-1(x a -
-
IJ.)
fl.),
where r = A''IT-1A (not necessarily diagonal). This estimator is uri"'biased and its covariance matrix is
(2) by the usual generalized least squares theoi)' [Bartlett (1937b),(1938)]. It is the minimum variance unbiased linear estimator of fa' If xa is normal, the estimator is also maximum likelihood. When fa is considered random [Thomson (1951)], we suppose Xa and fa have a joint normal distribution with mean vector (fl.', 0')' and covariance matrix
(3)
c(X) = ('IT+A«PA' f «PA'
A«P ) «P .
Then the regression of f on X (Section 2.5) is
( 4)
cS'(J1X) = «PA'( 'IT + A«PA') -\ x - fl.)
= «P( «P + «P r «P) -I «P A ''11- 1(x - fl.). The estimator or predictor of fa is
(5) If «P = J, the predictor is
(6) When r is also diagonal, the jth element of (6) is Yj/(1 + Yj) times the jth element of (1). In the conditional distribution of Xc< given fa (for «P = I)
(7)
cS'(f:lfa) = (J + r) -I rfa'
(8)
C(i:lfa) = (I + r) -lr(I + r)-\
593
PROBLEMS
(9)
$ [(
f: - fer) (l* - fa)' If" 1 = ( I + f) -
1
(f + fa f~ ) (I + f) - 1 •
$(/: - fa )(/;' -fa)' = (/+ f)-l.
(10)
This last matrix, describing the mean squared error, is smaller than (2) describing the unbiased estimator. The estimator (5) or (6) is a Bayes estimator and is appropriate when fer is treated as random.
rROBLEMS 14.1. (Sec. 14.2)
identijication by U's. Let
A--
0 ( A(I)
c=
(C
II
C:!l
where C is nonsingular. Show that
implies
if and only if
14.2. (Sec. 14.3)
NI)
is of rank m - I.
For p = 3, m = 1. and A = A. prove I e;}1 = n;~ I( Ai jI]JJ.
14.3. (Sec. 14.3) The EM algorithm.
f and U are normal and f and X are observed. show that the likelihood function based on (xp.fl), ... ,(x.Ii,!,v) is
(3) If
•
1
I
I
(27r)''''II:
. [ -.,'f'(f.... .... ·· 'f exp
-
I
,
l}
.
594
FACfORANALYSIS
lb) Show that when the factor scores are included as data the sufficient set of statistics is i, j, C u = C,
1 CfJ=f\i
N
_
_
L (f.-f)(f,,-f)'. u=1
lc) Show that the conditional expectations of the covariances in (b) given X=(xl, ... ,X N ), A, «1>, and 'if are
C:, =.f( C"IX, A, «1>, 'if) C:[ =
cC( C,[IX,
=
Cu
'
A, «1>, 'if) = CxxC'if
CfJ = cC( CrrlX, A, «1>, 'if) =
+ A«I>A') -I A «I> ,
«I> A '('if + A «I> A ') -IC,,('if + A«I>A') -I A«I>
+ «I> - «I> A '('if + A«I>A') - I A«I>. ld) Show that the maximum likelihood estimators of A and 'if given «I> = I are
CHAPTER 15
Patterns of Dependence; Graphical Models
15.1. INTRODUCTION
An emphasis in multivariate statistical analysis is that several measurements on a number in individuals or objects may be correlated, and the methods developed in this book take Recount of that dependence. The amonnt of association between two variables may be measured by the (Pearson) correlation of them (a symmetric measure); the association between one variable and a set may be quantified by a multiple correlation; and the dependence between one set and another set may be studied by criteria of independence such as studied in Chapter 9 or by canonical correlations. Similar measures can be applied in conditional distributions. Another kind of dependence (asymmetrical) is characterized by regression coefficients and related measures. In this chapter we study models which involve several kinds of dependence or more intricate patterns of dependence. A graphical model in statistics is a visual diagram in which observable variables are identified with points (vertices or nodes) connected by edges and an associated family of probability distributions satisfying some independences specified by the visual pattern. Edges may be undirected (drawn as line segments) or directed (drawn as arrows). Undirected edges have to do with symmetrical dependence and independence, while directed edges may reflect a possible direction of action or sequence in time. These independences may come from a priori knowledge of the subject matter or may derive from these or other data. Advantages of the graphical display include
An Introduction to Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0-471-36091-0 Copyright © 2003 John Wiley & Sons, Inc.
595
596
PAITERNS OF DEPENDENCE; GRJ\PHICJ\L MODELS
ease of comprehension, particularly of complicated patterns, ease of elicitation of expert opinion, and ease of comparing probabilities. Use of such diagrams goes back at least to tile work of the geneticist Sewall Wright (1921),(1934), who used the term "path analysis." An elaborate algebra has been developed for graphical models. Specification of independences reduces the number of parameters to be determined. Some of these independences are known as Markov properties. In a time series analysis of a Markov process (or order 1), for example, the future of the process is considered independent of the past when the present is given; in such a model the correlation between a variable in the past and a variable in the future is determined by the correlation between the present variable and the variable of the immediate future. This idea is expanded in several ways. The family of probability distributions associated with a given diagram depends on the properties of the distribution that are represented by the graph. These properties for diagrams consisting of undirected edges (known as undirected graphs) will be described in Section 15.2; the properties for diagrams consisting entirely of directed edges (known as directed graphs) in Section 15.3; and properties of diagrams with both types of edges in Section 15.4. The methods of statistical inference will he given in Section 15.5. In this chapter we assume that the variables have a joint nonsingular normal distribution; hence, the characterization of a model is in terms of the covariance matrix and its inverse, and functions of them. This 'issumption implies that the variables are quantitative and have a positive density. The mathematics of graphical models may apply to discrete variables (contingency tables) and to nonnormal quantitative variables, but we shall not develop the theory necessary to include them. There is a considerable social science literature that has followed Wright's original work. For recent reviews of this writing see, for example, Pearl (2000) and McDonald (2002).
15.2. UNDIRECTED GRAPHS
A graph is a set of vertices and edges, G == (V, E). Each vertex is identified with a random vector. In this chapter the random variables have a joint normal distribution. Each undirected edge is a line connecting two vertices. It is designated by its two end points; (u, v) is the same as (v, u) in an undirected graph (but not in directed graphs). Two vertices connected by an edge are called adjacent; if not connected by an edge, they are called nonadjacent. In Figure 15.l(a) all vertices are
15.2
597
UNDIRECTED GRAPHS
./b
.b
a. c (a)
a
c
L Ii a
c
c
(c)
(b)
(d)
Figure 15.1
nonadjacent; in (b) a and b a1 e adjacent; in (c) the pair a and b and the pair a and c are adjacent; in (d) every pair of vertices are adjacent. The family of (norma\) distributions associated with G is defined by a set of requirements on conditional distributions, known as Markov properties. Since the distributions considered here are normal, the conditions have to do with the covariance matrix l: and its inverse A = l: -I , which is known as the concentration matrix. However, many of the lemmas and theorems hold for nonnormal distributions. We shall consider three definitions of Markov and then show that they are equivalent. Definition 15.2.1. The probability distributioll 011 a graph is pairwise Markov with respect to G if for every pair of vertices (u, v) that are not a((iacent Xu and Xu are independent conditional on all the other variables in the graph. In symbols
(1)
Xu JL XulX V\(u. uj'
where JL means independence and V\ (u, v) indicates the set V with II and v deleted. The definition of pairwise Markov is that Puu,V\l'''''i = 0 for all pairs for which (u, v) $. E. We may also write u JL vi V\ (u, v). Let l: and A = l: -1 be partitioned as
(2)
A = [AAA ABA
A'III}
ABB '
where A and B are disjoint sets of vertices. The conditional distribution of X A given X B is
(3) The condition II covariance matrix is
(4)
598
PATIERNS OF DEPENDENCE; GRAPHICAL MODELS
If A = (1,2) and B = (3, ... , p), the covariance of XI and X 2 given X 3 , ••• , Xp is UI~';p in LAB = (ujn ... p)' This is 0 if and only if AI2 = 0; that is, I A . B is diagonal if and only if A A A is diagonal.
Theorem 15.2.1. (i, j) $.
If a distribution on a graph is pai/Wise Mari:ov,
Aij
= 0 for
v.
Definition 15.2.2. The boundary of a set A, termed bd(A), consists of those vertices not in A that are adjacent to A. The closure of A, termed d(A), i:; Au bd(A). Definition 15.2.3. A distribution on a graph is locally Markov if for every vertex v the variable Xu is independent of the variables not in deB) conditional on the boundary of v: in notation,
Theorem 15.2.2.
The conditional independences
X lL Y1z,
(6)
X lL ZIY
hold if and only if
(7)
XlL(Y,Z).
Proof The relations (6) imply that the density of X, Y, and Z can be written as f(x,y, z) = f(xlz)g(ylz)h(z)
(8)
= k(xly)l(zly)m(y). Since g(ylz)h(z) = n(y, z) = l(zly )m(y), (8) implies f(xlz) = k(xly), which in turn implies f(xlz) = k(xly) = p(x). Hence (9)
f(x, y, z)
= p(x)n(y, z),
which is the density generating (7). Conversely, (9) can be written as either • form in (8), implying (7). Corollary 15.2.1.
(10) hold if and only
( 11)
The relations X lL
Y1Z, W,
XlLZIY,W
If
x lL(Y,Z)IW.
15.2
599
UNDlRECfED GRAPHS
The relations in Theorem 15.2.2 and Corollary 15.2.1 are sometimes called the block independence theorem. They are based on positive densities, that is, nonsingular normal distributions. Theorem 15.2.3. Markov.
A locally Markov distribution on a graph is pabwise
Proof Suppose the graph is locally Markov (Definition 15.2.3). Let u and v be nonadjacent vertices. Because v is not adjacent to u, it is not in bd(u); hence, (12)
The relation (12) can he written (13)
XII
lL {Xv, X V\[II.v.hd(II»)} Ibd( u).
Then Corollary 15.2.1 (X=X II , Y=X v' Z=ZV\[d(u).v), W=Xbd(u» implies
•
(14) ~'
11,
r
Theorem 15.2.4. Markov.
A pabwise Markov distribution on a graph is locally
Proof Let V\cI(u) = (15)
u lL vJlbd(u)
U
U ... U
VI
v2 U
Vn •
Then
... U vn '
which by Corollary 15.2.1 implies
(16) Further, (16) and
(17) imply
(18) This procedure leads to (19)
u lL
VI
U ... U
v"lbd(u).
•
A third notion of Markov, namely, global, requires some definitions.
600
PATIERNS OF DEPENDENCE; GRAPHICAL MODELS
Definition 15.2.4. A path from 8 to C is a sequence vo, VI' v2 '· •• , Vn of adjacent vertices with Vo E Band vn E C. Definition 15.2.5. A set S separates sets Band C if S, B, and Care disjoint and every path from B to C intersects S. Thus S separates Band C if for every sequence of vertices VO' VI'··.' vn with Vo E Band Vn E C at least one of VI' ... ' Vn _I is a vertex in S. Here B and/or Care nonempty, but S can be empty. Definition 15.2.6. A distribution on a graph is globally Markov if for every triplet of disjoint sets S, 8, and C such that S separates Band C the vector variables X B and Xc are independent conditional on Xs. In the example of Figure IS.1(c), a separates band c. If Pbc.a = 0, that is, = 0, the distribution is globally Markov. Note that a set of vertices is identified with a vector of variables. The global Markov property puts restrictions on the possibie (normal) distributions, and that implies fewer parameters about which to make inferences. Suppose V = A u BUS, where A, B, and S are disjoint. Partition I and A = I -I, the concentration matrix, as Pbc - Pba Pac
(20)
The conditional distribution of (X~, matri'<:
(21)
[IAA I(A.R).S = IBA
X~)'
given Y s is normal with covariance·
IAB] _ [IAS ]I-1[I IBR I BS S5 SA
r
I SB ]
l
= [AAA
ABA
AAB ABB
Theorem 15.2.5. If S separates A and B in a gr
15.2
601
UNDIRECTED GRAPHS
A and B without intersecting S. The globally Markov property is that XI
and X B are uncorrelated in the conditional distribution, implying that I, o'f. Ii 'os is block diagonal and hence that AAB = O. • Theorem 15.2.6. Markov.
A distribution on a globally Markov graph is paim'ise
Proof Let the set B be i, the set C be j not adjacent to i. and the set A the rest of the variables. Any path from B to C must include elements of A Hence i is independent of j in the distribution conditioned on the other variables. • 0
Theorem 15.2.7. locally Markov.
A globally Markov family of distributions on a graph is
Proof The boundary of a set B separates Band V\ d( B).
Theorem 15.2.8. globally Markov.
•
A pairwise Markov family of distributions on a graph is
Proof Let A, B, and S be disjoint sets in a pairwise Markov graph such that S separates A and B. Let #(S) and #(V) denote the numbers of vertices in S and V, respectively. If #(V) = #(S) + 2, that is, V = Au BuS. then there must be one vertex in each of A and B, and the pairwise Markm property is exactly the globally Markov property. The rest of the proof is a backward induction on #(S). Suppose #( V) - #(S) > 2 and V = A U BUS. Then either A or B or both have more than one vertex. Suppose A has more than one vertex, and let u EA. Then S U u separates A \u and B, and SuA separates u and B. By the induction hypothesis o
(22) By Corollary 15.2.1
(23) Now suppose A U BUS c V. Let u E V\ (A U BuS). Then S U II separates A and B. By the induction hypothesis
(24) Also, either Au S separates u and B or BUS separates A and II. (Otherwise there would be a path from B to u and from u to A that would
602
PATTERNS OF DEPENDENCE; GRAPHICAL MODELS
not intersect S.) If AU S scpa 'atcs
II
and B,
(25) Then Corollary 15.2.1 applied to (19) and (20) implies
(26) from which we derive X~ lL X/JIXs '
•
Theorems 15.2.3, 15.2.5, and 15.2.6 show that the three Markov properties are equivalent: anyone implies the other two. The proofs here hold fairly generally, but in this chapter a nonsingular mult;variate normal distribution is assumed: thus all densities are positive. Definition 15.2.7. A graph G = (V, E) is complete if and only if every two vertices ill V are adjacem. The definition implies that the graph specifies no restnctlon on the covariance matrix of the multivariate normal distribution. A subset A <:;; V induces a subgraph G.\ = (A, E/I), where the edge set EA includes all edges (Ll, v) of G with (Ll, v) E E, where u E A and v EA. A subset of a graph is complete if and only if every two vertices in A are adjacent in E,\. Definition 15.2.S.
A clique is a maximal complete set of vertices .
.. Maximal" means that if another vertex from V is added to the set, the set will no longer be complete. A clique can be constructed by starting with one vertex. say VI' If it is not adjacent to any other vertex, VI alone constitutes a clique. If [J2 is adjacent to VI [(VI' V2) E EJ, continue constructing a clique with VI and V 2 in it until a maximal complete subset is obtained. Thus every vertex is a member of at least one clique, and every edge is included in at least one clique. Lemma 15.2.1.
scI
If Ih" dislriiJlllio// of XI' is Markov, it is detemlined by the
or murgi//ul dislrihllliolls of all cliques.
In Figure 15.1(a) each of a, b, (' is a clique; in (b) each of (a, h) and c is a clique: in (c) each of(a, b) and (a, c) is a clique; in (d) (a, h, c) is a clique. Definition 15.2.9. The density f(X v ) factorizes with respect to G if there are //onnegalive jilllctiolls gc(Xc ) depending on the complete sLlbgraphs such that
(27)
f(Xv) =
n
Ccompll'h:
gdXc)'
15.2
UNDIRECTED GRAPHS
603
Since it suffices to consider only cliques, an alternative factorization is
(28) These functions gc(Xc ) and gc'(Xc') are not necessarily densities or conditional densities. The problems of statistical inference may be reduced to the problems of the complete sub graphs or cliques. Definition 15.2.10. A decomposition of a graph is formed by three disjoint sets A, B, S if V = A U BuS, S separates A and B, and S is complete.
In this definition one or mort. of the sets A, B, and S may be empty. If both A and Bare nonempty, the decomposition is termed proper. DefinUion 15.2.11. A graph is decomposable if it is complete or if there is a proper decomposition (A, B, S) into decomposable subgraphs GA U s and G BuS ' Theorem 15.2.9. Suppose A, B, S decomposes G = (V, E). Then the density of X v factorizes with respect to G if and only if its marginal densities fA U sCXA U s) and fB U SCx BUS) factorize and the densities satisfy
(29) Proof Suppose that f vex v) factorizes as
(30) Because A, B, S decomposes G, every clique is either a subset of A uS or a subset of BuS. Let d denote the cliques that are subsets of A U S, and f?6 those that are subsets of B. T1en fv(xv)=h(xAus)k(xBus), where
(31) (32) Integration of (30) with respect to Xc for C E .'1B\s1 gives
(33)
604
PATIERNS OF DEPENDENCE; GRAPHICAL MODELS
grade 3
-----~
difficulty 1
recommendation 4
.------+1.
2. - - - - - - - - + 1 • 5 IQ SAT Figure \5.2
where
•
(34) In turn
fAUS(X AUS )
and
fBUix BUS )
can be factorized, leading to (28).
15.3. DIRECTED GRAPHS We now include relations with a direction; the measurement represented by one vertex u may precede the measurement represented by another vertex v. In the graph this directed edge is displayed as an arrow pointing from u to v; in notation it appears as (u, v), which is now distinguished from (v, u). The precedence may indicate the times of measurement, for example, the precipitation on two successive days, or may indicate possible causation. The difficulty of an examination Xl may affect the grade of a student x3; the grade is also affected by his/her IQ X 2 • In turn the grade of the student influences the quality of a letter of recommendation x 4 ; the IQ is a factor in performance on the SAT, xs. See Figure 15.2. (We shall draw figures so that the action proceeds from left to right.) A graph composed entirely of directed edges is called a directed graph. A cycle, such as 1 -> 2, 2 -> 3, 3 -> 1, is hard to interpret and hence is usually ruled out. A directed graph without a cycle is an acyclic directed graph (ADG or DAG), also known as an acyclic digraph. All directed graphs in this chapter are acyclic. An acyclic directed graph may represent a recursive linear system. For example, Figure 15.2 could represent
(1) (2)
15.3
605
D1RECfED GRAPHS
+ {332 X 2 + u 3 ,
(3)
X3 =
(4)
X4={343X3+U4'
(5)
X, = (3'2 X 2 + u"
{33I X I
where u l ' u 2 , u 3 ' u 4 , Us are mutually independent unobserved variahles. Wold (1960) called such models callsal chains. Note that the matrix of coefficients is lower triangular. In general Xi may depend on Xl.·.·. Xi .1' The recursive linear system (I) to (5) generates the recursive factorization
(6)
f12345 ( X I' X 2' X 3' X 4' X 5)
= fl( x I) f2( x 2 ) f3112 (x3lx I' X 2 )f41123 (x 4 lx 3 )f511234 (XSIX2)' A directed graph induces a partial order. Definition 15.3.1. A partial ordering of an acyclic directed graph defined by the existence of a directed path
(7)
u
=
Vo -> VI -> '"
-> VII
II :::;
v is
= V.
The partial ordering satisfies the conditions (i) reflexive: v:::; v: (ij) transitive: u :::; v and v:::; w imply u :::; J1.; and (iii) antisymmetric: u :::; v and v:::; II imply u = ,J. Further, U :::; v and U 7; V defines 11 < u. Definition 15.3.2.
Ifu
->
v, then
11
is a parent of v, termed u = pa(v), and
v is a child of u, termed v = ch(u). In symhols
(8)
pa(v) = {w
E
V\vlw
->
v},
(9)
ch ( u)
{w
E
V\ u Iu
->
w} .
=
In the graph displayed in Figure 15.2 we have (1,2)
= pa(3), 3·= pa(4).
2 = pa(5), 3 = ch(1, 2), 4 = ch(3), and 5 = ch(2). Definition 15.3.3.
If u < v, then v is a descendant of u,
(10)
de(u)
= {vlu < v},
an(v)
=
and u is an ancestor of v,
(11)
{ulu < v}.
The set of nondescendants of II is Nd(ll) = V\ de(lI), and the set of stria nondescendants is nd(u) = Nd(u)\u. Define An(A) = an(A) uA.
606
PATIERNS OF DEPENDENCE; GRAPHICAL MODELS
pa(v)
\
w
•
C<~(V) .w
Figure 15.3
Note that (12)
pa(v)
~an(v) ~nd(v).
In our study of undirected graphs we considered three Markov properties independently defined and then showed that a graph with one Markov property also has the other two. In the case of acyclic directed graphs we shall define three similar Markov properties, but the definitions are different because they take account of the direction of action or influence. Definition 15.3.4. A distribution on an acyclic directed graph G is pairwise Markov if fOl eve/)' v E Vand w E nd(v)\pa(v)
(13 )
vJLwlnd(v)\w.
In Clll11parison with Definition 15.2.1 for undirected graphs, note that attention is paid only to vertices in nd(v); since pa(v) is the effective boundary of v, the vertices wand v are nonadjacent. (See Figure 15.3.) Note also that the conditioning set inclues the parents of v, but not the children (which are descendants). Defmition 15.3.5. Markov if
(14)
A distribution on an acyclic directed graph is locally
vJL[nd(v)\pa(v)]lpa(v).
In the definition of locally Markov the conditioning is only on the parents of v. but in the definition of pairwise Markov the conditioning is on all of the other nondescendants. Thest. features correspond to Definitions 15.2.1 and 15.2.3 for undirected graphs. In Figure 15.2, we have lJL2,5, 3JL512, 4JL1,2,513, and 5JL1,3,412. In an undirecteu graph constructeu by replacing arrows in Figure 15.2 by
15.3
607
DIRECfED GRAPHS
lines (directed edges by undirected edges), a locally Markov distribution on the graph would include the conditional independences 1 JI. 213, 1,2 J1.413, 1,3,4 JI. 5. In the interpretation of the arrow indicating time sequence X 4 relates to the future of (X2 , X 3 ); the future cannot be conditioned on. As another example, consider an autoregressive time series Yo, YI'···' YT defined by
Yt = PYt-1 + Up
(15)
i=I,2, ... ,T,
where u l , ••• , u T are independ~nt N(O, (f' 2) variables and Yo has distri"oution N[O, (f' 2/(1 - p)2 )]. In this case given Yt' the future Yi + I, ... , YT is independent of the past Yo, ... ,Yt-l. Theorem 15.3.1. is pairwise Markov.
A locally Markov distribution on an acyclic directed graph
Proof The proof is the same as the proof of Theorem 15.2.3 for undi• rected graphs. Theorem 15.3.2. A pairwise Markov distribution on an acyclic directed graph is locally Markov.
Proof The proof is the same as the proof of Theorem 15.2.4.
•
Another Markov property is based on numbering the vertices in an order reflecting the direction of the action or the partial ordering induced. Definition 15.3.6.
numbered if i <j
An enumeration of the elements of V is coiled well<. Vi' or equivalently Vj < Vi => j < i.
=> Vj
Theorem 15.3.3.
A finite ordered set (V,::;) admits at least one well-
numbering. ,:~ =>
Definition 15.3.7. a* =b.
An element a*
E
V is maximal (or terminal) if a* ::; b
Lemma 15.3.1. A finite, partially ordered set (V,::;) has at least one maximal element a* .
Proof of Lemma. The proof is by induction with a* = a if #(V) = 1. Assume the lemma holds for #(V) = n, and consider #(V) = n + 1. Then V = aU (V\a) for any a E V. Since #(V\a) = n, V\a has a maximal element, say a. Then either a::; a and so a is maximal, or a <. a and so a is maximal. •
608
PAITERNS OF DEPENDENCE; GRAPHICAL MODELS
Proof of Theorem 15.3.3. We shall construct a well-numbering. Let v* be a maximal element; define vn ~ v*. In V\ VII let v* * be a maximal element; define vn_ 1 = v**. At the jth stage let v*** be a maximal element in V\ (vn, ... , vn_j + I ); define vn _j = v***, j = 3, ... , n - 1. Then VI = V\ (VI"'" Un_I)' This construction satisfies Definition 15.3.6. •
The well-numbering of V as v(l), ... , VCII) implies that in any directed path -. v(i") = V the indices satisfy io::s i l ::S •• ::S in' The well-numbering is not necessarily unique. Since V is finite, a maximal element can be found by comparing Vi and Vj for at most n(n - 1)/2 pairs.
u
= utio) -> V(il) -> '"
Definition 15.3.8. Let {VI"'" vn} be a well-numbering of the acyclic directed graph C. A distrihution 011 G is well-numbered Markov witli respect to this well-numbering if ( 16)
i = 3, ... , n.
Apparently the definition depends on the choice of well-numbering, but this is not the case, by Theorem 15.3.4. Theorem 15.3.4. A distribution on an acyclic directed graph that is we/!numbered Markov is locally Markov.
The definition of the global Markov property depends on relating the directed graph to a corresponding undirected graph. The moral graph G m of an acyclic directed graph G = (V. E) is the lin directed graph cOllstrtlctl'd by adding (1IIldirected) edges hetween parents of each vertex V E V and replacing every directed edge by an undirected edge.
Definition 15.3.9.
In the .iargon of graph theory, the parents of a vertex are "married." DefinitioT) 15.3.10. A distribution on {!n acyclic directed graph is globally Markov if All BIS for every A, B, and S such that S separates A and B in [GAn(A UP. U s]m. Theorem 15.3.5. A distribution on an acyclic directed graph that is globally Markov is locally Markov.
609
15.3 DIRECTED GRAPHS
Proof For any v E V let pa(v) = S in the definition of globally Markov. Let v =A and nd(v)\pa(v) = B. A vertex wE nd(v)\pa(v) is a vertex in An(A U BUS). Let 7T = W = V0, VI' ... , vn = V be a path from W to u in [GAn(AUBUS)]'" = [GNd(u)]m. If (Vn_I'Vn) corresponds to a directed edge (vn- J -> vn) in [GNd(u)]m, then vn- I E pa(v) = Sand pa(v) separates nd(v)\pa(v) and v. [The directed edge (vn_ 1 <- vn) implies vn_ 1 E de(v).]
•
Theorem 15.3.6. A distribution on an acyclic directed graph that is locally Markov is globally Markov. The proof is very lengthy and is omitted.
Recursive Factorization The recursive aspect of the acyclic directed graph permits a systematic factorization of the density. Use the construction of Theorem 15.3.4. Let n = I VI; then vn is a maximal element of V. Then (17) Thus (in the normal case) ctxu.1pa(vn) = an + BnX pa ("."
(18)
At the jth step let Vn _j + I be a maximal element of V\< v"' ... , Vn _j
+" ). Then
(20) Thus
(21) (22)
ctXu._i+,lpa(un-i+I)=all-j+1 cS'(XU._i+, -cS'XU._i+,)(XU._J +, -ctxu•. i+Y=I.n-i+I'
+B"-i+IXpa(I'._J")'
j=l, ... ,II-l.
The vector X U._i +' is independent of paC vn _j+ I)' The relations (18) to (22) can be written as generating equations. Let
(23) (24)
(25) (26)
Xn = all
+Bn(x't"",X:,_I)' + £""
610
PATIERNS OF DEPENDENCE; GRAPHICAL MODELS
where E I"'" En are independent random vectors with form (23) to (26) are
,g'E jEj
= l:j. In matrix
Bx=a+E, where
az
I -B 21
0 I
0 0
0 0
a,
-B,I
-B,2
I
0
a (28)
a=
l
B=
-8,,1
a"
-8,,2
-B",
['.
£2
,
'[
£ =
~.~
E"
and Bji = 0 if i <j-k j. Because the determinant of B is 1, (27) can be I solved for (29) The matrix
r-
I
is also lower triangular.
15.4. CHAIN GRAPHS A chain graph includes both directed and undirected edges; however, only certain patterns of vertices and edges are permitted. Suppose the set of vertices V of the graph G = (V, E) can be partitioned into subsets V = V(1) U '" U veT) so that within a subset the lertices are joined by umlirected edges and directed edges join vertices in different subsets. Let greG) be the set of vertices 1, ... , T and let rff(G) be the (directed) edge set such that T -> u if and only if there is at least one element U E V( T) and at least one element v E V(u) such that u -> v is in E, the edge set of G. Then 0(G)=[.Y(G), rP'(G)] is an acyclic directed graph; we can define pa!1/!(T), etc., for :/! (G). Let X T = (X.,I1I E V(T)}. Within a set the vcrtices form an undirected graph rdative to the probability distribution conditional on the past (that is, earlier sets). See Figure 15.4 [Lauritzen (1996)] and Figure 15.5. We now define the Markov properties as specified by Lauritzen and Wermuth (1989) and Frydenberg (1990):
(cn (1)
The distribution of X" T = 1, ... , T, is locally Markov with respect to the acyclic directed graph 0'J (G); that is, U
E nd ~ ( T) \ pa C0 ( T) .
15.4 CHAIN GRAPHS
611
V(l) V(3)
Figure 15.4. A chain graph.
V(1)~
/
V(3)
V(2) Figure 15.5. The corresponding induced acyclic directed graph on V = VO) u V(2) u V(3).
(C2) For each T the conditional distribution of X T given Xpa ",< T) is globally Markov with respect to the undirected graph on V( T). (C3)
Here bdG(U) = paG(U) U nbG(U). A distribution on the chain graph G that satisfies (C1), (C2), (C3) is LWF block recursive Markov. In Figure 15.6 pa !1/!(T) = {T - 1, T - 2} and nd q,(T )\pa q,(T) = {T- 3, T4, ... , I}. The set U = {u, w} is a set in V( T), and pa(,.(U) is the set in V( T - 1) U V( T - 2) that includes paG(u) for u E U; that is, paG(U) = {x, y}.
Vet-I) Figure 15.6. A chain graph.
612
PATIERNS OF DEPENDENCE; GRAPHICAL MODELS
1
2 Y(l)
3
4 Y(2)
Figure 15.7. A chain graph.
Andersson, Madigan, and Perlman (2001) have proposed an alternative Markov property (AMP), replacing (C3) by (C3*)
(3) In Figure IS.6, Xu for a vertex v in V( T -- 2) U V( T - 1) is conditionally independent of Xu [u E U ~ V(T)] when regressp.d on XpaG(U) = (Xx, Xy). The difference between (C3) and (C3*) is that the conditioning in (C3) is on bdG(U) = paG(U) U nbG(U), but the conditioning in (C3*) is on paG(U) only. See Figure IS.6. The conditioning in (C3*) is on variables in the past. Figure IS.7 [Andersson, Madigan, and Perlman (2001)] illustrates the difference between the L WF and AMP Markov properties:
(4)
LWF:
X I JLX 4 IX 2 ,X3 ,
X 2 JLX 3 IX l ,X4 ,
(S)
AMP:
XI JL X 4 1X 2 ,
X 2 JL X 3 IX l •
Note that in (S) Xl and X 4 are conditionally independent given X 2 ; the conditional distribution of X 4 depends on pa(v 2 ), but not X 3 • The AMP specification allows a block recursive equation formulation. In the example in Figure IS.7 the distribution of scalars Xl and X 2 [VI' v2 E VO)] can be specified as
(6) (7) where
(8 1,8 2 )
has an arbitrary (normal) distribution. Since X3 depends
15.5
6B
STATISTICAL INFERENCE
directly on Xl and X 4 depends directly on X 2 , we write
(8) (9) where (8 3 ,84 ) has an arbitrary distribution independent of (8 p 8~), and hence independent of (XI' X 2 ). In general the AMP model can be expressed as (26) of Section 15.3.
15.5. STATISTICAL INFERENCE 15.5.1. Normal Distribution Let XI"'" x N be N observations on X with distribution NCIJ., 1:). Let i = N-IL.~_IXa and S=(N_l)-IL.~_I(X,,-i)(x,,-i),=(N-J) 1[L:;~~IX"x:. - ~.xX'l. The likelihood is
The above form shows tl-Jat i and S are a pair of sufficient statistics for IJ. and :t, and they are independently distributed. The interest in this chapter is on the dependences, which depend only on the covariance matrix 1:, not IJ.. For the rest of this chapter we shall suppress the mean. Accordingly, we suppose that the parent distribution is N(O, 1:) and the sample is X I' ... , x,,), and S = (1jn)L.~_lx"x~, Tt e likelihood function can be written
(2) =exP[-''l'(A) -
~.t. \,1;;- LA;/;J]' 1=
I
I<j
where A=(A ij )=1:- I , T=(ti)=L.~_IX"X~, and 'l'(A)=~plllog(27T) - inloglAI. The likelihood is in the exponential famil; with canonical parameter .\ and statistic T. The maximum likelihood estimator of 1: with no restriction is l;=S=(1jn)T. Since A=1:- l is a 1-to-1 transformation of 1:, the maxi· mum likelihood estimator of A of A= l; - I .
614
PATIERNS OF DEP':NDENCE; GRAPHICAL MODELS
15.5.2. Covariance Selection Models In undirected graphs many of the models involve zero restrictions on elements of A. Dempster (1972) studied such models and introduced the term covariallce selection. When the (directed) graph satisfies the pairwise Markov condition, A.;j = 0 for Ci, j) $. E. We assume here th
(3)
(21T)
-p/2
1 P Aii+ IAI -n /2 exp ( 2L j=1
L
A.;jns ij
)
,
(i,ilEE
where A satisfies the condition Aij = 0, (i, j) $. E. In this form the canonical parameters are All"'" App and Aij , (i, j) E E. The canonical variables are SII""'Spp and Sij' (i,j)EE; these form a sufficient set of statistics. To maximize the likelihood function we differentiate (3) with respect to A.;i' i = L .... p, and A'j' (i, j) E E, to obtain the equations (4) and (5). Theorem 15.5.1. is given by
The maximum likelihood estimator of l: in the model (3)
(4) (5)
i=jor(i,j)EE, Aij=O,
i "" j and (i, j)
$.
E,
This result follows from the general theory of exponential families. See Lauritzen (996), Theorem 5.3 and Appendix D.l. Here we shall show that for a decomposable graph the equations (4) and (5) have a unique positive definite solution by developing G.n algorithm for its computation. We follow Speed and Kiiveri (1986). Theorem 15.5.2. Let Land M be p X P positive definite matrices. There exists a unique positive definite matrix K such that (6) (7)
i=jor (i,j) EE, i "" j and (i, j)
$.
E,
where (k ij ) = K- I and (m ij ) = M- I . The proof of Theorem 15.5.2 depends on several lemmas. In the maximum likelihood estimation L = S, M = I or any other diagonal matrix, and K = :t,
15.5
STATISTICAL INFERENCE
615
To develop this subject we use the Kullback information. For a pair of multivariate normal distributions N(O, P) and N(O, R) define
n(xIO,P) I( P IR ) -"'1 - 0p ogn(xIO,R)
(8)
=, -HlogIPR-11 +tr(l-PR- I )].
Lemma 15.5.1.
Suppose P and R are positive definite. Then:
(i) I(PIR) > 0, P
* R, and I(PIP) = 0.
(ij) If {Pn} and {Rn} are sequences of positive definite matrices such that I(Pnl~~n) -> 0, then PnR;; I -> I.
Proof 0) Let the roots of IP - sRI =
°be Sl :;:; ... :;:; sp. Then p
(9)
loglPR-11 + tr(l-PR- I ) =
L
(log Si'+ 1- Si) ;:: 0,
i-I
°
and (9) is if and only if SI = ... = sp = 1. (ij) Let the roots of IPn - sRnl = be sl(n):;:; ... :;:; sp(n). Then I(Pn IRn) -> 0 implies [sl(n), ... , s,,(n)]-> 0, ... ,1), which implies that PnR;;1 -> I .
°
• Lemma 1;;.5.2.
Let
(10) Then (j) The matrix
( 11)
( 12)
(ii) I( PI R) = I(PI Q) + I(QI R).
616
PATIERNS C;:>F DEPENDENCE; GRAPHICAL MODELS
Proof 0) Let (13) Then I=Q-IQ can be solved for S=PI"/ +R12(R22)-lR21; Q=(Q-l)-l follows from Theorem A.3.3. Then (ij) follows from
(14)
(15)
PQ-l
= PR-1 +.(
I-P R-l
~l
11
•
trPQ-1 +trQR- 1 =trPR- 1 +trlp'
Lemma 15.5.2 provides the solution to the problem of finding a matrix Q, given positive definite matrices P and R, such that
(16)
(i,n
( 17)
(i,j) f/. {1, ... ,t}.
E
{1, ... ,t},
We now develop an iterative method to find K to satisfy (6) and (7), thus proving Theorem 15.5.2. Suppose E = C 1 U ... U cm' where c 1, ••• , c m are the 1 m cliques of a decomposable graph G = (V, E). Let 1 = M- • Define recursively Kn = (ki/n» such that
Ko
(18) (19)
i,j
kij(n) =(j' ij
k ( n)
=
k ij (n - 1),
ECnmodm'
i,j f/.
cnmod m'
By Lemma 15.5.2, Kn is uniquely determined. (The algorithm cycles through the cliques.) By construction (20) Summation of (20) from 1 to q gives q
(21)
I( LIKo)
= I( LIKq) +
L I( KjIKj_d. j~l
Since I(LIKql ~ 0, LJ~l I(Kjl K j - 1) is bounded and I(KjIKj_1) --> 0 as n --> 00. The set (K-1II(LIK)::; I(LIKo)} is strictly convex. Consider the vector sequence (Krm+l, ... ,Krm+m) with index r (n =rm). It has a convergent subsequence (r(i)}; that is, (K mru )+ 1 ' " ' ' KmrU)+m) converges to (Kf, ... ,K!), say. Since I(KjIKj_1)--> 0, KjKj'!l-->l. Then the
15.5
617
STATISTICAL INFERENCE
Cl
Figure 15.8. Diagram of
("1
and
c, for
("1
U C2
=
E.
matrix Kmr(i)+jK;;';(i)+j-1 -> /, j = 2, ... , rn, which implies Ki = '" = K~ = K, say. Note that Ii, jli, j $ E) satisfies i, j $ Ci , i = 1, ... , m. Hence K" satisfies (7), n = 0, 1, ... , and K does too. Further, kJrnr(i) + t) satisfies (18) i,jEC, and K does, too. Figure 15.8 diagrams the sets for ("I = (i,j). i,j= 1, ... ,1, and cz=(i,j), i,j=u,u + 1, ... ,p, II
(22)
rffp 10gn(xIO,P)
= -~(plog27T-p-logIPI).
Note that I~I =nf=lO)RI, where R=(Pij)' Given that Uii=Sii' the selection of Pij to maximize the entropy of the fitted normal distribution satisfying the requirements also minimizes IRI [Demspter (1972)].
