Frontiers in Health Policy Research, Vol. 5

Frontiers in Health Policy Research 5

This page intentionally left blank

Frontiers in Health Policy Research 5

edited by Alan M. Garber

National Bureau of Economic Research Cambridge, Massachusetts The MIT Press Cambridge, Massachusetts London, England

Frontiers in Health Policy Research, 5,2002 ISSN: 1096-231x E-ISSN 1537-2634 ISBN: Hardcover 0-262-07234-3 Paperback 0-262-57156-0 Published annually by The MIT Press, Cambridge, Massachusetts 02142 An electronic, full-text version of Frontiers in Health Policy Research is available from MIT Press Journals when purchasing a subscription. Subscription Rates Hardcover/Print and Electronic: $58.00 Paperback/Print and Electronic: $25.00 Outside the U.S. and Canada add $10.00 for postage and handling. Canadians add 7% GST. Subscription and address changes should be addressed to: MIT Press Journals, Five Cambridge Center, Cambridge, MA 02142-1407, phone 617-253-2889; fax 617-577-1545; email: [email protected]. Claims will be honored free of charge if made within three months of the publication date of the issue. Claims may be submitted to [email protected]. Prices are subject to change without notice. In the United Kingdom, continental Europe, and the Middle East and Africa, send back volume orders and business correspondence to: The MIT Press, Ltd., Fitzroy House, 11 Chenies Street, London WC1E 7ET England, phone 44-020-7306-0603, fax 44-020-7306-0604, email [email protected] In the United States and for all other countries, send single copy and back volume orders to: The MIT Press, Five Cambridge Center, Cambridge, MA 02142, toll-free book orders 800-356-0343, fax 617-625-6660, email [email protected] Copyright Information Permission to photocopy articles for internal or personal use, or the internal or personal use of specific clients, is granted by the copyright owner for users registered with the Copyright Clearance Center (CCC) Transactional Reporting Service, provided that the fee of $10.00 per copy is paid directly to CCC, 222 Rosewood Drive, Danvers, MA 01923. The fee code for users of the Transactional Reporting Service is: 1096-231x/02 $10.00. For those organizations that have been granted a photocopy license with CCC, a separate system of payment has been arranged. © 2002 by the National Bureau of Economic Research and the Massachusetts Institute of Technology.

National Bureau of Economic Research

Officers Carl F. Christ, Chairman Michael H. Moskow, Vice Chairman Martin Feldstein, President and Chief Executive Officer Susan Colligan, Vice President for Administration and Budget and Corporate Secretary Robert Mednick, Treasurer Kelly Horak, Controller and Assistant Corporate Secretary Gerardine Johnson, Assistant Corporate Secretary

Directors at Large Peter C. Aldrich Elizabeth E. Bailey John H. Biggs Andrew Brimmer Carl F. Christ John S. Clarkeson Don R. Conlan George C. Eads Martin Feldstein Stephen Friedman Judith M. Gueron George Hatsopoulos Karen N. Horn Judy C. Lewent John Lipsky Michael H. Moskow Alicia H. Munnell Rudolph A. Oswald Robert T. Parry Peter G. Peterson Richard N. Rosett Kathleen P. Utgoff Marina v. N. Whitman Martin B. Zimmerman

Directors by University Appointment George Akerlof, California, Berkeley Jagdish Bhagwati, Columbia William C. Brainard, Yale Michael J. Brennan, California, Los Angeles Glen G. Cain, Wisconsin Franklin Fisher, Massachusetts Institute of Technology Saul H. Hymans, Michigan Marjorie B. McElroy, Duke Joel Mokyr, Northwestern Andrew Postlewaite, Pennsylvania Nathan Rosenberg, Stanford Michael Rothschild, Princeton Craig Swan, Minnesota David B. Yoffie, Harvard Arnold Zellner, Chicago Directors by Appointment of Other Organizations Mark Drabenstott, American Agricultural Economics Association Gail D. Fosler, The Conference Board A. Ronald Gallant, American Statistical Association Robert S. Hamada, American Finance Association Robert Mednick, American Institute of Certified Public Accountants Angelo Melino, Canadian Economics Association Richard D. Rippe, National Association for Business Economics John J. Siegfried, American Economic Association

vi

National Bureau of Economic Research

David A. Smith, American Federation of Labor and Congress of Industrial Organizations Josh S. Westeon, Committee for Economic Development Gavin Wright, Economic History Association

Directors Emeriti Thomas D. Flynn Lawrence R. Klein Franklin A. Lindsay Paul W. McCracken Bert Seidman Eli Shapiro Since this volume is a record of conference proceedings, it has been exempted from the rules governing critical review of manuscripts by the Board of Directors of the National Bureau (resolution adopted 8 June 1948, as revised 21 November 1949 and 20 April 1968).

Contents

Acknowledgments Introduction xi

ix

Alan M. Garber 1 CMS Payments Necessary to Support HMO Participation in Medicare Managed Care 1 John Cawley, Michael Chernew, and Catherine McLaughlin 2 The Effects of Medicare on Health Care Utilization and Outcomes 27 Frank R. Lichtenberg 3 Effects of Competition Under Prospective Payment on Hospital Costs Among High- and Low-Cost Admissions: Evidence from California, 1983 and 1993 53 David Meltzer and Jeanette Chung 4 Tax Credits, the Distribution of Subsidized Health Insurance Premiums, and the Uninsured 103 Mark V. Pauly, Bradley Herring, and David Song 5 Hospital Ownership Conversions: Defining the Appropriate Public Oversight Role 123 Frank A. Sloan

This page intentionally left blank

Acknowledgments

I am grateful to the NBER administrative staff for their help in organizing the conference and publications; staff members include Lita Kimble, Rob Shannon, and Helena Fitz-Patrick. The National Institute on Aging, the Agency for Healthcare Research and Quality, and the Robert Wood Johnson Foundation supported much of the research reported here.

This page intentionally left blank

Introduction

The fifth annual Frontiers in Health Policy Research Conference, held in Bethesda, Maryland, on June 7, 2001, brought together academic economists and health policy experts from Washington, including researchers, legislative staff, and government officials. The papers presented at that conference are gathered in this volume, which presents impartial, cutting edge research that is directly relevant to contemporary health policy debates. Three papers focused on critical issues facing the Medicare program. For many years, the adoption of capitated risk plans has been proposed as a solution to many of Medicare's current and future problems. Prominent Medicare reform proposals, such as the premium support plan proposed by the Medicare Reform Commission, build on a foundation of beneficiary choice among competing health plans. Capitated plans are particularly attractive because they are expected to reduce costs, coordinate care, and provide enhanced services to Medicare beneficiaries. Such plans have been available for several years, sometimes under the name of Medicare risk plans or Medicare+Choice plans. Despite the great hopes for such plans, the number of Medicare beneficiaries who have enrolled in the plans has fallen far short of expectations. In fact, in the wake of the Balanced Budget Act of 1997, enrollment has decreased instead of increased. Although the disappointing enrollment reflects in part a lack of demand by beneficiaries, the well-publicized withdrawals of Medicare risk plans have also impeded the growth of capitation. This impediment, according to Cawley, Chernew, and McLaughlin, results directly from inadequate reimbursement. They estimate the level of capitation payments necessary to make it profitable for Medicare risk plans to be offered in a county. Capitation payments may

xii

Introduction

need to be particularly high in sparsely populated counties, but they may also need to exceed current payment levels in other areas. According to the authors, premium payments in nearly 80 percent of the counties in their sample are now too low to support the availability of a single Medicare managed-care plan. Because it provides nearly universal health insurance for a large and vulnerable segment of the U.S. population, Medicare would be expected to have large effects on the health and welfare of elderly and disabled beneficiaries. Furthermore, because the largest group of beneficiaries becomes eligible for Medicare simply by reaching their sixty-fifth birthday, one would expect to see immediate changes in the utilization of health services at age 65. Lichtenberg asks whether the age profiles of utilization, morbidity, and mortality reveal that Medicare is having a great impact on health. His work reveals that the effects of Medicare eligibility are not only detectable but surprisingly large. The evidence he examines suggests that Medicare does increase the utilization of medical services, as expected, and that this increased use is associated with improved health outcomes. Because evidence of market failure is ubiquitous, health care markets are often cited as exceptions to general economic rules. One such exception may be competition among providers of health care. When there are more hospitals in an area, can we expect lower prices for their services? Since the hospitals may compete on quality as well as price, do we observe measurably higher hospital quality in markets characterized by a high level of competition? How would the answers to these questions change when hospitals are compensated for a fixed fee per admission (prospective payment), rather than receiving compensation for each service they provide? Meltzer and Chung note that earlier studies reported competition may increase hospital costs when hospitals are reimbursed on a fee-for-service basis, and it may have the opposite effect under prospective payment. These questions are critically important for policies regarding hospital competition, which might either raise or lower costs to Medicare and to consumers. The two authors raise the possibility that under prospective payment, competition might lower costs for patients who are unprofitable and raise costs for patients who are profitable. They address this question by examining data on hospital charges and cost-to-charge ratios from California in two different years, one just before implementation of Medicare's Prospective Payment System (1983), the other ten years later (1993). Classifying the degree of hospital competition within each county into four