618
PATIERNS OF DEPENDENCE; GRAPHICAL MODEU
15.5.3. Decomposition of Covariance Selection Models An undirected graph is decomposable if the graph is formed by three disjoint sets .4, B, C, where V = A u B U C, A and Bare non empty, C separates A and B, and C is complete. Then if Xv is globally Markov.vith respect to G, we have XA JL XBIX c ,
AAC]
(23)
ABc,
Acc
(24)
:lAA
:l(AB)·C
= ( :l
BA
and
(25) The maximum likelihood estimator of :l can be constructed from the maximum likelihood estimators of :lAA'C, :lAB.C, :lBB.e, P(AB)'C' and :lee. If there is no restriction on :l' the maximum likelihood estimator of :l is
S(AB)C]
(26)
S
ce
~ SABC'
where
(27)
SAA'C
S(AB)'C
= (S
BA'C
If the restriction (25) is imposed, the maximum likelihood estimator is (26) with SA B.C replaced by 0 to obtain
(28)
The matrix S(AB)C has the Wishart distribution W[:l(AB)'C' n - (PA + PB)]' where P"I and 1'8 arc the number of components of X A and X n , respectively
619
15.5 STATISTICAL INFERENCE
(Section 8.3). The matrix B(AB).e = S(AB)eSC~ conditional on (XCI"'" Xen) =Xe has a normal distribution, the covariance of which is given by Ae
(29)
-I . ] ([BA.e])'1 rff vec [ B BB.e vec BB.e See = Sec ®
[~AA 0
0]
~BB'
and See has the Wishart distribution W(~ee, n). The matrix S(AB).e and the matrix B(AB).e, are independent (Chapter 8). Consider testing the null hypothesis (25). This is testing the null hypothesis ~JlB.e = 0 against the alternative ~AB.e *- o. Th: determinant of (26) is I~nl = ISAB·el 'ISeel; the determinant of (28) is I~wI = ISAAI ·1 SBBI ·ISeel. The likelihood ratio criterion is
(30)
A) n /2 I~wl ( linl =
(I
S(AB)
I)
e
ISAA el'IS BB
n /2
el
Since the sample covariance matrix S(AB).e has the Wishart distribution W[~(AB)'C' n - (PA + PB)]' where PA and PB are the numbers of components of X A and X B (Section 8.2), the criterion is, in effect, upA,PB.n-(PA+PB)' studied in Sl!ctions 8.4 and 8.5. As another example, consider the graph in Figure 15.9. Note that node 4 separates (1, 2, 3) and (5,6); nodes 1,4 separate 2 and 3; and node 4 separates 5 and 6. These separations imply three conditional independences: (XI' X 2 , X 3 )JL(XS ' X 6 )IX4 , X 2 JL X 3 1(X 1 , X 4 ), and Xs JL X 6 1X 4 • In terms of covariances these conditional independences are (31) (32) (33)
~(123)(S6)'4 = ~(123)(S6) ~23.(l4) = ~23
-
-
~(123)4~441 ~4(56) = 0,
~2(14) ~(ll)(14) ~(14)3
~S6'4 = ~S6
=
0,
~S4 ~441 ~46 = O.
-
2
5 4
6
3 Figure 15.9
620
PAITERNS OF DEPENDENCE; GRAPHICAL MODELS
In view of (31) the restriction (32) can be written as
(34) It will be convenient to reorder the subvectors as X 2 , X 3 , XI' X s, X , X to 4 6 write
(35)
5=
=
5 26
[S;'
. 8 62
5 66
5 42
8 46
[["'.
S:' 1 864 8 44
Su'l ..
82~.4
5 66 . 4
+
[S.. 144 .
8
8 46 ]
8 46 ]
+ 5 (2 ... 6)4 5-44 18 4·(2 ... 6)
[;JI
8(2 ... 6)4} 5 44
84(2 ... 6)
The determinant of
1[842
864
[ 8 42
= [8(2 ... 6)(2 .. 6)·4
8
is
(36) If the condition (XI' X 2 , X 3 )lL(X5 , X 6 )IX 4 i~ imposed, the maximum likelihood estimator is (35) with 5(125)(56)-4 replaced by 0 to obtain
(S(231)~231).4
(37)
r
5(56:56) 4)
+ 5(2
.. 6)4 S';:/54(2 ... 6)
S4(2 ... 6)
The determinant of (37) is
(38)
18 (231)(231)-41.1 5 (56)(56).41./ 8 44 /.
15.5
621
STATISTICAL INFERENCE
The likelihood ratio criterion for
~(\23)(56)'4
= 0 is
i )11/2 lS (2"06)(2"06)-4 ( 1S(231)(231)041'1 S(56)(56)041
(39)
_ -
Un / 2
(231)(56)04'
Here Uc231)(56)04 has the distribution of UP ,+P,+P3.p,+p,.n-p (Section 8.4) since the distribution of S(20006)(2 00 06)'4 is W(~(20006)(20006H,/1-P4)' independent of S44' The first three rows and columns of S(20 06)(20006)04 constitute the matrix (40)
_ [(S22 -
0
0
14 S32-14
(
( r
S22014
=
S'1
4) 1
~~:o:
_' 0
4)1
S21 ( Sl1 4
S32 014
0
,
SIH
The determinant of (40) is
(41) The estimator of S(231)(231)'4 with X 2 replace by 0 to obtain
(42)
the determinant of which is
(43)
JL X31XP
X 4 imposed is (40) with
S23014
622
PATTERNS OF DEPENDENCE; GRAPHICAL MODELS
The lik.elihood ratio criterion for
l:23.14
= 0 is
(44) The statistic U2-'1 .. has the distribution of Up2 . p ,.n-(p,+P2+P3+P4) (Section 8.4) since S(23)(23)'14 has the distribution W[l:(23)(23)'14' n - (PI + pJ] independent of S(l4)(1")' The estimator of l:(S6)(56)'4 with l:S6'4 = 0 imposed is S(56)(56)'4 with S56.4 replaced by 0 to obtain (45) The Iik.elihood ratio criterion for testing
l:56.4
= 0 is
(46) The statistic rl sO . 4 has the distribution of u p ,.p,.n-(p4+p,+p,) since S(56)(56)'4 has the distribution W(l:(56)(56)'4, n - P4) independent of S44' The estimator of l: under three null hypotheses is (37) with S(231)(231)'4 replaced by (42) and S(56)(56)'4 replaced by (45). The determinant of this matrix is
The likelihood ratio criterion for testing the three null hypotheses is
( 48) When the null hypotheses are true, the factors L'<231)(56).4' U23 . 14 , and U56 ' 4 are independent. Their distributions are discussed in Sections 8.4 and 8.5. In particular the moments of these factors are given and asymptotic expansions of distributions are described.
15.5.4. Directed Graphs We suppose that the vertices are well-numbered, 1, ... , n; the N observations are made on X = (X;, ... , X p )'. The model is (22) to (25) or (26) of Section 15.3. Let x=N-IL~_IX(a) and S=(N_l)-IL~_I(X(a) xXx(a) -i)'. The model (26) consists of Xl = 0: 1 + ") and n -1 regressions
x(I) . . . • , X(N)
f
623
ACKNOWLEDGMENTS
0: I in XI = 0: I ~ E ~ i$ I!stimatl!J. by XI. if palL':) is vacuous, 0: 2 is estimated by x:; if pal D:) is not vacuous and X pa\, : \ = X I. then B 2 and 0: 2 are estimated by
(23) to (25). Thl! \"I!ctor
1
Ct~1 (X2(")-X2)(X!(U)-X!)'L~1 (X!«ll-X!)(X!(Ul-X!)'r ,
(49)
Ez =
(50)
&2 =X 2 + E2 x!.
In general N
(51)
Ej = L
[Xj(a)-Xj] [xpa(uj}(a)-xpa(Uj)]'
Ct~1
Conditional on
xpa(uj)(a)'
the distribution of these estimators is normal.
15.5.5. Chain Graphs The condition (Cl) of Section 15.4 specifies that Xu lL XulXpa "" for U E V( T) and for v E V( u ), where u E nd q, (T) \ pa q, ( T); that is, the past earlier than pa q,( T) is independent of the present. This condition corresponds to the Markov property in time series analysis. Thus Xu is in terms of deviations from the regression of X, on X pa '" (,) (g' X,IX pa ",(')
= 0:, + B,Xpa ,,(,).
The vector 0:, and the matrix B, are estimated as for directed graphs. The Markov property (C2) indicates the analysis in terms of deviations X, - 0:, - B,Xpa ",(,). The estimation of the structurc of dependence within V( T) is carried out as in Section 15.5.2. The Markov property (C3*) specifies Xu lLXulXpaG(u) for uEU~V(T) and v E pa q;(U) U nbG(U). The property is a restriction on the regression of X, on X pa ",(,). ACKNOWLEDGMENTS
The preparation of this chapter drew heavily on lecture notes of Michael Perlman. Thanks are also due Michael Perlman and Ingram Olkin for reading drafts of the manuscript.
APPENDIX A
Matrix Theory
A.1. DEFINITION OF A MATRIX AND OPERATIONS
ON MATRICES In this appendix we summarize some of the well-known definitions and theorems of matrix algebra. A number of results that are not always contained in books on matrix algebra are proved here. An m X n matrix A is a rectangular array of real numbers
( 1)
A=
which may be abbreviated (a i ) , i = 1,2, ... , m, j = 1,2, ... , n. Capit'l-l boldface letters will be used to denote matrices whose elements are the corresponding lowercase letters with appropriate subscripts. The sum of two matrices A and B of the same numbers of rows and columns, respectively, is defined by
(2) The product of a matrix by a real number A is defined by
(3)
An Introduction to Multivariate Statistical Analysis. Third Edition. By T. W. Anderson ISBN 0-471-36091-0 Copyright © 2003 John Wiley & Sons. Inc.
624
A.I
DEFINITION OF A MATRIX AND OPERATIONS ON MATRICES
625
These operations have the algebraic properties
(4)
A+B=B+A,
(5)
(A + B) + C = A + (B + C).
(6)
A+(-I)A=(O),
(7)
(A + /L)A = AA + /LA,
(8)
A(A +B) = AA + AB,
(9)
A( /LA) = (A/L)A.
The matrix (0) with all elements 0 is denoted as O. The operation A + (-IB) is denoted as A-B. If A has the same number of columns as B has rows, that is, A = (a ij ), i = 1, ... , t, j = 1, ... , m, B = (bjk ), j = 1, ... , m, k = 1, ... , n, then A and B can be multiplied according to the rule
(10)
AB
= (aij)(bjk ) =
(.f:
aijbjk ) ,
i=l .... ,I,
k=l, .... n;
J~I
that is, AB is a matrix with I rows and n columns, the element in the ith row and kth column being L~ 1 aijbjk' The matrix product has the properties (11)
(AB)C =A( BC),
(12)
A(B + C) =AB +AC,
(13)
(A +B)C =AC +BC.
The relationships (11)-(13) hold provided one side is meaningful (i.e., the numbers of rows and columns are such that the operations can be performed); it follows then that the other side is also meaningful. Because of (11) we can write
(14)
(AB)C=A(BC) =ABC.
The product BA may be meaningless even if AB is meaningful, and even when both are meaningful they are not necessarily equal. The transpas' of the t X m matrix A = (a;) is defined to be the m x I matrix A' which has in the jth row and ith column the element that A has in
626
MATRIX THEORY
the ith row and jth column. The operation of transposition has the properties
(15)
(AT =A,
(16)
(A +B)' =A' +B',
(17)
(AB)' =B'A',
again with the restriction (which is understood throughout this book) that at least one side is meaningful. A vector x with m components can be treated as a matrix with m rows and one column. Therefore, the above operations hold for vectors. We shall now be concerned with square matrices of the same size, which can be added and multiplied at will. The number of rows and columns will be taken to be p. A is called symmetric if A = A'. A particular matrix of considerable interest is the identity matrix
100
(18)
1=
o
1
0
0
0
000 where 8ij • the Kronecker delta, is defined by ( 19)
i
= j,
i=foj.
= 0, The identity matrix satisfies
(20)
LA=A/=A.
We shall write the identity as Ip when we wish to emphasize that it is of order p. Associated with any square matrix A is thc determinant IAI, defined
by (21 )
IAI
=
L ( -[ )I(II ...
p
jp)
na ;= I
,
ij r
where the summation is taken over all permutations UI> ... ' j p) of the set of integers 0, ... , p), and I(jl' ... ' jp) is the number of transpositions required to change (1, ... , p) into (j l, ... , j p). A transposition consists of interchanging two numbers, and it can be s~own that, although one can transform 0, ... , p) into (jl, ... ,jp) by transpositions in many different ways, the number of
A.1
627
DEFINITION OF A MATRIX AND OPERATIONS ON MATRICES
transpositions required is always even or always odd, so that (-Of(j,·" "".jp) is consistently defined. Then
IABI = IAI ·IBI.
(22) Also
IAI =IA'I.
(23)
A submatrix of A is a rectangular array obtained from A by deleting rows and columns. A minor is the determinant of a square submatrix of A. The minor of an element aij is the" determinant of the submatrix of a square matrix A obtained by deleting the ith row and jth column. The cofactor of aij , say A ij , if, (_1)i+j times the minor of aij • It follows from (21) that p
IAI =
(24)
p
L a;jA;j = L ajkA jk · j=l
j=l
*"
If IAI 0, there exists a unique matrix B such that AB = I. Then B is called the inverse of A and is denoted by A-I. Let a hk be the element of A-I in the hth row and kth column. Then
(25) The operation of taking the inverse satisfies
(26) since
Also r I = I and A -IA = I. Furthermore, since the transposition of (27) gives lA -I )'A' = I, we have (A -1)' = (A,)-l. A matrix whose determinant is not zero is called nonsingular. If IAI 0, then the only solution to
*"
(28)
Az =0
is the trivial one z = 0 [by multiplication of (28) on the left by A -I]. If IAI = 0, there is at least one nontrivial solution (that is, z 0). Thus an equivalent definition of A being nonsingular is that (28) have only the trivial solution. A set of vectors Zl"'" Z, is said to be line~rly independent if there exists no set of scalars C1,. .. ,C,' not all zero, such that L~=ICiZ;=O. A qXp
*"
628
MATRIX THEORY
matrix D is said to be of rank r if the maximum number of linearly independent columns is r. Then every minor of order r + 1 must be zero (from the remarks in the preceding paragraph applied to the relevant square matrix of order r + 1), and at least one minor of order r must be nonzero. Conversely, if there is at least one minor of order r that is nonzero, there is at least one set of r columns (or rows) which is linearly independent. If all minors of order r + 1 are zero, there cannot be any set of r + I columns (or rows) that are linearly independent, for such linear independence would imply a nonzero minor of order r + 1, but this contradicts the assumption. Thus rank r is equivalently defined by the maximum number of linearly independent rows, by the maximum number of linearly independent columns, or by the maximum order of nonzero minors. We now consider the quadratic form p
(29)
x'Ax=
L
aijxixj ,
i.j~l
where x' = (Xl"", Xp) and A = (a ij ) is a symmetric matrix. This matrix A and the quadratic form are called positive semidefinite if x'Ax;::: 0 for all x. If x' Ax> 0 for all x*' 0, then A and the quadratic form are called positive definite. In this book positive definite implies the matrix is symmetric.
Theorem A.I.I. If C with p rows and columns is positive definite, and if B with p rows and q columns, q :s'p, is of rank q, then B'CB is positive definite. Proof Given a vector y Tben
(30)
*' 0, let x = By . .Since B is of rank q, By = x*' O.
y'(B'CB)y = (Hy)'C(By) =x'Cx>O.
The proof is completed by observing that B'CB is symmetric. As a converse, we observe that B 'CB is positive definite only if B is of rank q, for otherwise there exists y*,O such that By = O. •
Corollary A.I.I. positive definite. Corollary A.I.2.
If C is positive definite and B is nonsingular, then B'CB is
If C is positive definite, then C- I is positive definite.
Proof C must be nonsingular; for if Cx = 0 for x*, 0, then x'Cx = 0 for this x, but that is contrary to the assumption that C is positive definite. Let
A.l
DEFINITION OF A MATRIX AND OPERATIONS ON MATRICES
629
B in Theorem A.Ll be C- I . Then B'CB=(C-I)'ce l =(C- I ),. Transpos• ing ce l =J, we have (el),c' =(el)'c=J. Thus e l =(e l ),.
Corollary A.1.3. The q X q matrix formed by deleting p - q rows of (/ positive definite matrix C and the corresponding p - q columns of C is positive definite.
Proof This follows from Theorem A.l.l hy forming B hy taking the p x p identity matrix and deleting the columns corresponding to those deleted • from C. The trace of a square matrix A is defined as tr A properties are v orified directly:
(:II)
tr( A + B) = tr A + tr B,
(32)
tr AB = tr BA.
= r.f~ I a ii'
The following
A square matrix A is said to be diagonal if aij = 0, i *- j. Then IAI = for in (24) IAI =aIlA Il , and in turn All is evaluated similarly. A square matrix A is said to be triangular if a ij = 0 for i > j or alternatively for i <j. If a ij = 0 for i > j, the matrix is upper triangular, and, if aij = 0 for i <j, it is lower triangular. The product of two upper triangular matrices A, B is upper triangular, for the i, jth term (i > j) of .4B is r.f~laikbkj=O since aik=O for kj. Similarly, the product of two lower triangular matrices is lower triangular. The determinant of a triangular matrix is the product of the diagonal elements. The inverse of a nonsingular triangular matrix is triangular in the same way. TIf~laii'
Theorem A.t.2. If A is nonsingular, there exists a nonsingular lower triangular matrix F such that FA = A* is nonsingular upper triangular.
Proof Let A =A 1• Define recursively Ag = (a\y) = Fg_ I A g _ 1 , g = 2, ... , p, where Fg_ I =
(33)
fj)g-I) = 1,
(34)
g-\) ri,g-I -
(35)
fi}g-I) = 0,
-
j= l.. .. ,p, a(g-I)
,.~-I
, a(g-I) g_ l • g _ 1
i=g .... ,p. otherwise.
630
MATRIX THEORY
Then
i=j+1, ... ,p,
j=l, ... ,g-l,
i=l, ... ,g-l,
j=l, ... ,p,
(38)
Note that F = Fp-I' ... ' FI is lower triangular and the elements of Ag in the first g - 1 columns below the diagonal are 0; in particular A* = FA is upper triangular. From \A\ '" 0 and IFg _ l \ = 1, we have IA g _ 1 \ '" O. Hencc a'N, ... , aig--/)g- 2 are different from 0 and the last p - g columns of A g_1 can be numbered so a~g_-ll)g_1 *- 0; then f/~-:"'\) is well defined. • The equation FA = A* can be solved to obtain A = LR, where R = A* is upper triangular and L = F- I is lower triangular and has l's on the main diagonal (because F is lower triangular and has l's on the main diagonal). This is known as the LR decomposition. Corollary A.1.4. If A is positive definite, there exists a lower triangular Ilollsillgztiar matrix F such that FAF' is diagonal and positive definite. Proof From Theorem A. 1.2, there exists a lower triangular nonsingular matrix F such that FA is upper triangular aT d nonsingular. Then FAF' is upper triangular and symmetric; hence it is diagonal. •
Corollary A.1.S.
The determinant of a positive definite matrix A is positive.
Proof From the construction of FAF', a(\)
0
0
0
0
a(2)
0
0
0
0
a(3) 33
0
0
0
0
alp) pp
II
(39)
FAF'=
is positive definite, and hence
\A\·\F\
=
\A\.
•
"
aW> 0, g = 1, ... , p, and 0 < \FAF'I = IFI·
631
A.2 CHARACTERISTIC ROOTS AND VECTORS
Corollary A.l.6. If A is positive definite, there exists a lower triangular matrix G such that GAG' = I. Proof Let FAF' = D2, and let D be the diagonal matrix whose diagonal elements are the positive square roots of the diagonal elements of D2. Then C = D -1 F serves the purpose. _
Corollary A.l.7 (Cholsky Decomposition). exists a unique lower triangular matrix T elements such that A = IT'.
(tjj
If A is positive definite, there
= 0, i <j) with positive diagonal
Proof From Corollary A. 1.6, A = G- 1(G,)-I, where G is lower triangular. _ Then T= G- 1 is lower triangular.
In effect this theorem was proved in Section 7.2 for A = W'.
A.2. CHARACTERISTIC ROOTS AND VECTORS
The characteristic roots of a square matrix B are defined as the roots of the characteristic equation
IB-AII =0.
(1)
Alternative terms are latent roots and eigenvalues. For example, with
B=(~
;),
we have
(2)
~
J-'\
1=25-4-10,\+,\2=,\2- 10 ,\+21.
The degree of the polynomial equation (1) is the order of the matrix Band the constant term is IBI. A matrix C is said to be orthogonal if c'e = I; it follows that CC' = 1. Let the vectors )"' = (Xl"'" Xp) and y' = (YI"'" Yp) represent two points in a p-dimensional Euclidean space. The distance squared between them is D(x, y) = (x - y)'(x - y). The transformation z = Cx can be thought of as a change of coordinate axes in the p-dimensionai space. If C is orthogona!, the
632
MATRIX THEORY
transformation is distance-preserving, for
(3)
D(Cx,cy) = (CY-Cx)'(cy-ex) = (y -x),C'C(y -x) = (y -.e)'(y -x) =D(x,y).
Since the angles of a triangle are determined by the lengths of its sides, the transformation z = Cx also preserves angles. It consists of a rotation together with a possible reflection of one or more axes. We shall denote RX by IIxll. Theorem A.2.1. matrix C such that
( 4)
Given any .Iymmetric matri.t B, there exists an orthogonal
dl
0
0
0
d2
0
0
0
dp
C'BC=D=
If B is positive semidefinite, then d i d i > 0, i = 1, ... ,p.
~
0, i
= 1, ... ,p; if B is positive definite, then
The proof is given in the discussion of principal components in Section 11.2 for the case of B positive semidefinite and holds for B symmetric. The characteristic equation (1) under transformation by C becomes
(5)
0 = IC'I'IB- AlI·ICI = IC'(B - AI)CI = IC'BC - All
=
ID - All
o o
p
=D(di-A). i~l
o
()
Thus the characteristic roots of B are the diagonal elements of the transformed matrix D. If Ai is a characteristic root of B, then a vector x; not icentically 0 satisfying
(6)
(B = AJ)x; = 0
is called a charactelistic vector (or eigenvector) of the matrix B corresponding to the characteristic root A;. Any scalar mUltiple of x; is also a characteristic.; vector. When B is symmetric, x;(B - AJ) = O. If the roots are distinct, xjBx; = 0 and xjx; = 0, i,,;: j. Let c; = (J /lIx;IDx; be the ith normalized
633
A.2 CHARACfERISTIC ROOTS AND VECtORS
characteristic vector, and let C = (c l , ••• , cpl. Then C'C = I and BC = CD. These lead to (4). If a characteristic root has multiplicity m, then a set of m corresponding characteristic vectors can be replaced by m linearly independent linear combinations of them. The vectors can be chosen to satisfy (6) and xjXj = 0 and xjBxj = 0, i j. A characteristic vector lies in the direction of the principal axis (see Chapter 11). The characteristic roots of B are proportional to the squares of the reciprocals of the lengths of the principal axes of the ellipsoid
"*
(7)
x'llx=l
since this becomes under the rotation y
=
Cx
p
(8)
L diy?
1 =y'Dy =
For a pair of matrices A (nonsingular) and B we shall also consider equations of the form IB- AAI =0.
(9)
The roots of such equations ar,~ of interest because of their invariance under certain transformations. In fac, for nonsingular C, the roots of IC'BC - A(C'AC)I = 0
(10)
are the same as those of (9) since (11)
IC'BC-AC'ACI =IC'(B-AA)CI =IC'I'IB-AAI'ICI
"*
and IC'I = ICI o. By Corollary A.1.6 we have that if A is positive definite there is a matrix E such that E' AE = I. Let E' BE = B*. From Theorem A.2.1 we deduce that there exists an orthogonal matrix C such that C' B* C = D, where D is diagonal. Defining EC as F, we have the follo\\ ing theorem:
Theorem A.2.2. Given B positive semidefinite and A positive definite, there exists a nonsingular matrix F such that
(12)
(13)
AI 0
0
0
A2
0
0
0
Ap
F'BF=
F'AF=/,
where Al ~ ... ~ Ap (~ 0) are the roots of (9). If B is positive definite, Ai i = 1, . .. ,p.
> 0,
634
MATRIX THEORY
Corresponding to each root Ai there is a vee tor Xi satisfying (14) and x;Axi = 1. If the roots fire distinct XjBxi = 0 and xjAx, = 0, i *" j. Then (Xl' .•. ' X p ). If a root has multiplicity m, t,1en a set of m linearly independent x/s can be replaced by m linearly independent combinations of them. The vectors can be chosen to satisfy (14) and XjBXi = 0 and xjAxi = 0, i *i.
F=
Theorem A.2.3 (The Singular Value Decomposition). Given an n X p matrix X, n ;;:.p, there exists an II X n orthogonal matrix P, a p X P orthogonal matrix Q, alld an n X p matrix D collsisting of a p X P diagonal positive semidefinite matrix and an (n - p) X P zero matrix such that
(15)
X=PDQ.
Proof From Theorem A.2.1, there exists a p xp orthogonal matrix Q and a diagonal matrix E such that
(16) where E I is diagonal and positive definite. Let XQ' = Y = (Y l Y2 ), where the number of columns of YI is the order of E l • Then YZY2 = 0, and hence Yz = o. Let PI = Yl Ell. Then P; PI = I. An Il X n orthogonal matrix P = (PI P) satisfying the theorem is obtained by adjoining P 2 to make P orthogonal. Then the upper left-hand corner of D is Elf, and the rest of u consists of zeros. •
Theorem A.2.4. Theil
Let A be positive definite and B be positive semidefinite.
(17) where Al and Ap are the largest alld smallest roots of (1), and (18) where Al and Ap are the largest and smallest roots of (9). Prooj: The inequalities (17) were essentially proved in Section 11.2, and can also be derived from (4). The inequalities (18) follow from Theorem A.2.2. •
A.3
635
PARTITIONED VECfORS AND MATRICES
A square matrix A is idempotent if A Z=A. If A satisfies IA - All = 0, there exists a vector x;60 such tilat Ax=Ax=Azx. However, A 2x=A(Ax) =AAx = AZx. Thus A2 = A, and A is either 0 or 1. The multiplicity of A = 1 is the rank of A. If A is p X p, then lp - A is idempotent of rank p - (rank A), and A and lp - A are orthogonal. If A is symmetric, there is an orthogonal matrix 0 such that
(19)
OAO' =
[~
:],
O(I-A)O'=[~ ~].
A.3. PARTITIONED VECTORS AND MATRICES Consider th( matrix A defined by (1) of Section A.I. Let i=l, ... ,p,
All = (au)'
(1 )
A 12 =(a j
J,
AZ! = (ajJ,
A22=(a jj ),
i=l, ... ,p,
j=q+1, ... ,n,
i=p+1, ... ,m, i=p+1, ... ,m,
i=l, ... ,q,
j=l, ... ,q,
j=q+1, ... ,n.
Then we can write
(2) We say that A has been partitioned into sub matrices A jj . Let B (m X n) be partitioned similarly into sub matrices B jj , i, j = 1,2. Then
(3) Now partition C (n X r) as (4) where C II and C l2 have q rows and C u and C2l have s columns. Then
~:: )
(5) = (AllC Il +A 1Z C 2l AZlC ll +A ZZ C 21
A U e l2 +A 12 C Z2 ) A 2l C lZ +A 2Z C 2Z .
636
MATRIX THEORY
To verify this, consider an element in the first p rows and first s columns of AC. The i, jth element is
(6)
i 5.p,
j 5.s.
This sum can be written n
q
L ajkc kj + L
(7)
k~l
aikc kj ·
k~q+l
The first sum is the i,jth element of AllC 11 , the second sum is the i,jth element of A 12 C21 , and therefore the entire sum (6) is the i,/th element of AIlC II +A 12 C21 • In a similar fashion we can verify that the other submatrices of AC can be written as in (5). We note in passing that if A is partitioned as in (2), then the transpose of A can be written A'=
(8)
A' 11 ( A'12
A'21) A'22 .
If A12 ,;" 0 and A2l = 0, then for A positive definite and All square,
(9) The matrix on the right exists because All and A22 are nonsingular. That the right-hand matrix is the inverse of A is verified by mUltiplication:
(10)
All
(o
(Au
0 ) A22 0
l
0 -1 A22
)
=
(I
0
0), I
which is a partitioned form of II'. We also note that
The evaluation of the first determinant in the middle is made by expanding according to minors of the last row; the only nonzero element in the sum is the last, which is 1 times a determinant of the same form with I of order one
A.3
637
PARTITIONED VECTORS AND MATRICES
less. The procedure is repeated until IA III is the minor. Similarly,
(12)
1: ~~:I=I~ :2JI~1l :121 = IA I·IA 22 1· 11
II
A useful fact is that if Al of q rows and p columns is of rank q, there exists a matrix A2 of p - q rows and p columns such that
(13) is nonsingular. This statement is verified by numbering the columns of A so that A 11 consisting of the first q columns of A I is nonsingular (at least one q X q minor of AI is different from zero) and then taking A2 as (0 1): then IAI2[=IA II
(14)
I,
which is not equal to zero.
Theorem A.J.t.
Let the square matrix A be partitioned as in (2) so that A22
is square. If An is nonsingular, let (15) Then
(16)
(17)
BAC
=
All -AI,A:;-,IA'I (
-
o
--
-
If A is symmetric, C = B'.
Theorem A.3.2.
Let the square matrix A be partitioned as in (2) so that A 22
is square. If A22 is nonsingular, (18)
638
MATRIX THEORY
Proof Equation (18) follows from (16) because IBI = 1. Corollary A.3.1.
(19)
A"
•
For C nonsingular
I~', '~I=IC-yy'l =I~ ~1=ICI(I-Y'C-Iy).
Theorem A.3.3. Let the nonsingular matrix A be partitioned as in (2) so that is square. If A12 is nonsingular, let A 11 .2 =A11 -A!2A2iIA21' Then
(20)
Proof From Theorem A3.1,
o
(~l)
A22
)C-
I.
Hence
(22)
A-I=C(A~1 = (
A:J -I B
-A~IA21
~)( A~12
)(
Multiplication gives the desired result.
Corollary A.3.2.
If x'
= (X(I),
X(2)'),
-A~A22 ),
A~21 ~ • then
(23)
Proof From the theorem (24) x' A -IX
= x(l)' A1/ 2 x(l) -
x(l)'AII12AI2A221 x(::)
I A-I A A- i +A- 1 ) (2) -x (2)'A-22 A 21 A-I 11·2 X(I) + X(2)'(A-:4 22' 21 11·2 12 22 22 X ,
which is equal to the right-hand side of (23).
•
A.4
639
SOME MISCELLANEOUS RESULTS
Theorem A.3.4. Let the nonsingular matrix A be partitioned as in (2) so that is square. If A22 is nonsingular,
A22
(25)
(A22 - A 2I Aj"/A 12
r l = A 2"21A21( All - A12A2"2IA21 r l A12A2"l + A;;I.
Proof The lower right-hand corner of A -I is the right-hand side of (25) by Theorem A.3.3 and is also the left-hand side of (25) by interchange of 1 and 2.
•
Theorem A.3.S. and
Let U be p X m. The conditions for Ip - UU', 1m - U'U,
(26) to be positive definite are the same.
Proof We have (27)
(v'
w')
(~, ~)(:) = v'v + v'Uw + w'U'v + w'w = v'(Im - UU')v + (U'v + w)'( U'v + w).
The second term on the right-hand side is nonnegative; the first term is positive for all v"* 0 if and only if 1m - U'U is positive definite. Reversing the roles of v and w shows that (26) is positive definite if and only if Ip - UU' is positive definite. •
,
A.4. SOME MISCELLANEOUS RESULTS Theorem A.4.1. Let C be p Xp, positive semidefinite, and of rank r (~p). Theil there is a nonsingular matrix A such that
(1 )
ACA' =
(~
:).
Proof Since C is of rank r, there is a (p - r) X p matrix A2 such that
(2) Choose B (r X p) such that
(3)
640
MATRIX THEORY
is nonsingular. Then
(4)
A'2 ) --
BCB' (0
~) .
This matrix is of rank r, and therefore BCB' is nonsingular. By Corollary A.1.6 there is a nonsingular matrix D such that D(BCB')D' = I r • Then
(5) is a nonsingular matrix such that (1) holas.
•
Lemma A.4.1. If E is P X p, symmetric, and lIonsingular, there is a nonsingular matrix F such that
(6)
FEF'=(OI
0)
-I '
where tlie order of I is the number of positive characteristic roots of E and the order of -I is the number of negative characteristic roots of E. Proof From Theorem A.2.1 we know there is an orthogonal matrix G such that
(7)
o o
GEG'=
where hI ~ ... ~ hq > 0 > hq+1 ~ ... ~ hp are the characteristic roots of E. Let
(8) K=
1/..jh;
0
0
0
0
1/ VJI;
0
0
0
0
1/~
0
0
0
0
1/Fh:
A.4
641
SOME MISCELLANEOUS RESULTS
Then
(9)
KGEG'K' = (KG)E(KG)' =
(~ _~).
-
COl'ollary A.4.1. Let C be p X p, sym11lerric, alld of rank r (:5, p). Theil there is a nonsingular matrir A such thaI
(10)
ACA' =
o
(~
-I
o
where the order of 1 is the number of positive characteristic roots of C and the order of -lis the number of negative characteristic roots, the slim of the orders being r. Proof The proof is the same as that of Theorem A.4.1 except that Lemma _ A.4.1 is used instead of Corollary A.I.6. Let A he 11 X 111 (11 > m) such that
Lemma A.4.2... ( 11 )
There exists an n
A 'A X
=
I,,, .
(n - m) matru B such that (A B) is orthogonal.
Proof Since A is of rank m, there exists an 11 X (11 - m) matrix C such that (A C) is nonsingular. Take D as C-AA'C; then D'A=O. Let E [en - m) X (n - m)] be sllch that E' D' DE = I. Then B can be taken as DE.
Lemma A.4.3. Let x be a vector of orthogonal matrix 0 such that
11
-
components. Then there exists an
(12)
where c = {i';. Proof Let the first row of 0 be (J Ic)x'. The other rows may be chosen in any way to make the matrix orthogonal. _
Lemma A.4.4.
(13)
Let R = (b ij ) be a p
X
P mrtrir. Theil
i,j=l .... ,p.
642
MATRIX THEORY
Proof The expansion of IBI by elements of the ith row is p
IBI =
(14)
L bjhBjh · h~
Since
Bih
I
does not contain b jj , the lemma follows.
•
Lemma A.4.S. Let bjj = f3j/CI>"" c n ) be the i,jth element of a p Xp matrix B. Then for g = 1, ... , n, (15) Theorem A.4.2.
If A =A',
(16) alAI = 2A-. aa jj 'I'
( 17)
i =foj.
Proof Equation (16) folluws from the expansion of IAI according to elements of the ith row. To prove (17) let b jj = bjj = a jj , i, j = 1, ... , p, i 5.j. Then by Lemma A4.5,
(18) Since IAI = IBI and B jj = B jj =A jj =A jj , (17) follows.
•
Theorem A.4.3.
(19)
:x (x'Ax) = 2Ax,
where a/ax' denotes taking partial derivatives with respect to each component of x and arranging the partial derivatives in a column. Proof Let h be a column vector of as many components as x. Then
(20)
(x + h)'A(x + h) =x' Ax + h'Ax +x'Ah + h'Ah =x'Ax+ 2h'Ax +h'Ah.
The partial derivative vector is the vector multiplying h' in the second term on the right. • Definition A.4.1. Let A = (a jj ) be a p X m matrix and B = (ball) be a q X ~ matrix. The pq X mn matrix with ajjball as the element in the i, ath row and the
A.4
SOME MISCELLANEOUS RESULTS
643
j, 13th column is called the Kronecker or direct product of A and B and is denoted by A ® B; that is,
(21)
A®B=
a11B a 21 B
a l2 B a 22 B
almB a 2m B
aplB
a p2 B
apmB
Some properties are the following when the orders of matrices permit the indicated operations:
(22) (23)
(A ®B)(C ®D) = (AC) ® (BD), (A ®B)-l =A- 1 ®B- 1 •
Theorem A.4.4. Let the ith characteristic root of A (p xp) be it; and the corresponding characteristic vector be Xj = (Xli"'" xPy, and let the ath root of B (q X q) be Va and the corresponding characteristic vector be Ya , a = 1, ... , q. Then the i, a th root ofA ® B is Ai va' and the corresponding characteristic vector is Xi ®Ya = (XliY~"'" XpIY~)" i = 1, ... , p, a = 1, ... , q.
Proof
(24)
(A®B)(Xi®Ya) =
= Ai va AiXpjBYa
• XpjYa
644
MATRIX THEORY
Theorem A.4.S
(25) Proof The determinant of any matrix is the product of its roots; therefore
•
(26)
Definition A.4.2. (a;, ... , a:n)'.
If the p X m matrix A = (a l , ... , am), then vee A =
Some properties of the vee operator [e.g., Magnus (1988)] are
(27) (28)
vee ABC = (C' ®A)veeB, vee xy' = y ®x.
Theorem A.4.6. The Jacobian of the transformation E = y- 1 (from E to Y) 2p , where p is the order of E and Y.
is IYI-
Proof From EY = I, we have
(29) where
(30)
Then
(31) If 0 = Ya {3' then
(32)
( aoa) E = - E (a) ao¥ E = - Y -1( aoa Y )Y -1 .
A.4
645
SOME MISCELLANEOUS RESULTS
where Ea/3 is a p xp matrix with all elements 0 except the element in the ath row and 13th column, \\hich is 1; and e' a is the ath column of E and e/3' is its 13th row. Thus aeij/ aYa/3 = -e ja e/3j' Then the Jacobian is the determinant of a p2 x p2 matrix
• Theorem A.4.7. Let A and B be symmetric matrices with characteristic roots a l ~ a 2 ~ ... ~ a p and b l ~ b 2 ~ ... ~ b,,, respectively, and let H be a p X P orthogonal matrix. Then p
(34)
p
max tr HAB'B = L ajb j , H
j=l
minHA'H'B= L ajb p +l _ j ' H
j=1
Proof Let A =HQDQH~ and B = HbDbHi, where HQ and Hb are orthogonal and DQ and Db are diagonal with diagonal elements a l , ... , a p and b l , ... , bp respectively. Then
(35)
max trH*AH*'B = maxtrH*H D H'H*'H b D b H'b H* H* a a a
= maxtr HDnH'Db' H
where H=HiH*HQ' We have p
(36) tr HDaH'Db = L (HDQH');;b; ;=1 p-I
= L
p
j
L (HDQH')jj(bj-b j+l ) +b p L (HD"H')jj
;=1 j=1 p-l
:::; L
i
j=l p
Laj(bj-bi+I)+bpLaj
;=\ j=\
j=1
by Lemma AA.6 below. The minimum in (34) is treated as the negative of the • maximum with B replaced by - B [von Neumann (1937)].
646
MATRIX THEORY
Lemma A.4.6. r.f~1 P,)
Let P = (Pi) be a doubly stochastic matrix (Pij ~ 0, ~Y2 ~ ... ~Yp' Then
= 1, r.Y_1 Pij = 1). Let Yl k
n
k
LYi~ L LPijYj'
(37)
i= I
i~1
k= 1, ... ,p.
j= I
Proof k
(38)
L
p
p
L PijYj = L gjYj'
i=1 j=1
j=1
where gj = r.7~ 1Pij' j = 1, ... , P (0::5, gj (39)
: 5, 1,
r.Y~1 gj = k). Then
j~1 gjYj - j~/j = - J~ Yj + Yk ( k - j~ gj) + j~1 gjYj k
p
= L(Yj-Yd(gj-l)+
(Yj-Yk)gj
j=k+l
j~1
::5,0.
L
•
Corollary A.4.2. Let A be a symmetric matrix with characteristic roots a 1 ~a2 ~ ... a p' Then k
(40)
max tr R'AR =
R'R~lk
L ai •
i=1
Proof In Theorem A.4.7 let
•
(41) Theorem A.4.8.
11 + xCI =1+xtrC+O(x 2 ).
(42)
Proof The determinant (42) is a polynomial in x of degree P; the coefficient of the linear term is the first derivative of the determ~ant evaluated at x = O. In Lemma A.4.5 let n = 1, c 1 =x, f3ih(X) = 8ih +XCih ' where 0ii = 1 and 0ih = 0, i h. Then df3ih(x)/dx = Cih ' B jj = 1 for x = 0, and Bih = 0 for x = 0, i h. Thus
*"
(43)
*"
dIB(X)lj = ~ .. dx _ L..c". x-o i~1
•
A.S
647
ORTHOGONALIZATION AND SOLUTION OF LINEAR EQUATIONS
A.S. GRAM-SCHMIDT ORTHOGONALlZATION AND THE SOLUTION OF LINEAR EQUATIONS A.S.I. Gram-Schmidt Orthogonalization
The derivation of the Wishart density in Section 7.2 included the Gram-Schmidt orthogonalization of a set of vectors; we shall review that development here. Consider the p linearly independent n-dim.!Ilsional vectors VI'" .,~'p (p $;11). Define WI = l'I'
i=2, ... ,p.
(1) Then W;Wj
Wi
*' 0, *'
i = 1, ... , p, because
VI"'" vp
are linearly independent, and
= 0, .i j, as was proved by induction in Section 7.2. Let
Ui
= (l/liwiIDwi,
i = 1, ... ,p. Then u 1 , ... , up are Ol1hononnal; that is, they are orthogonal and of unit length. Let V = (u 1 , ••• , u). Then V'V = I. Define tu = IIwill (> 0), £'~w·
(2)
tij
=
ilw;ITJ = viu, j , I
j=I, ... ,i-I,
and tij = 0, j = i + 1, ... , p, i = 1, ... , P - 1. Then T = lar matrix. We can write (1) as i-I
(3)
Vi
= IlwillU i +
L
(t ij )
i=2, ... ,p,
is a lower triangu-
i
(v;uj)U j
=
j=!
L tijU j ,
i=I, ... ,p,
j=1
that is,
(4) Then A = V'V= TV'VT' = TT'
(5)
as shown in Section 7.2. Note that if V is square, we have decomposed an arbitrary nonsingular matrix into the product of an orthogonal matrix and an upper triangular malrix wilh positive diagonal elements; this is sometimes known as the QR decomposition. The matrices V and T in (4) are unique. These operations can be done in a different order. Let V = (v\O), .. . , v~O». For k = 1, ... , p - 1 define recursively ___ 1_
(6)
Uk -
k
(7) (8)
tjk
= VY-I) 'Uk'
vYl=vY-ll
-tjku k ,
(k-Il _
II V (k_I)II Vk
-
J..t
kk
(k-I) Vk ,
j=k+I, ... ,p, j=k+l, ... ,p.
648
MATRIX THEORY
Finally tpp = IIV?-I)II and up = (1/tpp)V~p-I). The same orthonormal vectors ul, ... ,u p and the same triangular matrix (tij) are given by the two procedures. The numbering of the columns of V is arbitrary. For numerical stability it is usually best at any given stage to select the largest of IIVY-I)II to call t kk • Instead of constructing Wi as orthogonal to wl> ... ,Wi_l, we can equivalently construct it as orthogonal to VI' .•. ' Vi-I. Let WI = VI' and define i-1
(9)
Wi
L IjjVj
= Vi +
j=1
such that i-1
o= V~Wj = V~Vj + L Ijjv~vj
(10)
j=1 i-I
= ahi +
L ah{h,
h=I, ... ,i-l.
j=1
Let
F
= (Jij), where
Iii
= 1 and
lij
= 0, i < j. Then
(11) Let D, be the diagonal matrix with IIwjll = tj} as the jth diagonal element. Then U = WD,-I = VF' D,-I. Comparison with V = UT' shows that F = DT- I. Since A = TT', we see that FA =DT' is upper triangular. Hence F is the matrix defined in Theorem A.1.2. There are other methods of accomplishing the QR decomposition that may be computationally more efficient or more stable. A Householder matrix has the form H = In - 20(0(', where 0('0( = 1, and is orthogonal and symmetric. Such a matrix H1 (i.e., a vector O() can be selected so that the first column of HIV has O's in all positions except the first, which is positive. The next matrix has the form
(12)
0)_2(0)(0 In_1 0(
O(')=(~
The (n - I)-component vector 0( is chosen so that the second column of HIV has all O's except the first two components, the second being positive. This process is continued until
(13)
A.5
ORTHOGONALIZATION AND SOLUTION OF LINEAR EQUATIONS
649
where T' is upper triangular and 0 is (n - p) xp. Let (14) where H(1) has p columns. Then from (13) we obtain V= H(1lT'. Since the decomposition is unique, H(1) = U. Another procedure uses Givens matrices. A Givens matrix G;j is I except for the elements gii = cos () = gjj and gij = sin () = - gji' i j. It is orthogonal. Multiplication of V on the left by such a matrix leaves all rows unchanged except the ith and jth; () can be chosen so that the i, jth element of G}" is O. Givens matrices G 21 , ... ,G"1 can be chosen in turn so G,,1 ... G 21 V has all D's in the first column except the first element, which is positive. Next G 32 , ••• ,G"2 can be selected in turn so that when they are applied the resulting matrix has D's in the second column except for the first two elements. Let
*"
(15) Then we obtain
(16) and
G(I)
= u.
A.S.2. Solution of Linear Equations In the computation of regression coefficients and other statistics. we need to solve linear equations
(17)
Ax=y,
where A is p X P and positive definite. One method of solution is Gaussian elimination of variable~, or pivotal condensation. In the proof of Theorem A.1.2 we constructed a lower triangular matrix F with diagonal elements 1 such that FA = A* is upper triangular. If Fy = y*, then the equation is
(18)
A*x = y*.
In coordinates this is p
(19)
L a;jx j = y(. j-l
650
MATRIX THEORY
Let art = arjla;" y;** =
yt lat, j = i, i + 1, ... , p, i = 1, ... , p. Then p
( 20)
L
xi=yi*-
lIj~'*Xi;
j·=i+ I
these equations are to be solved successively for x p' X p - 1" ' " Xl' The calculation of FA = A* is known as the !olWard solution, and the solution of (18) as the backward solution. Since FAF' =A* F' = n2 diagonal, (20) is A** x = y**, where A** = n- 2A* and y** = D-~ y*. Solving this equation gives
X=A**-ly"* =F'y**.
(21 ) The computation is (22)
The multiplier of y in (22) indicates a sequence of row operations which yields AThe operations of the forward solution transform A to the upper triangular matrix A*. As seen in Section A.S.l, the triangularization of a matrix can be done by a sequence of Householder transformations or by a sequence of Givens transformations. From FA = A*, we obtain J
•
p
( 23)
IAI =
na~;), i=l
\vhich is the product of the diagonal elements of A*, resulting from the forward solution. We also have (24)
=y*'y**. The forward solution gives a computation for the quadratic form which occurs in T~ and other statistics. For more on matrix computations consult Golub and Von Loan (1989).
t
APPENDIX B
Tables
TABLEB.1 WILKS'LIKELlHooD CRITERION: FACI'ORS C(p, m, M) TO AoruSTTO WHERE M=n -p + 1
X;'m,
5% Significance Level
p-3 IO
M\m
2
4
6
8
12
14
16
1 2 4 5
1.295 1.109 1.058 1.036 1.025
1.422 1.174 1.099 1.065 1.046
1.535 1.241 1.145 1.099 1.072
' .. 632 1.302 1.190 1.133 1.100
1.716 1.359 1.232 1.167 1.127
U91 lAIO 1.272 U99 U54
1.857 1.458 1.309 1.229 1.179
1.916 1501 1.344 1.258 1.204
1.971 1.542 1.377 1.286 1.228
6 7 8 9 10
1.018 1.014 1.011 1.009 1.007
1.035 1.027 1.022 1.018 1.015
1.056 1.044 1.036 1.030 1.025
1.078 1.063 1.052 1.043 1.037
1.101 1.082 1.068 1.058 1.050
LI23 LI01 1.085 1.073 1.063
1.145 1.121 1.102 1.088 1.076
1.167 1.139 1.119 1.102 1.089
1.188 1.158 U35 LI17 U03
12 15 20 30 60
1.005 1.003 1.002 1.001 1.000 1.000
1.011 1.008 1.004 1.002 1.001 1.000
1.019 1.008 1.004 1.001 1.000
1.028 1.020 1.012 1.006 1.002 1.000
1.038 1.027 1.017 1.009 1.002 1.000
1.048 1.035 1.022 1.011 1.003 1.000
1.059 1.043 1.028 1.015 1.004 1.000
1.070 1.052 1.034 1.018 1.006 1.000
1.081 1.060 1.040 1.021 1.007 1.000
12.5916
21.0261
28.8693
36.4150
43.7730
50.9985
00
X; ..
LOB
58.1240 65.1708
IH
72.1532
An Introduction to Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0-471-36091-0 Copyright © 2003 John Wiley & Sons, Inc.
651
TABLE B.1
(Continued)
. 5% Significance Level
p=4
p=3 20
22
2
4
6
8
10
12
14
1 2 3 4 5
2.021 1.580 l.408 1.313 1.251
2.067 1.616 l.438 1.338 l.273
l.407 1.161 1.089 1.057 l.040
l.451 1.194 1.114 1.076 l.055
1.517 1.240 1.148 1.102 1.076
1.583 1.286 1.183 1.130 1.099
l.644 1.331 1.218 1.159 1.122
1.700 1.373 1.252 1.186 1.145
1.751 1.413 1.284 l.213 1.168
6 7 8 9 10
1.208 1.176 1.151 1.132 1.116
l.227 1.193 1.167 1.147 1.129
1.030 l.023 l.018 1.015 l.012
1.042 l.033 1.027 1.022 1.018
l.r59 l.047 l.038 1.032 1.027
l.078 1.063 1.052 1.044 1.038
l.097 1.080 l.067 1.057 1.049
l.118 l.097 1.082 l.070 l.061
1.137 l.115 1.097 l.084 1.073
12 15 20 30 60 00
1.092 l.069 l.046 1.025 l.008 1.000
1.103 l.078 1.052 l.029 1.009 1.000
1.009 l.006 1.003 1.002 1.000 1.000
l.014 l.009 1.006 1.003 1.001 1.000
l.020 1.014 1.009 1.004 1.001 l.ooo
l.029 l.020 1.013 1.006 1.002 1.000
1.038 l.027 l.017 l.009 1.003 1.000
l.047 l.035 l.022 l.01l l.003 l.ooo
l.058 l.042 l.027 l.014 l.004 l.ooo
X~m
79.0819
85.9649
15.5073
26.2962
36.4150
65.1708
74.468
M\m
46.1943 55.7585
TABLE B.1 (Continued) 5% Significance Level
p-4 M\m
16
18
p-5 20
4
6
8
10
12
1.483 l.216 1.130 l.089 1.065
1.514 l.245 1.154 1.108 1.081
1.556 1.280 1.182 1.131 1.100
HOO
1.233
l.503 1.209 1.120 1.079 1.056
1.315 l.211 1.155 1.120
1.643 1.350 l.24O 1.179 1.141
1.176 1.149 1.128 1.111 l.098
1.194 1.165 1.143 1.125 l.110
1.042 1.033 l.026 l.022 l.018
1.050 1.040 l.032 l.027 l.023
l.063 1.051 1.042 1.035 1.030
UJ79 1.065 1.054 l.046 1.039
1.097 1.080 l.067 1.057 l.050
1.114 1.095 l.081 l.070 1.061
l.078 1.058 l.039 1.021 1.007 1.000
l.088 1.066 1.045 1.024 l.008 1.000
1.013 1.009 1.005 1.002 l.001 1.000
1.017 l.01l 1.007 l.003 1.001 1.000
l.023 1.016 1.010 1.005 1.001 1.000
1.030 1.021 1.013 l.007 1.002 1.000
1.038 1.028 l.018 1.009 1.003 1.000
1.047 1.034 1.022 1.012 1.004 1.000
55.7585
67.5048
1 2 3 4 5
l.799 1.450 1.314 l.239 1.190
l.843 1.485 1.343 l.264 l.212
1.884 1.518 1.371 1.288
6 7 8 9 10
1.157 1.132 1.113 l.098 l.086
12 15 20 30 60 00
l.068 l.050 l.033 l.018 l.005 l.ooo
X~m
83.6753
652
2
92.8083 101.879
18.3070
31.4104 43.7730
79.081
TABLE B.l
(Continued)
5% Significance Level p~5
p=7
p=6
M\m
14
16
2
6
8
10
12
2
4
1 2 3 4 5
1.683 1.383 1.267 1.203 1.161
1.722 1.415 1.294 1.226 1.181
1.587 1.254 1.150 1.100 1.072
1.520 1.255 1.163 1.116 1.088
1.543 1.279 1.184 1.134 1.103
1.573 1.307 1.208 1.154 1.120
1.605 1.335 1.232 1.175 1.138
1.662 1.297 1.178 1.121 1.089
1.550 1.263 1.165 1.116 1.087
6 7 8 9 10
1.132 1.111 1.095 1.082 1.072
1.150 1.127 1.109 1.095 1.083'
1.055 1.04\ 1.03; 1.029 1.024
1.069 1.056 1.046 1.039 1.034
1.082 1.068 1.057 1.048 1.042
1.097 1.081 1.068 1.059 1.051
1.113 1.095 1.081 1.070 1.061
1.068 1.054 1.044 1.036 1.031
1.068 1.055 1.045 1.038 1.032
12 15 20 30 60
1.057 1.042 1.027 1.014 1.004 1.000
1.066 1.049 1.033 1.018 1.006 1.000
1.018 1.012 1.007 1.003 1.001 1.000
1.025 1.018 1.011 1.006 1.002 1.000
1.032 1.023 1.014 1.007 1.002 1.000
1.040 1.029 1.018 1.010 1.003 1.000
1.048 1.035 1.023 1.012 1.004 1.000
1.023 1.016 1.0lO 1.005 1.001 1.000
1.024 1.017 1.011 1.005 1.001 1.000
79.0819
92.8083
23.6848
41.337
00
X;m
90.5312 101.879
21.0261
50.9985 65.1708
TABLE B.1
(Continued)
5% Significance Level
p=7
p
p= 8
=
P = 10
9
6
8
10
2
8
2
4
6
2
1 2 3 4 5
1.530 1.266 1.173 1.124 1.095
1.538 1.282 1.189 1.139 1.108
1.557 1.303 1.208 1.155 1.122
1.729 1.336 1.206 1.142 1.105
1.538 1.288 1.195 1.144 1.113
1.791 1.373 1.232 1.162 1.121
1.614 1.309 1.201 1.144 1.110
1.558 1.293 1.196 1.144 1.112
1.847 1.40S 1.257 1.182 1.137
6 7 8 9 10
1.075 1.062 1.051 1.043 1.037
1.086 1.071 1.060 1.051 1.044
1.099 1.083 1.070 1.060 1.053
1.081 1.065 1.053 1.044 1.038
1.091 1.076 1.064 1.055 1.048
1.094 1.076 1.062 1.052 1.045
1.088 1.071 1.060 1.050 1.043
1.090 1.074 1.062 1.053 1.046
1.107 1.087 1.072 1.061 1.052
12 15 20 30 60
1.029 1.020 1.013 1.006 1.002 1.000
1.034 1.024 1.016 1.008 1.002 1.000
1.042 1.031 1.019 1.010 1.003 1.000
1.028 1.019 1.012 1.006 1.001 1.000
1.038 1.027 1.017 1.009 1.003 1.000
1.034 1.023 1.014 1.007 1.002 1.000
1.033 1.023 1.015 1.007 1.002 1.000
1.035 1.025 1.016 1.008 1.002 1.000
1.039 1.028 1.017 1.009 1.002 1.000
58.1240
74.4683
83.6753
28.8693
50.9985
72.1532
31.4104
M\m
00
X;m
~
90.5312 26.2962
653
TABLE B.l
(Continued)
1 % Significance Level p=3 4
6
10
12
14
5
1.356 1.131 1.070 1.043 1.030
1.514 1.207 1.116 1.076 1.054
1.649 1.282 1.167 1.113 1.082
1.763 1.350 1.216 1.150 1.112
1.862 1.413 1.262 1.187 1.141
1.949 1.470 1.306 1.221 1.170
2.026 1.523 1.346 1.254 1.198
2.095 1.571 1.384 1.285 1.224
6 7 8 9 10
1.022 l.016 l.013 1.010 1.009
1.040 l.031 1.025 1.021 1.017
1.063 l.050 1.041 1.034 1.028
1.087 l.070 1.058 1.048 1.041
1.112 1.091 1.075 1.055
1.136 1.11l l.093 1.080 1.069
1.159 1.132 1.111 1.095 1.082
1.182 1.152 1.129 1.111 1.097
12 15 20 30 00
1.006 1.004 1.002 1.001 1.000 1.000
1.012 1.009 1.005 l.002 1.001 1.000
1.021 1.014 1.009 l.004 1.001 1.000
1.031 1.021 1.013 1.007 1.002 1.000
l.042 1.030 1.019 1.009 1.003 1.000
1.053 l.038 1.024 l.012 1.004 1.000
1.064 1.047 l.030 l.016 1.005 1.000
1.076 1.056 1.036 l.019 1.006 1.000
X;",
16.8119
26.2170
34.8053
42.9798
50.8922
58.6192
M\m 2 3 4
60
TABLE B.l
1.064
16
66.2062 73.6826
(Continued)
1 % Significance Level 18
p=3 20
22
2
4
6
8
10
1 2 3 4 5
2.158 1.616 l.420 1.315 1.249
2.216 1.657 1.453 1.344 1.274
2.269 l.696 1.485 1.371 1.297
1.490 1.192 1.106 l.068 1.047
1.550 1.229 1.132 l.088 1.063
1.628 1.279 1.168 1.115 1.085
1.704 1.330 1.207 1.146 1.109
1.774 1.379 1.244 1.176 1.134
6 7 8
1.204 1.171 1.146 1.127 1.111
1.226 1.190 1.163 1.142 1.125
l.246 l.209 1.180 1.157 1.139
l.035 1.027 1.021 1.017 l.014
1.048 l.037 l.030 1.025 1.021
1.066 1.052 l.043 1.036 1.030
1.086 1.070 1.053 1.048 l.041
1.107 l.088 l.073 l.062 1.054
1.087 1.065 1.043 1.023 1.007 1.000
1.099 1.074 1.049 1.027 1.009 l.ooo
1.110 1.083 l.056 1.031 1.010 1.000
1.010 1.007 1.004 1.002 1.000 1.000
1.015 1.010 1.006 1.003 1.001 1.000
1.023 1.016 l.01O 1.005 1.001 1.000
1.031 1.022 l.014 1.007 l.002 1.000
1.041 1.029 1.019 1.009 1.003 1.000
81.0688
88.3794
95.6257
20.0902
31.9999
42.9798
Q
10 12 15 20 30
60 00
X~m
654
p-4
M\m
53.4858 63.6907
TABLE B.1 (Continued) 1 % Significance Level p-4 18 16 20
2
p-5
12
14
4
6
1 2 3 4 5
1.838 1.424 1.280 1.205 1.159
1.896 1.467 1.314 1.234 1.183
1.949 1.507 1.347 1.261 1.207
1.999 1.545 1.378 1.287 1.230
2.045 1.580 1.408 1.313 1.252
1.606 1.248 1.141 1.092 1.065
1.589 1.253 1.150 1.101 1.074
1.625 1.284 1.175 1.121 1.090
6 7 8 9 10
1.128 1.106 1.089 1.076 1.066
1.149 1.124 1.105 1.091 1.079
1.169 1.142 1.121 1.105 1.092
1.189 1.160 1.137 1.119 1.105
1.208 1.177 1.153 1.133 1.118
1.049 1.038 1.031 1.025 1.021
1.056 1.044 1.036 1.030 1.025
1.070 1.056 1.046 1.039 1.033
12 15 20 30 60 00
1.051 1.017 1.024 1.012 1.004 1.000
1.062 1.045 1.029 1.015 1.005 1.000
1.073 1.053 1.035 1.019 1.006 1.000
1.083 1.062 1.041 1.022 1.007 1.000
1.094 1.071 1.047 1.026 1.008 1.000
1.015 1.010 1.006 1.003 1.001 1.000
1.019 1.013 1.008 1.004 1.001 1.000
1.025 1.017 1.011 1.005 1.001 1.000
X;m
73.6826
83.5134
M\m
93.2168 102.8168 112.3292 23.2093
37.5662 50.8922
TABLE B.1 (Continued)
--
I % Significance Level p-5
p=6
1\
10
12
14
16
2
6
8
1 2 3 4 5
1.672 1.321 1.204 1.145 1.110
1.721 1.359 1.235 1.171 1.131
1.768 1.396 1.265 1.196 1.153
1.813 1.431 1.294 1.221 1.174
1.855 1.465 1.323 1.245 1.196
1.707 1.300 1.175 1.116 1.084
1.631 1.294 1.183 1.129 1.097
1.656 1.319 1.205 1.148 1.113
6 7 8 9 10
1.087 1.071 ,1.059 1.050 1.043
1.105 1.087 1.073 1.062 1.054
1.124 1.103 1.087 1.075 1.065
1.143 1.119 1.102 1.088 1.077
1.161 1.136 1.116 1.101 1.089
1.063 1.050 1.040 1.033 1.028
1.076 1.061 1.051 1.043 1.037
1.090 1.074 1.062 1.052 1.045
12 15 20 30 00
1.033 1.023 1.015 1.007 1.002 1.000
1.041 1.030 1.019 1.010 UlO3 1.000
1.051 1.037 1.024 1.012 1.004 1.000
1.060 1.044 1.029 1.015 1.005 1.000
1.070 1.052 1.034 1.019 1.006 1.000
1.021 1.014 1.008 1.004 1.001 1.000
1.028 1.020 1.012 1.006 1.002 1.000
1.035 1.024 1.015 1.008 1.002 1.000
X;m
63.6907
76.1539
M\m
60
88.3794 100.425 ~ 112.32SIl
26.2170
58.6192 73.6826
655
TABLE B.1
(Continued)
1 % Significance Level
p=7
p-6
M\m
10
12
2
4
6
8
10
1 2 3 4 5
1.687 1.348 1.230 1.169 1.131
1.722 1.378 1.255 1.191 1.150
1.797 1.348 1.207 1.140 1.102
1.667 1.305 1.188 1.130 1.097
1.642 1.306 1.194 1.138 1.105
1.648 1.321 1.210 1.152 1.117
1.66'; 1.342 1.229 1.169 1.132
6 7 8 9
1.106 1.087 1.074 1.063 1.055
1.122 1.102 1.086 1.075 1.065
1.078 1.062 1.050 1.042 1.035
1.076 1.061 1.050 1.042 1.036
1.083 1.067 1.056 1.047 1.041
1.094 1.077 1.065 1.055 1.048
1.107 1.089 1.075 1.065 1.056
1.042 1.030 1.019 1.010 1.003 1.000
1.051 1.037 1.024 1.013 1.004 1.000
1.026 1.018 1.011 1.005 1.001 1.000
1.027 1.019 1.012 1.006 1.002 1.000
1.031 1.022 1.014 1.007 1.002 1.000
1.037 1.025 1.017 1.00'1 1.003 1.000
1.044 1.032 1.020 1.011 1.003 1.000
29.1412
48.2782
66.2062
10 12 15 20 30 60 00
X~m
88.3794 102.816
83.5134 100.425
TABLE B.1 (Continued) 1 % Significance Level 2
8
2
4
6
p -10 2
1.879 1.394 1.238 1.163 1.120
1.646 1.326 1.215 1.158 1.123
1.953 1.436 1.267 1.185 1.138
1.740 1.355 1.226 1.161 1.122
1.671 1.333 1.218 1.158 1.122
2.021 1.476 1.296 1.207 1.155
i.092 1.074 1.060 1.050 1.043
1.099 1.082 1.069 1.059 1.051
1.107 1.086 1.070 1.059 1.050
1.096 1.078 1.065 1.055 1.047
1.098 1.080 1.067 1.058 1.050
1.121 1.098 1.081 1.068 1.058
1.032 1.022 1.013 1.007 1.002 1.000
1.040 1.028 1.018 1.009 1.003 1.000
1.038 1.026 1.016 1.008 1.002 1.000
1.036 1.026 1.016 1.008 1.002 1.000
1.037 1.027 1.017 1.009 1.003 1.000
1.044 1.031 1.019 1.010 1.003 1.000
34.8053
58.6192
81.0688
375662
p-9
p=8
M\m 1 2 3 4 5 6 7 ~
9 10 12 15 20 30 60 00
X~m
656
31.9999 93.2168
TABLE B.2 TABLES OF SIGNIFICANCE POINTS FOR TIlE LAWLEy-HOTELLlNG TRACE TEST Pr{ -;;; W <= X. } -
IX
5% Significance Level
p-2 n\m
4
10
12
15
9.859· 58.428 23.999 15.639 12.175 10.334 9.207
10.659· 58.915 23.312 14.864 11.411 9.594 8.488
11.098· 59.161 22.918 14.422 10.975 9.169 8.075
11.373· 59.308 22.663 14.135 10.691 8.893 7.805
11.562· 59.407 22.484 13.934 10.491 8.697 7.614
11.804· 59.531 22.250 13.670 10.228 8.440 7.361
11.952' 59.606 22.104 13.504 10.063 8.277 7.201
12.052" 59.655 22.003 13.391 9.949 8.164 7.090
12.153' 59.705 21.901 13.275 9.832 8.048 6.975
10 12 14 18 20
7.909 7.190 6.735 6.193 6.019
7.224 6.528 6.090 5.571 5.405
6.829 6.146 5.717 5.209 5.047
6.570 5.894 5.470 4.970 4.810
6.386 5.715 5.294 4.798 4.640
6.141 5.474 5.057 4.566 4.410
5.984 5.320 4.905 4.416 4.260
5.875 5.212 4.798 4.309 4.154
5.761 5.100 4.686 4.198 4.042
25 30 35 40
5.724 5.540 5.414 5.322
5.124 4.949 4.829 4.742
4.774 4.604 4.488 4.404
4.542 4.374 4.260 4.178
4.374 4.209 4.096 4.014
4.147 3.983 3.872 3.791
3.998 3.835 3.724 3.643
3.892 3.729 3.618 3.538
3.780 3.617 3.505 3.425
50 60 70 80
5.198 5.118 5.062 5.020
4.625 4.549 4.496 4.457
4.290 4.217 4.165 4.127
4.066 3.994 3.944 3.907
3.904 3.833 3.783 3.747
3.682 3.611 3.562 3.526
3.535 3.465 3.416 3.380
3.429 3.359 3.310 3.274
3.315 3.245
100 200
4.963 4.851
4.403 4.298
4.075 3.974
3.856 3.757
3.696 3.598
3.476 3.380
3.330 3.234
3.224 3.127
3.109 3.012
00
4.744
4.197
3.877
3.661
3.504
3.287
3.141
3.035
2.918
2 4 5 6 7
2
"Multiply by
6
3.196 3.159
\0 2•
657
TABLE B.2
(Continued)
1% Significance Level
p-2 n\m
2
10
12
15
6 7 8
VI67 t 2.985" 74.275 38.295 26.118 20.388 17,152
2.667t 2.990· 71.026 35.567 23.794 18.326 15.268
2.776 t 2.992· 69.244 34.070 22.517 17.191 14.229
2.8441 2.994· 68.116 33.121 21.706 16.469 13.567
2.891t 2.995" 67.337 32.465 21.143 15.%7 13.106
2.9521 2.996" 66.332 31.615 20.413 15.313 12.504
2.989 t 2.997· 65.712 31.088 19.958 14.905 12.127
3.014t 2.997· 65.290 30.729 1\'.648 14.626 11.868
3.039 t 2.998" 64.862 30.364 19.332 14.341 11.603
10 12 14 18 20
13.701 11.920 10.844 9.617 9.236
12.038 10.388 9.399 8.278 7.932
11.120 9.541 8.597 7.533 7.206
10.531 8.996 8.082 7.053 6.736
10.121 8.615 7.720 6.714 6.406
9.582 8.113 7.242 6.265 5.966
9.243 7.796 6.939 5.979 5.685
9.0lJ 7.577 6.729 5.780 5.489
8.769 7.351 6.511 5.572 5.284
25 30 35 40
8.604 8.219 7.959 7.773
7.360 7.013 6.780 6.613
6.666 6.339 6.120 5.964
6.217 5.903 5.692 5.542
5.899 5.593 5.389 5.243
5.476 5.180 4.982 4.841
5.204 4.914 4.720 4.582
5.013 4.726 4.535 4.398
4.813 4.529 4.339 4.204
50 60 70 80
7.523 7.363 7.252 7.171
6.389 6.247 6.148 6.075
5.754 5.621 5.529 5.461
5.341 5.214 5.125 5.061
5.048 4.924 4.838 4.775
4.653 4.534 4.451 4.391
4.397 4.280 4.199 4.140
4.216 4.100 4.020 3.961
4.023 3.908 3.829 3.770
100 200
7.059 6.843
5.976 5.785
5.369 5.191
4.972 4.803
4.690 4.525
4.308 4.150
4.059 3.903
3.881 3.727
3.691 3.538
00
6.638
5.604
5.023
4.642
4.369
4.000
3.757
3.582
3.393
4
tMultiply by 10', 'Multiply by 10'.
658
4
6
TABLE B.2 (Continued) 5% Significance Level
n\m
~.
6
p=3 8
3
4
10
12
15
20
4 5 6 7
25.930· 1.18842.474 25.456 18.752 15.308
26.9961.193" 41.764 24.715 18.056 14.657
27.6651.196" 41.305 24.235 17.605 14.233
28.125" 1.198" 40.983 23.899 17.288 13.934
28.7121.200' 40.562 23.458 16.870 13.540
29.073" 1.20240.300 23.182 16.608 13.290 .
29.3161.203" 40.120 22.992 16.427 13.118
29.5611.20439.937 22.799 16.241 12.941
29.809' 1.205' 39.750 22.600 16.051 12.758
10 12 14 16 18 20
11.893 10.229 9.255 8.618 8.170 7.838
11.306 9.682 8.736 8.118 . 7.685 7.365
10.921 9.323 8.394 7.788 7.364 7.051
10.649 9.068 8.149 7.553 7.135 6.826
10.287 8.727 7.822 7.236 6.825 6.522
10.057 8.509 7.612 7.031 6.624 6.325
9.897 8.357 7.465 6.887 6.483 6.185
9.732 8.198 7.311 6.736 6.334 6.038
9.560 8.033 7.150 6.577 6.177 5.882
25 30 35 40
7.294 6.965 6.745 6.588
6.841 6.524 6.313 6.162
6.539 6.231 6.025 5.878
6.323 6.020 5.818 5.673
6.029 5.732 5.534 5.393
5.837 5.543 5.348 5.208
5.700 5.409 5.214 5.076
5.556 5.265 5.072 4.934
5.401 5.112 4.919 4.781
50 60 70 80
6.377 6.243 6.150 6.082
5.961 5.832 5.744 5.679
5.682 5.558 5.471 5.408
5.481 5.359 5.274 5.212
5.205 5.086 5.003 4.943
5.022 4.904 4.823 4.763
4.891 4.774 4.693 4.634
4.750 4.633 4.553 4.493
4.597 4.480 4.399 4.339
100
200
5.989 5.810
5.590 5.419
5.322 5.156
5.128 4.965
4.860 4.702
4.682 4.525
4.552 4.397
4.413 4.257
4.258 4.102
00
5.640
5.256
4.999
4.812
4.552
4.377
4.250
4.110
3.954
"Multiply by 10 2•
659
TABLE B.2 (Continued) 1% Signi/ica Ice Level
n\m
p=3 8
4
5
6.484t
6.750 t
6.917 t
4 5 6 7 8
5.9901.27459.507 37.994 28.308
5.9951.24257.032 35.993 26.599
5.9981.22255.462 34.721 25.511
7.031 6.0001.20854.377 33.840 24.755
10 12 14 16 18 20
19.737 15.973 13.905 12.610 11.729 11.091
18.355 14.765 12.803 11.581 10.751 10.152
17.471 13.990 12.096 10.918 10.120 9.545
25 30 35 40 50 60 70 80
10.075 9.479 9.087 8.811 8.448 8.220 8.063 7.948
9.201 8.644 8.280 8.023 7.686 7.474 7.329 7.224
100 200
7.793 7.498
00
7.222 4
tMultiply by 10 • 'Multiply by 10 2•
660
10
12
15 7.389 t
6.002' 1.190' 52.973 32.695 23.771
7.267t 6.0031.17952.102 31.984 23.157
7.328t 6.005' 1.172" 51.509 31.498 22.737
7.451t 6.006' 6.007' 1.164- 1.156' 50.906 50.292 31.002 30.496 22.308 21.868
16.855 13.448 11.599 10.452 9.676 9.117
16.050 12.737 10.945 9.836 9.087 8.549
15.544 12.288 10.530 9.441 8.712 8.186
15.197 11.978 10.243 9.172 8.450 7.932
14.840 11.659 9.946 8.890 8.178 7.668
14.472 11.328 9.638 8.596 7.893 7.390
8.634 8.102 7.755 7.511 7.189 6.988 6.850 6.750
8.233 7.718 7.3'82 7.146 6.836 6.642 6.509 6.412
7.699 7.205 6.883 6.65(, 6.360 6.174 6.047 5.955
7.356 6.874 6.560 6.339 6.050 5.870 5.746 5.656
7.115 6.641 6.332 6.115 5.831 5.653 5.531 5.443
6.803 6.395 6.091 5.877 5.597 5.422 5.302 5.215
6.596 6.135 5.834 5.623 5.346 5.172 5.053 4.967
7.081 6.808
6.614 6.356
6.281 6.032
5.830 5.593
5.534 5.304
5.323 5.096
5.096 4.873
4.850 4.627
6.554
6.116
5.801
5.373
5.089
4.885
4.664
4.419
6 t
7.178t
20
TABLE B.2 (Continued) 5% Significance Level
p-4
n\m
10
12
15
20
25
4 S 6 7 8
49.964· 1.996" 6S.71S 37.343 26.S16
5l.204" 2.001· 64.999 36.629 2S.868
52.054· 2.005· 64.497 36.129 25.413
53.142· 2.00963.841 35.474 24.814
53.808" 2.011· 63.432 35.064 24.437
54.258" 2.013· 63.151 34.782 24.178
54.71" 2.0lS· 62.866 34.495 23.912
55.17· 2.016· 62.573 34.200 23.639
55.462.017· 62.396 34.019 23.471
10 14 16 18 20
17.875 14.338 12.455 11.295 10.512 9.950
17.326 13.848 12.002 10.868 10.104 9.556
16.938 13.500 11.680 10.563 9.812 9.274
16.424 13.037 11.248 10.154 9.419 8.893
16.098 12.741 10.972 9.890 9.165 8.645
15.872 12.535 10.778 9.705 8.986 8.471
15.640 12.321 10.577 9.512 8.798 8.287
15.399 12.099 10.366 9.309 8.600 8.093
15.250 11.961 10.234 9.181 8.475 7.970
25 30 35 40 SO 60 70 80
9.059 8.538 8.197 7.957 7.640 7.442 7.305 7.206
8.688 8.182 7.852 7.619 7.313 7.120 6.988 6.892
L422 7.927 7.603 7.375
7.826 7.350 7.040 6.821 6.535 6.356 6.232 6.143
7.659 7.188 6.880 6.664
7.075 6.887 6.758 6.665
8.062 7.578 7.263 7.041 6.750 6.568 6.443 6.351
6.380 6.203 6.081 5.992
7.482 7.015 6.710 6.495 6.214 6.038 5.917 5.829
7.293 6.829 6.526 6.313 6.033 5.858 5.738 5.650
7.l73 6.710 6.408 6.195 5.916 5.740 5.620 5.532
100 200
7.071 6.814
6.762 6.514
6.537 6.295
6.228 5.993
6.021 5.791
5.872 5.644
5.710 5.484
5.531 5.305
5.413 5.186
00
6.S74
6.282
6.069
5.774
5.576
5.431
5.272
5.094
4.974
12
4
6
-Multiply by 10'-
661
TABLE B.2 ( Continued) 1 % Significance Level
n\m
4
p=4 10
6
6 7 8
12.491+ 9.9991.93885.053 51.991
12.800+ 10.0041.906· 82.731 50.178
13.012! 10.0081.88581.125 48.921
13.2831 10.0121.85779.047 47.290
10 12 14 16 18 20
29.789 21.965 18.142 15.916 14.473 13.466
28.478 20.889 17.199 15.059 13.674 12.710
27.566 20.138 16.539 14.457 13.112 12.177
25 30 35 40
11.924 11.055 10.499 10.114
11.237 10.409 9.880 9.514
50 60 70 80
9.614 9.305 9.095 8.944
100
8.739 8.354 8.000
4
200 00
tMultiply by 10". "MUltiply by 10'.