Introduction

xiii

categories, and focusing on the twelve highest-volume diagnostic categories, they report that increased competition led to an increase in cost growth in 1983 among the high-cost patients within these diagnoses, but had the opposite effect among these high-cost Medicare beneficiaries in 1993. They also find that cost reductions are largest for the most expensive patients. Under many plans to extend health insurance coverage to the uninsured, including the Bush administration's proposal, subsidies would be used to enable the poor and the near-poor to purchase private health insurance. The subsidies would be administered in the form of refundable tax credits. One of the most controversial aspects of the tax credit approach is the size of the tax credit that would be needed to achieve a substantial increase in the number of Americans with health insurance. According to some experts, only prohibitively large subsidies would have the desired effect, but other work has shown that tax credits large enough to cut health insurance premiums in half would also cut in half the number of uninsured. Pauly, Herring, and Song address the effects of tax credits by asking how a flat tax credit of $1,000 would affect net premiums (individual market health insurance premiums minus the subsidy) and the uptake of health insurance. Their work uses several measures of net premiums to approximate more closely the premiums that the target population of insurance nonpurchasers face, and they estimate the distribution of insurance purchases based on the resulting net premiums. In part because their measures of premium costs are lower than those used in prior analyses, they find that the $1,000 tax credit would result in a surprisingly large increase in insurance purchases. A flat tax credit would do less for high-risk individuals than a risk-adjusted tax credit, and there are many questions about the risk profile of the individuals who would begin to purchase insurance under such a program. Depending on the risk profiles, as perceived by health plans, the net premiums might be either lower or higher than Pauly and his colleagues estimate. The roles of for-profit and nonprofit institutions in health care continue to be hotly debated. Critics of for-profit hospitals and insurers argue that for-profit corporations provide lower quality care, shun the most vulnerable patients, and raise costs of health care. Critics of nonprofits argue that they are less efficient and provide lower quality of care; another point of view holds that market competition forces nonprofits and for-profits to behave in similar ways. To the extent that for-profits and nonprofits behave differently, conversions between the

xiv

Introduction

two categories can have important welfare implications. Sloan argues that recent increases in the for-profit share of hospitals, resulting from hospital closings, mergers, and ownership changes, have the potential to alter hospital performance. After reviewing the literature on the relationship between hospital ownership and behavior, he analyzes data on hospital conversions from 1988-1996 to determine whether for-profit conversion affects the quality of care or costs. By examining utilization and inpatient mortality for selected diseases, Sloan finds that for-profit conversions are associated with reduced lengths of stay, but mortality remains unchanged. He also finds that pneumonia complication rates became more common after for-profit conversion. Whether this finding is a signal of general problems with such conversions or whether this is an isolated result remains uncertain. However, the study finds little other evidence of major effects of for-profit conversions on outcomes.

1 CMS Payments Necessary to Support HMO Participation in Medicare Managed Care John Cawley, Cornell University andNBER Michael Chernew, University of Michigan andNBER Catherine McLaughlin, University of Michigan

Executive Summary In recent years, many health maintenance organizations (HMOs) have exited the market for Medicare managed care; since 1998, the number of participating plans has fallen from 346 to 174. As a result of this reduced participation by HMOs, hundreds of thousands of Medicare beneficiaries have been involuntarily disenrolled from the program at the end of each year from 1998 to 2001. This paper estimates the Centers for Medicare and Medicaid Services (CMS)1 capitation payments that are necessary to support the participation of various numbers of HMOs in Medicare managed care per county market. This paper does not make a normative statement about how many HMOs should be supported in this program; rather, it makes a positive statement about the levels of payment necessary to support various numbers of HMOs. The identification strategy is to observe how the number of participating HMOs varies over counties and time in response to CMS payment, while controlling for estimated costs. This paper studies the period 1993-2001 and focuses in particular on the variation in payment, independent of costs, that occurred as a result of the Balanced Budget Act of 1997, which dramatically changed the way that HMOs are paid in this program. In light of the fact that it may not be cost-effective for CMS to support HMO participation in relatively rural or unpopulated counties, the sample used in this paper is limited to the 60 percent of U.S. counties with the largest populations of Medicare beneficiaries. The ordered probit results presented in this paper indicate that, to support one Medicare HMO in 2001 in half of the counties in the sample, CMS would have to pay $682.08 per average enrollee per month in the marginal county. To support one Medicare HMO in 2001 in every county in the sample, CMS would need to pay $1,008.25 per enrollee per month in the maximum-payment county. For comparison, the maximum monthly payment paid by CMS to any county in 2001 was $833.55. This paper finds that 79.3 percent of counties in the sample received a CMS payment in 2001 that was less than what was necessary to support a single HMO in Medicare managed care. Compared to those counties that received a payment exceeding the estimated threshold for HMO participation, these

2

Cawley, Chernew, and McLaughlin

counties are, on average, more rural and less populated, with citizens who are less wealthy and less educated. The relative disadvantage of rural and unpopulated counties persists three years after the Balanced Budget Act of 1997, designed in part to eliminate such disparities, took effect.

I. Introduction This paper studies how the equilibrium number of health maintenance organizations (HMOs) participating in county Medicare managed care markets varies with the Centers for Medicare and Medicaid Services (CMS) capitation payment. The number of HMOs participating in Medicare managed care markets is of interest for several reasons. The participation of a single HMO in a Medicare managed care market offers Medicare beneficiaries in that market an alternative to fee-forservice care. The participation of multiple HMOs in a market creates competition for enrollment, which results in greater benefits and/or lower costs for managed care enrollees.2 This paper does not take a position on how many HMOs should be supported in this program in different areas of the United States; that question is left for policy makers. Instead, this paper seeks to provide the best estimate of the levels of payment necessary to support various numbers of HMOs. The identification strategy of this paper is to examine how the number of participating HMOs in this program varies over counties and time in response to CMS payment, controlling for estimated costs. In particular, variation in payment independent of costs occurs because of the Balanced Budget Act of 1997, which dramatically changed the way that HMOs are paid in this program. Those eligible for Medicare Part A (Hospital Insurance) and enrolled in Medicare Part B (Supplementary Medical Insurance) may enroll in a Medicare managed care plan, if one is available.3 Figure 1.1 depicts the number of Medicare managed care enrollees from 1985 to 2001, a period during which enrollment grew from 0.44 million in 1985 to 6.35 million in 1999, before falling to 5.6 million in 2001.4 In 2001,15 percent of all Medicare beneficiaries chose managed care.5 The continuous growth in enrollment between 1985 and 1999 masks considerable change in the number of HMO plans participating in Medicare managed care. Figure 1.2 shows that the number of participating plans rose considerably during the early and mid-1990s, but fell from 346 to 174 between 1998 and 2001. This fall in plan participation

CMS Payments

3

Figure 1.1 Medicare managed care enrollment

coincides with the period when the provisions of the Balanced Budget Act of 1997 were in effect. As a result of the reduced participation of HMOs, many Medicare beneficiaries have been involuntarily disenrolled from the program. At the end of 1998, 407,000 (or 7 percent of all) Medicare HMO enrollees were involuntarily disenrolled, and 327,000 (5.3 percent) were involuntarily disenrolled at the end of 1999.6 It is estimated that 934,000 enrollees (15.1 percent) were disenrolled at the end of the year 20007 Beneficiaries involuntarily disenrolled from a Medicare managed care plan are forced either to find another HMO in their county with a risk contract from Medicare or to return to traditional fee-for-service Medicare. Laschober et al. (1999) surveyed Medicare beneficiaries whose HMO had recently withdrawn from Medicare. They found that one-third experienced a decline in benefits, 39 percent reported higher monthly premiums, and one in seven lost prescription drug coverage. It may not be cost-effective for CMS to support HMO participation hi Medicare managed care in relatively rural or unpopulated counties; for this reason, the sample used in this paper is limited to the 60 percent of U.S. counties with the largest population of Medicare beneficiaries. Our estimates indicate that 79.3 percent of counties in this

4

Cawley, Chernew, and McLaughlin

Figure 1.2 Number of Medicare managed care plans

sample received a CMS payment less than what was necessary to support a single HMO in the Medicare managed care market. In particular, CMS appears to underestimate the payment necessary to support HMOs in rural and less populous areas. Section II of this paper outlines the methodology for examining the relationship between CMS payment rates and the extent of HMO participation. Section III describes the data used in this study; Section IV presents the results of the empirical work, and the final section presents our conclusions. II. Methodology To illustrate why higher payments may lead to a larger number of participating HMOs, suppose that the Medicare managed care market is in equilibrium, and then CMS raises the payment to HMOs while costs remain constant. The payment has been raised above the marginal cost of caring for additional Medicare beneficiaries, so HMOs will compete to increase enrollment and therefore profits. HMOs compete for enrollment by increasing benefits (and, therefore, marginal and average costs). The provision of additional benefits raises the cost curves; in particular, the average cost curve will rise to equal the new, higher payment. The provision of additional benefits makes Medicare managed

CMS Payments

5

care more attractive relative to fee-for-service Medicare; this shifts the demand curve for Medicare managed care and, as a result, the new equilibrium will be associated with a higher quantity of enrollment. Since marginal costs are rising in enrollment, the higher enrollment may be associated with a larger number of participating HMOs and cannot be associated with fewer participating HMOs. Congress, in the Balanced Budget Act (BBA) of 1997, changed CMS's formula for setting payment levels effective in 1998. Prior to 1998, county CMS payments were set according to the 1982 Tax Equity and Fiscal Responsibility Act (TEFRA). Under TEFRA, HMOs were paid 95 percent of the projected average fee-for-service costs of Medicare beneficiaries in that county, multiplied by a risk-adjustment factor based on the enrollee's age, sex, Medicaid eligibility, institutional status, and working status. The TEFRA payment formula was criticized for overpaying HMOs. Despite the strategy of paying HMOs 95 percent of projected average fee-for-service costs, several studies concur that it cost CMS more to enroll beneficiaries in managed care than if they had remained in fee-for-service Medicare. The reason is that enrollees in Medicare managed care have proven to be systematically healthier than fee-forservice Medicare beneficiaries. As a result, the medical expenses of the Medicare managed care enrollees were far lower than 95 percent of average fee-for-service costs.8 The TEFRA payment formula was also criticized for creating disparities in payments across counties; in particular, few HMOs entered rural counties. It was argued that tying managed care payments to local fee-for-service charges rewarded counties that were inefficient at providing fee-for-service care and counties with high reimbursements for graduate medical education, which are included in the fee-for-service costs. Concerned about the rising cost of caring for Medicare beneficiaries, Congress passed the BBA of 1997, which created the Medicare + Choice program (M+C) and changed the way that HMOs are reimbursed for risk contracts.9 Under M+C, CMS, beginning in 1998, pays HMOs the greatest of the following three rates10: 1. A blend of an input-price adjusted national rate and an area-specific rate; however, if total projected payments exceed a budget limit, this blended rate is reduced. The blend is intended to reduce the variation in payments across counties by increasing the lowest rates and decreasing the highest rates.