662
13.449+ 10.0141.840·
12
15
20
25
77.759 46.276
13.561+ . 13.67+ 10.016- 10.018·1.828· 1.816· 76.882 75.989 45.583 44.877
13.87+ 10.021.804- 1.797· 75.082 74.522 44.156 43.715
26.376 19.154 15.670 13.662 12.368 11.470
25.632 18.534 15.121 13.157 11.894 11.018
25.121 18.108 14.742 12.807 11.564 lO.703
24.597 17.668 14.349 12.444 11.221 10.374
24.060 17.215 13.943 12.066 10.863 lO.030
23.731 16.936 13.691 11.831 lO.639 9.814
10.751 9.951 9.440 9.087
lO.103 9.338 8.851 8.514
9.687 8.943 8.470 8.142
9.395 8.665 8.200 7.879
9.089 8.372 7.915 7.600
8.766 8.060 7.611 7.301
8.562 7.363 7.418 7.110
9.040 8.747 8.549 8.405
8.631 8.319 8.158 8.020
8.079 7.311 7.630 7.498
7.720 7.460 7.284 7.157
7.465 7.2lO 7.037 6.912
7.194 6.943 6.774 6.651
6.902 6.655 6.488 6.367
6.713 6.468 6.301 6.181
8.211 7.848 7.513
7.833 7.484 7.163
7.321 6.990 6.686
6.985 6.664 6.369
6.744 6.429 6.140
6.486 6.176 5.892
6.204 5.898 5.616
6.019 5.714 5.432
13.79+ 10.02·
TABLE B.2
(Continued)
5% Significance Level
p=5 8
10
12
83.352· 3.014* 93.042 50.646
85.093* 3.020· 92.102 49.739
86.160 t 3.024t 91.515 49.170
86.88 t 3.027 t 91.113 48.780
3.029t 90.705 48.382
3.032 t 90.29 47.973
90.04 47.723 47.35
10 27.667 12 20.169 14 16.643 16 14.624 18 13.326 20 12.424
27.115 19.701 16.224 14.239 12.963 12.078
26.387 19.079 15.666 13.722 12.476 11.612
25.927 18.683 15.309 13.389 12.161 11.310
25.610 18.409 15.059 13.157 11.939 11.097
25.284 18.124 14.800 12.914 11.708 10.874
24.947 17.830 14.530 12.659 11.463 10.637
24.740 17.647 14.361 12.499 11.310 10.488
'24.422 17.365 14.100 12.250 11.068 10.252
25 30 40
11.046 10.270 9.774 9.429
10.728 9.969 9.484 9.147
10.297 9.559 9.088 8.761
10.016 9.291 8.828 8.507
9.817 9.099 8.642 8.325
9.606 8.896 8.444 8.130
9.381 8.679 8.230 7.919
9.239 8.539 8.093 7.783
9.010 8.314 7.869 7.561
50 60 70 80
8.982 8.706 8.517 8.381
8.711 8.441 8.257 8.124
8.339 8.077 7.899 7.770
8.Q92 7.836 7.661 7.535
7.915 7.662 7.489 7.365
7.725 7.474 7.304 7.181
7.518 7.269 7.100 6.978
7.383 7.135 6.967 6.845
7.161 6.912 6.743 6.621
100 200
8.197 7.850 7.531
7.945 7.607 7.295
7.597 7.271 6.970
7.365 7.045 6.750
7.197 6.881 6.590
7.014 6.702 6.414
6.813 6.503 6.217
6.680 6.370 6.084
6.455 6.142 5.850
6 7 8
3S
00
81.991· 3.009* 93.762 51.339
15
20
25
6
n\m
40
tMultiply by 10'. "Multiply by 10 2.
663
TABLEB.2 (Continued) 1 % Significance Level
p=5
n\m
6
8
10
12
15
20
25
40
5 6 7 8
20.495· 15.014· 2.735· 1.150"
20.834· 15.019* 2.704* 1.128"
21.267· 15.025* 2.665* 1.099·
21.53* 15.029* 15.033* 15.03* 15.06" 2.640* 2.623* 2.606* 2.590* 1.081· 1.069· 1.057* 1.044*
10 12 14 16 18 20
48.048 31.108 24.016 20.240 17.929 16.380
46.670 30.065 23.145 19.472 17.228 15.727
44.877 28.701 22.001 18.459 16.302 14.862
43.758 27.846 21.279 17.817 15.713 14.310
42.992 27.257 20.781 17.373 15.304 13.925
42.210 26.653 20.268 16.913 14.878 13.525
41.408 26.031 19.736 16.435 14.435 13.105
40.921 25.648 19.408 16.138 14.159 12.843
25 30 35 40
14.107 12.880 12.115 11.593
13.529 12.345 11.607 11.105
12.759 11.629 10.926 10.448
12.265 11.167 10.486 10.022
11.918 10.842 10.174 9.720
11.555 10.500 9.845 9.401
11.172 10.136 9.494 9.058
10.930 10.547 9.906 9.538 9.271 8.911 8.839 8.484
50 60 70 80
10.928 10.523 10.251 10.055
10.465 10.076 9.814 9.626
9.841 9.471 9.223 9.045
9.434 9.076 8.835 8.663
9.144 8.794 8.559 8.390
8.836 8.493 8.263 8.097
8.504 8.167 7.941 7.779
8.290 7.956 7.732 7.571
7.940 7.609 7.386 7.225
100 200
9.793 9.306 8.863
9.374 8.907 8.482
8.806 8.363 7.961
8.432 8.004 7.615
8.164 7.745 7.365
7.876 7.465 7.093
7.561 U57 6.790
7.355 6.953 6.588
7.009 6.606 6.236
00
·Multiply by 10 2•
664
2.579* 1.036* . -
25.06 18.90 15.678 13.727 12.431
TABLE B.2 (Continued) 5% Significance Level
n\m
10
12
p=6 15
20
25
30
35
10 12 14 16 18 20
45.722 28.959 22.321 18.858 16.755 15.351
44.677 28.l21 21.600 18.210 16.157 14.788
44.019 27.590 21.141 17.795 15.772 14.424
43.567 27.223 20.821 17.505 15.501 14.168
43.103 26.843 20.489 17.202 15.218 13.899
42.626 26.451 20.144 16.886 14.921 13.615
42.334 26.209 19.929 16.688 14.735 13.436
42.136 26.044 19.783 16.553 14.607 13.313
41.993 25.925 19.677 16.455 14.513 13.223
25 30 35 40
13.293 12.180 11.484 11.009
12.786 11.705 11.031 10.571
12.456 11.395 10.733 10.282
12.222 11.173 10.520 10.075
11.975 10.939 10.293 9.853
11.711 10.687 10.049 9.614
11.544 10.526 9.892 9.460
11.428 11.343 10.414 10.331 9.782 9.700 9.351 9.270
50 60 70 80
10.402 10.031 9.781 9.601
9.983 9.625 9.383 9.209
9.706 9.355 9.118 8.948
9.507 9.160 8.927 8.759
9.293 8.951 8.720 8.555
9.060 8.721 8.494 8.330
8.908 8.572 8.345 8.l82
8.801 8.465 8.239 8.076
8.721 8.385 8.l59 7.996
100 200 500 1000
9.360 8.910 8.659 8.579
8.976 8.542 8.300 8.223
8.720 8.295 8.059 7.983
8.534 8.115 7.882 7.808
8.333 7.919 7.689 7.616
8.110 7.701 7.473 7.400
7.963 7.555 7.328 7.255
7.857 7.449 7.222 7.149
7.777 7.369 7.140 7.067
00
8.500
8.146
7.908
7.734
7.543
7.328
7.183
7.077
6.994
6
TABLE B.2 (Continlled) 1% Significance Level
p-6 n\m
8
10
12
15
20
25
10 86.397 12 46.027 14 32.433 16 25.977 18 22.292 20 19.935
83.565 44.103 30.918 24.689 21.146 18.886
81.804 42.899 29.%6 23.875 20.418 18.217
80.602 42.073 29.309 23.311 19.913 17.752
79.376 41.227 28.634 22.729 19.389 17.267
78.124 40.359 27.936 22.126 18.844 16.761
77.360 39.826 27.507 21.753 18.505 16.445
76.845 39.466 27.215 21.498 18.273 16.229
76.474 39.206 27.004 21.314 18.105 16.071
25 16.642 30 14.944 35 13.913 40 13.223
15.737 14.118 13.138 12.482
15.156 13.586 12.635 12.000
14.749 13.211 12.281 11.659
14.324 12.816 11.906 11.298
13.875 12.398 11.506 10.911
13.592 12.133 11.252 10.663
13.397 11.949 11.074 10.490
13.254 11.814 10.943 10.361
50 60 70 80
12.358 11.839 11.493 11.246
11.661 11.169 10.841 10.607
11.206 10.730 10.413 10.187
10.882 10.417 10.107 9.886
10.538 10.083 9.779 9.563
10.167 9.721 9.424 9.212
9.927 9.486 9.192 8.983
9.759 9.320 9.028 8.819
9.633 9.196 8.905 8.697
100 200 500 1000
10.917 10.312 9.980 9.874
10.295 9.723 9.409 9.308
9.886 9.333 9.030 8.933
9.592 9.052 8.755 8.661
9.276 8.748 8.458 8.365
8.930 8.412 8.128 8.037
8.703 8.190 7.907 7.817
8.541 8.030 7.747 7.657
8.419 7.908 7.625 7.534
9.770
9.210
8.838
8.568
8.274
7.948
7.728
7.568
7.446
00
666
6
30
35
~ ~
TABLE B.2 (Conlin/lcd)
n\m
(;
10
5% Significance Level p=7 12 15 20 25
10 12 14 16 18 20
85.040 84.082 83.426 42.850 42.126 41.627 29.968 29.373 28.')61 24.038 23.519 23.158 20.692 20.222 19.893 18.561 18.125 17.819
25 30 35 40
15.587 14.049 13.113 12.485
82.755 41.113 28.534 22.781 19549 17.498
30
35
81.648 81.364 81.159 40.257 27.817 22.145 18.964 16.947
40.037 27.631 21.978 18.809 16.800
39.877 27.49:> 21.857 18696 16.694
15.202 14.930 14.642 14.337 14.143 14.009 13.911 13.693 13.440 13.172 12.884 12.701 12.573 12.478 12.776 12.535 12.278 12.002 11.825 11.700 11.608 12.160 11.927 11.679 11.411 11.237 11.115 11.025
50 11.695 . 60 11.219 70 10.901 80 10.674
11.386 11.165 10.927 10.921 10.706 10.475 10.610 10.400 10.173 10.388 10.181 9.957
100 10.371 200 9.812 500 9.504 1000 9.405
10.091 9.545 9.244 9.148
9.889 9.350 9.054 8.959
9.308
9.053
8.866
00
82.068 40.583 28.091 22.389 19.189 17.159
10.668 10.221 9.923 9.710
10.500 10.056 9.760 9.548
9.669 9.l38 8.846 8.753
9.426 8.902 8.613 8.521
9.265 8.744 8.456 8.365
9.150 8.629 8.342 8.250
9.062 8.542 8.254 8.162
8.661
8.431
8.275
8.160
8.072
10.381 10.292 9.938 9.850 9.643 9.555 9.432 9.344
667
TABLEB.2 (Continued) 1% Significance Level p~7
n\m
8
10
12
10 185.93 182.94 180.90 12 71.731 69.978 68.779 14 44.255 42.978 42.099 16 33.097 32.057 31.339 18 27.273 26.374 25.750 20 23.757 22.949 22.388 25 30 35 40
19.117 16.848 15.512 14.634
178.83 67.552 41.197 30.599 25.105 21.804
20
25
30
35
176.73 175.44 174.57 173.92 66.296 65.528 65.010 64.636 40.269 39.698 39.311 39.032 29.834 29.361 29.039 28.806 24.435 24.019 23.735 23.529 21.195 20.816 20.556 20.367
18.440 17.965 17.469 16.947 16.619 16.392 16.227 16.239 15.810 15.360 14.882 14.580 14.370 14.216 14.945 14.544 14.121 13.670 13.383 13.183 13.036 14.095 13.713 13.309 12.876 12.599 12.405 12.262
50 a553 13.049 12.691 60 12.914 12.432 12.088 70 12.492 12.024 11.690 80 12.193 11.736 11.408
668
15
12.310 1l.899 1l.634 11.720 11.323 1l.065 11.332 10.942 10.689 1l.056 10.673 10.422
100 200 500 1000
11.797 1l.077 10.685 10.561
11.353 11.034 10.658 10.356 10.230 9.987 9.869 10.l60
10.691 10.028 9.668 9.553
00
10.439
10.043
9.755
9.441
10.316 10.070 9.667 9.427 9.314 9.078 9.202 8.966 9.092
8.857
11.448 10.882 10.509 10.244
11.309 10.746 10.374 10.110
9.894 9.254 8.906 8.795
9.761 9.123 8.774 8.663
8.686
8.555
TABLE B.2 (Continued)
8
10
5% Significance Level p-= 8 15 20 25 12
14 16 18 20
42.516 31.894 26.421 23.127
41.737 31.242 25.847 22.605
41.198 40.641 40.066 39.711 30.788 30.318 29.829 29.525 25.446 25.028 24.591 24.319 22.239 21.856 21.454 21.201
25 30 35 40
18.770 16.626 15.356 14.518
18.324 18.009 16.221 15.934 14.977 14.707 14.156 13.898
50 60 70 80
13.482 12.866 12.459 12.169
13.142 12.540 12.142 11.858
100 200 500 1000
11.785 11.084 10.701 10.579
11.483 11.264 11.026 10.798 10.589 10.362 9.999 10.423 10.221 10.304 10.104 9.884
00
10.459
10.188
n\m
17.677 17.325 15.629 15.303 14.418 14.109 13.621 13.322
12.898 12.636 12.305 12.051 11.912 11.665 11.634 11.390
9.989
9.771
30
35
39.470 29.318 24.132 21.028
39.296 29.167 23.996 20.902
17.102 16.947 16.834 15.095 14.950 14.843 13.910 13.771 13.668 13.129 12.994 12.893
12.351 12.165 11.774 11.593 11.393 11.215 11.122 10.946
12.034 11.465 11.088 10.820
11.936 11.368 10.992 10.725
10.763 10.590 10.465 10.370 9'.939 10.108 9.816 9.722 9.751 9.584 9.461 9.367 9.470 9.348 9.254 9.637 9.526
9.360
9.238
9.144
669
TABLE B.2 (Continued)
n\m
10
35
59.019 39.753 30.882 25.924
58.639 39.456 30.629 25.697
64.035 62.828 61.592 43.633 42.707 41.754 34.146 33.373 32.573 28.808 28.129 27.425
60.323 40.771 31.745 26.691
25 23.001 30 19.867 35 18.077 40 16.924
22.212 21.661 21.085 19.173 18.686 18.173 17.440 16.991 16.516 16.324 15.900 15.451
20.480 20.100 19.838 19.647 17.631 17.288 17.051 16.876 16.011 15.690 15.466 15.301 14.970 14.662 14.447 14.288
100 200 500 1000 00
15.528 14.715 14.184 13.810
59.545 40.164 31.232 26.235
30
14 65.793 16 44.977 18 35.265 20 29.786
50 60 70 80
6711
8
1 % Significance Level p = 8 12 15 20 25
14.975 14.190 13.677 13.315
14.582 14.163 13.711 13.815 13.414 12.980 13.313 12.925 12.502 12.960 12.580 12.165
13.420 13.216 13.063 12.698 12.499 12.351 12.226 12.031 11.885 11.894 11.701 11.556
13.317 12.839 12.429 11.983 11.951 11.521 11.800 11.375
12.496 12.127 11.722 11.660 11.311 10.925 11.210 10.871 lO.495 11.067 10.732 lO.359
11.457 11.267 11.124 10.669 10.484 10.343 10.244 10.061 9.921 9.927 9.787 10.109
11.652
10.928
11.233
10.597
10.227
9.978
9.796
9.656
TABLE B.2 (Continlled) 5% Significance Level p -10
10
12
15
20
14 16 18 20
98.999 58.554 43.061 35.146
98.013 57.814 42.454 34.620
97.002 57.050 41.824 34.071
95.963 56.260 41.169 33.497
95.326 55.772 40.762 33.140
94.9 55.44 40.485 32.895
94.6 55.20 .40.284 32.716
25 30 35 40
26.080 22.140 19.955 18.569
25.660 21.773 19.618 18.252
25.219 21.384 19.260 17.914
24.753 20.970 18.876 17.550
24.458 20.706 18.630 17.316
24.255 20.523 18.458 17.151
24.107 20.388 18.331 17.029
50 16.913 60 15.960 70 15.341 80 14.907
16.622 15.684 15.074 14.647
16.309 15.385 14.786 14.365
15.969 15.059 14.469 14.055
15.748 14.847 14.261 13.851
15.592 14.695 14.113 13.705
15.476 14.582 14.002 13.595
100 14.338 SOO 12.774 1000 12.602
14.087 13.085 12.548 12.379
13.814 12.828 12.301 12.134
13.513 12.542 12.023 11.859
13.313 12.351 11.836 11.674
13.170 12.212 11.699 11.538
13.061 12.106 11.594 11.432
12.434
12.214
11.972
11.700
11.515
11.380 11.275
n\m
~
200 13.319
00
25
30
35
671
l
l'
TABLEB.2 (Continued)
n\m 14 16 18 20 25 30 35
12
180.90 178.28 175.62 172.91 89.068 87.414 85.270 83.980 59.564 58.328 57.055 55.742 45.963 44.951 43.905 42.821
171.24 82.91 54.933 42.150
30
35
170 82.2 81.7 54.384 53.990 41.693 41.362
31.774 26.115 23.116 21.267
31.029 25.489 22.556 20.749
30.253 24.832 21.966 20.201
29.440 24.139 21.338 19.615
28.932 23.701 20.939 19.241
28.583 23.399 20.663 18.980
28.328 23.177 20.459 18.787
50 60 70 80
19.114 17.901 17.124
18.148 16.992 16.252 15.738
17.611 16.484 15.762
17.266 16.154 15.443
16.583
18.646 17.462 16.703 16.175
15.260
14.948
17.023 15.922 15.216 14.726
16.842 15.748 15.046 14.559
100 200 500 1000
15.881 14.641 13.986 13.780
15.490 14.280 13.641 13.441
15.069 13.889 13.266 13.070
14.608 13.457 12.848 12.658
14.305 13.169 12.569 12.381
14.088 12.962 12.366 12.179
13.'125 12.803 12.210 12.023
00
13.581
13.246
12.881
12.472
12.198
11.997 11.842
40
672
10
1 % Significance Level p = 10 15 20 25
TABLE B.3 TABLES OF SIGNIFICANCE POll ITS FOR THE BARTLETT-NANDA-PILLAI TRACE TEST
n+mV~ Pr{ -m--
"
I~ 13 15 19 23 27
33
.05
43 63 83 123 243 00
I
.01
13 15 19 23 27
33
43 63 83 123 243 00
1
2
3
4
5
Xa
}= a
6
7
B
9
10
15
20
5.499 5.567 5.659 5.718 5.759 5.801 5.8015 5.891 5.914 5.938 5.962 5.991
4.250 4.310 4.396 4.453 4.495 4.539 4.586 4.635 4.661 4.688 4.715 4.7«
3.730 3.7B2 3.858 3.911 3.950 3.992 4.037 4. 0B6 4.112 4.139 4.168 4.197
3.430 3.476 3.546 3.595 3.632 3.672 3.716 3.764 3.790 3.81B 3 8016 3.877
3.229 3.271 3.336 3.383 3.418 3.456 3.499 3.547 3.573 3.601 3.630 3.661
3.082 3.122 3.183 3.22B 3.261 3.299 3.341 3.389 3.415 3.443 3.472 3.504
2.970 3.008 3.066 3.109 3.141 3.17B 3.219 3.266 3.293 3.321 3.351 3.384
2.881 2.917 2.972 3.013 3.045 3.081 3.122 3.169 3.195 3.223 3.254 3.287
2.808 2.8012 2.895 2.935 2.966 3.001 3.041 3.088 3.114 3.143 3.174 3.208
2.747 2.779 2.831 2.869 2. B99 2.934 2.974 3.020 3.046 3.075 3.106 3.141
2.545 2.572 2.616 2.650 2.677 2.709 2.746 2.791 2. B1B 2.8017 2.880 2.918
2.431 2.455 2.493 2.524 2.548 2.578 2.613 2.657 2.6B3 2.713 2.748 2.788
7.499 7.710 8.007 8.206 8.349 8.500 8.660 B.831 8.920 9.012 9.108 9.210
5.409 5.539 5.732 5.868 5.970 6.080 6.201 6.333 6.404 6.476 6.556 6.63B
4.810 4.671 4.824 4.935 5.019 5.111 5.214 5.329 5.392 5.459 5.529 5.604
4.094 4.180 4.312 4.409 4.483 4.566 40659 4.764 4.823 4. B85 4.951 5.023
3.780 3.857 3.976 4.064 4.131 4.207 4.294 4.393 4.449 4.508 4.572 4.642
3.555 3.625 3.734 3.815 3. B7B 3.950 4.032 4.127 4.1Bl 4.238 4.301 4.369
3.383 3.448 3.550 3.627 3.686 3.754 3.833 3.925 3.977 4.033 4.095 4.163
3.248 3.309 3.405 3.478 3.534 3.600 3.675 3.765 3.B15 3. B71 3.932 4.000
3.138 3.196 3.287 3.356 3.410 3.473 3.547 3.634 3.684 3.739 3.800 3.867
3.047 3.101. 3.188 3.255 3.307 3.368 3.439 3.525 3.574 3.62B 3.689 3. 757
2.751 2.795 2.867 2.923 2.968 3.021 3.085 3.163 3.210 3.263 3.323 3.393
2.587 2.625 2.686 2.735 2.775 2.823 2.881 2.955 3.000 3.052 3.113 3.185
15
20
p-3
"
.05
~ 14 16 20 24 2B 34 44 64 801 124 244 00
.01
14 16 20 24 2B 34 44 64 801 124 2« 00
1
2
3
4
5
6
7
B
9
10
6.129 6.168 6.209 6.251 6.296
5.019 4.6B4 5.082 4.738 5.177 4.822 5.245 4. BB3 5.295 4.929 5.351 4.980 5.412 5.037 5.480 5.101 5.517 5.137 5.556 5.174 5.597 5.214 5.640 5.257
4.458 4.507 4.5B3 4.639 4.682 4.730 4.7801 4.8016 4. B80 4.917 4.957 4.999
4.293 4.33B 4.409 4.461 4.501 4.547 4.599 4.660 4.693 4.730 4.769 4. B12
4.165 4.207 4.274 4.323 4.362 4.406 4.457 4.516 4.549 4.585 4.624 4.667
4.063 4.103 4.166 4.213 4.250 4.293 4.342 4.400 4.433 4.469 4.508 4.552
3.979 4.017 4.077 4.122 4.158 4.200 4.248 4.305 4.33B 4.374 4.413 4.457
3. 90S 3.944 4.002 4.046 4.081 4.1'1 4.169 4.225 4.257 4.293 4.333 4.377
3.672 3.702 3.751 3.790 3.821 3.857 3.901 3.955 3.986 4.022 4.063 4.110
3.537 3.563 3.606 3.640 3.668 3:702 3.743 3.795 3.826 3.862 3.904 3.954
6.855 7.006 7.236 7.403 7.528 7.667 7.821 7.994 8.088 8.188 8.294 8.406
5.970 6.083 6.258 6.387 6.486 6.598 6.724 6.867 6.947 7.032 7.124 7.222
5.457 5.551 5.698 5.808 5.893 5.990 6.101 6.230 6.301 6.379 6.463 6.554
5.112 5.195 5.326 5.424 5.501 5.588 5.690 5.809 5.876 5.948 6.028 6.116
4.862 4.937 5.056 5.146 5.217 5.298 5.393 5.505 5.569 5.639 5.716 5.801
4.669 4.738 4.8019 4.933 4.999 5.076 5.167 5.274 5.335 5.403 5.478 5.562
4.516 4.581 4.6B4 4.764 4,827
4.390 4.451 4.549 4.625 4.6B5 4.756 4.839 4.939 4.998 5.063 5.136 5.218
4.285 4.343 4.436 4.509 4.567 4.635 4.715 4.813 4.871 •. 935 5.007 5.089
3.939 3.986 4.063 4.124 4.174 4.233 4.305 4.39. 4.448 •. 510 4.581 4.664
3.743 3.783 3.850 3.903 3.948 4.001 4.067 4.151 4.203 4.263
6.989 7.095 7.243 7.341 7.410 7.482 7.559 7.639 7.681 7.724 7.768 7 815
5.595 5.673 5.787 5.866 5.925 5.987
8.971 9.245 9.639 9.910 10.106 10.317 10.545 10.790 10.920 11.056 11. 196 11.345
6.055
4.900 4.986 5.090 5.150 5.216 5.290 5.372
4.334 4.419
673
TABLE B.3 (Continued) ---~--~--
--------------.
-.-~-----.----
p-4
.
,ml
I~
In I
I
15
20
5.041 5.080 5.143 5.191 5.230 5.275 5.329 5.395 5.432 f.566 5.475 5.613 5.522 5.£67 5.576
4.779 4.8\1 4.864 4.906 4.939 4.980 5.029 5.090 5.127 5.168 5.216 5.')]2
4.6')]
5.479 5.539 5.638 5.716 5.779 5.853 5.942 6.117 6.190 6.273 6.369
5.095 5.144 5.225 5.290 5.344 5.408 5.487 5.586 5.646 5.715 5.796 5.892
4.874 4.916 4.987 5.044 5.091 5.149 5.221 5.314 5.371 5.439 5.519 5.616
10
15
20
10
"'-
I
05 1
6.859 6.952 7.091 7.190 7.263 7.343 7.431 7.528 7.580 7.635 7.693 7.754
6.245 6.318 6.429 6.510 6.571 6.640 6.716 6.802 6.849 6.899 6.952 7.009
5.885 5.642 5.941 5.696 6.043 5.782 6.114 5.846 6.168 5.896 6.229 5.952 6.298 6.017. 6.378 6.092 6.42' 6.134 6.46V 6.179 6.519 6.228 6.574 6.282
5.462 5.512 5.591 5.650 5.696 5.749 5.811 5.883 5.923 5.968 6.016 6.069
5.323 5.369 '5.443 5.498 5.542 5.593 5.652 5.721 5.761 5.804 5.852
5.; 12 5.; 15 5. ,24 5.377 5.418 5.467 5.524 5.592 5.631 5.674 5.721 5.905 5.774
5.1\9 5.160 5.225 5.276 5.316 5.363 5.418 5.485 5.523
10.293 8.188 10.619 8.360 11.095 8.625 11.428 8.818 1\.672 8.966 11.938 9.131 12.228 9.318 12.545 9.529 12.715 9.645 12.893 9.769 13.080 9.902 13.277 10.045
7.276 7.401 7.598 7.744 7.858 7.987 8.135 8.306 8.402 8. S05 8.617 8.739
6.737 6.840 7.003 7.126 7.222 7.332 7.460 7.610 7.695 7.787 7.889 8.000
6.105 6.184 6.313 6.411 6.490 6.581 6.688 6.816 6.890 6.971 7.062 7.163
5.898 5.971 6.089 6.180 6.253 6.338 6.439 6.561 6.632 6.710 6.798 6.897
5.594 5.658 5.762 5.844 5.909 5.986 6.079 6.192 6.258 6.333 6.417 6.513
16 18 22 26 30 36 46 66 86 126 246
9.589 9.761 10.007 10.176 10.298 10.429 10.571 10.714 10.805 10.890 10.978 11.071
8.071 8.179 8.340 8.457 8.5« 8.641 8.748 8.868 8.933 9.002 9.076 9.154
7.430 7.512 7.639 7.732 7.803 7.883 7.974 8.077 8.134 8.195 8.261 8.332
7.052 6.795 6.605 6.457 6.338 6.239 6.155 5.873 5.706 7.120 7.228 7.308 7.370 7.440 7.521 7.615 7.667 7.724 7.786 7.853
6.854 6.949 7.021 7.077
16 18 22 26 30 36 46 66 86 126 246
11.534 1\.902 12.449 12.837 13.125 13.442 13.790 14.176 14.385 14.606 14.839 15.086
9.451 8.521 7.966 9.642 8.658 8.077 9.939 8.876 8.255 10.159 9.040 8.390 10.328 9.168 8.497 10.518 9.314 8.621 10.735 9.483 8.765 10.984 9.681 8.936 1\.\22 9.793 9.034 11.270 9.914 9.142 1\.431 10.047 9.260 1\.605 10.193 9.392
15 17 21 25 29 35 45 65 85 125 245
8.331
8.412
I
8.671
8.805
8.901 9.004 9.113 9.229 9.291 9.354 9.419 9.488
15 17 21 25 29 35 45 65 85 125 245
.01
6.373 6.462 6.604 6.712 6.798 6.897 7.012 7.149 7.227 7.313 7.408 7.513
5.731 5.799 5.909 5.995 6.064 6. 145 6.241 6.358 6.426 6.503 6.588 6.686
6.052
4.65-4
4.701 4.738 4.768
4.805 4.851 4.910 4.945 4.987 5.IYl5 5.094
p-5 a
I~\ I
.05
I
J
I
I
I
I
1/i7~
7.216 7.303 7.353 7.407 7.466 7.531
6.659 6.745 6.810 6.862 6.922 6.992 7.075 7.122 7.174 7.232 7.296
6.384 6.458 6.516 6.562 6.616 6.680 6.757 6·802 6.851 7.052 6.907 7.115 6.970
6.282 6.352 6.407 6.451 6.503 6.565 6.640 6.684 6.733 6.788 6.851
7.587 7.682 7.835 7.954 8.048 8.158 8.287 8.442 8.531 8.630 8.740 8.863
7.306 7.391 7.528 7.635 7.720 7.820 7.939 8.083 8.167 8.260 8.364 8.482
7.088 7.165 7.291 7.389 7.468 7.561 7.673 7.808 7.888 7.977 8.077 8.192
6.767 6.833 6.943 7.030 7.100 7.183 7.284 7.409 7.483 7.567 7.663 7.773
7.141
6.506 6.586 6.647 6.696 6.752 6.819 6.899 6.944 6.995
6.912 6.983 7.100 7.192 7.266 7.354 7.460 7.589 7.666 7.752 7.849 7.961
&.196 6.263 6.316 6.358 6.408
5.906 5.961 6.006 6.042 6.086 6.140 6.208 6.248 6.295 6.350 6.414
5.735 5.784 5.823 5.856 5.896 5.945 6.009 6.049 6.095 6. ISO 6.217
6.230 6.281 6.366 6.434 6.491 6.559 6.644 6.752 6.818 6.895 7.506 6.985 7.615 7.093
5.989 6.033 6.106 6.166 6.216 6.277 6.355 6.455 6.518 6.592 6.681 6.790
6. .0168 6.541 6.584 6.633 6.688 6.7SO 6.6« 6.707 6.810 6.893 6.961} 7.040 7.137 7.258 7.330 7.412
•'" 'J
TABLE B.3 (Continued)
al~1 .05
10
15
20
17 19 23 27 31 37 47 67 87 127 247
10.794 10.993 11.282 11.483 11.630 11.790 11.964 1'-.154 12.255 12.362 12.474 12.592
9.247 9.367 9.550 9.684 9.784 9.897 10.024 10.166 .10.245 10.328 10.417 10.513
8.193 8.268 8.386 8.475 8.545 8.624 8.716 8.824 8.885 8.951 9.024 9.104
7.926 7.990 8.093 8m 8.234 8.306 8.390 8.490 8.547 8.609 8.678 8.755
7.728 7.785 7.878 7.950 8.007 8.073 8.151 8.245 8.299 8.359 8.425 8.500
7.573 7.625 7.711 7.m 7.830 7.892 7.966 8.056 8.108 8.165 8.230 8.303
7.448 7.49" 7:576 7.638 7.688 7.747 7.817 7.903 7.954 8.010 8.074 8.146
7.344 7.389 7.464 7.523 7.570 7.627 7.694 7.n8 7.827 7.882 7.945 8.017
7.256 7.299 7.369 7.425 7.471 7.525 7.590 7.m 7.720 7.n4 7.836 7.908
6.956 6.990 7.048 7.095 7.134 7.181 7.239 7.312 7.357 7.409 7.470 7.543
6.n8 6.808 6.858 6.899 6.934 6.976 7.029 7.099 7.142 7.193 7.254 7.328
17 19 23 27 31 37 47 67 87 127 247
12.722 13.126 13.736 14.173 14.501 14.865 15.270 15.723 15.970 16.233 16.513 16.812
10.664 9.721 9.157 10.874 9.873 9.277 11.202 10.111 9.469 11.446 10.292 9.617 11.635 10.433 9.734 11.850 10.596 9.871 12.097 10.787 10.032 12.382 11.011 10.224 12.542 11.138 10.335 12.715 11.27810.457 12.903 11.432 10.593 13.108 11.60210.745
8.767 8.869 9.034 9.162 9.264 9.384 9.527 9.700 9.800 9.912 10.037 10.178
8.478 8.567 8.714 8.828 8.921 9.030 9.160 9.319 9.413 9.517 9.635 9. no
8.252 8.332 8.465 8.570 8.655 8.756 8.878 9.027 9.115 9.215 9.328 9.458
8.069 8.143 8.266 8.363 8.442 8.537 8.652 8.794 8.878 8.974 9.084 9.210
7.917 7.986 8.100 8.192 8.267 8.356 8.466 8.602 8.683 8.n6 8.883 9.008
7.788 7.853 7.961 8.048 8.119 8.204 8.309 8.440 8.520 8.610 8.715 8.838
7.351 7.403 7.490 7.561 7.621 7.693 7.783 7.899 7.971 8.055 8.154 8.274
7.093 7.137 7.213 7.275 7.328 7.392 7.474 7.581 7.649 7.729 7.827 7.948
10
15
20
18 20 24 28 32 38 48 68 88 128 248
11.961 12.184 12.513 12.744 12.915 13.102 13.308 13.534 13.657 lZ.786 13.923 14.067
10.396 10.528 10.731 10.880 10.994 11.123 11.267 11.433 11.524 11.623 11.728 11.842
9.719 9.817 9.972 10.088 10.178 10.281 10.399 10.537 10.614 10.698 10.790 10.890
9.040 9.109 9.m 9.306 9.374 9.453 9.547 9.658 9.722 9.793 9.872 9.960
8.835 8.896 8.996 9.073 9.135 9.208 9.294 9.398 9.459 9.526 9.602 9.687
8.675 8.730 8.821 8.892 8.950 9.017 9.098 9.197 9.255 9.320 9.394 9.4n
8.545 8.596 8.680 8.746 8.800 8.864 8.941 9.036 9.092 9.155 9.226 9.309
8.437 8.484 8.563 8.626 8.676 8.737 8.811 8.902 8.956 9.018 9.089 9.170
8.345 8.390 8.464 8.523 8.572 8.630 8.701 8.789 8.842 8.903 8.m 9.053
8.031 8.0S7 8.127 8.176 8.2)(, 8.266 8.328 8.407 8.456 8.513 8.580 8.661
7.843 7.874 7.926 7.969 8.001. 8.049 8.106 8.180 8.226 8.;:81 8.348 8.431
13.~74
11.841 12.069 12.426 12.694 12.902 13.141 13.416 13.737 13.919 14.116 14.333 14.571
10.&95 10.321 9.923 9.627 11. 056 10.448 10.031 9.721 11.314 10.655 10.207 9.876 11.51010.81510.344 9.999 11. 665 10.943 10.455 10.098 11.845 11 ..092 10.586 10.215 12.056 11.269 10.742 10.357 12.306 11.482 10.932 10.532 12.449 11. 605 11. 043 10.635 12.607 11. 743 11. 168 10.751 12.78211.89711.308 10.883 12.m 12.070 lU68ll.034
9.049 8.915 9.121 8.982 9.240 9.095 9.337 9.186 9.416 9.261 9.511 9.351 9.628 9.463 9.n6 9.605 9.865 9.691 9.967 9.790 10.085 9.906 10.22310.043
8.460 8.512 8.602 8.676 8.738 8.814 8.909 9.033 9.110 9.201 9.309 9.441
8.188 8.233 8.310 8.374 8.429 8.496 8.582 8.696 8.768 8.854 8.960 9.092
.
.01
.
.05
..
.01
18 20 24 28 32 38 48 68 88 128 248
.
14.310 14.974 15.456 15.822 16.230 16.688 17.206 17.491 17.796 18.124 18.475
8.5a5 8.676 8.817 8.922 9.003 9.094 9.199 9.319 9.387 9.45.9 9.538 9.623
9.316 9.396 9.525 9.622 9.699 9.787 9.890 10.012 10.082 10.158 10.242 10.334
9.395 9.206 9.479 9.283 9.619 9.412 9.731 9.515 9.821 9.599 9.930 9.700 10.061 9.824 10.224 9.978 10.321 Ill. 070 10.431 10.176 10.557 10.298 10.70310.439
----
675
TABLE BJ
.~ .05
19 21 25 29 33 39 49 69
89
129 249
..
.01
19 21 25 29 33 39 49 69 £9 129 249
..
.~ .05
21 23 27 31 35 41 51 71 91 131
..
251
.01
21 23 27 31 35 41 51 71 91 131 251
..