6

Cawley, Chernew, and McLaughlin

2. A minimum or "floor" payment, adjusted annually, intended to increase rates in historically lower-rate counties where Medicare managed care plans generally have not been offered. 3. A minimum increase over the previous year's payment, which is intended to protect high payment areas. For 1998,1999, and January and February of 1999, the minimum increase over the previous year's payment was 2 percent. Since March 2001, the minimum increase is 3 percent. Since the BBA took effect, the budget limits have typically been binding, forcing reductions in the blended rate. These reductions have been so great that only in the year 2000 did any county receive the blended payment. The BBA also affected HMO profits by increasing their administrative burdens and charging them user fees (which amounted to $95 million in both 1998 and 1999), the proceeds of which are used to inform Medicare beneficiaries about their managed care options. There is one final component of CMS payments to HMOs. The Balanced Budget Refinement Act of 1999 mandates that CMS, starting in the year 2000, pay bonuses of 5 percent the first year and 3 percent the second year to HMOs that offer Medicare+Choice in previously unserved counties.11 Three studies have modeled the decisions of individual HMOs to participate in the Medicare managed care market (Adamache and Rossiter 1986, Porell and Wallack 1990, and Abraham et al. 2000). Each of these studies used HMO-level data, which entails two complicated problems, neither of which is addressed by the three referenced studies. The first problem is that, in counties with noncompetitive Medicare managed care markets, the entry decision of each firm is a function of the entry decisions of all potential participants in that market. Complicating the problem is that some potential participants are not observed because they chose not to enter. The second problem inherent in the use of HMO-level data to study this problem is the likelihood of multiple equilibria. For example, a county may be able to support two HMOs in its Medicare managed care market, but it may be random which two HMOs participate. Bresnahan and Reiss (1991 a) show that multiple equilibria occur in simultaneous-move models under very general conditions. In this paper, we study the aggregate number of HMOs participating

CMS Payments

7

at the county level. This avoids the problems of simultaneity and multiple equilibria because we are concerned only with the number of firms that can be supported in the county, not the identities of the individual HMOs. In our focus on the number of firms that can be supported in distinct geographic markets, our paper is similar to an earlier literature that includes Bresnahan and Reiss (1987,1990,1991b); Dranove, Shanley, and Simon (1992); Kronick, Goodman, Wennberg, and Wagner (1993); and Erasure, Stearns, Norton, and Ricketts (1999).12 However, we differ from this literature because our regressor of interest is not the market size but the market "price." We follow the methodology developed in Bresnahan and Reiss (1987, 1990,1991b) in using a latent profit variable to motivate the use of an ordered probit to study the number of firms that can be supported in a geographic market. We assume that profit has an additively separable unobserved component, represented by an error term. It is assumed that the error term is normally distributed, independent across markets and independent of the regressors. We assume that all HMOs in the same market have the same unobserved profit. These assumptions permit the use of the ordered probit to estimate entry thresholds. The dependent variable is the number of HMOs participating in Medicare managed care in a county. We estimate the latent profit function using a reduced form approach. Cameron and Trivedi (1998) conclude that when the data generating process is a continuous latent variable (in our case unobserved profits), an ordered model should be used in place of a count data regression model.13 Accordingly, we estimate our model using an ordered probit regression. The number of participating HMOs in a given county in a given year is regressed on payment and the factors that affect variable costs, market size, the probability of enrollment, and fixed costs. Ordered probit regression will provide us with threshold values of CMS payments for HMO participation. If bp represents the ordered probit coefficient on CMS payment, b represents the vector of all other ordered probit coefficients, and X represents the set of regressors other than the CMS payment, then PN , the minimum CMS payment needed to support the participation of N HMOs, is:

8

Cawley, Chernew, and McLaughlin

where lN is the cutoff in the ordered probit regression associated with N HMOs.14 We predict that a higher CMS payment, controlling for observable factors that affect costs, will be associated with the participation of a greater number of HMOs. III. Data This section explains how we control for each component of the profit function introduced in the previous section. The data used in this paper come from two sources. CMS is the source for data on Medicare managed care enrollment, Medicare managed care contracts with HMOs, CMS payments by county, and input price indices. The second major source of data for this paper is the Area Resource File (ARF), which provides medical and demographic data at the county level.15 The unit of observation in this paper is the county. A market has traditionally been defined as a region in which a single price prevails for a homogenous good.16 By this definition, counties represent distinct markets for Medicare managed care; CMS sets Medicare managed care payments on a county-by-county basis. Furthermore, CMS requires separate contracts from HMOs for each county in which they wish to offer Medicare managed care. For the purposes of this study, a risk plan is defined as participating in a county Medicare managed care market if CMS market penetration files indicate that the plan has enrolled at least 0.5 percent of the county's Medicare-eligible residents.17 We exclude plans that have enrolled less than 0.5 percent of eligible residents because plans with such low county enrollment may not actually be operating in the county. CMS market penetration files list enrollees by their county of residence instead of the county in which they have enrolled in an HMO; as a result, many plan enrollees are found in counties where the plan does not have a contract to operate. The number of HMOs participating in a county, by year, is shown in table 1.1. Table 1.1 indicates that the number of counties with zero HMOs participating in Medicare managed care fell every year from 1993 to 1999, but rose from 1999 to 2001. It may not be cost-effective for CMS to support HMO participation in Medicare managed care in relatively rural or unpopulated counties. Table 1.2 lists the percentage of counties with at least one HMO participating in Medicare managed care, by the quintile of its 1990 popu-

Table 1.1 Number of counties with a given number of HMOs participating in Medicare managed care, by yeara Number of HMOs in county participating in Medicare managed care

Year

1993

1994

1995

1996

1997

1998

1999

2000

2001

0 1 2 3 4 5 6 7 8 9 10

2,816

2,728

2,569

2,401

2,289

2,230

2,210

2,273

2,415

166 55 21 4 7 4 1 0 0 0

202 81 30 22 3 5 1 2 0 0

281 114 56 28 14 5 2 4 1 0

309 155 101 48 43 9 3 3 2 0

317 188 105 84 51 23 11 5 1 0

329 205 126 80 57 25 17 2 3 0

387 199 126 74 34 27 9 6 2 0

366 189 116 70 27 20 7 3 2 1

336 183 88 23 14 6 6 2 1 0

Total number of counties

3,074

3,074

3,074

3,074

3,074

3,074

3,074

3,074

3,074

a

Data: HFCA Medicare managed care market penetration files, 1993-2001.

10

Cawley, Chernew, and McLaughlin

lation of Medicare beneficiaries. The table shows that counties in the fifth (most populous) quintile are several times more likely to have a participating HMO than are counties in the first quintile (least populous). Many counties are too rural or unpopulated ever to support HMO participation, so we exclude these counties from the sample. Thus, their history of nonparticipation does not influence the payment thresholds estimated for other counties. The sample used in this paper consists of counties whose population of Medicare beneficiaries is in the top three quintiles; in other words, its 1990 Medicare population was at least 2,783. In addition, all counties in Alaska and Hawaii are excluded. The sample contains observations of these counties for each year from 1993 to 2001. Plan-county data are aggregated to the HMO level and HMO-level data are aggregated to the county level.18 The dependent variable used in this paper is the number of HMOs participating in a county in a given year. In ordered probit regressions, this dependent variable is top-coded at six or more.19 The HMO Profit Function The profit function for all HMOs in a market is: P = [P - AVC]dS -rF + where P is the CMS payment, AVC is the average variable cost function, d is the probability of enrollment in Medicare managed care of the representative Medicare eligible, S is the number of Medicare eligibles, r is the interest rate, F is the fixed cost of entry, and represents unobserved profits. Listed below are the variables we use to proxy for each of the components of the profit function. P: Payment The regressor of interest is the CMS per-enrollee, per-month payment specific to the county. We enter the CMS payment directly and interact it with an indicator for the BBA regime (1998-2001), which allows the effect of the CMS payment to vary before and after the BBA of 1997 took effect. These payment variables include bonuses, paid only in 2000 and 2001, which are equal to 5 percent of the per-enrollee payment for the first year, and 3 percent of the per-enrollee payment for the second year, that an HMO operates in a previously unserved county.20

11

CMS Payments

Table 1.2 Percentage of counties with at least one active Medicare managed care HMO, by quintile of Medicare beneficiaries in 1990a Quintile of Medicare beneficiaries in 1990 Year

1

2

3

4

5

1993 1994 1995 1996 1997 1998 1999 2000 2001

4.7 5.4 7.6 9.8 9.9 10.9 7.6 6.5 4.2

2.4 3.6 4.9 7.2 9.3 10.4 12.8 12.4 9.4

3.9 6.0 9.8 13.0 17.1 18.9 20.3 17.1 13.8

7.5 10.7 15.3 22.4 27.8 31.4 33.3 31.1 22.6

23.5 30.6 44.6 57.2 63.7 65.8 66.4 63.4 57.2

615

615

615

615

614

14

1,482

2,783

4,714

9,718

1,479

2,781

4,708

9,680

877,581

Number of counties Minimum number of Medicare beneficiaries in quintile Maximum number of Medicare beneficiaries in quintile a

Data: CMS market penetration files, 1993-2001, and Area Resource File.

Although in practice the per-capita payments of CMS to HMOs are adjusted to take into account the demographic and (more recently) risk factors associated with the enrollee, we do not make these adjustments. Thus, the payment used in our empirical work represents the payment for the average enrollee.21 Summary statistics of the CMS per-enrollee monthly payments are listed in table 1.3 in nominal dollars. Table 1.3 indicates that the average CMS county monthly payment per enrollee rose each year from 1993 to 2001. The variance in the county payments rose until 1997, when the BBA was passed in part to reduce disparities in payments across counties. Since 1997, the variance in payments across counties has fallen each year. CMS payments to HMOs are constant during a calendar year; the exception to this rule is 2001, when payments were raised effective March 2001 by the Medicare, Medicaid, and State Children's Health Insurance Program (SCHIP) Benefits Improvement and Protection Act of 2000. We use the March payment rate for 2001 because the dependent variable in 2001 is also created using March data.

12

Cawley, Chemew, and McLaughlin

Table 1.3 Summary statistics of monthly per-enrollee CMS payments, by yeara Year

Mean

Standard deviation

Minimum

Maximum

1993 1994 1995 1996 1997 1998 1999 2000 2001 (Jan.-Feb.) 2001 (Mar.-Dec.)

301.86 314.72 332.43 372.13 394.78 417.09 427.33 449.78 460.39 498.82

55.46 58.29 62.99 70.58 76.69 62.99 62.69 56.85 56.66 41.70

168.15 171.07 177.32 207.31 220.92 367.00 379.84 401.52 414.88 475.00

598.65 653.44 678.90 881.35 767.35 782.70 798.35 809.28 825.46 833.55

a

Figures are in nominal dollars. The BIPA of 2000 raised payments to HMOs effective March 2001. Payments do not include bonuses for operating in previously unserved counties during 2000 and 2001. Source: HCFA Medicare managed care historical payment files, 1993-2001.

AVC: Average Variable Costs

We do not observe the average variable costs of HMOs; we estimate these costs in the following way. We assume that average variable costs in county c in year i, denoted AVCc,t , have the following structure:

where Ac,1991 is the average Medicare Part A (Hospital Insurance) reimbursement per enrollee in county c in 1991. This amount is multiplied by the percentage change in Part A costs since 1991, as measured by the CMS Hospital Input Price Index, which is represented in the equation above by PA,t.22 Likewise, Bc,1991 is the average Medicare Part B (Supplementary Medical Insurance) reimbursement per enrollee in county c in 1991. This amount is multiplied by the percentage change in Part B costs since 1991, as measured by the CMS Medicare Economic Index, which is represented in the equation above by PB,t.23 The change in costs observed over time is due to prices, not necessarily utilization. Also note that the Hospital Input Price Index and the Medicare Economic Index are nationwide indices, and therefore all of the difference across counties in costs is due to the baseline difference in costs in 1991. In the

CMS Payments

13

regression model, Part A and Part B costs will be entered separately. HMOs may be better able to control one type of costs than the other, and therefore costs in the two areas may have different effects on the likelihood that HMOs will participate. In the average variable costs equation listed above, Xc is a vector of county characteristics that may affect costs, specifically, the number of general practitioners in 1990, the number of registered nurses in 1990, the number of hospitals in 1993, and median rent in 1990.24 We also include as regressors population density and the percentage of the population that is urban because geographically dispersed populations may be more costly to serve. Finally, year-specific costs are captured by It, an indicator variable that equals 1 if the observation is for year t. S: Size of the Market

Although the sample is limited to relatively populous counties, even within that group, HMOs may prefer to enter more populous counties. We control for the size of the county market using the number of Medicare beneficiaries in the county in 1990.25 We also include the percentage change in this number from 1980 to 1990 to account for the fact that HMOs may prefer to enter growing markets. F: Fixed Costs of Entry

We control for two factors that Brown and Gold (1999) suggest affect the fixed costs of entry into the Medicare managed care market. The first is whether the HMO already operates in the commercial market in the county. This may affect the fixed costs of entering Medicare managed care for two reasons: (1) the HMO would have already sunk the costs of establishing a network of health care providers in the county (that is, there may be economies of scope to participating in multiple managed care markets in the same county), (2) CMS limits participation in the Medicare managed care market to HMOs participating in the county's commercial market. HMOs that historically participated in the commercial market of the county may face lower barriers to entering the Medicare managed care market. We do not simply control for the number of HMOs participating in the county's commercial managed care market. Because an HMO could enter a county's commercial market for the purpose of subsequently entering its Medicare managed care market, current participation in the commercial market may be endogenous. Instead, we control for the

14

Cawley, Chernew, and McLaughlin

number of HMOs in the county in 1980, before the TEFRA of 1982 created the modern Medicare managed care market.26 We also control for the likelihood of HMOs participating in the county commercial market using the percentage of the workforce in manufacturing or white-collar jobs in 1990. The presence of these types of workers proxies for the presence of employers likely to demand commercial managed care for its employees. The second factor that affects the fixed cost of entering a county Medicare managed care market is whether an HMO participates in nearby counties. It may be cheaper for an HMO to enter a county adjacent to its current service area because the HMO may already be familiar with local providers and have acquired information about the local market. To proxy for the likelihood of participating in adjacent counties, we control for the total number of Medicare beneficiaries in 1990 in all adjacent counties and its percentage growth from 1980 to 1990. d: Probability That Medicare Eligibles Will Enroll in Medicare Managed Care

It has been found repeatedly that relatively healthy Medicare beneficiaries are the most likely to enroll in managed care.27 To capture cross-county differences in the proportion of healthy beneficiaries (and therefore demand for Medicare managed care), we control for per capita income, the poverty rate among the county's elderly, the percentage of adults with a high school diploma, and the percentage of adults with a college degree.28 Each of these variables was measured in 1990. Summary statistics for the sample used in this paper appear in table 1.4. We acknowledge that characteristics of the individual HMOs participating in the market may affect variable or fixed costs, or the triggers at which the HMO will enter or exit. For example, certain model types may be more efficient at providing care and the exit trigger may be lower for nonprofit than for for-profit HMOs. We ignore the characteristics of the individual participating HMOs for two reasons. First, these characteristics are endogenous. An HMO may change its model type or profit status to suit the characteristics of the markets in which it participates. Second, as mentioned earlier, HMO entry into Medicare managed care is an example of a multiple-agent discrete-move game. It is likely that multiple equilibria exist and that the number of firms participating is determined, but the identity of the individual HMOs that participate is to some extent random.

Table 1.4 Summary statistics Variable

Year(s) of data

N

Mean

Standard deviation

Minimum

Maximum

Number of HMOs active in Medicare managed care CMS payment (per enrollee, per month) Average annual Medicare Part A costs Average annual Medicare Part B costs Number of general practitioners Number of registered nurses Number of hospitals Number of HMOs active in commercial market Per capita income Poverty rate among elderly Median rent Percentage of adults who are high school graduates Percentage of adults who are college graduates Number of Medicare beneficiaries Percentage growth in Medicare beneficiaries Medicare beneficiaries in neighboring counties Percentage growth of Medicare beneficiaries in neighboring counties Percentage of population that lives in urban areas Population density Percentage of workers in manufacturing Percentage of workers in white collar jobs

1993-2001 1993-2001 1993-2001 1993-2001 1990 1990 1993 1980 1993 1990 1990 1990 1990 1990 1980-1990 1990

16,596 16,596 16,596 16,596 16,596 16,596 15,741 1,134 16,596 16,596 16,596 16,596 16,596 16,596 16,596 16,314

66 401.95 2,247.67 1,247.10 34.28 987.82 3.25 1.82 17,289.8 .15 352.76 71.02 14.75 17,071.2 .37 81,258.8

1.31 90.65 457.25 259.16 94.37 2,502.60 6.16 1.41 3,654.60 .07 97.55 9.56 6.97 39,845.4 .23 109,335

0 187.14 1,107.55 482.03 1 11 1 1 6,306 .04 175 31.6 4.6 2,783 -.06 2,015

10 833.55 5,658.37 2,910.01 2,605 52,780 148 11 52,277 .53 834 92.9 52.3 877,581 2.97 1,452,320

1980-1990 1990 1994 1990 1990

16,314 16,029 16,596 16,596 16.596

.35 50.44 334.42 20.71 49.02

.16 24.80 1,847.7 9.61 9.04

-.05 .1 1.8 2.7 29.5

1.35 100 53,801.1 52 79.2

16

Cawley, Chernew, and McLaughlin

IV. Empirical Results The results of the ordered probit regression of the number of HMOs participating in Medicare managed care at the county level are presented in table 1.5. In all the results reported in this paper, standard errors are cluster-corrected to account for the dependence in errors within each county over time. The coefficients on CMS payment and CMS payment interacted with the BBA regime are positive and statistically significant at the 1 percent level, which is consistent with our hypothesis that, controlling for costs, a higher payment is associated with the participation of more HMOs. As described in Section II, the coefficients presented in table 1.5 can be used to calculate the CMS payments necessary to support a given number of HMOs in the market. Each county has unique thresholds needed to support given numbers of HMOs in this program. Rather than report the thresholds associated with over 2,000 counties, table 1.6 lists the payment thresholds associated with counties at the 25th, 50th, 75th, and 100th percentiles for payment threshold. Table 1.6 indicates that to support a single HMO in the median county in the sample, it is necessary for CMS to pay $682.08 per average enrollee per month in the median county. To support a single HMO in every county of the sample, CMS would have to pay $1,008.25 per average enrollee per month in the maximum-payment county. Table 1.6 also lists the CMS payment thresholds necessary to support multiple HMOs in county Medicare managed care markets. CMS may desire multiple HMOs in each market because the competition between the HMOs for market share leads to lower out-of-pocket costs and additional benefits for enrollees. Table 1.6 suggests that, conditional on two HMOs already participating, CMS must pay roughly $115 more per enrollee per month to support each additional Medicare HMO. Although we report our estimated thresholds to the cent, we do not claim absolute precision about the estimates. The exact threshold is determined in part by assumptions, such as the functional form of regression. The standard errors, which appear in parentheses below the thresholds in table 1.6, in some cases imply large confidence intervals. Derivation of standard errors for the thresholds is difficult because the thresholds are nonlinear functions of several random variables. Accordingly, we calculate bootstrap standard errors. Specifically, bootstrap samples of size equal to the overall sample are formed by

Table 1.5 Ordered probit regression of number of HMOs in county on county characteristics Variable Payment CMS payment CMS payment * indicator for 1998-2001

Coefficienta

Z Scoreb

.0034 .0016

4.04 4.61

.2402 .5610 .8232 1.0187 .3526 .2861 .0050 -.5497

8.57 13.23 12.25 12.52 2.05 1.62 0.03 -2.45

-.00002 .0002 .0030 -.0002 .0067 .0034 -.00002 .0013

-0.88 1.41 3.09 -2.87 0.65 6.66 -1.85 0.84

Indicator variables for year

1994 1995 1996 1997 1998 1999 2000 2001 Variables affecting average variable costs Average Medicare Part A costs Average Medicare Part B costs Number of general practitioners Number of registered nurses Number of hospitals Median rent Population density Percentage of population in urban areas Measures of the size of the market Number of Medicare beneficiaries Percentage of growth in Medicare beneficiaries