676
1
2
3
4
(Continued)
5
6
7
8
9
13.101 13.346 13.710 13.970 14.163 14.3n 14.614 14.m 15.021 15.173 15.335 15.507
11.524 11.667 11.889 12.054 12.180 12.323 12.487 12.674 12.779 12.892 13.015 13.148
10.835 10.423 10.141 9.930 9.766 9.632 9.521 10.941 10.509 10.214 9.995 9.824 9.685 9.570 11.109 10.647 10.333 10.101 9.920 9.n4 9.652 11. 235 10.753 10.425 10.184 9.996 9.844 9.718 11.334 10.837 10.499 10.250 10.057 9.902 9.m 11.448 10.934 10.585 10.329 10.130 9.970 9.837 11.580 11.048 10.688 10.423 10.218 10.053 9.917 11.73411.18310.81110.53810.32610.15610.016 11.822 11.261 10.883 10.605 10.390 10.218 10.075 11.918 11.347 10.962 10.68D 10.462 10.288 10.143 12.023 11.442 11.051 10.765 10.5« 10.367 10.221 12.138 11.549 11.152 10.862 10.638 10.459 10.312
14.999 15.463 16.m 16.700 17.100 17.549 18.058 18.640 18.962 19.310 19.684 20.090
12.m 13.235 13.620 13.910 14.137 14.398 14.702 15.058 15.261 15.484 15.729 16.000
12.043 11.463 11. 06D 10.758 12.215 11.598 11.174 10.857 12.491 11.81911.360 11.021 12.703 11.991 11.50711.151 12.871 12.12811.62611.256 13.06712.290 11.766 11.383 13.297 12.482 11.935 11.536 13.57312.71612.143 11.725 13.733 12.853 12.265 11.838 13.9D9 13.005 12.403 11.965 14.106 13.177 12.559 12.112 14.32713.371 12.738 12.280
1
2
3
4
5
6
10.521 10.328 10.610 10.409 10.757 10.543 10.87410.651 10.970 10.740 11.086 10.848 11.227 10.980 11.403 11.146 11.509 11.247 11. 63D 11. 362 11.769 11.495 11.930 11.652
)
8
10
15
20
9.426 9.472 9.550 9.612 9.663 9.725 9.801 9.897 9.955 10.021 10.098 10.188
9.100 9.136 9.198 9.249 9.292 9.344 9.409 9.494 9.547 9.609 9.682 9.m
8.904 8.935 8.988 9.032 9.070 9.116 9.176 9.254 9.304 9.363 9.436 9.526
10.167 10.030 10.241 10.099 10.366 10.216 10.46710.310 10.550 10..389 10.651 10.485 10.n6 10.604 10.934 10.756 11.030 10.849 11.142 10.956 11.271 11.083 11.424 11.233
9.558 9.612 9.704 9.780 9.844 9.924 10.024 10.155 10.238 10.335 10.453 10.597
9.275 9.321 9.399 9.465 9.521 9.591 9.681 9.801 9.m 9.970 10.083 10.227
15
20
9
10
15.322 15.604 16.033 16.344 16.580 16.843 17.140 17.476 17.662 17.861 18.076 18.307
13.733 13.897 14.154 14.347 14.497 14.669 14.868 15.100 15.231 15.375 15.532 15.705
13.027 12.603 12.311 12.093 11.922 11.782 11.666 11.566 11.222 11.013 13.14712.700 12.39312.16411.98511.840 11.71911.616 11.260 11.045 13.340 12.85712.52612.28212.09111.93711.80811.69911.32511.100 13.487 12.978 12.631 12.375 12.176 12.015 11- 881 11. 768 11.38D 11. 146 13.603 13.075 12.716 12.451 12.245 12.079 11.941 11.821 11.426 11.186 13.73713.188 12.816 12.541 12.328 12.156 12.014 11.893 11.482 11.236 13.89513.32312.936 12.651 12.430 12.251 12.104 11.979 11.555 11.301 14.083 13.486 13.083 12.786 12.556 12.37112.21812.08911.650 11.388 14.19113.58113.16912.866 12.632 12.443 12.288 12.156 11.710 11.443 14.31013.68713.266 12.957 12.718 12.526 12.36812.234 11.78111.511 14.44313.806 13.376 13.061 12.81812.622 12.461 12.325 11.867 11.594 14.591 13.940 13.501 13.180 12.933 12.735 12.572 12.434 11. 972 11. 700
17.197 17.707 18.505 19.101 19.562 20.088 20.692 21.394 21.790 22.221 22.692 23.2D9
15.234 15.507 15.941 16.273 16.535 16.839 17.196 17.623 17.868 18.141 18.444 18.783
14.284 13.698 14.476 13.849 14.788 14.096 15.029 14.290 15.222 14.447 15.448 14.632 15.718 14.855 16.04515.130 16.236 15.292 16.450 15.475 16.690 15.683 16.964 15.923
1'3.288 12.980 12.736 12.537 12.371 12.22811.733 11.432 13.413 13.088 12.832 12.624 12.449 12.301 11.789 11.478 13.621 13.268 12.993 12.769 12.584 12.426 11.885 11.559 13.786 13.41313.123 12.888 12.693 12.528 11.965 11.628 13.920 13.53113.230 12.986 12.785 12.614 12.034 ]1.687 14.080 13.674 13.359 13.106 12.897 12.720 12.119 11.761 14.274 13.848 13.519 13.255 13.037 12.852 12.229 11. 858 14.51614.06713.72213.44513.21613.02412.37411.989 14.66D 14.199 13.845 13.561 13.327 13.130 12.466 12.074 14.824 14.350 13.986 13.695 13.456 13.254 12.sn 12. m 15.012 14.525 14.152 13.853 13.608 13.402 12.712 12.307 15.231 14.730 14.346 14.041 13.791 13.581 12.881 12.472
TABLE B.4 TABLES OF SIGNIFICANCE POINTS FOR THE ROY MAXIMUM ROOT TEST
m+n
Pr { --m- R ~
XQ
}
=
a
p=2 a
.05
~~~ 13 15 19 23 27 33 .3 63 83 123 2.3 QQ
.01
13 15 19 23 27 33 .3 63 83 123 2.3 QQ
I
2
5.499 5.56, 5.659 5.718 5.759 5.801 5.8.5 5.891 5.91. 5.938 5.962 5.991
3.736 3.807 3.905 3.971 4.018
7.499 7.710 8.007 8.206 8.3.9 8.500 8.660 8.831 8.920 9.012 9.108 9.210
3
4
7
8
9
10
15
20
1. 823 1.869 1. 940 1.993 2.033 2.079 2.131 2.190 2.223 2.259 2 298 2.3'0
2.157 2.211 2.293 2.352 2.396 2.445 2.498 2.558 2.591 2.626 2 663 2.702
2.018 2.069 2.148 2.204 2.2.7 2.294 2.347 2.407 2.440 2.476 2.513 2.55.
1.910 1.959 2.033 2.087 2.129 2.175 2.228 2.288 2.321 2.356 2.395 2.436
1. 752 1.796 1. 86' 1. 915 1.95. 1.998 2.049 2.109 2.1'2 2.178 2.217 2.261
1. 527 1.562 1618 1. 661 1696 1.736 1.783 1.840 1. 873 1909 1 951 1.998
1407 1436 1. .8. 1521 1.552 1.588 1.631 1. 686 1.718 1.755 1.797 18.7
2.681 2.782 2.936 3.048 3.133 3.228 3.33. 3.45. 3.520 3.591 '.100 3.666 4.182 3.7.7
2 .• 32 2.523 2.66' 2.768 2.847 2.936 3.037 3.153 3.217 3.285 3.360 3.440
2.2'9 2.333 2.463 2.559 2.634 2.718 2.815 2.926 2.989 3.056 3.130 3209
2.109 1.997 1.907 2.186 2.069 1. 973 2.307 2.182 2.080 2.397 2.268 2.161 2.468 2.335 2.225 2·548 2.• 12 2.299 2.6.1 2.501 2.386 2.7.9 2.607 2.488 2.810 2.666 2.5.7 2.877 2.732 2.612 2.950 2.80. 2.683 3.029 2.884 2.763
1. 625 1676 1. 758 1. 823 1876 1. 938 2.013 2.105 2.160 2.222 2.292 2.373
1.• 78 1.519 1.587 1. 6.1 1.686 17.0 1. 807 1.891 1.943 2.002 2.072 2.15.
'.265 •. 297
2.605 2.668 2.759 2.822 2.868 2.918 2.973 3.032 3.064 3.097 3.132 3.169
•. 675 •. 834 5.06. 5.223 5.339 5.•65 5.600 5.747 5.825 5.906 5.991 6.080
3.610 3.7.2 3.937 •. 07. '.176 '.287 '.409 •. 543 4.616 4.692 4.772 •. 856
3.0.0 3.154 3.325 3.448 3.540 3.642 3.755 3.881 3.950 4.022
•. 120 '.176 4.205 4.235
6
2.342 2.• 01 2.• 87 2.548 2.593 2.643 2.697 2.757 2.789 2.823 2.859 2.897
3.011 3.078 3.173 3.239 3.286 3.336 3.391 3.449 3 .•80 3.512 3.545 3.580
•. 068
5
p=3
I. 16 20 2.
28 .05
34
••
6. 8. 12. 2.4 QQ
.01
1. 16 20 24 28 34 44 &.4 84 124 244 QQ
6.989 7.095 7.243 7.341 7.• 10 7.482 7.559 7.639 7.681 7.724 7.768 7.815
'.517 •. 617 4.760
8.971 9.245 9.639 9.910 10.106 10.317 10.545 10.790 10.920
5.416 5.613 5.905 6.111 6.2&.4 6.431 6.6U 6.815 6.923 7.037 7.157 7.284
11.056
11. 196 11.346
•. 858 •. 929 5.00. 5.086 5.173 5.220 5.268 5.318 5.370
3.544 3.63. 3.767 3.859 3.927 ..001 •. 081 '.169 •. 216 •. 265 4.317 4.371
3.010 3.092 3.215 3.302 3.367 3.• 39 3.517 3.604 3.651 3.701 3.75. 3.810
2.669 2.745 2.859 2.942 3.00. 3.073 3.1.9 3.235 3.282 3.332 3.386 3.443
2.• 30 2.501 2.608 2.686 2.7.6 2.812 2.887 2.972 3.019
2.25. 2.319 2.•20 2.495 2.552 2.616 2.689 2.773 2.820 3.069 2.870 3.123 2.92. 3.181 2.983
2.117 2.178 2.27. 2.345 2.• 00 2 .• 62 2.53. 2· 616 2.663 2.713 2.768 2.828
2.008 2.065 2.156 2.224 2.277 2.338 2.• 08 2.489 2.535 2.586 2.6.1 2.701
10
15
20
1.919 1. 973 2.059 2.12' 2.176 2.234 2.303 2.383 2.• 29 2 .•79 2.535 2.596
1.639 1.682 1. 751 1.805 1.849 1. ?01 1.962 2.038 2.082 2.132 2.189 2.253
l.m 1.526 1.585 1.631 1.669 1.715 1.771 1. 8.2 1.885 1.934 1. 991 2.059
•. 106 ;, .• 12 2.978 2.680 2.462 2.295 2.163 2.055 1.724 1. 552 •. 265 •. 507 4.681 4.811 •. 955 5.116 5.296 5.393 5.•9 ' 5.601 5.72.;
3.098 3.284 3.• 22 3.528 3.647 3.784 3.9.0 •. 027 '.658 '.120 •. 763 '.221 •. 875 4.331
3.548 3.757 3.910 4.026 '.156 4.303 •. 469 •. 560
2.787 2.956 3.082 3.180 3.292 3 .• 20 3.568 3.652 3.742 3.841 3.948
2.559 2.1U 2.831 2.922 3.027 3.148 3.291 3.372 3.• 60 3.556 3.663
2.384 2.527 2.636 2.722 2.821 2.938 3.075 3.153 3.239 3.334 3.440
2.245 2.378 2.481 1.562 2.657 2.768 2.901 2.977
2.132 2.257 2.354 2.431 2.521 2.628 2.757 2.832 3.062 2.915 3.155 3.008 3.260 3.112
1.782 1. 877 1. 954 2.016 2.091 2.182 2.295 2.363 2.441 2.530 2.634
1.598 1.676 1.7.0 1.792 1.857
1.937 2.040 2.103 .2·177 2.2&.4 2.369
677
678
TABLE BA
a
I'~
.05
17 19 23 27 31 37 47 67 87 127 247
.01
17 19 23 27 31 37 47 67 87 127 247
DO
DO
a
.05
I~ 18 20 24 28 32 38 48 68 88 128 248
14.865 15.270 15.723 15.970 16.233 16.513 16.812
1 11.961 12.184 12.513 12.744 12.915 13.102 13.308 13.534 13.657 13.786 13.923
14.067 13.874 14.310 14.974
32 38 48 68 88 128 248
15.822 16.230 16.688 17.206 17.491 17. /96 18.124 18.475
15.~
4
5
6
7
8
9
10
15
20
4.005 3.468 3.098 2.827 2.620 2.455 2.322 1.908 1. 693 4.128 3.579 3.199 2.920 2.705 2.535 2.396 1. 965 1.740 4.320 3. 753 3.359 3.068 2.844 2.665 2.519 2.062 1.819
7.188 7.334 7.495 7.675 7.774 7.878 7.989 8.107
3.884 3.985 4.101 4.235 4.390 4.477 4.573 4.676 5.405 4.790
3.481 3.576 3.686 3.813 3.963 4.048 4.141 4.244 4.357
3.182 3.272 3.376 3.498 3.643 3.727 3.818 3.920 4.033
2.951 3.036 3.136 3.253 3.394 3.476 3.566 3.667 3.780
2.767 2.615 2.139 1.884 2.848 2.693 2.203 1. 939 2.943 2.785 2.280 2.006 3.057 2.894 2.375 2.090 3.194 3.028 2.495 2.200 3.274 3.107 2.568 2.268 3.363 3.195 2.652 2.348 3.463 3.294 2.749 2.444 3.576 3.408 2.864 2.561
7.296 7.570 7.992 8.303 8.541 8.808 9.112 9.458 9.650 9.856 10.079 10.319
5.360 5.574 5.912 6.164 6.360 6.583 6.839 7.136 7.303 7.484 7.682 7.897
4.352 4.531 4.817 5.034 5.204 5.400 5.628 5.895 6.047 6.213 6.395 6.596
3.306 3.444 3.667 3.841 3.980 4.142 4.334 4.565 4.699 4.847 5.012 5.198
2.998 3.122 3.325 3.484 3.612 3.763 3.943 4.161 4.289 4.431 4.591 4.772
2.764 2.877 3.063 3.210 3.329 3.470 3.640 3.848 3.97.1 4.108 4.264 4.442
2.580 2.683 2.855 2.993 3.104 3.237 3.399 3.598 3.716 3.850 4.002 4.177
12.722 13.126 13.736 14.173
14.501
3 4.861 5.001 5.216 5.372 5.491 5.625 5.774 5.944 6.038 6.138 6.246 6.362
6.470 6.634 6.880
DO
DO
2
10.794 10.993 11. 282 11.483 11.630 11. 790 11.964 12.154 12.255 12.362 12.474 12.592
18 20 24
28
.01
1
(Colltillued)
7.056
2
3
4.462 4.571 4.695 4.836 4.997 5.088 5.185 5.291
4 4.304 4.437 4.647 4.803 4.925 5.063 5.222 5.407 5.511 5.624 5.747 5.882
3.730 3.885 4.135 4.328 4.480 4.657 4.864 5.111 5.252 5.408 5.580 5.772
5
6
7.063 7.243 7.516 7.714 7.863 8.030 8.216 8.426 8.541 .8.665 8.797 8.938
5.258 5.411 5.647 5.821 5.954 6.104 6.275 6.471 6.579 6.697 6.824 6.961
7.872 8.164 8.619 8.957 9.218 9.514 9.852 10.243 10.461 10.697 10.954 11.233
5.744 4.640 3.960 3.498 5.971 4.829 4.124 3.642 6.332 5.133 4.389 3.879 6.605 5.367 4.595 4.065 6.818 5.551 4.759 4.214 7.063 5.765 4.952 4.390 7.346 6.015 5.179 4.600 7.679 6.313 5.452 4.855 7.867 6.483 5.610 5.003 8.073 6.670 5.785 5.169 8.298 6.878 5.980 5.356 8.546 7.108 6.200 5.567
3.708 3.827 4.016 4.160 4.272 4.401 4.551 4.727 4.827 4.937 5.058 5.191
7
3.298 2.999 3.406 3.098 3.580 3.258 3.713 3.382 3.817 3.481 3.939 3.596 4.081 3.732 4.251 3.895 4.348 3.990 4.455 4.095 4.573 4.211 4.705 4.343 3.163 3.292 3.507 3.676 3.814 3.977 4.173 4.413 4.554 4.713 4.894 5.099
8
9
2.431 2.526 2.687 2.816 2.921 3.047 3.201 3.393
3.507
3.637 3.786 3.959
1.974 2.044 2.165 2: 264 2.347 2.448 2.576 2.739 2.840 2.958 3.097 3.264
1.739 1.795 1.892 1. 974 2.043 2.128 2.238 2.383 2.475 2.584 2.717 2.882
20
10
15
2.589 2.674 2.814 2.924 3.012 3.117 3.243 3.396 3.486 3.588 ~. 702 3.833
2.442 2.522 2.653 2.757 2.842 2.942 3.063 3.213 3.301 3.401 3.514 3.645
1.989 2.049 '2. 151 2.235 2.304 2.388 2.492. 2.625
2.707 2.816 2.997 3.143 3.506 3.262 3.659 3.406 3.843 3.581 4.072 3.799 4.207 3.929 4.360 4.078 4.535 4.248 4.736 4.446
2.545 2.646 2.814 2.951 3.063 3.199 3.366 3.575 3.701 3.846 4.012 4.207
2.050 1.797 2.124 1.855 2.250 I. 956 2.355 2.042 2.443 2.115 2.551 2.206 2.688 2.324 2.866 2.481 2.976 2.581 3.106 2.700 ·3.260 2.847 3.447 3.030
2.770 2.861 3.011 3.127 3.220 3.330 3.461 3.619 3.711 3.814 3.929 4.060 2.908 3.026 3.222 3.379
1.753 1.802 1. 887 1.956 2.015 2.088 2.180 2.300 2.706 2.376 2.800 2.465 2.910 2.573 3.041 2.705
679
TABLE B.4 . (Continlle~J) p-8 10 19 21 25 29 33 39 49 69 69 129 249
.OS
I
.01
00
19 21 25 29 33 39 49 69 69 129 249 00
13.101 13.346 13.710 13.970 14.163 14.377 14.614 14.877 15.021 15.173 15.335 15.507
7.640 7.834 6.132 6.350 6.515 6.701 6.912 9.151 9.263 9.426 9.579 9.745
5.645 5.808 6.063 6.253 6.399 6.566 6.757 6.977 7.101 7.235 7.361 7.541
4.594 4.737 4.962 5.131 5.264 5.416 5.593 5.600 5.917 6.046 6.167 6.342
14.999 15.463 16.177 16.700 17.100 17.549 16.056 16.640 16.962 19.310 19.684 20.090
6.435 6.743 9.226 9.589 9.671 10.194 10.565 10.996 II. 242 11.508 11.796 12.117
6.119 6.357 6.739 7.030 7.259 7.524 7.633 6.199 6.408 6.638 6.691 9.173
4.921 5.119 5.439 5.687 5.885 6.115 6.367 6.713 6.901 7.109 7.341 7.601
3.941 4.067 4.270 4.425 4.547 4.686 4.854
.OS
00
.01
21 23 27 31 35 41 51 71
2.916 3.012 3.171 3.295 3.396 3.515 3.656 3.632 3.935 4.050 4.180 4.329
2.719 2.808 2.956 3.074 3.169 3.263 3.420 3.589 3.689 3.602 3.931 4.076
2.559 2.643 2.762 2.693 2.984 3.092 3.224 3.386
5.163 5.267 5.424 5.577
3.166 3.270 3.441 3.574 3.660 3.806 3.955 4.136 4.241 4.358 4.491 4.640
2.067 2.130 2.236 2.326 2.0100 2.490 2.603 2.747 3.~ 2.636 3.597 2.940 3.725 3.063 3.872 3.210
1.812 1.863 1.952 2.025 2.068 2.165 2.264 2.395 2.478 2.576 2.695 2.643
4.185 4.355 4.634 4.653 5.026 5.234 5.460 5.778 5.952 6.146 6.364 6.610
3.685 3.636 4.084 4.260 4.439 4.627 4.654 5.131 5.294 5.476 5.685 5.922
3.323 3.458 3.662 3.861
3.048 3.171 3.376 3.541 3.676 3.639 4.037 4.284 4.432 4.601 4.794 5.019
2.632 2.944 3.134 3.267 3.414 3.566 3.754 3.990 4.132 4.295 4.463 4.703
2.656 2.762 2.936 3.081 3.200 3.344 3.523 3.749 3.886 4.044 4.226 4.445
1.653 1.913 2.016 2.107 2.184 2.280
5. OSO
4.001
4.161 4.392 4.653 4808 4.963 5.163 5.413
9~
131 251 00
6811
I
15.322 15.604 16.033 16.344 16.580 16.843 17.140 17.476 17.662 17.861 16.076 16.307
6.761 6.979 9.320 9.573 9.769 9.992 10.247 10.543 10.709 10.690 11.087 n.303
6.395 6.577 6.864 7.082 7.252 7.446 7.676 7.944 8.097 8.265 6.450 8.654
5.158 5.315 5.566 5.759 5.912 6.090 6.299 6.548 6.691
4.392 4.531 4.756 4.931 5.071 5.234 5.429 5.663 5.600 6.850 5.952 7.027 6.122 7.224 6.315
3.869 3.994 4.199 4.359 4.486 4.641 4.624 5.047 5.176 5.324 5.490 5.6,79
3.469 3.602 3.790 3.939 4.060 4.203 4.376 4.590 4.716 4.656 5.021 5.207
3.199 3.303 3.477 3.616 3.730 3.865 4.030 4.235 4.358 4.496 U56 4.840
17.197 17.707 16.505 19.101 19.562 20.068 20.692 21.394 21.790 22.221 22.692 23.209
9.534 9.867 10.399 10.805 11.125 11.495 11.928 12.441 12.735
6.851 7.107 7.523 7.846 8.103
5.470 5.682 6.029 6.302 6.522 6.762 7.093 7.473 7.695 7.944 6.225 8.545
4.051 4.211 4.476 4.693
3.636 3.779 4.021 4.216 4.376 4.570 4.808 5.107 5.267 5.494 5.732 6.010
3.322 3.452 3.672 3.651 4.000 4.180 4.403 4.686 4.657
8.4OS
6.761 9.190 9.439 13. OS9 9.716 13.417 10.025 13.616 10.373
20
3.493 3.607 3.792 3.935 4.049 4.161 4.338 4.526 4.634 4.755 4.1>89 5.040
10 21 23 27 31 35 41 51 71 91 131 251
15
4.624 4.806 5.107 5.346 5.541
4.668
5.772 5.076 6.OS2 5.335 6.397 5.654 6.601 5.645 6.832 6.062 7.094 6.310 7.395 6.596
2.125 2.201 2.332 2.442 2.535 2.649 2.795 2.986
2.4OS
2.573
3. lOS 2.660
3.246 2.610 3.415 2.970 3.622 3.171
15
20
4.070 4.206 4.363 4.546
2.785 2.675 3.026 3.151 3.253 3.376 3.527 3.719 3.834 3.967 4.122 4.304
2.217 1.925 2.285 1.980 2.403 2.075 2.500 2.156 2.581 2.225 2:682 2.311 2.610 2.423 2977 2.572 3.081 2.668 3.204 2.783 3.350 2.925 3.529 3.102
3.075 3.194 3.397 3.564 3.702 3.671 4.081 4.350 4.515 5.055 4. 70s 5.285 4.928 5.556 5.193
2.677 2.967 3.175 3.330 3.460 3.619 3.619 4.076 4.234 4.418 4.636 4.695
2.271 2.351 2.491 2.608 2.709 2.835 2.996 3.211 3.347 3.510 3.707 3.952
2.970 3.067 3.229 3.360 3.467 3.596 3.754 3.95~
1.962 2.025 2.137 2.233 2.315 2.420 2.558 2.745 2.867 3.016 3.201 3.436
TABLE B.5 SIGNIFICANCE POINTS FOR THE MODIFIED LIKELIHOOD RATIO TEST Of EQUALITY OF COVARIANCE MATRICES BASED ON EQUAL SAMPLE SIZES Pr{ -21ogX· ~ X} = 0.05 ng \q
2
4
5
6
7
8
9
10
----------------------~-----.--------
3 4 5 6 7 8 9 10
p=2 12.18 18.70 24.55 30.09 35.45 40.68 45.81 10.70 16.65 22.00 27.07 31.97 36.76 41.45 9.97 15.63 20.73 25.56 30.23 34.79 39.26
15.02 14.62 14.33 14.11 13.94
19.97 19.46 19.10 18.83 18.61
24.66 24.05 23.62 23.30 23.05
19.2
30.5
41.0
p=3 51.0 60.7
6 7 8 9 10
17.57 16.59 15.93 15.46 15.11
28.24 38.06 47.49 56.68 65.69 26.84 36.29 45.37 54.21 62.89 25.90 35.10 43.93 52.54 60.99 25.22 34.24 42.90 51.34 59.62 24.71 33.59 42.11 50.42 58.58
11
12 13
14.83 24.31 14.61 23.99 14.43 23.73
33.08 41.50 49.71 32.67 41.01 49.13 32.33 40.60 48.66
6 7 8 9 10
30.07 27.31 25.61 24.46 23.62
65.91 82.6 60.90 76.56 57.77 72.78 55.62 70.17 54.05 68.27
11 12 13 14 15
22.98 38.41 22.48 37.67 22.08 37.08 21.75 36.59 21.47 36.17
9.53 9.24 9.04 8.88 8.76
29.19 33.61 28.49 32.82 27.99 32.26 27.62 31.84 27.33 31.51
70.3
-
50.87 55.87 46.07 50.64 43.67 48.02
37.95 37.07 36.45 35.98 35.61
42.22 4l.26 40.57 40.06 36.65
46.45 45.40 44.65 44.08 43.64
79.7
89.0
98.3
74.58 83.37 92.09 71.45 79.91 88.29 69.33 77.56 85.72 67.79 75.86 83.86 66.62 74.57 82.45
57.76 65.71 57.11 64.97 56.57 64.37
73.56 81.35 72.75 80.46 72.08 79.72
p-4 48.63 44.69 42.24 40.56 39.34
52.85 51.90 51.13 50.50 49.97
98.9 91.89 87.46 84.42 82.19
115.0 107.0 101.9 98.45 95.91
66.81 80.49 65.66 79.14 64.73 78.04 63.96 77.14 63.31 76.38
93.95 92.41 91.16 90.12 89.25
13l.0 12l.9 137.0 152.0 116.2 130.4 144.6 112.3 126.1 139.8 109.5 122.9 136.3 107.3 105.5 104.1 103.0 102.0
120.5 118.5 117.0 115.7 114.6
133.6 131.5 129.7 128.3 127.1
.----------
681
- - . - . - .•. ilK
\q
TABLE B.5
_._-
2
4
(Colltilllled)
--.-\q 2
6
39.29 36.70 34.92
65.15 61.40 58.79
89.46 84.63 81.25
p=6 113.0 107.2 103.1
129.3 124.5
151.5 145.7
11 12 13 14 15
33.62 32.62 31.83 31.19 30.66
56.86 55.37 54.19 53.24 52.44
78.76 76.83 75.30 74.06 73.02
100.0 97.68 95.81 94.29 93.03
120.9 118.2 116.0 114.2 112.7
141.6 138.4 135.9 133.8 132.1
16
30.21
5l.77
72.14
91.95
111.4
130.6
682
4
----_._----------------
p = 5
8 9 10
..- - - -
Ilg
-.-.-.-------------.-----------~
10
49.95
11 12 13 14 15
47.43 45.56 44.11 42.96 42.03
16 17 18 19 20
41.25 40.59 40.02 39.53 39.11
84.43
117.0
80.69
112.2 108.6 105.7 103.5 101.6
77.90 75.74 74.01 72.59
71.41 100.1 70.41 98.75 69.55 97.63 68.80 96.64 68.14 95.78
142.9 138.4 135.0 132.2 129.9 128.0 126.4 125.0 123.8 122.7
TABLE B.6 CORRECTI0N FACTORS fOR SIGNIfiCANCE POINTS fOR Hili SPHERICITY TEST
5% Significance Level
n\p
3
4
4 5 6 '7 8 9 10
1.217 1.074 1.038 1.023 1.015 1.011 1.008
1.322 1.122 1.066 1.041 1.029 1.021
12 14 16 18 20
1.005 1.004 1.003 1.002 1.002
24 28 34 42 50 100 2
X
5
6
7
8
1.088 1.057 1.040
1.420 1.180 1.098 1.071
1.442 1.199 1.121
1.455 1.214
1.0l3 1.008 1.006 1.005 1.004
1.023 1.015 1.011 1.008 1.006
1.039 1.024 1.017 1.012 1.010
1.060 1.037 1.025 1.018 1.014
1.093 1.054 1.035 1.025 1.019
1.001 1.001
1.002 1.002
1.004 1.003
1.006 1.004
1.009 1.006
1.012 1.008
1.000 1.000 1.000 1.000
1.001 1.001 1.000 1.000
1.002 1.001 1.001 1.000
1.003 1.002 1.001 1.000
1.004 1.002 1.002 1.000
1.005 1.003 1.002 1.000
11.Q70S
16.WO
23.6848
31.4104
40.1133
49.8018
1.383
1.155
683
TABLE B.6 (Continued) 1% Significance Level
n\p
3
4
5
6
7
8
4 5 6 7 8 9 10
1.266 1.091 1.046 1.028 1.019 1.013 1.010
1.396 1.148 1.079 1.049 1.034 1.025
1.471 1.186 1.103 1.067 1.047
1.511 1.213 1.123 1.081
1.542 1.234 1.138
1.556 1.250
12 14 16 18 20
1.006 1.004 1.003 1.002 1.002
1.015 1.010 1.007 1.005 1.004
1.027 1.018 1.012 1.009 1.007
1.044 1.028 1.019 1.014 1.011
1.068 1.041 1.028 1.020 1.015
1.104 1.060 1.039 1.028 1.021
24 28
1.001 1.001
1.003 1.002
1.005 1.003
1.007 1.005
1.010 1.007
1.013 1.009
34 42 50 100
1.001 1.000 1.000 1.000
1.001 1.001 1.001 1.000
1.002 1.001 1.001 1.000
1.003 1.002 1.001 1.000
1.004 1.003 1.002 1.000
1.006 1.003 1.002 1.001
15.0863
2l.6660
29.l412
37.5662
46.9629
57.3421
2
X
684
TABLE B.7t SIGNIFICANCE POINTS FOR THE MODIFIED LIKELIHOOD RATIO TEST
Pr{ - 2 log hi ~ x} = 0.05
n
5%
1%
n
5%
1%
n
p=2
6 7 8 9 10
8.94 8.75 8.62 8.52 8.44
p=3 19.95 15.56 14.l3 13.42 13.00 12.73 12.53 12.38 12.26
4 5
1%
n
25.6 22.68
6 15.81 7 15.19 8 14.77 9 14.47 10 14.24
21.23 20.36 19.78 19.36 19.04
11
14.06 13.92
14 15
13.80 13.70 13.62
24 26 28 30 32 34 36 38 40
12
p=4
13
25.8 24.06 23.00 22.28
30.8 29.33 28.36 27.66
11 21.75 12 21.35 13 21.03 14 20.77 15 20.56
27.13 26.71 26.38 26.10 25.87
7 8 9 10
p=7
32.5 31.4
40.0 38.6
11
14 15
30.55 29.92 29.42 29.02 28.68
37.51 36.72 36.09 35.57 35.15
18.80 18.61
16 17
28.40 28.15
34.79 34.49
18.45 18.31 18.20
18 19 20
27.94 27.76 27.60
34.23 34.00 33.79
58.4 57.7 57.09 56.61
67.1 66.3 65.68 65.12
28 30
70.1 69.4
56.20 55.84 55.54 55.26 55.03
64.64 64.23 63.87 63.55 63.28
13
p=9
p=8
18 19 20 21 22
48.6 48.2 47.7 47.34 47.00
56.9 56.3 55.8 55.36 54.96
24 26 28 30 32 34
46.43 45.97 45.58 45.25 44.97 44.73
54.28 53.73 53.27 52.88 52.55 52.27
p=6
9 10
12
Io
5% 1% __ ._--------
p=5
18.8 16.82
=
._-------
5%
- - - - I - --.
2 l3.50 3 10.64 4 9.69 5 9.22
I
79.6 78.8
12
40.9 13 40.0 14 39.3 15 38.7
49.0 47.8 47.0 46.2
16 17 18 19 20 21
38.22 37.81 37.45 37.14 36.87 36.63
45.65 45.13 44.70 44.32 43.99 43.69
22 24 26 28 30
36.41 36.05 35.75 35.49 35.28
43.43 42.99 42.63 42.32 42.07
P = 10 34 (82.3) (92.4) 36 81.7 91.8 38 81.2 91.2 40 80.7 90.7
32 68.8 78.17 34 68.34 77.60 36 (67.91) (77.08) 45 38 (67.53) (76.65) 50 40 67.21 76.29 55 60 45 66.54 75.51 65 50 66.02 74.92 55 65.61 74.44 70 60 65.28 74.06 75
79.83 79.13 78.57 78.13 77.75
89.63 88.83 88.20 87.68 87.26
77.44 86.89 77.18 86.59
tElltries in parentheses have been interpolated or extrapolated into Korin's table. p - number of variates; N = number of observations; n = N - I. "i = n loglIol np - n loglSI + n tr(SIil 1 ), where S is the sample covariance matrix.
685
References
At the end of each reference in brackets is a list of sections in which that reference is used. Abramowitz, Milton, and Irene Stegun (1972), Handbook of Mathematical Functions with Formulas, Graphs, and Mat/lematical Tables, National Bureau of Standards. U.S. Government Printing Office, Washington, D.C. [8.5) Abruzzi, Adam (1950), Experimental Procedures and Criteria for Estimating and Evaluating Industrial Productivity, doctoral dissertation, Columbia University Library. [9.7, 9.P) Adrian, Robert (1808), Research concerning the probabilities of the errors which happen in making observations, etc., The Analyst or Mathematical Museum, 1, 93-109. [1.2) Abn, S. K., and G. C. Reinsel (1988), Nested reduced-rank autoregressive models for multiple time series, Journal of American Statistical Association, 83, 849-856. [12.7) Aitken, A. C. (1937), Studies in practical mathematics, II. The evaluation of the latent roots and latent vectors of a matrix, Proceedings of the Royal Society of Edinburgh, 57, 269-305. [11.4) Amemiya, Yasuo, and T. W. Anderson (1990), Asymptotic chi-square tests for a large class of factor analysis models, Annals of Statistics, 18, 1453-1463. [14.6) Anderson, R. L., and T. A. Bancroft (1952), Statistical Theory in Research, McGrawHill, New York. [S.P) Anderson, T. W. (1946a), The non-central Wishart distribution and certain problems of multivariate statistics, Annals of Mathematical Statistics, 17,409-431. (Correction, 35 (1964), 923-924.) [14.4) Anderson, T. W. (1946b), Analysis of multivariate variance, unpublished. [10.6)
An Introduction to Multivariate Statistical Analysis, Third Edition. By T. W. Anderson ISBN 0-471-36091-0 Copyright © 2003 John Wiley & Sons, Inc.
687
688
REFERENCES
Anderson, T. W. (1950), Estimation of the parameters of a single ~quation by the limited-information maximum-likelihood method, Statistical Inference in Dynamic Economic Models (Tjalling C. Koopmaris, ed.) John Wiley & Sons, Inc., New York. [12.8] Anclerson, T. W. (1951a), Classification by multivariate analysis, Pyschometrika, 16, 31-50. [6.5, 6.9] Anderson, T. W. (1951b), Estimating linear restrictions on regression coefficients for multivariate normal distributions, Annals of Mathematical Statistics 22, 327-351. (Correction, Annals of Statistics, 8 (1980), 1400.) [12.6, 12.7] Anderson, T. W. (1951c), The asymptotic distribution of certain characteristic roots and vectors, Proceedings of the Second Berkeley Symposium on Mathematical Statistics and Probability (Ierzy Neyman, ed.), University of California, Berkeley, 105-130. [12.7] Anderson, T. W. (1955a), Some statistical problems in relating experimental data to predicting performance of a production process, Journal of the American Statistical Association, 50, 163-177. [8.P] Anderson, T. W. (1955b), The integral of a symmetric unimodal function over a symmetric convex set and some probability inequalities, Proceedings of the American Mathematical Society, 6, 170-176. [8.10] Anderson, T. W. (1957), Maximum likelihood estimates for a multivariate normal distribution when some obervations are missing, Journal of the American Statistical Association, 52, 676-687. [4.P] Anderson, T. W. (1963a), Asymptotic theory for principal component analysis, Annals of Mathematical Statistics, 34, 122-148. [11.3, 11.6, 11.7, 13.5] Anderson, T. W. (1963b), A test for equality of means when covariance matrices are unequal, Annals of Mathematical Statistics, 34, 671-672. [5.P] Anderson, T. W. (1965a), Some optimum confidencL bounds for roots of determinantal equations, Annals of Mathematical Statistics, 36, 468-488. [11.6] Anderson, T. W. (1965b), Some properties of confidence regions and tests of parameters in multivariate distributions (with discussion), Proceedings of the IBM Scientific Computing Symposium in Statistics, October 21-23, 1963, IBM Data Processing Division, White Plains, New York, 15-28. [11.6] Anderson, T. W. (1969), Statistical inference for covariance matrices with linear structure, Multivariate Analysis 1/ (P. R. Krishnaiah, ed.), Academic, New York, 55-66. [3.P] Anderson, T. W. (1971), The Statistical Analysis of Time Series, John Wiley & Sons, Inc., New York. [8.11] Anderson, T. W. (1973a), Asymptotically efficient estimation of covarian.:e matrices with linear structure, Annals of Statistics, 1, 135-141. [14.3] Anderson, T. W.
REFERENCES
689
Anderson, T. W. (1974), An asymptotic expansion of the distribution of the limited information maximum likelihood estimate of a coefficient in a simultaneous equation system, Journal of the American Statistical Association, 69. 565-573. (Correction, 71 (1976), 1010.) [12.7] Anderson, T. W. (1976), Estimation of linear functional relationships: approximate distributions and connections with simultaneous equations in econometrics (with discussion), Journal of the Royal Statistical Society B, 38, 1-36. [12.7] Anderson, T. W. (1977), Asymptotic expansions of the distrinutions of estimates in simultaneous equations for alternative parameter sequences, Econometrica, 45, S09-S1R. [12.7]
Anderson, T. W. (1984a), Estimating linear statistical relationships, Annals of Statistics, 12, 1-45. []0.6, ]2.6, 12.7, 14.1] Anderson, T. W. (]984b), Asymptotic distribution of an estimator of linear functional relationships, unpublished. [12.7] Anderson, T. W. (1987), Multivariate linear relations, Proceedings of the Second International Tampere Conference in Statistics (Tarmo Pukkila and Simo Puntanen. eds.), Tampere, Finland, 9-36. [12.7] Anderson, T. W. (1989a), The asymptotic distribution of the likelihood ratio criterion for testing rank in multivariate components of variance, Journal of Multivariate AnalysiS, 30, 72-79. [10.6, 12.7] Anderson, T. W. (1989b), Th ~ asymptotic distrinution of characteristic roots and vectors in multivariate components of variance, Contributions to Probability and Statistics: Essays in Honor of Ingram Olkin (Leon Jay Gieser, Michael D. Perlman, S. James Press, and Allan R. Sampson, eds.), Springer Verlag, New York, 177-196. [13.6] Anderson, T. W. (1989c), Linear latent variable models and covariance structures. Joumal of Econometrics, 41, 91-119. [Correction 43 (1990), 395.] [12.8] Anderson, T. W. (1993), Nonnormal multivariate distributions: Inference based on elliptically contoured distributions, Multivariate Analysis: Fl/tlIre Directions (c. R. Rao, ed.), North-HOlland, Amsterdam, 1-25. [3.6] Anderson, T. W. (1994), Inference in linear models, Proceedings of the Internationa! Symposium on Multivariate Analysis and its Ayplications (T. W. Anderson, K. T. Fang, and 1. Olkin, eds.), Institute of Mathematical Statistics, ]-20. [12.7, ]2.8, 14.3] Anderson, T. W. (1999a), Asymptotic theory for canonical correlation analysis. JOl/rnal of Multivariate Analysis, 70,1-29. [13.7] Anderson, T. W. (1999b). Asymptotic distribution of the reduced rank regression estimator under general conditions, Annals of Statistics, 27, 1141-1154. [13.7] Anderson, T. W. (2002), Specification and misspecification in reduced rank regression, Sankhya, 64A, 1-13. [12.7] Anderson, T. W., and Yasuo Amemiya (j 988a), The asymptotic normal distrihution of estimators in factor analysis under general conditions. Annals of Swti,'tics. 16. 759-771. [14.6]
690
REFERENCES
Anderson, T. W., and Yasuo Amemiya (1988b), Asymptotic distributions in factor analysis and linear structural relations, Proceedings of the Intemati, 'nal Conference on Advances in Multivariate Statistical Analysis (S. Das Gupta and 1. K. Ghosh, eds.), Indian Statistical Institute, Calcutta, 1-22. [14.6) Anderson, T. W., and R. R. Bahadur (1962), Classification into two multivariate normal distlibutions with different covariance matrices, Annals of Mathematical Statistics. 33, 420-431. [6.10) Anderson, T. W., and Kai-Tai Fang (1990a), On the theory of multivariate elliptically contoured distributions and their applications, Statistical Inference in Elliptically COl/toured alld Related Distributions (Kai-Tai Fang and T. W. Anderson, eds.), Allerton Press, Inc., New York, 1-23. [4.5) Anderson. T. W., and Kai-Tai Fang (1990b), Inference in multivariate elliptically contoured distributions based on maximum likelihood, Statistical Inference in Elliptical(y Contoured and Related Distributions (Kai-Tai Fang and T. W. Anderson. eds.). Allerton Press, Inc., New York, 201-216. [3.6) Anderson, T. W., and Kai-Tai Fang (1992), Theory and applications of elliptically contoured and related distributions, The Developmen/ of Statistics: Recent Contrihwiom from Chilla (X. R. Chen, K. T. Fang, and C. C. Yang, eds.), Longman Scientific and Technical, Harlow, Essex, 41-62. [10.11) Anderson. T. W.o Kai-Tai Fang, and Huang Hsu (1986), Maximum likelihood estimates and likelihood-ratio criteria for multivariate elliptically contoured distributions, Canadian Journal of Statistics, 14, 55-59. [Reprinted in Statistical Inference ill Elliptically Contoured and Related Distributions (Kai-Tai Fang and T. W. Anderson, eds.), Allerton Pres.;, Inc., New York, 1990,217-223. [3.6) Anderson. T. W., Somesh Das Gupta, and George P. H. Styan (1972), A Bibliography of MlIltivariate Statistical Analysis, Oliver & Boyd, Edinburgh. (Reprinted by Robert E. Krieger, Malabar, Florida, 1977.) [Preface) Anderson. T. W.o Naoto Kunitomo, and Takamitsu Sawa (1983a), Evaluation of the distribution function of the limited information maximum likelihood estimator, Econometrica, 50, 1009-1027. [12.7J Anderson, T. W., Naoto Kunitomo, and Takamitsu Sawa (1983b), Comparison of the densities of the TSLS and LlMLK estimators, Global Econometrics, Essays in Honor of Lawrence R. Klein (F. Gerard Adams and Bert Hickman, eds.), MIT, Cambridge, MA, 103-124. [12.7) Anderson. T. W., and I. Olkin (1985), Maximum likelitood estimation of the parameters of a multivariate normal distribution, Linear Algebra and Its Applications, 70, 147--171. [3.2J Anderson, T. W., and Michael D. Perlman (1987), Consistency of invariant tests for the multivariate analysis of variance, Proceedings o.f the Second International Tampere Conference in Statistics (Tarmo Pukkila and Simo Puntaner., eds.), Tampere, Finland, 225-243. [8.10J Anderson, T. w.. and Michael D. Perlman (1993), Parameter consistency of invariant tests for MAN OVA and related multivariate hypotheses, Statistics and Probability: A Raghu Raj Bahadur Festschi/t (]. K. Ghosh, S. K. Mitra, K. R. Parthasarathy, and B. L. S. Prakasa Rao, eds.), Wiley Eastern Limited, 37-62. [8.10)
REFERENCES
691
Anderson, T. W., and Herman Rubin (1949), Estimation of the parameters of a single equation in a complete system of stochastic equations, Annals of Mathematical Statistics, 20, 46-63. [Reprinted in Readings in Econometric Theory (1. Malcolm Dowling and Fred R. Glahe, eds.), Colorado Associated University, 1970, 358-375.] [12.8] Anderson, T. W., and Herman Rubin (1950), The asymptotic properties of estimates of the parameters of a single equation in a complete system of stochastic equations, Annals of Mathematical Statistics, 21, 570-582. [Reprinted in Readings in Econometric Theory (J. Malcolm Dowling and Fred R. Glahe, eds.), Colorado Associated University, 1970,376-388.] [12.8] Anderson, T. W., and Herman Rubin (1956), Statistical inference in factor analysis, Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability (1erzy Neyman, ed.), Vol. V, University of California, Berkeley and Los
Angeles, 111-150. [14.2, 14.3, 14.4, 14.6] Anderson, T. W., and Takamitsu Sawa (1973), Distributions of estimates of coefficients of a single equation in a simultaneous system and their asymptotic expansiuns, Econometrica, 41, 683-714. [12.7] Anderson, T. W., and Takamitsu Sawa (1977), Tables of the distribution of the maximum likelihood estimate of the slope coefficient and approximations, Technical Report No. 234, Economics Series, Institute for Mathematical Studies in the Social Sciences, Stanford University, April. [12.7] Anderson, T. W., and Takamitsu Sawa (1979), Evaluation of the distribution function of the two-stage least squares estimate, Econometrica, 47, 163-182. [12.7] Anderson, T. W., and Takamitsu Sawa (1982), Exact and approximate distributions of the maximum likelihood estimator of a slope coefficient, Journal of the Royal Statisti~al Society B, 44, 52-62. [12.7] Anderson, T. W., and George P. H. Styan (1982), Cochran's theorem, rank additivity and tripotent matrices, Statistics and Probability: Essays in Honor of C. R Rao (G. Kallianpur, P. R. Krishllaiah, and J. K. Ghosh, edsJ, North-Holland, Amsterdam, 1-23. [7.4] Anderson, T. W., and Akimichi Takemura (1982), A new proof of admissibility of tests in multivariate analysis, Journal of Multivariate Analysis, 12, 457-468. [8.10] Andersson, Steen A., David Madigan, and Michael D. Perlman (2001), Alternative Markov properties for chain graphs, Scandinavianlournal of Statistics, 28, 33-85. [15.2] Barnard, M. M. (1935), The secular variations of skull characters in four series of Egyptian skulls, Annals of Eugenics, 6, 352-371. [8.8] Barndorff-Nielson, O. E. (1978), Information and Exponential Families in Statistical Theory, John Wiley & Sons, New York. [15.5] Barnes, E. W. (1899), The theory of the gamma function, Messenger of Mathematics, 29, 64-129. [8.5] Bartlett, M. S. (1934), The vector representation of a sample, Proceedings of the Cambridge Philosophical Sodety, 30, 327-340. [8.3] Bartlett, M. S. (1937a), Properties of sufficiency arid statistical tests, Proceedings of the Royal Society of London A, 160, 268-282. [10.2]
692
REFERENCES
Bartlett, M. S. (1937b), The statistical conception of mental factors, British Jvurnal vf Psychvlvgy, 28, 97-104. [14.7] Bartlett, M. S. (1938), Further aspects of the theory of multiple regression, Prvceedings vf the Cambridge Philvsvphical Svciety, 34,33-40. [14.7] Bartlett, M. S. (1939), A note on tests of significance in multivariate analysis, Prvceedings vf the Cambridge Philosvphical Svciety, 35, 180-185. [7.2, 8.6] Bartlett, M. S. (1947), Multivariate analysis, Jvurnal vf the Rvyal Statistical Svciety, Supplement, 9, 176-197. [8.8] Bartlett, M. S. (J 950), 1 csts of significance in factor analysis, BritiJh Journal vf Psychvlvgy (Statistics Sectivn), 3, 77-85. [14.3] Basmann, R L. (I 957), A generalized classical method of linear estimation of coefficients in a structural equation, Ecvnvmetrica, 25, 77-83. [12.8] Basmann, R L. (1961), A note on the exact finite sample frequency functions of generalized classical linear estimators in two leading overidentified cases, Jvurnal of the American Statistical Assvciativn, 56, 619-636. [12.8] Basmann, R L. (1963), A note on the exact finite sample frequency functions of generalized classical linear estimators in a leading three-equation case, Jvurnalvf the American Statistical Assvciativn, 58, 161-171. [12.8] Bennett, B. M. (1951), Note on a solution of the generalized Behrens-Fisher problem, Annals vf the Institute vf Statistical Mathematics, 2, 87-90. [5.5] Berger, J. O. (1975), Minimax estimation of location vectors for a wide class of densities, Annals vf Statistics, 3, 1318-1328. [3.5] Berger, J. O. (1976), Admissibility results for generalized Bayes estimators of coordinates of a location vector, Annals vf Statistics, 4, 334-356. [3.5] Berger, J. O. (1980a), A robust generalized Bayes estimator and confidence region for a multivariate normal mean, Annals vf Statistics, 8, 716-761. [3.5] Berger, J. O. (1980b), Statistical Decisivn Thevry, Fvundativns, Concepts and Methvds, Springer-Verlag, New York. [3.4, 6.2, 6.7] Berndt, Ernst R, and Eugene Savin (1977), Conflict among criteria for testing hypotheses in the multivariate linear regression model, Ecvnometrica, 45, 1263-1277. [8.6] Bhattacharya, P. K. (1966), Estimating the mean of a multivariate normal population with general quadratic loss function, Annals vf Mathematical Statistics, 37, 1819-1824. [3.5] Birnbaum, Allan (1955), Characterizations of complete classes of tests of some multiparametric hypotheses, with applications to likelihood ratio tests, Annals vf Mathematical Statistics, 26, 21-36. [5.6, 8.10] Bjorck, A., and G. Golub (1973), Numerical methods for computing angles between liuear subspaces, Mathematics of Cvmputativn, 27, 579-594. [12.3] Blackwell, David, and M. A. Girshick (1954), Thevry vf Games and Statistical Decisivns, John Wiley & Sons, New York. [6.2, 6.7] Blalock, H. M., Jr (ed.) (1971), Causal Models in the So:ial Sciences, A1dine-Atheston, Chicago. [15.1] Bonnescn, T., and W. Fenchel (1948), Thevrie der Kvnuexen Korper, Chelsea, New York. [8.10]
693
REFERENCES
Bose, R. C. (1936a), On the exact distribution and moment-coefficients of the D 2-statistic, Sankya, 2, 143-154. [3.3] Bose, R. C. (1936b), A note on the distribution of differences in mean values of twO samples drawn from two multivariate normally distributed populations. and the definition of the D 2 -statistic, Sankhya, 2, 379-384. [3.3] Bose, R. C .. and S. N. Roy (j 938), The distribution of the studentised D "-statistic. Sankhya, 4, 19-38. [5.4] Bowker, A. H. (1960), A representation of Hotelling's T2 and Anderson's classification statistic W in terms of simple statistics, COlllributiolls 10 Probability alld Statistics 0. Olkin, S. G. Ghurye, W. Hoeffding, W. G. Madow and H. B. Mann. eds,), Stanford University, Stanford, California, 142-149. [5.2] Bowker, A. H., and R. Stigreaves (1961), An asymptotic expansion for the distribution function of the W-classification statistic, Studies in Item Analysis and Prediction (Herbert Solomon, ed.), Stanford University, Stanford, California, 193-310. [6.6] Box, G. E. P. (1949), A general distribution theory for a class of likelihood criteria, Biometrika, 36, 317-346. [8.5, 9.4, 10.4, 10.5] Bravais, Auguste (1846), Analyse mathematique sur les probabilites des erreurs de situation d'un pOint, Memoires Presenres par Divers Savants I'Academie Royale des Sciences de l'Institut de France, 9, 255-332. [1.2]
a
Brown, G. W. (1939), On the power of the Ll test for equality of several variances, Annals of Mathematical Statistics, 10, 119-128. [10.2] Browne, M. W. (1974), Generalized least squares estimates in the analysis of covariance structures, South African Statistical loumal, 8, 1-24. [14.3] Chambers, JlJhn M. (1977), Computational Methods for Data Ana(vsis, John Wiley & Sons, New York. [12.3] Chan, Tony F., Gene H. Golub, and Randall J. leVeque (1981), Algorithms for computing the sample variance: analysis and recommendations, unpublished. [3.2] Chernoff, Herman (1956), Large sample theory: parametric case, Annals of Mathematical Statistics, 27, 1-22. [8.6] Chernoff, Herman, and N. Divinsky (1953), The computation of maximum likelihood estimates of linear structural equations, Studies in Econometric Method (W. C. Hood and T. C. Koopmans, eds.), John Wiley & Sons, Inc .. New York. [11.8] Chu, S. Sylvia, and K. C. S. Pillai (1979), Power comparisons of two-sided tests of equality of two covariance matrices based on six criteria. Annals of the [nstitwe of Statistical Mathematics, 31, 185-205. [10.6] Chung, Kai Lai (1974), A Course in Probability Them)', 2nd ed., Academic. New York. [2.2] Clemm, D. S., P. R. Krishnaiah, and V. B. Waikar (973), Tables for the extreme roots of the Wishart matrix, loumal of Statistical Computation and Simulation. 2. 65-92. [10.8] Clunies-Ross, C. W., and R. H. Riffenburgh (960), Geometry and linear discrimination, Biometrika, 47, 185-189. [6.10] Cochran, W. G. (1934), The distribution of quadratic forms in a normal system. with applications to the analysis of covariance, Proceedings of the Cambridge Philosophical Society, 30, 178-191. [7.4]
694
REFERENCES
Constantine. A. G. (1963). Some non-central distribution problems in multivariate analysis. Annals of Mathematical Statistics, 34, 1270-1285. [12.4] Constantine. A. G. (J 966). The distribution of HoteHing's generalised T,r. Annals of Mathematical Statistics, 37, 215-225. [8.6] Consul. P. C. (1966). On the exact distributions of the likelihood ratio criteria for testing linear hypotheses about regression coefficients, Annals of Mathematical Statistics. 37, 1319-1330. [8.4] Consul. P. C. 0967a), On the exact distributions of likelihcod ratio criteria for testing independence of sets of variates under the nuH hypotllesis, Annals of Mathematical Statistics, 38, 1160-1169. [9.3] Consul, P. C. (l967b), On the exact distributions of the criterion W for testing sphericity and in a p-variate normal distribution, Annals of Mathematical Statistics, 38, 1170-1174. [10.7] Courant. R.. and D. Hilbert (1953), Methods of Mathematicr.! Physics, lnterscience, New York. [9.10] Cox. D. R.. and N. Wermuth (1996). Multivariate Dependencies, Chapman and Halt, London. [15.1] Cramer. H. (1946), Mathematical Methods of Statistics, Princeton University, Princeton. [2.6. 3.2, 3.4. 6.5. 7.2] Dahm. P. Fred. and Wayne A. Fuller (l98!). Generalized least squares estimation of the functional multivariate linear errors in variables model, unpublished. [14.3] Daly. J. F. (1940). On the unbiased character of likelihood-ratio tests for independence in normal systems, Allnals of Mathematical Statistics, 11, 1-32. [9.2] Darlington. R. B .. S. L. Weinberg, and H. 1. Walberg (1973), Canonical variate analysis and related techniques, Review of Educational Research, 43, 433-454. [12.2] Das Gupta. Somesh (1965). Optimum classification rules for classification into two multivariate normal populations, Annals of Mathematical Statistics, 36, 1174-1184. [6.6] Das Gupta. S., T. W. Anderson, and G. S. Mudholkar (1964), Monotonicity of the power functions of some tests of the multivariate linear hypothesis, AnnaL~ of Mathematical Statistics. 35, 200-205. [8.10] David, F. N. (937), A note on unbiased limits for the correlation coefficient, Biometrika, 29, 157-160. [4.2] David. F. N. (1938). Tables of the Ordinates and Probability Integral of the Distribution of the Correlation Coefficient in Small Samples, Cambridge University, Cambridge. [4.2] Davis. A. W. (1968). A system of linear differential equations for the distribution of HoteHing's generalized T6, Annals of Mathematical Statistics, 39, 815-832. [8.6] Davis. A W. (l970a), Exact distributions of Hotelling's generalized T6, Biometrika, 57. 187-191. [Preface, 8.6] Davis, A W. (1970b), Further applications of a differential equation for Hotelting's generalized T6. Annals of the Institute of Statistical Mathematics, 22, 77-87. [Preface. 8.6] Davis. A W. (I971). Percentile approximations for a class of likelihood ratio criteria, Biometrika, 58, 349-356. [10.8, 10.9]
REFERENCES
695
Davis, A. W. (J972a), On the marginal distributions of the latent roots of the multivariate beta matrix, Annals of Mathematical Statistics, 43, 1664-1670. [S.6] Davis, A. W. (l972b), On the distributions of the latent roots and traces of certain random matrices, Journal of Ihullivariate Analysis, 2, 189-200. [8.6] Davis, A. W. (19S0), Further tabulation of Hotelling's generalized TJ, Communications in Statistics, 89, 321-336. [Preface, S.6] Davis, A. W., and J. B. F. Field (1971), Tables of some multivariate test criteria, Technical Report No. 32, Division of Mathematical Statistics, C.S.I.R.O., Canberra, Australia. [lO.S] Davis, Harold T. (1933), Tables of the Higher Mathematical Functions, Vol. I, Principia Press, Bloomington, Indiana. [S.5] Davis, Harold T. (1935), Tables of the Higher Mathematical Functions, Vol. II, Principia Press, Bloomington, Indiana. [S.5] Deemer, Walter L., and Ingram Olkin (1951), The Jacobians of certain matrix transformations useful in multivariate analysis. Based on lectures of P. L. Hsu at the University of North Carolina, 1947, Biometrika, 38, 345-367. [13.2] De Groot, Morris H. (I 970), Optimal Statistical Decisions, McGraw-Hill, New York. [3.4, 6.2, 6.7] Dempster, A. P. (1972), Covariance selection, Biomellics,28, 157-175. [15.5] Dempster, A. P., N. M. Laird, and D. B. Rubin (1977), Maximum likelihood from incomplete data via the EM algorithm (with discussion). Journal of the Royal Statistical Society B, 39, 1-38. [14.3] Diaconis, Persi, and Bradley Erron (I9S3), Computer-intensive methods in statistics, Scientific American, 248, 116-130. [4.3] Eaton, M. L., and M. D. Perlman (1974), A monotonicity property of the power functions of some invariant fests for MANOVA, Annals of Statistics, 2, 10:22-1028.
[S.IO] Edwards, D. (1995), Introduction to Graphical Modelling, Springer-Verlag, New York. [15.1] Efron, Bradley (1982), The Jackknife, the Bootstrap, and Other Resampling PlaT'S, Society for Industrial and Applied Mathematics, Philadelphia. [4.2] Efron, Bradley, and Carl Morris (J 977), Stein's paradox in statistics, Scientific American, 236, 119-127. [3.5] Elfving, G. (1947), A simple method of deducing certain distributions connected with multivariat,~ sampling, SkandinaviskAktuarietidskrift, 30, 56-74. [7.2] Fang, Kai-Tai, Samuel Kotz, and Kai-Wang Ng (1990), Symmetric Multivariate and Related Distributions, Chapman and Hall, New York. [2.7] Fang, Kai-Tai, and Ya6-Ting Zhang (1990), Generalized Multivariate Analysis, Springer-Verlag, New York. [2.7, 3.6, 10.11,7.9] Ferguson, Thomas Shelburne (1967), Mathematical Statistics: A Decision Theoretic Approach, Academic, New York. [3.4,6.2,6.7] Fisher, R. A. (1915), Fr.:quency distribution of the values of the correlation coefficient in samples from an indefinitely large population, Biometrika, 10, 507-521. [4.2,7.2] Fisher, R. A. (1921), On the "probable error" of a coefficient of correlation deduced from a small sample, Metron, I, Part 4, 3-32. [4.2]
696
REFERENCES
Fisher, R. A (1924), The distribution of the partial correlation coefficient, Metron, 3, 329-332. [4.3] Fisher, R. A (1928), The general sampling distribution of the multiple correlation coefficient, Proceedings of the Royal Society of London, A, 121, 654-673. [4.4] Fisher, R. A (1936), The use of multiple measurements in taxonomic problems, Annals of Eugenics, 7, 179-188. [5.3, 6.5, 11.5] Fisher, R. A. (1939), The sampling distribution of some statistics obtained from non-lmear equations, Annals of Eugenics, 9, 238-249. [13.2] Fisher, R A (1947a), The Design of Experiments (4th ed.), Oliver and Boyd, Fdint urgh. [8.9] Fisher, R. A (1947b), The analysis of covaria'1ce method for the relation between a part and the whole, Biometrics, 3, 65-68. [3.P] Fisher, R. A., and F. Yates (1942), Statistical Tables for Biological, Agricultural and Medical Research (2nd ed.), Oliver and Boyd, Edinburgh. [4.2] Fog, David (1948), The geometrical method in the theory of sampling, Biometrika, 35, 46-54. [7.2] Foster, F. G. (1957), Upper percentage points of the generalized beta d:stribution, II, Biometrika, 44, 441-453. [8.6] Foster, F. G. (1958), Upper percentage points of the generalized beta distribution, III, Biometrika, 45, 492-503. [8.6] Foster, F. G., and D. H. Rees (1957), Upper percentage points of the generalized beta distribution, I, Biometrika, 44, 237-247. [8.6] Frets, G. P. (1921), Heredity of head form in man, Genetica, 3, 193-384. [3.P] Frisch, R. (1929), Correlation and scatter in statistical variables, Nordic Statistical Journal, 8, 36-102. [7.5] Frydenberg, M. (1990), The chain graph Markov property, Scandinavian Journal of StatistiCS, 17,333-353. [15.2] Fujikoshi, Y. (1973), Monotonicity of the power functions of some tests in general MANOY A models, Annals of Statistics, 1, 188-391. [8.6] FUjikoshi, Y. (1974), The likelihood ratio tests for the dimensionality of regression coefficients, Journal of Multivariate Analysis, 4, 327-340. [12.4] Fujikoshi, Y., and M. Kanazawa (1976), The ML classification statistic in covariate discriminant analysis and its asymptotic expansions, Essays in Probability and Statistics, 305-320. [6.6] Fuller, Wayne A, Sastry, G. Pantula, and Yasuo Amemiya (1982), The covariance matrix of estimators for the factor model, unpublished. [14.4] Gabriel, K. R. (1969), Simultaneous test procedures-some theory of multiple comparisons, Annals of Mathematical Statistics, 40, 224-250. [8.7] Gajjar, A Y. (1967), Limiting distributions of certain transformations of multiple correlation coefficients, Metron, 26, 189-193. [4.4] Galton, Francis (1889), Natural Inheritance, MacMillan, London. [1.2, 2.5] Gauss, K. F. (1823), Theory of the Combination of Observations, Gottingen. [1.2] Giri, N. (1977), Multivariate Statistical Inference, Academic, New York. [7.2, 7.7] Giri, N., and 1. Kiefer (1964), Local and asymptotic minimax properties of multivariate tests, Annals 0/ Mathematical Statistics, 35, 21-35. [5.6]
REFERENCES
697
Giri, N., J. Kiefer, and C. Stein (1963), Minimax character of HoteHing's T' test in the simplest case, Annals of Mathematical Statistics, 34, 1524-1535. [5.6] Girshick, M. A. (1939), On the sampling theory of roots of determinantal equations. Annals of Mathematical Statistics, 10, 203-224. [l3.2] Gieser, Leon Jay (1981), Estimation in a multivariate "errors in variables" regression model: large sample results, Annals of Statistics, 9, 24-44. [12.7] Glynn, W. 1., and R. J. Muirhead (1978), Inference in canonical correlation analysis. Journal of Multivariate Analysis, 8, 468-478. [12.4] Golub, Gene H., and Franklin T. Luk (1976), Singular value decomposition: applications and computations, unpublished. [12.3] Golub, Gene H., and Charles F. Van Loan (1989), Matrix Computatiol1S (2nd ed.l. Johns Hopkins University Press, Baltimore. [11.4, 12.3, A.5] Grubbs, F. E. (1954), Tables of 1% and 5% probability levels of Hotclling's generalized T2 st!ltistics, Technical Note No. 926, Ballistic Research Laboratory. Aberdeen Proving Ground, Maryland. [8.6] Gupta, Shanti S. (1963), Bibliography on the multivariate normal integrals and related topics, Annals of Mathematical Statistics, 34, 829-838. [2.3] Guriand, John (1968), A relatively simple form of the distribution of the multiple correlation coefficient, Journal of the Royal Statistical Society B. 30, 276-283. [4.4J Gurland, J., and R. Milton (1970), Further consideration of the distribution of the mqltiple correlation coefficient, Journal of the Royal Statistical Society B. 32. 381-394. [4.4] Haavelmo, T. (1944), The probability approach in econometrics. Economerrica, 12. Supplement, 1-118. [12.7] Haff, L. R. (1980), Empirical Bayes estimation of the multivariate normal coyariancc matrix, Annals of Statistics, 8, 586-597. [7.8] Halmos, P. R. (1950), Measure Theory. D. van Nostrand. New York. [45. LUJ Harris, Bernard, and Andrew P. Soms (1980), The use of the tetrachoric series for evaluating multivariate pormal probabilities, Journal of Multivariate Analysis. 10. 252-267. [2.3] Hayakawa, Takesi (1967), (In the distribution of the maximum latent root of a positive definite symmetric random matrix, Annals of the Institute of Statistical Mathematics, 19, 1-17. [8.6] Heck, D. L (1960), Charts of some upper percentage points of the distribution of the largest characteristic root, Annals of Mathematical Statistics, 31, 625-642. [8.6] Hickman, W. Braddock (1953), The Volume of Corporate Bond Financing Since 1900. Princeton University, Princeton, 82-90. [10.7] Hoel, Paul G. (1937), A significance test for component analysis, Annals of Mathematical Statistics, 8, 149-158. [7.5] Hooker, R. H. (1907), The correlation of the weather and crops. Journal of rhe Royal Statistical Society, 70, 1-42. [4.2] HotelIing, Harold (1931), The generalization of Student'S ratio. Annals of .'vIarhemarical Statistics, 2, 360-378. [5.1, 5.P] HoteHing, Harold (1933), Analysis of a complex of statistical variables into principal components, Journal of Educational Psychology, 24, 417-441. 498-520. [11.1. 14.3]
698
REFERENCES
Hotelling. Harold (1935). The most predictable criterion, Journal of Educational P5ychology. 26. 139-142. [12.2] Hotelling, Harold (1936). Relations between two sets of variates, Biometrika, 28, 321-377. [12.1] Hotelling. Harold (1947). Multivariate quality control, illustrated by the air testing of sample bombsights. Techniques of Statistical Analysis (c. Eisenhart, M. Hastay, and W. A. Wallis, edsJ. McGraw-Hili, New York, III-IR4. [8.6] Hotelling. Harold ([951l. A generalized T test and measure of multivariate dispersion. Proceedings of the Second Berkeley Symposium on Mathematical Statistics and ProbabililY (Jerzy Neyman. ed.), University of California, Los Angeles and Berkeley. 23-41. [8.6. 10.7] Hotelling. Harold (1953). New light on the correlation coefficient and its transforms (with discussion), Journal of the Royal Stalistical Society B, 15, 193-232. [4.2] Howe. W. G. (1955). Some Colltribwions 10 FaClOr Ana(ysis, U.S. Atomic Energy Commission Report. Oak Ridge National Laboratory, Oak Ridge, Tennessee. [14.2, 14.6] Hsu. P. L. (1938). Notes on Hotelling's generalized T, Annals of Ar'athematical Statistics. 9, 231-·243. [5.4] Hsu. P. L. (1939a), On the distribution of the roots of certain determinantal equations, Annals of Eugenics, 9, 250-258. [13.2] Hsu. P. L. (1939b), A new proof of the jOint product moment distribution, Proceedings of the Cambridge Philosophical Society, 35, 336-338. [7.2] Hsu, P. L. (945), On the power functions for the E 2-test and the T 2-test, Annals of Mathematical Statistics, 16, 278-286. [5.6] Hudson. M. (1974), Empirical Bayes estimation, Technical Report No. 58, NSF contract GP 3071 IX-2, Department of Statistics, Stanford University. [3.5] Immer, F. Roo H. D. Hayes, and LeRoy Powers (1934), Statistical determination of barIey varietal adaptation, Journal of the American Society of Agronomy, 26, 403-407. [8.9] Ingham, A. E. (1933), An integral which occurs in statistics, Proceedings of the Cambridge Philosophical Society, 29, 271-276. [7.2] Ito. K. (1956). Asymptotic formulae for the distribution of Hotelling's generalized T~ statistic. Annals of Mathematical Statistics, 27, 109 I -1105. [8.6] Ito. K. (1960), Asymptotic formulae for the distribution of Hotelling's generalized T~ statistic. II. Al/IllIls of Mathl'lI1atiCClI Statistics, 31, 114R-IIS3. [R.ti] Izenman, A. J. (1975), Reduced-rank regression for the multivariate linear model, Journal of MlIltivariateAnalysis, 5, 248-264. [12.7] [zenman, Alan Julian (I 980), Assessing dimensionality in multivariate regression, Ana(vsis of Variance, Handbook of Statistics. Vol. I (P. R. Krishnaiah, ed.), North-Holland, Amsterdam, 571-591. [3.P] J ames, A. T. (1954), Normal multivariate analysis and the orthogonal group, Annals of Mathematical Statistics, 25, 40-75. [7.2] J ames, A. T. (1964), Distributions of matrix variates and latent roots derived from normal samples, Annals of Mathematical Statistics, 35. 475-501. [8.6]
REFERENCES
699
James, W., and C. Stein (1961), Estimation with quadratic loss, Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability (Jerzy Neyman, ed.), Vol. I, 361-379. University of California, Berkeley. [3.5, 7.8] Japanese Standards Association (1972). Statistical Tables and Formulas with Computer Applications. [Preface] Jennrich, Robert I., and Dorothy T. Thayer (1973), A note on Lawley's formulas for standard errors in maximum likelihood factor analysis, Psychometrika, 38, 571-580. [14.3] Johansen, S. (1995), Likelihood-based Inference in Cointegrated Vector Autoregressive . Models, Oxford University Press. [12.7] Jolicoeur, Pierre, and J. E. Mosimann (1960), Size and shape variation in the painted turtle, a principal component analysis, Growth, 24, 339-354. Also in Benchmark Papers in Systematic and Evolutionary Biology (E. H. Bryant and W. R. Atchley, eds.), 2 (1975), 86-101. [11.P] Joreskog, K. G. (1969) A general approach to confirmatory maximum likelihood factor analysis, Psychometrika, 34, 183-202. [14.2] Joreskog, K. G., and Arthur S. Goldberger (1972), Factor analysis by generalized least squares, Psychometrika, 37, 243-260. [14.3] Kaiser, Henry F. (1958), The varimax criterion for analytic rotation in factor analysis, Psychometrika, 23, 187-200. [14.5] Kanazawa, M. (1979), The asymptotic cut-off point and comparison of error probabilities in covariate discriminant analysis, Journal of the Japan Statistical Society, 9, 7-17. [6.6] Kelley, T. L. (1928), Crossroads in the Mind of Mall, Stanford University, Stanford. [4.P,9.P] Kendall, M. G., and Alan Stuart (1973), The Advanced Theory of Statistics (3rd ed.), Vol. 2, Charles Griffin, London. [12.6]
l
Kennedy, William J., Jr., and James E. Gentle (1980), Statistical Computing, Marcel Dekker, New York. [12.3] Khatri, C. G. (1963), Joint estimation of the parameters of multivariate normal populations, Journal of Indian Statistical Association, 1, 125-133. [7.2] Khatri, C. G. (1966), A note on a large sample distribution of a transformed multiple corrcl?tion coefficient, Annals of the Institute of Slatisticai Mathematics, 18, 375-380. [4.4] Khatri, C. G. (1972), On the exact finite series distribution of the smallest or the largest root of matrices in three situations, Journal of Multivariate Analysis, 2, 201-207. [8.6] Khatri, C. G., and K. C. Sreedharan Pillai (1966), On the moments of the trace of a matrix and approximations to its non-central distribution, Annals of Mathematical Statistics, 37, 1312-1318. [8.6] Khatri, C. G., and K. C. S. Pillai (1968), On the non-central distributions of two test criteria in multivariate analysis of variance, Annals of Mathematical Statistics, 39, 215-226. [8.6] Khatri, C. G., and K. V. Ramachandran (1958); Certain multivariate distribution problems, I (Wishart's distribution), Journal of the Maharaja Sayajairo, University of Baroda, 7, 79-82. [7.2]
700
REFERENCES
Kiefer, J. (1957), Invariance, minimax sequential estimation, and continuous time processes, Annals of Mathematical Statistics, 28, 573-601. [7.8) Kiefer, J. (1966), Multivariate opiimality results, Multivariate Analysis (Parachuyi R Krishnaiah, ed.), Academic, New York, 255-274. [7.8) Kiefer, J., and R Schwartz (1965), Admissible Bayes character of T2_, R2_, and other fully invariant tests for classical multivariate normal problems, Annals of Mathematical Statistics, 36, 747-770. [5.P, 9.9, 10.10) Klotz, Jerome, and Joseph Putter (1969), Maximum likelihood estimation of the multivariate covariance components for the balanced one-way layout. Annals of Mathematical Statistics, 40,1100-1105. [10.6] Kolmogorov, A. (1950), Foundations of the Theory of Probability, Chelsea, New York. [2.2) Konishi, Sadanori (1978a), An approximation to the distribution of the sample correlation coefficient, Biometrika, 65, 654-656. [4.2) Konishi, Sadanori (J 978h), Asymptotic expansions for the distributions of statistics based on a correlation matrix, Canadian Journal of Statistics, 6, 49-56. [4.2] Konishi, Sadanori (1979), Asymptotic expansions for the distributions of functions of a correlation matrix, Journal of Multivariate Analysis, 9, 259-266. [4.2) Koopmans, T. c., and Olav Reicrs\'ll (1950), The idcntification of structural characteristics, Annals of Mathrmatical Statistics, 21, 165-181. [14.2) Korin, B. P. (1968), On the distribution of a statistic used for testing a covariance matrix, Biometrika, 55, 171-178. [10.8) Korin, B. P. (1969), On testing the equality of k covariance matrices, Biometrika, 56, 216-218. [10.5) Kramer, K. H. (1963), Tables for constructing confidence limits on the multiple correlation coefficient, Journal of the American Statistical Association, 58, 1082-1085. [4.4) Krishnaiah, P. R. (1978), Some recent developments on real multivariate distributions, Development in Statistics (P. R. Krishnaiah, ed.), Vol. 1, Academic, New York, 135-169. [8.6) Krishnaiah, P. R. (1980), Computations of some multivariate distributions, Analysis of Variance, Handbook of Statistics, Vol. 1 (P. R. Krishnaiah, ed.), North-Holland, Amsterdam, 745-971. Krishnaiah, P. R, and T. C. Chang (1972), On the exact distributions of the traces of SMI + S2)-1 and SISil, Sankhya, A, 34,153-160. [8.6) Krishnaiah, P. R, and F. J. Schuurmann (1974), On the evaluation of some distributions that arise in simultaneous tests for the equality of the latent roots of the covariance matrix, Journal of MultivariateAnalysis, 4, 265-282. [10.7) Kshirsagar, A. M. (1959), Bartlett decomposition and Wishart distribution, Annals of Mathematical Statistics, 30, 239-241. [7.2) Kudo, H..(1955), On minimax invariant estimates of the transformation parameter, Natural Science Repol1, 6, 31-73, Ochanomizu University, Tokyo, Japan. [7.8) Kunitomo, Naoto (1980), Asymptotic expansions of the distributions of estimators in a linear functional relationship and simultaneous equations, Journal of the American Statistical Association, 75, 693-700. [12.7)