.000007 .3024

2.36 2.07

-.0202 -.0042 .0276 .000003

-0.18 -1.12 3.12 7.35

.4744

2.24

-.00002 -1.1877 .0223 -.0400

-1.90 -1.55 3.36 - 4.18

Variables affecting fixed costs of entry Number of HMOs in county in 1980 Percentage of workforce in manufacturing Percentage of workforce who are white collar Number of Medicare beneficiaries in all adjacent counties Percentage growth in Medicare beneficiaries in all adjacent counties Variables affecting the probability of enrollment Per capita income Poverty rate among elderly Percentage of adults with high school diploma Percentage of adults with college degree Number of observations Log likelihood a

16,596 -12,618.9

Coefficients on indicator variables for missing values are omitted. Z scores reflected cluster-corrections of standard errors by county.

b

Cawley, Chernew, and McLaughlin

18

Table 1.6 Estimated monthly payments necessary to support given numbers of HMOs in Medicare managed care per county in the year 2001a, b Monthly CMS payment necessary ($) Desired number of HMOs / county 1 2 3 4 5

6 or more

25th percentile 568.75 (19.23) 710.18 (59.26) 832.03 (96.42) 947.45 (132.63) 1,053.46 (162.61) 1,163.66 (194.46)

Median 682.08 (55.70) 823.51 (99.33) 945.36 (136.95) 1,060.78 (173.36) 1,166.79 (203.38) 1,277.00 (235.47)

75th percentile

Maximum

764.21 (85.61) 905.64 (129.38) 1,027.49 (167.03) 1,142.90 (203.46) 1,248.91 (233.48) 1,359.12 (265.47)

1,008.25 (147.02) 1,149.68 (189.77) 1,271.53 (226.79) 1,386.95 (262.99) 1,492.96 (292.95) 1,603.17 (324.60)

a

Calculated using coefficients reported in table 1.5. Sample consists of counties with Medicare population in top three quintiles. Bootstrap standard errors appear in parentheses. Payments calculated using ordered probit coefficients.

b

randomly selecting, with replacement from the overall sample, all observations of a particular county. The standard errors are calculated from the variance observed in the thresholds calculated using the bootstrapped samples. We follow the recommendation of Efron and Tibshirani (1993) and conduct 200 replications to estimate standard errors. Table 1.7 compares the mean characteristics of two groups of counties: those in which CMS payments in the year 2001 were more than the estimated payment necessary for one HMO to participate in the county, and those in which CMS payments were less than that threshold. The table also lists the difference in means and the t statistic associated with the test of the hypothesis that the means are equal across the two groups of counties. In the year 2001,381 counties in the sample were assigned CMS payments that exceeded the estimated payment necessary to support one HMO, while 1,463 counties were assigned payments less than the single-HMO threshold. Table 1.7 indicates that counties assigned payments greater than the estimated single-HMO threshold have both higher CMS payments and higher Part A and B Medicare costs than the

Table 1.7 Difference in mean characteristics between counties with actual payments above and below estimated payment threshold for one HMO to participate in Medicare managed care in the year 2001 County characteristic Number of participating HMOs, 2001 Monthly CMS payment, 2001 Average Medicare Part A costs, 1991 Average Medicare Part B costs, 1991 Number of general practitioners, 1990 Number of hospitals, 1993 Per capita income, 1993 Poverty rate among elderly, 1990 Percentage of adults who are high school graduates, 1990 Percentage of adults who are college graduates, 1990 Number of Medicare beneficiaries, 1990 Percentage of population that lives in urban areas, 1990 Population density, 1994 Number of counties

Mean for counties with actual payment > threshold

Mean for counties with actual payment < threshold

Difference in means

t statistics for equality of means

1.93 563.78 2,104.95 1,320.42 104.85 8.32 21,207.23 .10

.26 510.11 1,826.53 1,056.78 15.87 1.93 16,269.64 .17

1.65 53.67 278.42 263.65 88.98 6.39 4,937.59 -.072

18.97 18.10 11.48 19.32 9.20 10.15 19.16 -24.92

78.35

69.11

9.24

22.06

21.16 51,889.71

13.08 8,003.63

8.08 43,886.08

18.64 11.12

74.73 1,257.56 381

43.88 94.02 1,463

30.85 1,163.54

23.08 5.78

20

Cawley, Chernew, and McLaughlin

counties assigned payments less than the threshold. In addition, the counties with above-threshold payments have many more hospitals and general practitioners and, in general, have better educated and wealthier populations. Each of these differences is statistically significant at the 1 percent significance level. Perhaps the most dramatic difference is in the size of the Medicare population: counties assigned payments greater than the estimated single-HMO threshold have on average a Medicare beneficiary population of almost 51,900, whereas counties assigned payments less than that threshold have on average a Medicare bene- ficiary population of only about 8,000. If a below-threshold current payment can be interpreted as an underestimate by CMS of costs in that county, then our results suggest that CMS tends to underestimate the costs of HMO participation in sparsely populated counties. Several studies noted that, under the TEFRA payment scheme that was used prior to 1998, rural counties were particularly unlikely to be served by HMOs.29 Passage of the BBA was intended to eliminate such disparities by raising payments more quickly in low-payment than in high-payment counties. We find that even three years after the BBA took effect, counties with CMS payment insufficient to support HMO participation tend to be far less populous than counties that receive what we estimate to be sufficient payment. This pattern is found in a sample that includes only those counties with a Medicare population large enough to be considered viable for HMO activity in this program. V.

Conclusion

At the end of 1998, 1999, and 2000, HMO exits from Medicare managed care markets resulted in the involuntary disenrollment of hundreds of thousands of elderly and disabled Americans from a program that was intended to generate additional benefits for beneficiaries and savings for Medicare. This paper estimates the CMS payments necessary to support the participation in Medicare managed care of a given number of HMOs per county market. Ordered probit estimates suggest that, to support one Medicare HMO in half of U.S. counties in our sample in 2001, CMS would have to pay $682.08 per average enrollee per month in the marginal county. To support one Medicare HMO in every county in the sample in the year 2001, CMS would need to pay $1,008.25 per enrollee per month in the maxi-

CMS Payments

21

mum-payment county. In the year 2001, actual CMS payments range from $475.00 to $833.55. Competition among Medicare HMOs generates additional services at lower cost for enrollees. If CMS desires multiple HMOs to participate in county markets, our estimates suggest that even greater payments are required. If two HMOs are already participating, roughly an extra $115 per enrollee per month is necessary to support each additional Medicare HMO. We find that 79.3 percent of all counties in our sample received less than the estimated amount necessary to support an HMO in this market. Compared to counties that received more than the estimated threshold for HMO participation, the counties receiving an insufficient payment are on average more rural and are less populated with citizens who are wealthy and educated. The relative disadvantage of rural and unpopulated counties persists three years after the BBA 1997, which was designed to eliminate such disparities, took effect. This pattern is found in a sample that includes only those counties with a Medicare population large enough to be considered viable for HMO activity in this program. Notes We thank the following people for their helpful comments and suggestions: Scott Adams, David Colby, Julie Cullen, Rachel Dunifon, Alan Garber, Hanns Kuttner, David Meltzer, Katie Merrell, and participants at the NBER Frontiers in Health Policy Research Conference held June 7, 2001, in Bethesda, Maryland. We thank Phil DeCicca for his expert research assistance. Please email comments to [email protected]. 1. In 2001, the Health Care Financing Administration was renamed the Center for Medicare and Medicaid Services. For the sake of consistency, the agency is referred to throughout this paper as CMS. 2. HMOs competing for market share in the Medicare managed care market tend to lower their premia or offer additional benefits to enrollees; see Physician Payment Review Commission (1996) and General Accounting Office Report 97-133 (1997c). 3. Medicare beneficiaries may enroll only in those HMOs with a risk contract from CMS to serve the beneficiary's county of residence. Medicare beneficiaries suffering from end-stage renal disease are not eligible for Medicare managed care. 4. CMS Medicare Managed Care Contract Reports are the source of the data shown in figures 1.1 and 1.2. The data for each year are from the December report, except the data for 2001, which are from the January report. 5. Health Care Financing Administration Medicare Managed Care January Contract Report (2001).

22

Cawley, Chernew, and McLaughlin

6. Laschober et al. (1999). 7. Health Care Financing Administration (2000b). 8. Studies of data prior to 1990 find that the health care costs of Medicare managed care enrollees were 20-42 percent lower than fee-for-service beneficiaries with the same demographic characteristics. Studies of post-1990 data find that the health care costs of Medicare managed care enrollees were 12-37 percent lower than comparable fee-for-service Medicare beneficiaries; see the review in General Accounting Office Report 97-16 (1997b). This has held true even after passage of the BBA. It is estimated that in 1998 HMOs were paid on average $1,000 more per enrollee than CMS would have paid had the enrollees remained in fee-for-service Medicare; see General Accounting Office Report 00-161 (2000). This favorable selection occurred even though HMOs are prohibited by law from selecting enrollees on the basis of health status. 9. Some provisions of the BBA were amended by the Balanced Budget Refinement Act of 1999 and the Medicare, Medicaid, and SCHIP Benefits Improvement and Protection Act of 2000. 10. In addition, the BBA requires CMS to adjust payments by the health status of plan enrollees. The risk adjustment will be phased in; payments in 2001 are 10 percent risk adjusted and 90 percent adjusted only for demographic factors. The full amount of the payment will be risk-adjusted by 2004. 11. The bonus is paid to the first HMO to enter a previously unserved county; if several HMOs enter on the same date, they each receive the bonus. 12. These papers did not study the market for Medicare managed care. Bresnahan and Reiss (1987, 1990, 1991b) studied markets for retail and professional service industries; Dranove, Shanley, and Simon (1992) studied hospitals; and Erasure, Stearns, Norton, and Ricketts (1999) studied physicians. Using a different methodology, Kronick, Goodman, Wennberg, and Wagner (1993) estimated the metropolitan area population necessary to support three HMOs in the commercial managed care market. 13. Cameron and Trivedi (1998), p. 86. 14. If the dependent variable in an ordered probit regression has M categories, the cutoffs represent fitted values above which the model predicts that the dependent variable will equal m for m = 1,..., M. 15. The Area Resource File (ARF) is a compilation of data from various sources. Unless otherwise noted, the original source of data taken from the ARF is the 1990 Census of Population and Housing. 16. See, for example, Marshall (1920), Book V, Chapter 1. 17. The enrollment data used to determine HMO participation is that for December for 1993-1997 and 2000, October for 1998-1999, and March in 2001. December reports are not used for 1998 and 1999 because the figures listed in those December reports are actually from the following January. 18. A plan is a uniform set of benefits and premiums. Each HMO may offer multiple plans. In our data, we find only thirty-seven counties in which a single HMO offers two plans. 19. We top-code the dependent variable because it can be difficult to estimate an ordered probit for values of the dependent variable that appear rarely in the data.