REFERENCES
701
Lachenbruch, P. A., and M. R. Mickey (1968), Estimation of error rates in discriminant analysis, Technometrics, 10, 1-11. [6.6] Laplace, P. S. (1811), Memoire sur les integrales definies et leur application aux probabilites, Memoires de llnstitut Imperial de France, Anm£e 1810, 279-347. [1.2] Lauritzen, Steffen L. (1996), Graphical Models, Clarendon Press, Oxford. [15.5] Lauritzen, Steffen L., and :'1. Wermuth (1989), Graphical models for associations between variables some of which are qualitative and some quantitative, Annals of Statistics, 17, 31-57. [15.2] Lauter, lUrgen, Ekkehard Glimm, and Siegfried Kropf (I996a), New multivariate tests for data with an inherent structure, Biometrics Journal, 38. 5-23. (Correction: 40. (1998), 1015.] [5.7] Lauter, liirgen, Ekkehard Glimm, and Siegfried Kropf (1996b), Multivariate tests based on left-spherical1y distributed linear scores, Annals of Statistics. 26. 1972-1988. [5.7] Uillter, Jiirgen, Ekkehard Glimm, and Siegfried Kropf (I996c). Exact stahle multivariate tests for applications in clinical research. ASA Proceedings of the Riol'itamJaceutical Section, 46-55. [5.7] Lawley, D. N. (1938), A generalization of Fisher's z test, Biometrika, 30, 180-187. [8.6) Lawley, D. N. (940), The estimation of factor loadings hy the method of maximum likelihOOd, Proceedings of the Royal Society of Edinburgh, Sec. A, 60, 64·-82. [14.3] Lawley, D. N. (1941), Further investigations in factor estimation, Proceedings of the Royal Society of Edinburgh, Sec. A, 61, 176-185. [14.4) Lawley, D. N. (1953), A modified method of estimation in factor analysis and some large sample results, Uppsala Symposium on Psychological FaClOr Analysis. 17·-19 March 1953, Uppsala, Almqvist and Wiksell, 35-42. [14.3) Lawley, D. N. (1958), Estimation in factor analysis under various initial assumptions. British Journal of Statistical Psychology, 11, 1-12. [14.2, 14.6] Lawley, D. N. (1959), Tests of significance in canonical analysis, Biometrika. 46. 59-66. [12.4] Lawley, D. N. (1967), Some new results in maximum likeliho)d factor analysis. Proceedings of the Royal Society of Edinburgh, Sec. A, 87, 256-264. [143] Lawley, D. N., and A. E. Maxwel1 (1971), Factor Analysis as a Sratisrical Method Cnd ed.), American Elsevier, New York. [14.3, 14.5] Lee, Y. S. 1I971a), Asymptotic formulae for the distribution of a multivariate test statistic: power comparisons of certain multivariate tests. Biometrika. 58, 647-651. [8.6) Lee, Y. S. (1971b), Some results on the sampling distribution of the multiple correlation coefficient, Journal of the Royal Statistical Society B, 33. 117-·\30. [4.4] Lee, Y. S. (1972), Tables of upper percentage points of the multiple correlation coefficient, Biometrika, 59, 175-189. [4.4] Lehmann, E. L. (1959), Testing Statistical Hypotheses, lohn Wiley & Sons, New York. [4.2,5.6] Lehmer, Emma (1944), Inverse tables of probabilities of errors of the second kind. Annals of Mathematical Statistics, 15,388-398. [5.4]
702
REFERENCES
Loeve. M. (\977). Probability Theory I (4th ed.), Springer-Verlag, New York. [2.2] Loeve, M. (978), Probability Theory II (4th ed.), Springer-Verlag, New York. [2.2] Mado\\'. W. G. (1938). Contributions to the theory of multivariate statistical analysis, TransactiollS of the American Mathematical Society, 44, 454-495. [7.2] ;"1agnus, Jan R. (1988), Linear Stmctures, Charles Griffin and Co., London. [A.4] Magnus, J. R, and H. Neudecker (1979), The commutation matrix: some properties and applications, The Annals of Statistics, 7, 381-394. [3.6] Mahalanobis, P. C (1930), On tests and measures of group divergence, Journal and Proceedings of the Asiatic Society of Bengal, 26, 541-588. [3.3] Mahalanobis. P. C, R. C. Bose, and S. N. Roy (1937), Normalisation of statistical variates and the use of rectangular co-ordinates in the theory of sampling distributions, Sankhya, 3, 1-40. [7.2] Mallows, C L. (1961), Latent vectors of random symmetric matrices, Biometrika, 48, 133-149. [11.6] Mardia, K. V. 0970), Measures of multivariate skewness and kurtosis with applications. Biometrika, 57. 519-530. [3.6] Mariano, Roberto S.. and Takamitsu Sawa (1972). The exact finite-sample distribution of the limited-information maximum likelihood estimator in the case of two included endogenous variables, Journal of the American Statistical Association, 67, 159-163. [12.7] Maronna, R. A. (1976), Robust M-estimators of multivariate location and scatter, Annals of Statistics, 4. 51-67. [3.6] Marshall, A. W., and 1. Olkin (1979), Inequalities: TheoiY of Majorization and Its Applications, Academic, New York. [8.10] Mathai, A. M. (1971), On the distribution of the likelihood ratio criterion for testing linear hypotheses on regression coefficients, Annals of the Institute of Statistical Mathematics. 23.181-197. [8.4] Mathai, A. M., and R. S. Katiyar (979), Exact percentage points for testing independence, Biometrika, 66. 353-356. [9.3] Mathai. A. M., and P. N. Rathie (J 980), The exact non-null distribution for testing equality of covariance matrices, Sankhya A, 42, 71:!-87. [10.4] Mathai. A. M .. and R. K. Saxena (1973), Generalized Hypergeometric Functions with Applications in Statistics and Physical Sciences, Lecture Notes No. 348, SpringerVerlag, New York. [9.3] Mauchly. J. W. (1940). Significance test for sphericity of a normal n-variate distribution. Annals of Mathematical Statistics, 11, 204-209. [10.7] Mauldon, J. G. (1955), Pivotal quantitics for Wishart's and related distributions, and a paradox in fiducial theory, Journal of the Royal Statistical Society B, 17, 79-85. [7.2] McDonald, Roderick P. (2002), What can we learn from path equations: identifiability, constraints, equivalence, Psychometrika, 67, 225-249. [15.1] McLachlan, O. J. (J 973), An asymptotic expansion of the expectation of the estimated error rate in discriminant analysis. Australian Journal of Statistics, 15, 210-214. [6.6]
703
REFERENCES
M:Lachlan, G. 1. (l974a), An as)mptotic unbiased technique for estimating the error rates in discriminant analysis. Biometrics, 30. 239-249. [6.6] McLachlan, G. J. (l974b). Estimation of the errors of mischlssific:ttilln on the criterion of asymptotic mean square error, Technometrics, 16, 255-260. [6.6] McLachlan, G. J. (1974c), The asymptotic distributions of the conditional error rate and risk in discriminant analysis, Biometrika, 61, 131-135. [6.6] McLachlan, G. J. (1977), Constrained sample discrimination with the studentized classification statistic W, Communications in Statistics-Theory and Methods, A6, 575-583 [6.6] Memon, A. Z., and M. Okamoto (1971), Asymptotic expansion of the distribution of The Z statistic in discriminant analysis, Journal of Multivariate Analysis, 1, 294-307. [6.6] s ) of s Roots (s = 1 - 50), The Mijares, T. A. (1964), Percentage Points of the Sum Statistical Center, University of the Philippines, Manila. [8.6] Mikhail, N. N. (1965), A comparison of tests of the Wilks-Lawley hypothesis in multivariate analysis, Biometrika, 52, 149-156. [8.6] Mood, A. M. (1951), On the distribution of the characteristic roots of normal second-moment matrices, Annals of Mathematical Statistics, 22, 266-273. [13.2] Morris, Blair, and Ingram Olkin (1964), Some estimation and testing problems for factor analysis models, unpublished. [10.6] Mudholkar, G. S. (1966), On confidence bounds associated with multivariate analysis of variance and non-independence between two sets of variates, Annals of Mathematical Statistics, 37,1736-1746. [8.7] Mudholkar, Govind S., and Madhusuclan C. Trivedi (1980), A normal approximation for the distribution of the likelihood ratio statistic in multivariate analysis of variance, Biometrika, 67, 485-488. [8.5] Mudholkar, Govind S., and Madhusudan C. Trivedi (1981), A normal approximation for the multivariate likelihood ratio statistics, Statistical Distributions in Scientific Work (c. Taillie et aI., eds.), Vol. 5, 219-230, D. Reidel Publishing. [8.5] Muirhead, R. 1. (1970), AsymptotiC distributions of some multivariate tests, Annals of Mathematical Statistics, 41, 1002-1010. [8.6] Muirhead, R. 1. (1980), The eftects of elliptical distributions on some standard procedures involving correlation coefficients: a review, Multivariate Statistical Analysis (R. P. Gupta, ed.), 143-159. [4.5] Muirhead, Robb J. (1982), Aspects of Multivariate Statistical Theory, John Wiley and Sons, New York. [2.7, 3.6, 7.7] Muirhead, R. J., and C. M. Waternaux (1980), Asymptotic distributions in canonical correlation analysis and other multivariate procedures for nonnormal populations, Biometrika, 67, 31-43. [4.5] Nagao, Hisao (1973a), On som~ test criteria for covariance matrix, Annals of Stati.7tics, I, 700-709. [9.5, 10.2, 10.1, 10.8] Nagao, Hisao (1973b), Asymptotic expansions of the distributions of Bartlett's test and sphericity test under the local alternatives, Annals of the Institute of Statistical . Mathematics, 25, 407-422. [10.5, 10.6] Nagao, Hisao (1973c), Nonnull distributions of two test criteria for incepend.::nce under local alternatives, Journal of Multivariate Analysis, 3, 435-444. [9.4]
Vr
704
REFERENCES
Nagarsenker, B. N., and K. C. S. Pillai (1972), The Distribution of the Sphericity Test Criterion, ARL 72-0154, Aerospace Research Laboratories. [Preface) Nagarsenker, B. N., and K. C. S. Pillai (1973a), The distribution of the sphericity test criterion, Journal of Multivariate Analysis, 3, 226-235. [10.7) Nagarsenker, B. N., and K. C. S. Pillai (1973b), Distribution of the likelihood ratio criterion for testing a hypothesis specifying a covariance matrix, Biometrika, 60, 359-364. [10.8) Nagarsenker, B. N., and K. C. S. Pillai (1974), Distribution of the likelihood ratio criterion for testing l: = l:o, IL = ILo, Journal of MultivUlit te Analysis, 4, 114-122. [10.91 Nanda, D. N. (1948), Distribution of a root of a determinantal equation, Annals of Mathematical Statistics, 19, 47-57. [8.6) Nanda, D. N. (1950), Distribution of the sum of roots of a determinantal equation under a certain condition, Annals of Mathematical Statistics, 21, 432-439. [8.61 Nanda, D. N. (1951), Probability distribution tables of the larger root of a determinantal equation with two roots, JOllmal of the !lIdia/l Society of AgriclIl;ural Statistics, 3, 175-177. [8.6) Narain, R. D. (I948), A new approach to samplirg distributions of the multivariate normal theory, I, Joumal of the !lIdiall Society of Agricullllral Statistics, I, 59-69. [7.21 Nanin, R. D. (I950), On the completely unbiased character of tests of independence in multivariate normal systems, Annals of Mathematical Statistic~', 21, 293-298. [9.21 National Bureau of Standards, United States (1959), Tables of the Bivariate Normal Distribution Function and Related FUnctions, U.S. Government Printing Office, Washington, D.C. [2.3) Neveu, Jacques (1965), Mathematical Foundations of the Calculus of Probability, Holden-Day, San Francisco. [2.21 Ogawa, J. (1953), On the sampling distributions of classical statistics in multivariate analysis, Osaka Mathematics Journal, 5, l3-52. [7.2) Okamoto, Masashi (1963), An asymptotic expansion for the distribution of the linear discriminant function, Annals of Mathematical Statistics, 34, 1286-l301. (Correction, 39 (I968), l358-l359.) [6.6) Okamoto, Masashi (1973), Distinctness of the eigenvalues of a quadratic form in a multivariate sample, Annals of Statistics, I, 763-765. [13.2) Olkin, Ingram, and S. N. Roy (1954), On multivariate distribution theory, Annals of Mathematical Statistics, 25, 329-339. [7.2) Olson, C. L. (1974), Comparative robustness of six tests in multivariate analysis of variance, Journal of the American Statistical Association, 69, 894-908. [8:6) Pearl, Judea (2000), Causality: Models, Reasoning, and !nferencC'., Cambridge University Press, Cambridge. [15.1) Pearson, E. S., and H. O. Hartley (1972), Biometrika Tables for Statisticians, Vol. II, Cambridge (England), Published for the Biometrika Trustees at the University Press. [Preface, 8.4) Pearson, E. S., and S. S. Wilks (J 933), Methods of statistical analysis appropriate for k samples of two variables, Biometrika, 25, 353-378. [10.5)
REFERENCES
70S
Pearson, K. (J896), Mathematical contributions to the theory of evolution----1I1. Regression, heredity and panmixia, Philosophical Transactions of the Royal Socier;.' of London, Series A, 187,253-318. [2.5, 3.2] Pearson, K. (1900), On the criterion that a given system of deviations from the probable in the case of a correlated system of variables is such that it can be reasonab'y supposed to have arisen from random sampling, Philosophical Magazine, 50 (fifth series), 157-175. [3.3] Pearson, K. (901), On lines and planes of closest fit to systems of points in space, Philosophical Magazine, 2 (sixth series), 559-572. [11.2] Pearson, K. (1930), Tables for Statisticians and Biometricians, Part I (3rd ed.), Cambridge University, Cambridge. [2.3] Pearson, K. (1931), Tables for Statisticians and Biometricians. Part II, Cambridge University, Cambridge. [2.3, 6.8] Perlman, M. D. (1980), Unbiasedness of the likelihood ratio tests for equality of several covariance matrices and equality of several multivariate normal populations, Annals of Statistics, 8, 247-263. [10.2, 10.3] Perlman, M. D., and 1. Olkin (1980), Unbiasedness of invariant tests for MANOVA and other multivariate problems, Annals of Statistics, 8, 1:\26- 1:\4 L [lU 0] Phillips, P. C. B. (1980), The exact distribution of instrumental variable estimators in an equation containing n + 1 endogenous variables, Economellica, 48, 861-878. [12.7] Phillips, P. C. B. (1982), A new approach to small sample theory. unpublished, Cowles Foundation for Research in Economics, Yale University. [12.7] Pillai, K. C. S. (1954), On some distribution problems in multivariate analysis. Mimeo Series No. 54, Institute of Statistics, University of North Carolina, Chapel Hill, North Carolina. [8.6] Pillai, K. C. S. (1955), Some new test criteria in multivariate analysis. Annals of Mathematical Statistics, 26, 117-121. [8.6] Pillai, K. C. S. (1956), On the distribution of the largest or the smallest root of a matrix in multivariate analysis, Biometrika, 43, 122-127. [8.6] Pillai, K. C. S. (1960), Statistical Tables for Tests of MlIitivUliare Hvpotheses. Statistical Center, University of the Philippines. Manila. [8.6] Pillai, K. C. S. (1964), On the moments of elementary symmetric functions of the roots of two matrices, Annals of Mathematical Statistics, 35. 1704-1712. [8.6] Pillai, K. C. s. (1965), On the distribution of the largest characteristic root of a matrix in multivariate analysis, Biometrika, 52, 405-412. [8.6] Pillai,K. C. S. (1967), Upper percentage points of the largest root of a matrix in multivariate analysis, Biometroka, 54, 189-194. [8_6] Pill ai, K. C. S., and A. K. Gupta (1969), On the exact distribution of Wilks' criterion. Biometrika, 56, 109-118. [8.4] Pillai, K. C. S., and Y. S. Hsu (1979), Exact robustness studies of the test of independence based 011 four multivariate criteria and their distribution prohlems under violations, Ann lis of the Institute of Statistical Marhematics. 31. Pan A. 85-101. [8.6] Pillai, K. C. S., and K. 1ayachandran (1967), Power comparisons of tests of two multivariate hypotheses based on four criteria, Biomerrika. 54. 195--~1O_ [8_6]
706
REFERENCES
Pillai. K. C. S.. and K. Jayachandran (1970), On the exact distri'>ution of PiHai's criterion. JOllrnai of Ih(' American Statistical Association, 65, 447-454. [8.6]
V(s)
Pillai, K. C. S .. and T. A. Mijares (1959), On the moments of the trace of a matrix' and approximations to its distribution, Annals of Mathematical Statistics, 30, 1135-1140. [S.6J Pillai, K. C. S., and B. N. Nagarsenker (197!), On the distribution of the sphericity test criterion in classical and complex normal populations having unknown covariance matrices, Annals of Mathematical Statistics, 42, 764-767. [10.7] Pillai. K. C. S .. and P. Samson, Jr. (1959), On HoteHing's generalization of T2, Biometrika. 46. 160-168. [8.6] Pill ai, K. C. S., and T. Sugiyama (1969), Non-central distributions of the largest latent roots of three matrices in multivariate analysis, Annals of the Institute of Statistical Mathematics, 21, 321-327. [8.6] Pillai. K. C. S., and D. L. Young (1971), On the exact distribution of Hotelling's generalized T~, Journal of Multivariate Analysis, 1,90-107. [8.6] Plana, G. A. A. (1813), Memoire sur divers probJ(:mes de probabilite, Memoires de i'Acadc,nit! Impeliale de Tlirin, pour les Anlll!es 1811-1812, 20, 355-408. [1.2] Polya, G. U949). Remarks on computing the probability integral in one and two dimensions, Pm~eedings of the Berkeley Symposium on Mathematical Statistics and Prohahilit\" U. Neyman. ed'). 63-7S. [2.P] Rao, C. R. 1194Sa), The utilization of multiple ml:asurements in problems of biological classification, JOllrnal of the Royal Statistical Society B, 10, 159-193.[6.9] Ral). C. R. (194gh). Tests of significance in multivariate analysis, Biometrika, 35, 5S-79. [5.3J Rao. C. Radhakrishna (1951), An asymptotic expansion of the distribution of Wilks's criterion, Blilletin of the International Statistical Institute, 33, Part 2, 177-180. [8.5] Rao, C. R. (1952>. Advanced Statistical Methods in Biometric Research, John Wiley & Sons, New York. [12.5] Rao. C. R. (1973). Linear Statistical Inference and Its Applications (2n<j ed.), John Wiley & Sons, New York. [4.2] Rasch, G. (1948), A functional equation for Wishart's distribution, Annals of Mathematical Statistics, 19, 262-266. [7.2] Reiers~l,
Olav (1950), On the identifiability of parameters in Thurstone's multiple factor analysis. Psychometn'ka, 15, 121-149. [14.2]
Reinsel, G. C. and R. P. Velu (1998), Multivariate Reduced-rank Regression, Springer, New York. [12.7] Richardson, D. H. ([968), The exact distribution of a structural coefficient estimator, JOllrnal of the American Statistical Association, 63, 1214-1226. [12.7] Rothenberg, Thomas J. (1977), Edgeworth expansions for multivariate test statistics, IP-255, Center for Research in Management Science, University of California, Berkeley. [8.6] Roy. S. N. (1939). p-statistics or some generalisations in analysis of variance appropriate to multivariate problems, Sankhya, 4, 381-396. [13.2] Roy, S. N. ([945), The individual sampling distribution of the maximum, the minimum and any intermediate of the p-statistics on the null-hypothesis, Sankhya, 7, 133-158. [8.6]
REFERENCES
707
Roy, S. N. (1953), On a heuristic method of test construction and its use in multivariate analysis, Annals of Mathematical Statistics, 24, 220-238. [8.6, 10.6] Roy, S. N. (1957), Some Aspects of Multivariate Analysis, John Wiley & Sons, New York. [8.6, 10.6, 10.8] Ruben, Harold (]966), Some new :·esults on the distribution of the sample correlation coefficient, Journal of the Royal Statistical Society B, 28, 513-525. [4.2] Rubin, Donald B., and Dorothy T. Thayer (1982), EM algorithms for ML factor analysis, Psychometrika, 47, 69-76. [14.3] Ryan, D. A. or., J. J. Hubert, E. M. Carter, J. B. Sprague, and J. Parrot (1992), A reduced-rank multivariate regression approach to jOint toxicity experiments, Biometrics, 48, 155-162. [12.7] Salaevskil, O. v. (1968), The minimax character of Hotelling's T2 test (RJssian), Doklady Akademii Nauk SSSR, 180, 1048-1050. [5.61 Salaevskfi, (Shalaevskii), O. V. (1971), Minimax character of Hotelling's T2 test. 1. Investigations in Classical Problems of Probability Theory and Mathematical Statistics, V. M. Kalinin and O. V. Shalaevskii (Seminar in Mathematics, V. 1. Steklov Institute, Leningrad, Vol. 13), Consultants Bureau, New York. [5.6] Sargan, J. D., and W. M. Mikhail (1971), A general approximation to the distribUlion of instrumental variables estimates, Econometrica, 39, 131-169. [12.7] Sawa, Takamitsu (1969), The exact sampling distribution of ordinary least squlfes and two-stage least squares estimators, lorunal of the American Statistical Association, 64, 923-937. [12.7] . Schatzoff, M. (1966a), Exact distributions of Wilks's likelihood ratio criterion, Biometrika, 53, 347-358. [8.4] Schatzoff, M. (1966b), Sensitivity comparisons among tests of the general linear hypothesis, Journal of the American Statistical Association, 61, 415-435. [8.6] Scheffe, Henry (1942), On the ratio of the variances of two normal populations, Annals of Mathematical Statistics, 13, 371-388. [10.2] Scheffe, Henry (1943). On solutions of the Behrens-Fisher problem, based on the t-distribution, Annals of Mathematical Statistics, 14, 35-44. [5.5] Schmidli, H. (1996), Reduced-rank Regression, Physica, Berlin. [12.7] Schuurmann, F. J., P. R. Krishnaiah, and A. K. Chattopadhyay (1975), Exact percentage points of the distribution of the trace of a multivariate beta matrix, Journal of Statistical Computation and Simulation, 3, 331-343. [8.6] Schuurmann, F. J., and V. B. Waikar (1973), Tables for the power function of Roy's two-sided test for testing hypothesis l: = I in the bivariate case, Communications in Statistics, 1, 271-280. [10.8] Schuurmann, F. J., V. B. Waikar, and P. R. Krishnaiah (1975), Percentage points of the joint distribution of the extreme roots of the random matrix S l(S 1 + S2)-1, Journal of Statistical Computation and Simulation, 2, 17-38. [10.6] Schwartz, R. (1967), Admissible tests in multivariate analysis of variance, Annals of Mathematical Statistics, 38, 698-710. [8.101 Serfiing, Robert J. (1980), Approximation Theorems of Mathematical Statistics, John Wiley & Sons, New York. [4.2] Simaika, J. B. (1941), On an optimum property of two important statistical tests, Biometrika, 32, 70-80. [5.6]
708
llEFERENCES
Siotani, Minoru (1980), Asymptotic approximations to the conditional distributions of the classification statistic Z and its studentized form Z', Tamkang Journal of Mathematics, 11, 19-32. [6.6] Siotani, M., and R. H. Wang (1975), Further expansion formulae for error rates and comparison of the W- and the Z-procedures in discriminant analysis, Technical Report No. 33, Department of Statistics, Kansas State University, Manhattan, Kansas. [6.6] Siotani, M., and R. H. Wang (1977), Asymptotic Expansions for Error Rates and Comparison of the W-Procedure and the Z-Procedure in Discriminant Analysis, MuitivariateAnalysis W, North-Holland, Amsterdam, 523-545. [6.6] Sitgreaves, Rosedith (1952), On the distribution of two random matrices used in classification procedures, Annals of Mathematical Statistics, 23, 263-270. [6.5] Solari, M. E. (1969), The "maximum likelihood solution" of the problem of estimating a linear functional relationship, Journal of the Royal Statistical Society B, 31, 372-375. [14.4] Spearman, Charles (1904), "General~intelligence," objectively determined and measured, American Journal of Psychology, 15, 201-293. [14.2] Speed, T. P., and H. Kiiveri (1986), Gaussian Markov distributions over finite graphs, Annals of Statistics, 14, 138-150. [15.5] Srivastava, M. S., and C. G. Khatri (1979), An Introduction to Multivariate Statistics, North-Holland, New York. [10.9, 13.P] Stein, C. (1956a), The admissibility of Hotelling's T 2-test, Annals of Mathematical Statistics, 27, 616-623. [5.6] Stein, C. (1956b), Inadmissibility of the usual estimator for the mean of a multivariate normal distribution, Proceedings of the Third Berkeley Symposium on Mathematical and Statistical Probability (Jerzy Neyman, ed.), Vol. I, 197-206, University of California, Berkeley. [3.5] Stein, C. (1974), Estimation of the parameters of a multivariate normal distribution. I. Estimation of the means, Technical Report No. 63, NSF Contract GP ~0711X-2, Department of Statistic>, Stanford University. [3.5] Stoica, P., and M. Viberg (1996), Maximum likelihood parameter and rank estimation in reduced-rank multivariate linear regressions, IEEE Transaction Signal Processing, 44, 3069-3078. [12.7] Student (W. S. Gosset) (1908), The probable error of a mean, Biometrika, 6, 1-25. [3.2] Styan, George P. H. (1990), The Collected Papers ofT. W. Anderson: 1943-1985, John Wiley & Sons, Inc., New York. [Preface] Subrahmaniam, Kocherlota, and Kathleen Subrahmaniam (1973), Multivariate Analysis: A Selected and Abstracted Bibliography, 1957-1972, Marcel Dekker, New York. [Preface] Sugiura, Nariaki, and Hisao Nagao (1968), Unbiasedness of some test criteria for the equality of one or two covariance matrices, Annals of Mathematical Statistics, 39, 1686-1692. [10.8] Sugiyama, T. (1967), Distribution of the largest latent root and the smallest latent root of the generalized B statistic and F statistic in multivariate analysis, Annals of Mathematical Statistics, 38, 1152-1159. [8.6]
REFERENCES
709
Sugiyama, T., and K. Fukutomi (1966). On the distribution of the extreme characteristic roots of the matrices in multivariate analysis, Reports of Staristical AppliclItion Research, Union of Japanese Scientists and Enginem', 13. [H.6] Sverdrup, Erling (1947), Derivation of the Wishart distrihution of the second order sample moments by straightforward integration of a multiple integral. Skandinavisk Aktuarietidskrift, 30, 151-166. [7.2] Tang, P. C. (1938), The power function of the analysis of variance tests with tables and illustrations of their use, Statistical Resecrch Memoirs. 2, 126-157. [5.4] Theil, H. (assisted by J. S. Cramer, H. Moerman, and A. Russchen) (1961), Economic Forecasts and Policy, (2nd rev. ed.), North-Holland, Amsterdam. Contributions to Economic Analysis No. XV (first published 1958). [12.8] Thomson, Godfrey H. (193,-), Hotelling's method modified to give Spearman's "g:' Journal of Educationallsychology, 25, 366-374. [14.3] Thomson, Godfrey H. (1951), The Factorial Analysis of Human Ability (5th ed.t University of London, London. [14.7] Thurstone, L. L. (1947), Multiple-Factor Analysis, University of Chicago. Chicago. [14.2, 14.5] Tsay, R. S., and G. C. Tiao (1985), Use of canonical analysis in time ,eric, model identification, Biometrika, 72, 299-315. [12.7] Tukey, J. W. (1949), Dyadic anova, an analysis of variance for vectors. Humall Biology, 21, 65-110. [8.9] Tyler, David E. (1981), Asymptotic inference for eigenvectors. Annals of Statistics. 9. 725-745. [11.7] Tyler, David E. (1982), Radial estimates and the test for sphericity. BiO.'llt'rn"/"r. 69. 429-436. [3.6] Tyler, David E. (1983a); Robustness and efficiency properties of scatter matrices. Biometrika, 70, 411-420. [3.6] Tyler, David E. (1983b), The asymptotic distribution of principal component roots under local alternatives to multiple. roots, Annals of Statistics, 1 L 1232-12·e. [11.7] Tyler, David E. (1987), A distribution free M-estimator of multivariate scatter. Allnals of Statistics, 15, 234-251. [3.6] . Velu, R. P., G. C. Reinsel, and D. W. Wichern (1986), Reduced rank models for multiple time series, Biometrika, 73, 105-118. [12.7] von Mises, R. (1945), On the classification of observation data into distinct groups. Annals of Mathematical Statistics, 16, 68-73. [6.8] von Neumann, 1. (1937), Some matrix-inequalities and metrization of matric-space. Tomsk University Review, 1 286-300. Reprinted in JO/III VOIl Neumall Collected Works (A. H. Taub, ed.), 4 (1962), Pergamon. New York. 205-219. [AA] Wald, A. (1943), Tests of statistical hypotheses concerning several parameters when the number of observations is large, Transactions of tire Americall Mathematical Society, 54, 426-482. [4.2] Wald, A. (1944), On a statistical problem arising in the classification llf an individual into one of two groups. Annals of Mathematical Statistics. 15. 145-162. [6.4. 6.5] Wald, A. (1950), Statistical DeciSion Functions. John Wiley &. Sons. New York. [6.2. 6.7, 8.10]
710
REFERENCES
Wald, A, and R. Brookner (1941), On the distribution of Wilks' statistic for testing the independence of several groups of variates, Annals of Mathematical Statistics, 12. 137-152. [8.4, 9.3] . Walker, Helen M. (1931), Studies in the History of Statistical Method, Williams and Wilkins, Baltimore. [1.1] \Vdch, P. D., and R. S. Wimpress (1961), Two multivariate statistical C lmputer programs and their application to the vowel recognition problem, Joul7lal of the Acoustical Society of America, 33, 426-434. [6.10] Wermuth, N. (1980), Linear recursive equations, covariance selection and path analysis, Journal oJ' the American Statistical Association, 75, 963-972. [15.5] Whittaker. E. T., and O. N. Watson (1943), A Course of Modem Analysis, Cambridge University, Cambridge. [8.5] Whittaker, Joe (1990), Graphical Models in Applied Multivariate Statistics, John Wiley & Sons, Inc., Chichester. [15.11 Wijsman, Robert A. (1979), Constructing all smallest simultaneous confidence sets in a given class. with applications to MANOVA, Allnals of Statistics, 7, 1003-1018. [8.7] Wijsman. Robert A (1980), Smallest simultaneous confidence sets with applications in multivariate analysis. Multivariate Analysis, V, 483-498. [8.7] Wilkinson, James Hardy (1965), The Algebraic Eigenvalue Problem, Clarendon, Oxford. [11.4]
Wilkinson, J. H., and C. Reinsch (1971), Linear Algebra, Springer-Verlag, New York. [11.4] \Vilks. S. S. U 932), Certain generalizations in the' analysis of variance, Biometrika, 24, 471-494. [7.5, 8.3, 10.4] Wilks. S. S. (I'}34), Moment-generating operators for determinants of product moments in samplcs from a normal sy'tem, Annals of Mathf'matics, 35, 312-340. [8.3] Wilks, S. S. (1935), On the independence of k sets of normally distributed statistical v,lfiahles, Econometrica, 3, 309-326. [8.4, 9.3, 9.P] Wishart, John (1928), The generalised product moment distribution in samples from a normal multivariate population, Biometrika, 20A, 32-52. [7.2] Wishart, John (1948), Proofs of the distribution law of the second order moment statistics, Biometrika, 35, 55-57. [7.2] Wishart. John. and M: S. Bartlett (1933), The generalised product moment distribution in a normal system, Proceedings of the Cambridge Philosophical Society, 29, 260-270. [7.2] Wold, H. D. A. (1954), Causality and econometrics, Economdrica 22, 162-177. [15.1] Wold, H. D. A. (1960), A generalization of casual chain models, Econometrica, 28, 443-463. [15.1] Woltz, w. G., W. A Reid, and W. E. Colwell (1948), Sugar and nicotine in cured bright tobacco as related to mineral element composition, Proceedings of the Soil Sciences Society of America, 13, 385-387. [8.P] Wright, Sewall (1921), Correlation and causation, Journal of AgJicultural Research, 20, 557--585. [15.1]
REFERENCES
711
Wright, Sewall (1934), The method of path coefficients, Annals of Mathematical Statistics, S, 161-215. [15.1] Yamauti, Ziro (1977), Concise Statistical Tables, Japanese Standards Association. [Preface] Yates, F., and W. G. Cochran (1938), The analysis of groups of experiments, Journal of Agricultural Science, 28, 556. [8.9] Yule, G. U. (1897a), On the significance of Bravais' formulae for regression & c., in the case of skew correlation, Proceedings of the Royal Society of London, 60, 477-489. [2.5] Yule, G. U. (1897b), On the theory of correlation, Journal of the Royal Statistical Society, 60, 812-854. [2.5] Zehna, P. W. (1966), lnvariance of maximum likelihood estimators, Annals of Mathematical Statistics, 37, 744. [3.2]
It
Index
Absolutely c. .ntinuous distribution, 8 Additivity of Wishart matrices, 259 Admissible, definition of, 88, 210, 235 Admissible proc~dures, 235 Admissible test, definition of, 193 Stein's theorem for, 194 Almost invariant test, 192 Analysis of variance, random effects model,
429 likelihood ratio test of, 431 See also Multivariate analysis of variance Anderson, Edgar, Iris data of, III A posteriori density, H9 of 11-, Hl) of II- and E, 274 of 11-, given i and S, 275 of E, 273 A>ymptotic distribution of a function, 1.~2 Asymptotic expansions of distributions of likelihood ratio criteria, 321 of gamma function, 318 Barley yields in two years, 349 Bartlett decomposition, 257 Bartlett-Nanda-Pillai trace criterion, see Linear hypothesis Bayes estimator of covariance matrix, 273 Bayes estimator of mean vector, 90 Bayes procedure, 89 extended, 90 Bayes risk, 89 Bernoulli polynomials, 318 Best estimator
oi covariance matrix
invariant with respect to triangular linear transformations, 279, 281 proportional to sample covariance matrix. 277
Best linear predictor. 37. 497 and predictand, 497 See also Canonical correlations and variates Best linear unbiased estimator. 298 Beta distribution, 177 Bhattacharya's estimator of the mean, 99 Bivariate normal density, 21, 35 distribution. 21. 35 computation of. 23 Bootstrap method. 135 I 17 Canonical analysis of
rc~rl.!ssiun c~)cfficicnt:-..
50H
sampk.5lll Canonical correlation~ and variate~. -1.87. 49) asymptotic distribution of sample correlalions. ::;0,)
computation of. 501 distribution of sample, 545 invariance of, 496 maximum likelihood estimators of. 501 sample. 500 testing number of nonzero correlations. 504 use in testing hypotheses of rank of covariance matrix. 504 Causal chain. 605 Central limit theorem. multivariate. 86 Characteristic function. 41 continuity theorem for. 45 inversion of. 45 of the multivariate normal distribution. 43 Ch"ractcristic roots and vectors. 631. 632 asympll1lil: disLrihuliolls of. 5-1.5. 559 distribution of roots of a symmetric matrix.
542 distrihutioll of routs llf Wishart matrix. 540
713
714 Characteristic roots and vector (Continued) of Wishart matrix in the metric of another, 529 asymptotic distribution of, 550 distribution of, 536 See also Principal components Chi-squared distribution, 286 non central, 82 Cholesky decomposition, 631 Oassification into nonnal populations Bayes into one of several, 236 into one of two, 204 discriminant function, 218 sample, 220 example, 240 invariance of procedures, 226 likelihood criterion for, 224 unequal covariance matrices, 242 linear, for unequal covariance matrices, 243 admissible, 246 maximum likelihood rule, 226 minimax one of several, 238 one of two, 218 one of several, 237 one of two, 216 W-statistic, 219 asymptotic distribution of, 222 asymptotic expansion of misclassification probabilities, 227 Z-statistic, 226 asymptotic expansion of misciassification probabilities, 231 See also Classification procedures Classification procedures admissible, 210, 235 into several popula'ions, 236 into two populations, 214 a priori probabilities, 209 Bayes, 89, 210 and admissible, 214, 236 into several populations, 234 into two populations, 216 complete class of, 211, 235 essentially, 211 minimal, 211 costs of misclassification, definition of, 208 expected loss from misclassification, 210 minimax, 211 for two populations, 215 probability of misciassification, 210, 227 See also Classification into nonnal populations
INDEX
Cochran's theorem, 262 Coefficient of alienation, 400 Complete class of procedures, 88 essentially, 88 minimal,88 Completeness, definition of, 84 of sample mean and covariance matrix, 84,85 Complex normal distribution, 64 characteristic function of, 65 linear transfonna tion in, 65 maximum likelihood estimators for, 112 Complex Wishart distribution, 287 Components of variance, 429 Concentration ellipsoid, 58, 85 Conditional density, 12 nonnal,34 Conditional probability, 12 Conjugate family of distributions, 272 Consistency, definition of, 86 Contours of constant density, 22 Correlation coefficient canonical, see Canonical correlations and variates confidence interval for, 128 by use of Fisher's z, 135 distribution of sample, asymptotic, 133 bootstrap method for, 135 tabulation of, 126 when population is not zero, 125 when population is zero, 121 distribution of set of sample 272 Fisher'S z, 134 geometric interpretation of sample, 72 invariance of population, 21 invariance of sample, 111 likelihood ratio test, 130 maximum likelihood estimator of, 71 as measure of association, 22 moments of, 166 monotone likelihood ratio of, 164 multiple, see Multiple correlation coefficient partial, see Partial correlation coefficient in the population (simple, Pearson, product-moment, total), 20 sample (Pearson), 71, 116 test of equality of two, 135 test of hypothesis about, 126 by Fisher's z, 134 power of, 128 test that it is zero, 121 Cosine of angle between two vectors, 72. See also Correlation coefficient Covariance, 17
I
i
I
1
INDEX Covariance matrix, 17 asymptotic distribution of sample, 86 Bayes estimator of, 273 characterization of distribution of sample, 77 confidence bounds for quadratic form in, 442 consistency of sample as estimator of population, 86 distribution of sample, 255 estimation, see Best estimator of geometrical interpretation of sample, 72 with linear structure, 113 maximum likelihood estimator of, 70 computation of, 70 when the mean vector is known, 112 of normal distribution, 20 sample, 77 singular, 31 tests of hypotheses, see Testing that a covariance matrix is a given matrix; Testing that a covariance matrix is proportional to a given matrix; Testing that a covariance matrix and mean vector are equal to a given matrix and vector; Testing equality of covariance matrices; Testing e quality of covariance matrices and 'mean vectors; Testing independence of sets of variates unbiased estimator of, 77 Covariance selection models, 614 decomposition of, 618 estimation in, 614 Cramer-Rao lower bound, 85 Critical function, 192 Cumulants, 46 of a multivariate normal distribution, 46 Cumulative distribution function (000, 7 Decision procedure, 88 Degenerate normal distribution, 30 Density,7 conditional, 12 normal,34 marginal,9 normal, 27, 29 multivariate normal, 20 Determinant, 626 derivative, of, 642 symmetric matrix, 642 Dirichlet distribution, 290 Discriminant function, see Classification into normal populations Distance, 631
715 Distance between two populations, 80 Distribution, see Canonical correlations; Characteristic roots; Correlation coefficient; Covariance matrix; Cumulative distribution function; Density; Generalized variance, Mean vector; Multiple correlation coefficient; Multivariate normal density; Multivariate I-distribution; Noncentral chi-squared distribution; Noncentral F-distribution; Noncentral T 2-distribution; Partial correlation coefficient; Principal components; Regression coefficients; T 2-test and statistic; Wishart distribution Distribution of matrix of sums of squares and cross-products, see Wishart distribution Duplication formula for gamma function, 82, 125, 309 $, 9 Efficiency of vector estimate, definition of, 85 Ellipsoid of concentration of vector estimate, 58,85 Ellipsoid of constant density, 32 Elliptically contoured distribution, 47 characteristic function of, 53 characteristic roots and vectors, asymptotic distribution of, 482, 564 correlation coefficient, asymptotic distribution of, 159 covariance of, 50 , covariance of sample covarianc.., 101 covariance of sample mean, 101 cumulants of, 54 kurtosis of, 54 likelihood ratio criterion for equalit) of covariance matrices, asymptotic distribution of, 451 likelihood ratio criterion for ir, denendence of sets, asymptotic distribution of, 406 likelihood ratio criterion for linear hypotheses, asymptotic distribution of, 371 maximum likelihood estimator of parameters, 104 multiple correlation coefficient, asymptotic distribution of, 159 rectangular coordinates, asymptotic distribution of, 283 test for regression coefficients, 371 Elliptically contoured matrix distribution, 104 characteristic roots and vectors, distribution of, 483, 566 likelihood ratio criterion for equality of covariance matrices, distribution of, 454
716 Ellipti-:ally contoured matrix (Continued) likelihood .atio criterion for independence of sets, distribution of, 408 likelihood ratIo criterion for linear hypotheses, distribution of, 373 rectangular coordinates, distribution of, 285 stochastic representation, 160, 285 sufficient statistics, 160 T2_distribution of, 200 Equiangular linc, 72 Exp, 15 Expected value of complex-valued function, 41 Exponential family of distributions, 194 Extended region, 355 Factor analysis, 569 centroid method, 586 communalities, 581 confirmatory, 574 EM algorithm, 580, 593 exploratory, 574 general factor, 570 identification of structure in, 572 by specified zeros, 571, 593 loadings, 570 maximum likelihood estimators, 578 asymptotic distribution of, 582 in case of identification by zeros, 590 nonexistence for fixed factors, 587 for random factors, 583 minimum distance methods, 583 model, 570 oblique factors, 571 orthogonal factors, 571 principal component analysis, relation to, 584 rotation of factors, 571 scores, estimation of, 591 simple structure, 573 space of factors. 572 transformations, 588 tests of fit for, 581 units of measurement, 575 varimax criterion, modified, 589 Factorization theorem, 83 Fisher's z, 133 asymptotic distribution of, 134 moments of, 134 See also Correlation coefficient; Partial correlation coefficient
[/1),257 Gamma function, multivariate, 257
INDEX
Generalized analysis of variance, see Multivariate analysis of variance; Regression coefficients and function Generalized T2, see T 2-test and statistic Generalized variance, 264 asymptotic distribution of sample, 270 distribution of sample, 268 geometric interpretation of sample in N dimensions, 267 in p dimensions, 26K invariance of, 465 moments of sample, 269 sample, 265 General linear hypothesis, see Linear hypothesis, testing of; Regression coeff;cients and function Gram-Schmidt orthogonalization, 252, 647 Graphical modds, 595 adjacent, nonadjacent vertices, 596 AMP (Anderson-Madigan-Perlman) Markov chain, 612 ancestor, 605 boundary, 598 chain graph, 630 chi'd,605 clique, 602 closure, 598 complete, 602 decomposition, 603 descendant, llondescendant, 605 edges, 595 directed, undirected, 596 LWF (L lUritzen-Wermuth-Frydenberg) Markov chain, 610 Markov properties, 597 globally, 600 locally, 598 pairwise, 597 moral graph, liOB nodes, 595 parellt, (l05 partial ordering, 605 path, 600 recursive factorization, 609 separate, 600 vertices, 595 well-numbered, 607 Haar invariant distribution of orthogonal matrices, 162, 541 conditional distribution, 542 Hadamard's inequality, 61 Head lengths and breadths of brothers, 109 Hotelling's T2, see T 2-test and statistic Hypergeometric function, 126
INDEX Incomplete beta function, 329 Independence, 10 mutual,l1 of nonnal variables, 26 of sample mean vector and sample covariance matrix, 77 tests of, see Correlation coefficient; Multiple correlation coefficient; Testing independence of sets of variates Information matrix, 85 Integral of a symmetric unimodal function over a symmetric convex set, 365 Intraclass correlation, 484 Invariance, see Classification into normal populations; Correlation coefficient; Generalized variance; Linear hypothesis; MUltiple correlation coefficient; Partial correlation coefficient; T 2-test; Testing that a covariance matrix is a given matrix; Testing that a covariance ma trix is proportional to a given matrix; Testing equality of covariance matrices; Testing equality of covariance matrices and means vectors; Testing independence of sets of variates Inverted Wishart distribution 272 Iris, four measurements on, J 10, 180 Jacobian, 13 James-Stein estimator, 91 for arbitrary known covariance matrix, 97 average mean squared error of, 95 Kronecker delta, 75 Kronecker product of matrices, 643 characteristic roots of, 643 determinant of, 643 Kurtosis, 54 estimation of, 103 Latin square, 377 Lawley-Hotelling trace crit"rion, see Linear hypothesis Least squares estimator, 295 Likelihood, induced, 71 Likelihood function for sample from multivariate nonnal distribution, 67 Likelihood loss function for covariance matrix, 276 Likelihood ratio test, definition of, 129. See also Correlation coeffiCient; Linear hypothesis; Mean vector; Multiple correlation coefficient; Regression
717 coefficients; T 2-test; Testing that a covariance matrix is given matrix; Testing that a covariance matrix is proportional to given matrix; Testing that a covariance matrix and mean vector are equal to a given matrix and vector; Testing equality of covariance matrices; Testing equality of covariance matrices and mean vectors: Testing independence of sets of variates Linear combinations of normal variables. distribution of, 29 Linear equations, solution of, 606 by Gaussian elimination. 607 Linear functional relationship. 513 relation to simultaneous equations, 520 Linear hypothesis. testing of admissibility of. 353 necessary condition for. 363 Bartlett-Nanda-Pillai trace criterion, 331 admissibility of, 379 asymptotic expansion of distribution of,333 as Bayes procedure, 378 table of significance points of, 673 tabulation of power of, 333 canonical form of, 303 comparison of powers, 334 invariance of criteria, 327 Lawley-Hotelling trace criterion, 328 admissibility of, 379 asymptotic expansion of distribution of,330 monotonicity of power function of. 368 table of significance points of, 657 tabulation of, 328 likelihood ratio criterion. 300 admissibility of, 378 asymptotic expansion of distribution of. 321 as Bayes procedure. 378 distributions of. 306. 3 \0 F-approximation to distribution of. 326 geometric interpretation or. 302 moments of. 309 monotonicity of power function of. 368 nonnal approximation to distribution of,323 table of significance poin ts. 651 tabulation of distribution of. 314 Wilks' A, 300 monotonicity of power function of an invariant test, 363 Roy's maximum root criterion. 333 distribution for p = 2, 334 monotonicity of power function of. 368
718 Linear hypothesis (Conlinlled) table of significance points, 677 tablulation of distribution of. 333 step·down test, 314 See also Regre~sion coefficients and function Linearly independent vectors, 627 Linear transformation of a normal vector, 23, 29.31 Loss. 88 LR decomposition, 630 Mahalanobis distance, 80, 217 sample. 22H Majorization. 355 weak. 355 Marginal density, 9 distribution. 9 normal. 27 l'vlathematical expectation, 9 Ylatri.~. 624 bi diagonal. upper, 503 characteristic roots and vectors of, see Characteristic roots and vectors cofactor in, 627 convexity, 358 definition of. 624 diagonalization of symmetric, 631 doubly stochastic, 646 eigenvalue. see Characteristic roots and vectors Givens, 471, 649 Householder. 470, 650 idempotent. 635 identity. 626 inverse, 627 minor of. 627 nonsingular. 627 operations with, 625 positive definite, 628 positive semidefinite, 628 rank of. 628 symmetriC, 626 trace of. 629 transpose. b25 triangular. 629 tridiagonal. 470 Matrix of sums of squares and cross· products of deviations from the means, 68 Maximum likelihood estimators, see Canonical correlations and variates: Correlation coefficient: Covariance matrix; Mean vector: Multiple correlation coefficient; Partial correlation coefficient: Principal components: Regression coefficients; Variance
INDEX
Maximum likelihood estimator of function of parameters, 71 Maximum of the likelihood function, 70 Maximum of variance of linear combinations, 464. See also Principal components Mean vector, 17 asymptotic normality of sample, 86 completeness of sample as an estimator of' population, 84 confidence region for difference of two when common covariance matrix is known, 80
when covariance matrix is unknown. L80 consistency of sample as estimate of population, 86 distribution of sample, 76 efficiency of sample, 85 improved estimator when covariance matrix is unknown, 185 maximum likelihood estimator of, 70 sample, 67 simultaneous confidence regions for linear functions of, 178 testing equality of, in several distributions, 206 testing equality of two when common covariance matrix is known, 80 tests of hypothesis about when covariance matrix is known, 80 when covariance matrix is unk'lown, see T 2·test See also lames·Stein estimator Minimax, 90 Missing observations, maximum likelihood estimators, 168 Modulus. 13 Moments, 9. 41 factoring of. II from marginal distributions, 10 of normal distributions, 46 Monotone region, 355 in majorization, 355 Multiple correlation coefficient adjusted, 153 distribution of sample conditional, 154 when population correlation is not zero, 156 when population correlation is zero, 150 geometric interpretation of sample, 148 invariance of population, 60 invariancc of sample, 166 likelihood ratio test that it is zero, 151
719
INDEX as maximum correlation between one variable and linear combination of other variables, 38 maximum likelihood estimator of, 147 moments of sample, 156 optimal properties of, 157 population, 38 sample, 145 tabulation of distribution of, 1:7 Multivariate analysis of variancc (MANOVA), 346 Latin square, 377 one-way, 342 two-way, 346 See also Linear hypothesis, testing of Multivariate beta distribution, 377 Multivariate of gamma function, 257 Multivariate normal density, 20 distribution, 20 computation of, 23 Multivariate I-distribution, 276, 289 n(xllL, D, 2.0 N(IL, E), 20
Neyman-Pearson fundamental lemma, 248 Noncentral chi-squared distribution, 82 Noncentral F-distributio'1, 186 tables of, 186 Noncentral T 2-distribution, 186 O(N xp), 161 Orthonormal vedors, 647
Parallelotope, 266 volume of, 266 Partial correlation coefficient computational formulas for, 39, 40, 41 confidence intervals for, 143 distribution of sample, 143 geometric interpretation of sample, 138 invariance of population, 63 invariancc (,f sample, 166 maximum likelihood estimator of, 138 in the population, 35 recursion formula for, 41 sample, 138 tests about, 144 Partial covariancc, 34 estimator of, 137 Partial variance, 34 Partioning of a matrix, 635 addition of, 635 of a covariance matrix, 25 dctcrminant of. 637
inverse of, 638 multiplication of, 635 Partioning of a vector, 635 of a mean vector, 25 of a random vector, 24 Path analysis, 596. See also Graphical models Pearson correlation coefficient, see Correlation coefficient Plane of closest fit, 466 Polar coordinatcs, 285 Positive definite matrix, 628 Positive part of a function, 96 of the James-Stein estimator. Q7 Positive semidefinite matrix, 628 Precision matrix, 272 unbiased estimatllf of, 274 Principal axes of ellipsoids of constant density, 465. See also Principal components Principal components. 459 asymptotic distribution of sample, 473 computation of, 469 confidence region for, 475, 477 distribution of sample, 540, 542 maximum likelihood estimator of, 467 population, 464 testing hypotheses about, 478, 479, 480 Probability element, 8 Product-moment correlation coefficient, see Correlation coefficient QL algorithm, 471 QR algorithm, 471 decomposition, 647 Quadratic form, 628 Quadratic loss function for covariance matrix, 276 r,71 ,1/1 (real part), 257 Random matrix, 16 expected value of, 17 Random vector, 16 Randomized tcst, definition of, 192 Rectangular coordinates, 257 distribution of, 255, 257 Reduced rank rcgression, 514 estimator, asymptotic distribution of, 550 Regression codTicients and function, 34 confidence regions for, 339 distribution of sample, 297 geomctric intcrpretation of sample, 138 maximum likelihood estimator of, 294 partial correlation. connection with, 61
sample. 294
720 Regression coefficients (Continued) simultaneous confidence intervals for, 340, 341 testing hypotheses of rank of, 512 testing they are zero, in case of one dependent variable, 152 Residuals from regression, 37 Risk function, 88
INDEX
Selection of linear combinations, 201 Simple correlation coefficient, See Correlation coefficient Simultaneous equations, 513 estimation of coefficients, 518 least variance ratio (LVR), 519 limited information maximum likelihood (LIMU, 519 two stage least squares (TSLS), 522 identificatirln by zeros, 516 reduced form, 516 estimation of, 517 Singular normal distribution, 30, 31 Singular value decomposition, 498, 634 Spherical distribution, 105 left, 105 right, 105 vector, 105 Spherical normal distribution, 23 Spherically contoured distribution 47 stochastic representation, 49 uniform distribution, 48 Sphericity test, see Testing that a covariance matrix is proportional to a given matrix Standardized sum statistics, 201 Standardized variable, 22 Steifel manifold, 162 Stochastic convergence, 113 of a sequence of random matrices, 113 Sufficiency, definition of, 83 Sufficiency of sample mc:m vector and covariance matrix, 83 Surface area of unit sphere, 286 Surfaces 0' constant density. 22 Symmetric matrix, 626
optimal properties of, 190 power of, 186 tables of, 186 for testing equality of means when covariance matrices are different, 187 for testing equality of two mean vectors when covariance matrix is unknown, 179 for testing symmetry in mean vector, 182 as uniformly most powerful invariant test of mean vector, 191 Testing that a covariance matrix is a given matrix, 438 invariant tests of, 442 likelihood ratio criterion for, 438 modified likelihood ratio criterion for, 438 asymptotic expansion of distribution of, 442 moments of, 440 table of significance points, 685 Nagao's criterion, 442 Testing that a covariance matrix is proportional to a given matrix, 431 invariant tests, 436 likelihood ratio criterion for, 434 admissibility, 458 asymptotic expansion of distribution of, 435 moments of, 434 table of significance points, 682 Nagao's criterion, 437 Testing that a covariance matrix and mean vector are equal to a given matrix and vector, 444 likelihood ratio criterion for, 444 asymptotic expansion of distribution of, 446 moments of, 445 Testing equality of covariance matrices, 412 invarianct tests, 428 likelihood ratio criterion for. ·H3 invariance of, 414 modified likelihood ratio criterion for, 413 admissihility of, 449 asymptotic expansion of distribution of,
T 2-statistic, 176, See also T 2-test and statistic T 2-test and statistic, 173 admissibility of, 196 as Bayes procedure, 199 distribution of statistic, 176 geor.le'ric mterpretation of statistic, 174 invariance of, 173 as likelinood ratio test of mean vector, 176 limiting distribution of, 176 noncentral distribution of statistic, 186
distribution of, 420 moments of, 422 table of significance points, 681 Nagao's criterion for, 415 Testing equality of covariance matrices and mean vectors, 415 likelihood ratio criterion for, 415 asymptotic expansion of distribution of, 426 distribution of, 421
425
721
INDEX
moments of, 422 unbiasedness of,416 Testing independence of sets of variates, 381 and canonical correlations, 504 likelihood ratio criterion for, 384 admissibility of, 401 asymptotic expansion of distribution of, 390 distribution of, 388 invariance of, 386 moments of, 388 monotonicity of power function of, 404 unbiasedness of, 386 Nagao's test, 392 asymptotic expansion of distribution of. 392 stepdown tests, 393 Testing rank of regression matrices, 512 Tests of hypotheses, see Correlation coefficient; Generalized analysis of variance; Linear hypothesis; Mean vector; Multiple correlation coefficient; Partial correlation coefficient; Regression coefficients; T 2-test and statistic Tetrachoric functions, 23
Total correlation coefficient. see Correlation coefficient Lace of a matrix, 629 Transformation of variables. 12 Unbiased estimator, definition of. 77 Unibased test, definition of, 364 Uniform distribution on unit sphere, 48 on O(NXp), 162 Variance, 17 generalized, see Generalized variance maximum likelihood estimator of. 71 w(AI I:, n), 255 weI:, n), 255 w-'(Blw,m),272 w-'(W, m), 272 Wishart distribution, 256 characteristic function of. 258 geometric interpretation of, 256 marginal distributions of, 260 noncentral, 587 for p = 2, 124
z, see Fisher's z Zonal polynomials. 473
WILEY SERIES IN PROBABILITY AND STATISTICS ESTABLISHED BY WALTER A. SHEWHART AND SAMUEL S. WILKS
Editors: David 1. Balding, Noel A. C. Cress ie, Nicholas I. Fisher, lain M. Johnstone, 1. B. Kadane, Louise M. Ryan, David W. Scott, Adrian F. M. Smith, Joze! L. Teugels Editors Emeriti: Vic Barnett, J. Stuart Hunter, David G. Kendall The Wiley Series in Probability and Statistics is well established and authoritative. It covers many topics of current research interest in both pure and applied statistics and probability theory. Written by leading statisticians and institutions, the titles span both state-of-the-art developments in the field and classical methods. Reflecting the wide rangc of cllrrent research in statistics, the series encompasses applied, methodological and theoretical statistics, ranging from applications and new techniques made pOSSible by advances in computeri:
ABRAHAM and LEDOLTER' Statistical Methods for Forecasting AGRESTI' Analysis of Ordinal Categorical Data AGRESTI' An Introduction to Categorical Data Analysis AGRESTI' Categorical Data Analysis, Second Edition ANDEL· Mathematics of Chance ANDERSON' An Introduction to Multivariate Statistical Analysis, Third Edition "ANDERSON' The Statistical Analysis of Time Series ANDERSON, AUQUIER, HAUCK, OAKES, VANDAELE, and WEISBERG' Statistical Methods for Comparative Studies ANDERSON and LOYNES . The Teaching of Practical Statistics ARMITAGE and DAVID (editors) . Advances in Biometry ARNOLD, BALAKRISHNAN, and NAGARAJA . Records "ARTHANARI and DODGE' Mathematical Programming in Statistics "BAILEY' The Elements of Stochastic Processes with Applications to the Natural Sciences BALAKRISHNAN and KOUTRAS . Runs and Scans with Applications BARNETT· Comparative Statistical Inference, Third Edition BARNETT and LEWIS· Outliers in Statistical Data, Third Edition BARTOSZYNSKI and NIEWIADOMSKA-BUGAJ . Probability and Statistical Inference BASILEVSKY' Statistical Factor Analysis and Related Methods: Theory and Applications BASU and RIGDON· Statistical Methods for the Reliability of Repairable Systems BATES and WATTS, Nonlinear Regression Analysis and Its Applications BECHHOFER, SANTNiR, and GOLDSMAN . Design and Analysis of Experiments for Statistical Selection, Screening, and Multiple Comparisons BELSLEY . Conditioning Diagnostics: Collinearity and Weak Data in Regression BELSLEY, KUH, and WELSCH· Regression Diagnostics: Identifying Influential Data and Sources of Collinearity BENDA T and PIERSOL' Random Data: Analysis and Measurement Procedures, Third Edition
"Now available in a lower priced paperback edition in the Wiley Classics Librpry.
BERRY, CHALONER, and GEWEKE . Bayesian Analysis in Statistics and Econometrics: Essays in Honor of Arnold Zellner BERNARDO and SMITH' Bayesian Theory BHA T and MILLER' Elements of Applied Stochastic Processes, Third Edition BHATTACHARYA and JOHNSON' Statistical Concepts and Methods BHATTACHARY A and WAYMIRE· Stochastic Processes with Applications BILLINGSLEY' Convergence of Probability Measures, Second Edition BILLINGSLEY· Probability and Measure, Third Edition BIRKES and DODGE' Alternative Methods ofRegre~sion BLISCHKE AND MURTHY (editors) . Case Studies in Reliability and Maintenance BLlSCHKE AND MURTHY' Reliability: Modeling, Prediction, and Optimization BLOOMFIELD· Fourier Analysis of Time Series: An Introduction, Second Edition BOLLEN' Structural Equations with Latent Variables BOROVKOV . Ergodicity and Stability of Stochastic Processes BOULEAU . Numerical Methods for Stochastic Processes BOX' Bayesian Inference in Statistical Analysis BOX' R. A. Fisher, the Life of a Scientist BOX and DRAPER' Empirical Model-Building and Response Surfaces *BOX and DRAPER' Evolutionary Operation: A Statistical Method for Process Improvement BOX, HUNTER, and HUNTER' Statistics for Experimenters: An Introduction to Design, Data Analysis, and Model Building BOX and LUCENO . Statistical Control by Monitoring and Feedback Adjustment BRAND!MARTE . Numerical Methods in Finance: A MA TLAB-Based Introduction BROWN and HOLLANDER' Statistics: A Biomedical Introduction BRUNNER, DOMHOF, and LANGER· Nonparametric Analysis of Longitudinal Data in Factorial Experiments BUCKLEW' Large Deviation Techniques in Decision, Simulation, and Estimation CAIROLi and DALANG . Sequential Stochastic Optimizatior CHAN' Time Series: Applications to Finance CHATTERJEE and HAD!' Sensitivity Analysis in Linear Regression CHATTERJEE and PRICE' Regression Analysis by Example, Third Edition CHERNICK· Bootstrap Methods: A Practitioner's Guide CHERNICK and FRIIS' Introductory Biostatistics for the Health Sciences CHILES and DELFINER . Geostatistics: Modeling Spatial Uncertainty CHOW and LIU' Design and Analysis of Clinical T:ials: Concepts and Methodologies CLARKE and DISNEY' Probability and Random Processes: A First Course with Applications, Second Edition *COCHRAN and COX, Experimental Designs, Second Edition CONGDON' Bayesian Statistical Modelling CONOVER' Practical Nonparametric Statistics, Second Edition COOK' Regression Graphics COOK and WEISBERG' Applied Regression Including Computing and Graphics COOK and WEISBERG' An Introduction to Regression 'Jraphics CORNELL' Experiments with Mixtures, Designs, Models, and the Analysis of Mixture Data, Third Edition COVER and THOMAS, Elements of Information Theory COX, A Handbook of Introductory Statistical Methods *COX . Planning of Experiments CRESSIE . Statistics for Spatial Data, Revised Edition CSORG6 and HORVATH' Limit Theorems in Change Point Analysis DANIEL' Applications of Statistics to Industrial Experimentation DANIEL' Biostatistics: A Foundation for Analysis in the Health Sciences, Sixth Edition *Now available in a lower priced paperback edition in the Wiley Classics Library.
'DANIEL' Fitting Equations to Data: Computer Analysis of Multifactor Data. Second Edition DASU and JOHNSON· Exploratory Data Mining and Data Cleaning DA VID and NAGARAJ A . Order Statistics, Third Edition 'DEGROOT, FIENBERG, and KADANE . Statistics and the Law DEL CASTILLO' Statistical Process Adjustment for Quality Control DETTE and STUDDEN . The Theory of Canonical Moments with Applications in Statistics, Probability, and Analysis DEY and MUKERJEE . Fractional Factorial Plans DILLON and GOLDSTEIN' Multivariate Analysis: Methods and Applications DODGE· Alternative Methods of Regression 'DODGE and ROMIG' Sampling Inspection Tables, Second Edition 'DOOB . Stochastic Processes DOWDY and WEARDEN . Statistics for Research, Second Edition DRAPER and SMITH· Applied Regression Analysis, Third Edition DRYDEN and MARDIA . Statistical Shape Analysis DUDEWICZ and MISHRA . Modem Mathematical Statistics DUNN and CLARK' Applied Statistics: Analysis of Variance and Regression. Second Edition DUNN and CLARK' Basic Statistics: A Primer for the Biomedical Sciences, Third Edition DUPUIS and ELLIS' A Weak Convergence Approach to the Theory of Large Deviations 'ELANDT-JOHNSON and JOHNSON· Survival Models and Data Analysis ENDERS, Applied Econometric Time Series ETHIER and KURTZ· Markov Processes: Characterization and Convergence EV ANS, HASTINGS, and PEACOCK' Statistical Distributions, Third Edition FELLER' An Introduction to Probability Theory and Its Applications. Volume l. Third Edition, Revised; Volume II, Second Edition FISHER and VAN BELLE' Biostatistics: A Methodology for the Health Sciences 'FLEISS . The Design and Analysis of Clinical Experiments FLEISS . Statistical Methods for Rates and Proportions, Second Editioll FLEMING and HARRINGTON' Counting Pr:Jcesses and Survival Analysis FULLER· Introduction to Statistical Time Series, Second Edition FULLER' Measurement Error Models GALLANT' Nonlinear Statistical Models GHOSH, MUKHOPADHYAY, and SEN' Sequential Estimation GIFI . Nonlinear Multi"ariate Analysis GLASSERMAN and Y AO . Monotone Structure in Discrete-Event Systems GNANADESIKAN . Methods for Statistical Data Analysis of Multivariate Observations. Second Edition GOLDSTEIN and LEWIS' Assessment: Problems, Development, and Statistical Issues GREENWOOD and NIKULIN . A Guide to Chi-Squared Testing GROSS and HARRIS· Fundamentals of Queueing Theory, Third Edition 'HAHN and SHAPIRO, Statistical Models in Engineering HAHN and MEEKER· Statistical Intervals: A Guide for Practitioners HALO, A History of Probability and Statistics and their Applications Before 1750 HALO, A History of Mathematical Statistics from 1750 to 1930 HAMPEL' Robust Statistics: The Approach Based on Influence Functions HANNAN and DEISTLER' The Statistical Theory of Linear Systems HEIBERGER· Computation for the Analysis of Designed Experiments HEDA Y AT and SINHA' Design and Inference in Finite Population Sampling HELLER· MACSYMA for Statisticians HINKELMAN and KEMPTHORNE: . Design and Analysis of Experiments. Volume I: Introduction to Experimental Design 'Now available in a lower priced paperback edition in the Wiley Classics Library.
HOAGLIN. MOSTELLER. and TUKEY· Exploratory Approach to Analysis of Variance HOAGLIN. ~WSTELLER. and TUKEY . Exploring Data Tabl<:s, Trends and Shapes *HOAGLIN. MOSTELLER. and TUKEY . Understanding Robu:;t and Exploratory Data Analvsis HOCHBERG ;nd TAMHANE . MUltiple Comparison Procedures HOCKING· Methods and Applications of Linear Models: Regr,ession and the Analysis of Variance. Seculld EdiTion HOEL . Introduction to Mathematical Statistics, Fifth Edition HOGG and KLUGMAN' Loss Distributions HOLLANDER and WOLFE· Nonparametric Statistical Methods, Second Edition HOSMER and LEMESHOW . Applied Logistic Regression, Second Edition HOSMER and LEMESHOW . Applied Survival Analysis: Regression Modeling of Time to Event Data HOYLAND and RAUSAND . System Reliability Theory: M0dels and Statistical Method:; HUBER· Robust Statistics HUBERTY· Applied Discriminant Analysis H1..:1'\T and KENNEDY· Financial Derivatives in Theory and Practice HUSKOV A, BERAN, and DUPAC . Collected Works of Jaroslav Hajekwith Commentary I1vlAN and CONOVER· A Modem Approach to Statistics J.ACKSON . A User's Guide to Principle Components JOHN· Statistical Methods in Engineering and Quality Assurance JOHl'iSON . Multivariate Statistical Simulation JOHNSON and BALAKRISHNAN' Advances in the Theory and Practice of Statistics: A Volume in Honor of Samuel Kotz JUDGE. GRifFITHS. HILL. LUTKEPOHL. and LEE' The Theory and Practice of Econometrics. St'colld EdiTion JOHNSON and KOTZ . Distributions in Statistics JOHNSON and KOTZ (editors) . Leading Personalities in Statistical Sciences: From the Seventeenth Century to the Present JOHl'\SON. KOTZ. and BALAKRISHNAN' Continuous Univariate Distributions, Volume I. Second EdiTion JOHNSON. KOTZ, and BALAKRISHNAN' Continuous Univariate Distributions, Volume 2. Second EdiTion JOHNSON. KOTZ, and BALAKRISHNAN' Discrete Multivariate Distributions JOHNSON. KaTZ. and KEMP, Univariate Discrete Distributions, Second Edition JUREC KOV A and SEN' Robust Statistical Procedures: Aymptotics and Interrelations JUREK and MASON' Operator-Limit Distributions in Probability Theory KADANE . Bayesian Methods and Ethics in a Clinical Trial Design KADANE AND SCHUM' A Probabilistic Analysis of the Sacco and Vanzetti Evidence KAl.BFLEISCH and PRENTICE' The Statistical Analysis of Failure Time Data, Second EdiTion KASS and \"OS . Geometrical Foundations of Asymptotic Inference KAUFMA'J and ROUSSEEUW . Finding Groups in Data: An Introduction to Cluster Analysis KEDE~I and FOKlANOS . Regression Models for Time Series Analysis KENDAl.L. BARDEN. CARNE, and LE . Shape and Shape Theory KHURI . Advanced Calculus with Applications in Statistics, Second Edition KHURI. MATHEW. and SINHA' Statistical Tests for Mixed l.inear Models KLEIBER and KOTZ . Statistical Size Distributions in Economics and Actuarial Sciences KLUGMAN. PANJER, and WILLMOT' Loss Models: From Data to Decisions KLUGMAN. PANJER, and WILL MOT . Solutions Manual to Accompany Loss Models: From Data to Decisions *Now available in a lower priced paperback edition in the Wiley Classics Library.
KOTZ, BALAKRISHNAN, and JOHNSON' Continuous Multivariate Distributions, Volume. I, Second Edition KOTZ and JOHNSON (editors) . Encyclopedia of Statistical Sciences: Volumes I to 9 with Index KOTZ and JOHNSON (editors) . Encyclopedia of Statistical Sciences: Supplement Volume KOTZ, READ, and BANKS (editors) . Encyclopedia of Statistical Sciences: Update Volume I KOTZ, READ, and BANKS (editors) . Encyclopedia of Statistical Sciences: Update ~=2
.
KOVALENKO, KUZNETZOV, and PEGG' Mathematical Theory of Reliability of Time-Dependent Systems with Practical Applications LACHIN· Biostatistical Methods: The Assessment of Relative Risks LAD· Operational Subjective Statistical Methods: A Mathematical, Philosophical, and Historical Introduction LAMPERT! . Probability: A Survey of the Mathematical Theory, Second Edition LANGE, RYAN, BILLARD, BRILLINGER, CONQUEST, and GREENHOUSE· Case Studies in Biometry LARSON' Introduction to Probability Theory and Statistical Inference, Third Edition LAWLESS' Statistical Models and Methods for Lifetime Data, Second Edition LAWSON' Statistical Methods in Spatial Epidemiology LE . Applied Categorical Data Analysis LE . Applied Survival Analysis LEE and WANG' Statistical Methods for Survival Data Analysis, Third Edition LEP AGE and BILLARD' Exploring the Limits of Bootstrap LEYLAND and GOLDSTEIN (editors) . Multilevel Modelling of Health Statistics L1AO Statistical Group Comparison LINDVALL' Lectures on the Coupling Method LINH.!\RT and ZUCCHINI' Model Selection LITTLF. and RUBIN· Statistical Analysis with Missing Data, Second Edition LLOYD· The Statistical Analysis of Categorical Data MAGNUS and NEUDECKER' M1trix Differential Calculus with Appl;cations in Statistics and Econometrics, RevisedEdition MALLER and ZHOU . Survival Analysis with Long Term Survivors MALLOWS' Design, Data, and Analysis by Some Friends of Cuthbert Daniel MANN, SCHAFER, and SINGPURWALLA' Methods for Statistical Analysis of Reliability and Life Data MANTON, WOODBURY, and TOLLEY' Statistical Applications Using Fuzzy Sets MARDIA and JUPP . Directional Statistics MASON, GUNST, and HESS' Statistical Design and Analysis of Experiments with Applications to Engineering and Science, Second Edition McCULLOCH and SEARLE' Generalized, Linear, and Mixed Models McFADDEN . Management of Data in Clinical Trials McLACHLAN' Discriminant Analysis and Statistical Pattern Recognition McLACHLAN and KRISHNAN· The EM Algorithm and Extensions McLACHLAN and PEEL' Finite Mixture Models McNEIL' Epidemiological Research Methods MEEKER and ESCOBAR' Statistical Methods for Reliability Data MEERSCHAERT and SCHEFFLER' Limit Distributions for Sums oflndependent Random Vectors: Heavy Tails in Theory and Practice *MILLER . Survival Analysis, Second Edition MONTGOMERY, PECK, and VINING' Introduction to Linear Regression Analysis, Third Edition . MORGENTHALER and TUKEY . Configural Polysampling: A Route to Practical Robustness *Now available in a lower priced paperback edition in the Wiley Classics Library.
MUIRHEAD, Aspects of Multivariate Statistical Theory MURRAY, X-STAT 2.0 Statistical Experimentatior. Design Data Analysis, and Nonlinear Optimization MYERS and MONTGOMERY, Response Surface Methodology: Process and Product Optimization Using Designed Experiments, Second Edition MYERS, MONTGOMERY, and VINING, Generalized Linear Models. With Applications in Engineering and the Sciences NELSON, Accelerated Testing, Statistical Models, Test Plans, and Data Analyses NELSON· Applied Life Data Analysis NEWMAN, Biostatistical Methods in Epidemiology OCHI . Applied Probability and Stochastic Processes in Engineering and Physical Sciences OKABE, BOOTS, SUGIHARA, and CHiU . Spatial Tesselations: Concepts and Applications of Voronoi Diagrams, Second Edition OLIVER and SMITH, Influence Diagrams, Belief Nets and Decision Analysis PANKRATZ, Forecasting with Dynamic Regression Models PANKRATZ' Forecasting with Univariate Box-Jenkins Models: Concepts and Cases "PARZEN . Modern Probability Theory and lis Applieations PENA, TlAO, and TSA Y . A Course in Time Series Analysis PIANTADOSI . Clinical Trials: A Methodologic Perspective PORT, Theoretical Probability for Applications POURAHMADI . Foundations of Time Series Analysis and Prediction Theory PRESS, Bayesian Statistics: Principles, Models, and Applications PRESS' Subjective and Objrctive Bayesian Statistics, Second Edition PRESS and TANUR' The Subjectivity of Scientists and the Bayesian Approach PUKELSHEIM . Optimal Experimental Design PURl, VILAPLANA, and WERTZ· New Perspectives in Theoretical and Applied Statistics PUTERMAN . Markov Decision Processes: Discrete Stochastic Dynamic Programming "RAO . Linear Statistical Inference and Its Applications, Second Edition RENCHER, Linear Models in Statistics RENCHER, Methods of Multivariate Analysis, Second Edition RENCHER, Multivariate Statistical Inference with Applications RIPLEY, Spatial Statistics RIPLEY, Stochastic Simulation ROBINSON, Practical Strategies for Experimenting ROHA TGJ and SALEH' An Introduction to Probability and Statistics, Second Edition ROLSKI, SCHMIDLI, SCHMIDT, and TEUGELS . Stochastic Processes for Insurance and Finance ROSENBERGER and LACHIN . Randomization in Clinical Trials: Theory and Practice ROSS, Introduction to Probability and Statistics for Engineers and Scientists ROUSSEEUW and LEROY· Robust Regression and Outlier Detection RUBIN' MUltiple Imputation for Nonresponse in Surveys RUBINSTEIN· Simulation and the Monte Carlo Method RUBINSTEIN and MELAMED· Modem Simulation and Modeling RY AN . Modem Regression Methods RYAN, Statistical Methods for Quality Improvement, Second Edition SALTELLI, CHAN, and SCOTT (editors)' Sensitivity Analysis "SCHEFFE . The Analysis of Variance SCHIMEK, Smoothing and Regression: Approaches, Computation, and Application SCHOTT, Matrix Analysis for Statistics SCHUSS, Theory and Applications of Stochastic Differential Equations SCOTT, Multivariate Density Estimation: Theory, Practice, and Visualization "SEARLE, Linear Models SEARLE' Linear Models for Unbalanced Data "Now avail.,ble in a lower priced paperback edition in the Wiley Classics Library.
SEARLE· Matrix Algebra Useful for Statistics SEARLE, CASELLA, and McCULLOCH' Variance Components SEARLE and WILLETT. Matrix Algebra for Applied Economics SEBER and LEE· Line; r Regression Analysis, Second Edition SEBER' Multivariate Observations SEBER and WILD' Nonlinear Regression SENNOTT' Stochastic Dynamic Programming and the Control of Queueing Systems ·SERFLING . Approximation Theorems of Mathematical Statistics SHAFER and VOVK . Probability and Finance: It's Only a Game! SMALL and McLEISH' Hilbert Space Methods in Probability and Statistical Inference SRIV ASTAV A . Methods of Multivariate Statistics ST APLETON . Linear Statistical Models STAUDTE and SHEATHER· Robust Estimation and Testing STOY AN, KENDALL, and MECKE . Stochastic Geometry and Its Applications. Second Edition STOY AN and STOY AN . Fractals, Random Shapes and Point Fields: Methods of Geometrical Statistics STY AN . The Collected Papers ofT. W. Anderson: 1943-19R5 SUTTON, ABRAMS, JONES, SHELDON, and SONG' Methods for Meta-Analysis in Medical Research TANAKA' Time Series Analysis: Nonstationary and Noninvertible Distribution Theory THOMPSON' Empirical Model Building THOMPSON' Sampling, Second Edition THOMPSON' Simulation: A Modeler's Approach THOMPSON and SEBER' Adaptive Sampling THOMPSON, WILLIAMS, and FINDLA Y . Models for Investors in Rea I World Markets TlAO, BISGAARD, HILL, PENA. and STIGLER (editors) . Box on Quality and Discovery: with Design, Control, and Robustness TIERNEY' L1SP-ST AT: An Object-Oriented Environment for Statistical Computing and Dynamic Graphics TSA Y . Analysis of Financial Time Series UPTON and FINGLETON . Spatial Data Analysis by Example. Volume II: Categorical and Directional Data V AN BELLE' Statistical Rules of Thumb VIDAKOVIC· Statistical Modeling by Wavelets WEISBERG' Applied Linear Regression. Second Edition WELSH' Aspects of Statistical Inference WESTFALL and YOUNG' Resampling-Based Multiple Testing: Examples and Methods for p-Value Adjustment WHITTAKER' Graphical Models in Applied Multivariate Statistics WINKER' Optimization Heuristics in Economics: Applications of Threshold Accepting WOl\,\lACOTT and WONNACOTT' Econometrics, Second Edition WOODING' Planning Phannaceutical Clinical Trials: Basic Statistical Principles WOOLSON and CLARKE' Statistical Methods for the Analysis of Biomedical Data. Seconc-l Edition WU and HAMADA· Experiments: Planning, Analysis. and Parameter Design Optimization YANG' The Construction Theory of Denumerable Markov Processes *ZELLNER' An Introduction to Bayesian Inference in Econometrics ZHOU, OBUCHOWSKI, and McCLISH· Statistical Methods in Diagnostic Medicine
"Now available in a lower priced paperback edition in the Wiley Classics Library.