CMS Payments

23

20. We determine whether each county is eligible for a bonus by checking the Medicare Managed Care geographic service area reports to see whether any HMO had a risk contract with CMS to serve the county the previous calendar year. 21. Demographic and risk adjustments are uniform across counties. 22. The Hospital Input Price Index tracks changes in the prices of hospital inputs such as wages, salaries, benefits, professional fees, utilities, liability insurance, pharmaceuticals, food, chemicals, medical instruments, photographic supplies, rubber and plastics, paper products, apparel, machinery and equipment, and other inputs. 23. The Medicare Economic Index tracks changes in the prices of inputs to physician-provided care such as physician compensation, nonphysician compensation, office expenses, medical materials and supplies, liability insurance, medical equipment, and other expenses. 24. The source of the data on the number of doctors is the American Medical Association Physician Masterfile, and that for the number of hospitals is the American Hospital Association Survey of Hospitals. 25. The number of Medicare beneficiaries includes both elderly and disabled beneficiaries (both are eligible for managed care). In 1998, the elderly represented 87.06 percent of all Medicare beneficiaries. 26. The source of the data on commercial HMO historic participation is the National HMO Census of Prepaid Plans. 27. Chapter 15 of Physician Payment Review Commission (1996) summarizes the literature that finds Medicare beneficiaries who enroll in managed care, compared to those who remain in fee-for-service Medicare, tend to have had lower utilization and Medicare costs in the preceding few years. See also General Accounting Office Report 97-160 (1997a). A similar difference in prior utilization characterizes those who enroll in commercial managed care plans; see the summary in Glied (2000). Possible reasons that the relatively healthy are more likely to enroll in managed care are that they are less likely to have an established health care provider and that they may be less averse to the risk that HMOs may deny them certain treatments. 28. We assume that the per-capita income and education of Medicare beneficiaries track those of the entire adult population in the county. The source of data on the poverty rate among the elderly is the Small Area Income Poverty estimates from the Bureau of the Census and that for per-capita income is the U.S. Department of Commerce.

29. See, for example, Serrato, Brown, and Bergeron (1995). References Abraham, Jean, Ashish Arora, Martin Gaynor, and Douglas Wholey (2000). "Enter at Your Own Risk: HMO Participation and Enrollment in the Medicare Risk Market," Economic Inquiry 38(3):385-401. Adamache, Killard W, and Louis F. Rossiter (1986). "The Entry of HMOs into the Medicare Market: Implications for TEFRA's Mandate," Inquiry, 23:349-364. Erasure, Michelle, Sally C. Stearns, Edward C. Norton, and Thomas Ricketts (1999). "Competitive Behavior in Local Physician Markets," Medical Care Research and Review 56(4):395-414.

24

Cawley, Chernew, and McLaughlin

Bresnahan, Timothy E, and Peter C. Reiss (1987). "Do Entry Conditions Vary Across Markets?" Brookings Papers on Economic Activity, #3:833-881. Bresnahan, Timothy F ., and Peter C. Reiss (1990). "Entry in Monopoly Markets," Review of Economic Studies 57(4)531-553. Bresnahan, Timothy F., and Peter C. Reiss (1991a). "Empirical Models of Discrete Games," Journal of Econometrics 48(1/2)57-81. Bresnahan, Timothy P., and Peter C. Reiss (1991b). "Entry and Competition in Concentrated Markets," Journal of Political Economy 99(5):977-1009. Brown, Randall S., and Marsha R. Gold (1999). "What Drives Medicare Managed Care Growth?" Health Affairs, 18(6):140-149. Cameron, A. Colin, and Pravin K. Trivedi (1998). Regression Analysis of Count Data. NY: Cambridge University Press. Dranove, David, Mark Shanley, and Carol Simon (1992). "Is Hospital Competition Wasteful?" RAND Journal of Economics 23(2):247-262. Efron, Bradley, and Robert J. Tibshirani (1993). An Introduction to the Bootstrap. NY: Chapman & Hall. General Accounting Office (1997a). "Medicare: Fewer and Lower Cost Beneficiaries with Chronic Conditions Enroll in HMOs," Report 97-160. Washington, D.C.: United States General Accounting Office. General Accounting Office (1997b). "Medicare HMOs: CMS Can Promptly Eliminate Hundreds of Millions in Excess Payments," Report 97-16. Washington, D.C.: United States General Accounting Office. General Accounting Office (1997c). "Medicare Managed Care: HMO Rates, Other Factors Create Uneven Availability of Benefits," Report 97-113. Washington, D.C.: United States General Accounting Office. General Accounting Office (2000). "Medicare+Choice: Payments Exceed Cost of Fee-for-Service Benefits, Adding Billions to Spending," Report 00-161. Washington, D.C.: United States General Accounting Office. Glied, Sherry (2000). "Managed Care," in Anthony J. Culyer and Joseph P. Newhouse, eds., Handbook of Health Economics, Volume 1A. NY: Elsevier/North-Holland. Health Care Financing Administration (2000). "Protecting Medicare Beneficiaries After Medicare+Choice Organizations Withdraw," CMS Fact Sheet. Washington, D.C.: Department of Health and Human Services. Health Care Financing Administration (1990-2001). Medicare Managed Care Contract Reports. Washington, D.C.: Department of Health and Human Services. Kronick, Richard, David C. Goodman, John Wennberg, and Edward Wagner (1993). "The Demographic Limitations of Managed Competition," New England Journal of Medicine, 328(2):148-152. Laschober, Mary A., Patricia Neuman, Michelle S. Kitchman, Laura Meyer, and Kathryn M. Langwell (1999). "Medicare HMO Withdrawals: What Happens to Beneficiaries?" Health Affairs, 18(6):150-157.

CMS Payments

25

Marshall, Alfred (1920). Principles of Economics, Eighth Edition. Philadelphia, PA: Porcupine Press. Physician Payment Review Commission (1996). Annual Report to Congress. Washington, D.C.: Physician Payment Review Commission. Porell, Frank W., and Stanley S. Wallack (1990). "Medicare Risk Contracting: Determinants of Market Entry," Health Care Financing Review, 12(2):75-85. Serrato, Carl, Randall S. Brown, and Jeanette Bergeron (1995). "Why Do So Few HMOs Offer Medicare Risk Plans in Rural Areas?" Health Care Financing Review, 17(l):85-97.

This page intentionally left blank

2 The Effects of Medicare on Health Care Utilization and Outcomes Frank R. Lichtenberg, Columbia University andNBER

Executive Summary Medicare, which provides health insurance to Americans over the age of 65 and to Americans living with disabilities, is one of the government's largest social programs. It accounts for 12 percent of federal on- and off-budget outlays, and in fiscal year 1999, $212 billion in Medicare benefits were paid. The largest shares of spending are for inpatient hospital services (48 percent) and physician services (27 percent). In thirty years, the number of Americans covered by Medicare will nearly double to 77 million, or 22 percent of the U.S. population. Perhaps the most important question we can ask about the Medicare program is, What impact does it have on the health of the U.S. population? One feature of the Medicare program can be exploited to shed light on its impacts: its age specificity. Most people become eligible for Medicare suddenly, the day they turn 65. Consequently, the age profiles of health services utilization and health outcomes (morbidity and mortality) can provide revealing evidence about Medicare's impacts. My objective is to obtain precise estimates of medical utilization and outcomes, by single year of age, for ages close to age 65. The most precise estimates can be obtained by using information obtained from medical providers (hospitals and doctors) pooled over several years. Utilization of ambulatory care and, to a much smaller extent, inpatient care increases suddenly and significantly at age 65, presumably due to Medicare eligibility. The evidence points to a structural change in the frequency of physician visits precisely at age 65. Attainment of age 65 marks not only an upward shift but also the beginning of a rapid upward trend (up until age 75) of about 2.8 percent per year in annual visits per capita. The number of physician visits in which at least one drug is prescribed also jumps up at age 65. Reaching age 65 has a strong positive impact on the consumption of hospital services, but most of this impact appears to be the result of postponement of hospitalization in the prior two years. We also examine whether this increase in utilization leads to an improvement in outcomes—a reduction in morbidity and mortality—relative to what one would expect given the trends in outcomes prior to age 65. The estimates are consistent with the hypothesis that the Medicare-induced increase in health

28

Lichtenberg

care utilization leads to a reduction in days spent in bed of about 13 percent and to slower growth in the probability of death after age 65. Physician visits are estimated to have a negative effect on the male death rate, conditional on age and the death rate in the previous year. The short-run elasticity of the death rate with respect to the number of physician visits is -.095, and the long-run elasticity is —.497: a permanent or sustained 10 percent increase in the number of visits ultimately leads to a 5 percent reduction in the death rate. Data on age-specific death probabilities every 10 years since 1900, i.e., before as well as after Medicare was enacted, provide an alternative way to test for the effect of Medicare on longevity. They also provide strong support for the hypothesis that Medicare increased the survival rate of the elderly by about 13 percent.

I. Introduction Between 1965 and 1967, there was a huge (65 percent) increase in real per-capita public health expenditure (figure 2.1). Medicare, which today provides health insurance to Americans over the age of 65, accounted for more than half (57 percent) of the 1965-1967 increase in public health expenditure. Figure 2.2 reveals that this increase in public health expenditure was offset, to some extent, by a reduction in private health expenditure. I estimate that each additional dollar of public health expenditure "crowded out" about 43 cents of private spending.1 Nevertheless, enactment of Medicare and Medicaid led to significant increases in per-capita health expenditure. Perhaps the most important question we can ask about the Medicare program is, What impact has it had on the health of the U.S. population? Attempting to answer this question with either individual-level or aggregate data may be fraught with difficulties. At the individual level, there is often an inverse relationship between medical expenditures and health outcomes: people in poor health have higher medical expenditures. The expenditures may improve their health, but unless a person's health is observable both pre- and postexpenditure—which is usually not the case—the contribution of expenditure to health cannot be identified. The Health Care Financing Administration (2000) cites aggregate data to support its argument that "the average life expectancy of elderly Americans has increased, in part, because of Medicare." That claim seems plausible. Life expectancy at age 65 increased at a faster rate since Medicare than it did before Medicare: 2.0 years between 1970

Figure 2.1 Percentage increase from previous year in real per-capita public health expenditure

Figure 2.2 Percentage increase from previous year in real per-capita public and private health expenditure

30

Lichtenberg

and 1990 versus 1.3 years between 1950 and 1970, although data on life expectancy at age 65, by gender, reveal that only men experienced faster growth in life expectancy after Medicare than before Medicare (see figure 2.3). Other factors, such as changes in rates of public and private biomedical innovation and government income security programs, may also have contributed to the acceleration of life expectancy at age 65, making it difficult to isolate the contribution of Medicare from aggregate time-series data. One feature of the Medicare program can be exploited to shed light on its impacts: its age specificity. Most people become eligible for Medicare suddenly, the day they turn 65. Consequently, the age profiles of health services utilization and health outcomes (morbidity and mortality) can provide revealing evidence about Medicare's impacts. II. Changes in Utilization and Outcomes at Age 65 Most Americans become eligible for Medicare benefits upon reaching the age of 65. (In 1990, 90 percent of Medicare beneficiaries were elderly, as opposed to disabled or ESRD enrollees.) Consequently, comparisons of health utilization and outcomes just before and just after age 65 may shed light on the impact of Medicare. Some variables (for example, mortality rates) may exhibit a trend prior to age 65. In such cases, it is appropriate to examine whether there is a break in the trend at age 65, rather than to test for a pre- versus post-65 difference in levels. Medicare eligibility is not the only major event that many people experience at or around the age of 65. Another important event is retirement. Indeed, the intent of Medicare was evidently to ensure that people continued to have access to medical care after they retired and were no longer covered by employer-sponsored health insurance. From this perspective, if Medicare had accomplished its objectives exactly, one might expect to observe no difference between (or no shift in the trend in) utilization and outcomes pre- versus post-age 65. Suppose that, in the absence of Medicare, a person's medical expenditure would drop significantly upon retirement, assumed to occur at age 65. The objective of Medicare was simply to fill the gap left by the termination of employer-sponsored insurance. This scenario is depicted in figure 2.4.

Figure 2.3 Life expectancy at age 65,1950-1998, by gender

32

Lichtenberg

Figure 2.4 Hypothetical effect of Medicare on age/medical expenditure profile

Presumably, policy makers did not intend to induce an upward shift in the age-expenditure profile at age 65. If they believed that medical expenditure before age 65 was too low, they could have designed the program to provide at least some benefits to people younger than 65. If policy makers wanted people to consume about the same amount of medical services (for example, physician visits) at age 66 as they had done at age 64, they should have ensured that the out-of-pocket cost was higher at age 66 because the consumption of medical services requires two inputs: purchased medical services (for example, the physician's time) and the patient's time. The opportunity cost (foregone earnings) of the patient's time is much higher before than after retirement. Therefore, if out-of-pocket cost is the same, one would expect people to visit the doctor more after they have retired. If everyone retired at age 65, when they become eligible for Medicare, it would be almost impossible to distinguish between the effects of retirement and the effects of Medicare from the age profiles of utilization and outcomes. In practice, however, many people retire before reaching the age of Medicare eligibility. According to Social Security Administration data for December 2000,46 percent of workers retire by age 62,2 and 62 percent of workers retire by age 64. Hence, if there are abrupt changes in utilization and outcomes precisely at age 65, it is unlikely that they can be accounted for by retirement.

Effects of Medicare on Health Care Utilization and Outcomes

33

III. The Age-Utilization Profile My objective is to obtain precise estimates of medical utilization and outcomes, by single year of age, for ages close to age 65. Household surveys, such as the 1996 Medical Expenditure Panel Survey (MEPS) and its predecessors, contain comprehensive information, but the number of individuals of any given age is quite small, resulting in large sampling error. For example, the average number of people per single year of age is only 221 for ages 45-64 in MEPS. Much more precise estimates can be obtained by using information obtained from medical providers (hospitals and doctors) pooled over several years. Hospital Discharges I obtained data on hospital discharges, by age, from the National Hospital Discharge Survey, 1979-1992, Multi-Year Data File. The National Hospital Discharge Survey (NHDS) provides data on inpatient utilization of short-stay, nonfederal hospitals in the United States. The NHDS abstracts both demographic and medical information from the face sheets of the medical records of inpatients selected from a national sample of hospitals. Based on this information, national and regional estimates of characteristics of patients, lengths of stay, diagnoses, and surgical and nonsurgical procedures in hospitals of various bed sizes and types of ownership are produced. The 1979-1992 Multi-Year Data File contains records of about 2.8 million nonnewborn hospital discharges. The age profile of hospital discharges is shown in figure 2.5. There is a marked discontinuity in the profile at age 65. The yearly (by age) growth rate of hospital discharges is shown in figure 2.6. From age 50 to age 62, the number of discharges increases by about 3 percent per year of age. From age 62 to age 64, the number of discharges is essentially constant (it actually declines a little). Between age 64 and age 65, the number of discharges increases 9.5 percent. Between ages 65 and 74, it increases about 0.5 percent per year. This evidence indicates that reaching age 65 has a strong positive impact on the consumption of hospital services. However, much of this impact appears to be the result of postponement of hospitalization in the prior two years. The average annual growth rate from age 62 to 65 is 3.1 percent. In contrast, the average annual growth rate from age 50

Figure 2.5 Number of 1979-1992 hospital admissions, by single year of age

Figure 2.6 Percentage increase in number of hospital admissions from age t -1 to age t

Effects of Medicare on Health Care Utilization and Outcomes

35

to 62 is 2.3, and from 59 to 62 is 2.4 percent. Hence the "excess" growth from age 62 to 65 is 0.7 to 0.8 percent per year, or about 2.1 to 2.4 percent additional discharges by the age of 65. Physician Visits I computed the frequency of physician office visits, by single year of age, by pooling data from the National Ambulatory Medical Care Surveys (NAMCS) for each of the seventeen years during 1973-1998 in which the survey was conducted.3 The number of visits surveyed varies from year to year; the 1998 survey contains information from 24,715 patient visits. The pooled data set contains data on approximately 313,000 visits. Average number of physician office visits, per person per year by single year of age for ages 61-69, are shown in figure 2.7.4 As in the case of hospital discharges, the evidence points to a structural change in visit frequency precisely at age 65. The average annual number of physician visits is 9.5 percent higher for ages 65-69 than it is for ages 61-64. Once they are eligible for Medicare, people visit the doctor more often.5 Figure 2.8 displays data on Medicare and non-Medicare physician visits per person per year, using a wider age window. From age 50 to age 64, the number of annual visits per capita is flat, and even exhibits a tendency to decline from age 58 to age 64. Attainment of age 65 marks not only an upward shift but also the beginning of a rapid upward trend (up until age 75) of about 2.8 percent per year in annual visits per capita. Since physicians prescribe at least one drug in about two-thirds of office visits, one would expect the number of "drug visits"—visits in which at least one drug is prescribed—also to increase at age 65. Figure 2.9 (based on data for 1985 and 1989-1998) confirms that this is the case. The number of drug visits increases 11.3 percent from age 64 to age 65. The average annual number of drug visits is 19 percent higher among 65 to 72-year-olds than it is among 60 to 64-year-olds. Data from the 1996 Medical Expenditure Panel Survey, a household-based survey, also indicate a sharp increase in pharmaceutical use near the age of 65. As shown in figure 2.10, the median number of annual prescriptions (including refills and free samples) increases from 7.3 for people age 56 to 65 to 12.1 for people age 66 to 75.

Figure 2.7 Average number of physician visits per person per year, by age

Figure 2.8 Medicare and non-Medicare physician visits per person per year, by age

Figure 2.9 Number of physician visits in which at least one drug was prescribed, 1985 and 1989-1998

Figure 2.10 Median number of prescriptions, by age, 1996

40

Lichtenberg

IV. The Age-Outcomes Profile The evidence just presented indicates that utilization of ambulatory care and, to a much smaller extent, inpatient care increases suddenly and significantly at age 65, presumably due to Medicare eligibility. We now address the question, Does this increase in utilization lead to an improvement in outcomes—a reduction in morbidity and mortality—relative to what one would expect given the trends in outcomes prior to age 65? Bed Days Data on one important indicator of morbidity—mean number of days spent in bed in the last twelve months, by age—are available from the National Health Interview Survey (NHIS). The purpose of the NHIS is to obtain information about the amount and distribution of illnesses, their effect in terms of disability and chronic impairments, and the kinds of health services people receive. I calculated mean annual bed-days from NHIS person files for the five years 1987-1991. These files contain data on about 142,000 people between the ages of 50 and 80. Mean annual bed days, by five-year age groups, are shown in figure 2.11. Mean bed days increases by 0.62 from ages 50-54 to ages 55-59, and increases even more—by 1.63 days—from ages 55-59 to ages 60-64. However, mean bed days of 65 to 69-year-olds is slightly lower than that of 60 to 64-year-olds. If the pre-age-65 trend (14 percent average quinquennial growth rate) had continued, mean bed days of 65 to 69-year-olds would have been 15 percent higher—10.58 days as opposed to 9.21 days. Mean bed days of 70 to 74-year-olds and 75 to 80-year-olds would also have been about 15 percent higher. These estimates are consistent with the hypothesis that the Medicare-induced increase in health care utilization at age 65 leads to a reduction in days spent in bed of about 13 percent.

Mortality To examine the shape of the age-mortality profile, I will use data taken from the period life table. There are two types of life tables—the generation or cohort life table and the period life table. The generation life table provides a longitudinal perspective because it follows the mortality

Figure 2.11 Mean number of bed days in last 12 months, by age

42

Lichtenberg

experience of a particular cohort (all persons born in the year 1900, for example) from the moment of birth through consecutive ages in successive calendar years. Based on age-specific death rates observed through consecutive calendar years, the generation life table reflects the mortality experience of an actual cohort from birth until no lives remain in the group. To prepare just a single complete generation life table requires data over many years. It is not feasible to construct generation life tables entirely on the basis of actual data for cohorts born in this century. It is necessary to project data for the incomplete period for cohorts whose life spans are not yet complete. The better-known period life table may, in contrast, be characterized as cross-sectional. Unlike the generation life table, the current life table does not represent the mortality experience of an actual cohort. Rather, the current life table considers a hypothetical cohort and assumes that it is subject to the age-specific death rates observed for an actual population during a particular period. For example, a current life table for 1995 assumes that a hypothetical cohort is subject throughout its lifetime to the age-specific death rates prevailing for the actual population in 1995. The current life table may thus be characterized as rendering a "snapshot" of current mortality experience, and shows the long-range implications of a set of age-specific death rates that prevailed in a given year. Period life tables are produced annually by two different federal agencies: the National Center for Health Statistics (NCHS) and the Social Security Administration (SSA), Office of the Actuary. Wilkin (1981) discusses the methods used to construct both sets of life tables and their relative reliability. NCHS tables are based primarily on data obtained from death certificates. Misstatement of the age of the decedent on death certificates is known to be a serious problem. SSA life tables utilize administrative data from the Medicare program. As Wilkin observes, over the years, the Medicare program has accumulated a large quantity of reliable data on the mortality of the aged. The problem of misstatement of age is greatly reduced in this case because most of the data relate to individuals who have had to verify their dates of birth to become entitled to benefits under the program.6 The problem of underregistration of deaths is small because the availability of a small lump-sum death payments on insured workers' accounts encourages survivors and funeral directors to report deaths. The problem of underenumeration of population is negligible because the group under observation is defined by program records; thus, the data do not in-

Effects of Medicare on Health Care Utilization and Outcomes

43

elude deaths of unobserved persons. Further, the data are so extensive, covering nearly the entire aged population of the United States, that meaningful analyses can be done over relatively short periods of time (and, hence, trends through time can be detected accurately). Wilkin concludes that "the Medicare data appear to be more accurate by age and more consistent through time than the NCHS data." The trustees of the Social Security system base their projections of income and outlays on SSA life tables rather than NCHS life tables. Therefore I will examine data on age-specific mortality rates from the SSA period life table. In particular, I will use the 1995 SSA period life table. The table provides data on the probability of dying within one year ("death probability"), by exact age (age = 1, 2, . . . , 119) and gender. Death probabilities of men, by age, are shown in figure 2.12. It seems in this figure that the death probability increases smoothly from about 1 percent at age 55 to about 5 percent at age 75. However, the appearance of smoothly increasing death probabilities is deceptive. Figure 2.13 depicts the percentage increase in the male death probability from the previous year. From age 50 to age 65, the death probability increases at an increasing rate. Initially, the death rate increases about 8 percent a year, and the growth rate rises fairly steadily to about 10 percent by age 65. But between ages 65 and 69, the slope of the curve is quite negative. The probability of death continues to increase, but more slowly than it did up until age 65. As figure 2.14 reveals, there is a similar dramatic decline in growth in the probability of the death of women after age 65. Suppose that, instead of declining after age 65, the growth rate of the probability of death for men had continued to grow at the rate it had grown from age 50 to age 65. Then as figure 2.15 indicates, the probability that a 65-year-old man would live at least 10 more years would have been 63.5 percent, rather than the actual probability of 68.6 percent. The post-65 slowdown in death probability raised the odds of being able to celebrate one's 75th birthday by 5.1 percentage points.7,8 This evidence is consistent with the hypothesis that the Medicareinduced increase in health care utilization at age 65 leads to slower growth in the probability of death after age 65.I performed a formal test of this hypothesis using regression analysis. Using data for ages 51 to 75,I estimated the following regression equation: dj = -1.86 + .809 dj-1 - .095 visitsj + .030 hospj + .018j (t = 1.40) (9.68) (3.28) (0.63) (2.54)

Figure 2.12 1995 death probability of men, by age

Figure 2.13 Percentage increase from previous year in probability of death: Men Source: Author's calculations based on Social Security Administration, 1995 Period life table, http://www.ssa.gov/statistics/ Supplement/1998/Tables/PDF/t4c6.pdf

Figure 2.14 Percentage increase from previous year in probability of death: Women Source: Author's calculations based on Social Security Administration, 1995 Period life table, ment/1998/Tables/PDF/t4c6.pdf

http://www.sss.gov/statistics/Supple-

Figure 2.15 Actual versus predicted probabilities of survival from age 65 to age t: Males

48

Lichtenberg

where dj = the log of the male death rate at age j visitsj = the log of the number of physician visits at age j hospj = the log of the number of hospital discharges at age j The hospital coefficient is not statistically significant, but the visits coefficient is highly significant (p value = .004), indicating that physician visits have a negative effect on the male death rate, conditional on age and the death rate in the previous year. In the short run, the elasticity of the death rate with respect to the number of physician visits is -.095; a 10 percent increase in the number of visits leads to an immediate reduction in the death rate of 0.95 percent. In the long run, the elasticity of the death rate with respect to the number of physician visits is -.497 (= -.095/[1 - .809]) a permanent or sustained 10 percent increase in the number of visits ultimately leads to a 5 percent reduction in the death rate. Mortality: An Alternative Approach The analysis in the previous section was based on age-specific death probabilities in a single year (1995). But data on age-specific death probabilities are available from NCHS (Anderson (1997) every 10 years back to 1900, that is, before as well as after Medicare was enacted. Medicare, which began in 1966, primarily benefits people age 65 and over.9 Hence 70-year-olds in 1970 and 1980 benefited from the program, but 70-year-olds in 1960 did not, nor did 60-year-olds in any year. An alternative way to test for the effect of Medicare on longevity is to estimate models of the following form:

where Sit = the survival rate of age group i in year t(i = 1,5,10,15,..., 100; t = 1900,1910, . . ., 1990, 1997) and "shift" is defined in various ways to test for shifts in survival rates.10 This model allows for both a different mean survival rate and a different trend rate of increase for each of the twenty-one age groups. If Medicare resulted in an upward shift of the survival of people over 65 after 1966, then the appropriate definition of the shift variable is: shift = 1 if year > 1966 and age > 65 = 0 otherwise

Effects of Medicare on Health Care Utilization and Outcomes

49

When shift is defined in this way, the point estimate (t statistic) of b is 0.132 (8.28). This provides strong support for the hypothesis that Medicare increased the survival rate of the elderly, by about 13 percent. To ensure that this shift corresponds to Medicare as opposed to some other factor(s), we can change the definition of the shift term; that is, we can choose an earlier or later year or a different age. The results of this sensitivity analysis are shown in table 2.1. Lines 2 and 3 indicate that there is no evidence of a shift in the survival rate of people over 65 before 1966 (in either 1950 or 1960). There is stronger evidence of a shift in 1970 than there is of one in 1980 (line 4). Line 5 shows that there is no evidence of a shift in the survival rate of people between the ages of 40 and 65 after 1966. (Although the survival rates of people in this age group increased, there was no shift in the time trend after 1966, as there was for older people.) V. Summary Medicare, which provides health insurance to Americans over the age of 65 and to Americans living with disabilities, is one of the government's largest social programs. It accounts for 12 percent of federal onand off-budget outlays. In fiscal year 1999, $212 billion in Medicare benefits were paid. The largest shares of spending are for inpatient hospital services (48 percent) and physician services (27 percent). In thirty years, the number of Americans covered by Medicare will nearly double to 77 million, or 22 percent of the U.S. population. Perhaps the most important question we can ask about the Medicare program is, What impact does it have on the health of the U.S. population? One feature of the Medicare program can be exploited to shed light on its impacts: its age specificity. Most people become eligible for Medicare suddenly, the day they turn 65. Consequently, the age profiles of health services utilization and health outcomes (morbidity and mortality) can provide revealing evidence about the effects of Medicare. I have attempted to obtain precise estimates of medical utilization and outcomes, by single year of age, for ages close to age 65. The most precise estimates can be obtained by using information obtained from medical providers (hospitals and doctors) pooled over several years. I found that the utilization of ambulatory care and, to a much smaller extent, inpatient care increases suddenly and significantly at age 65, presumably due to Medicare eligibility. The evidence points to a

50

Lichtenberg

Table 2.1 Estimates of equation (2.1) with alternative definitions of shift variable (t statistics in parentheses) Year criterion

b

Age > 65

Year > 1970

2

Age > 65

Year > 1950

3

Age > 65

Year > 1960

4

Age > 65

Year > 1980

5

40 < Age < 65

Year > 1970

0.132 (8.28) 0.018 (0.91) 0.013 (0.66) 0.102 (6.57) 0.004 (0.2)

Line

Age criterion

1

structural change in the frequency of physician visits precisely at age 65. Attainment of age 65 marks not only an upward shift but also the beginning of a rapid upward trend (up until age 75) of about 2.8 percent per year in annual visits per capita. The number of physician visits in which at least one drug is prescribed also increases at age 65. Reaching age 65 has a strong positive impact on the consumption of . hospital services, but most of this impact appears to be the result of postponement of hospitalization in the prior two years. I also examined whether this increase in utilization leads to an improvement in outcomes—a reduction in morbidity and mortality—relative to what one would expect given the trends in outcomes prior to age 65. The estimates were consistent with the hypothesis that the Medicare-induced increase in health care utilization leads to a reduction in days spent in bed of about 13 percent and to slower growth in the probability of death after age 65. Physician visits are estimated to have a negative effect on the male death rate, conditional on age and the death rate in the previous year. The short-run elasticity of the death rate with respect to the number of physician visits is -.095, and the long-run elasticity is -.497; a permanent or sustained 10 percent increase in the number of visits ultimately leads to a 5 percent reduction in the death rate. Data on age-specific death probabilities every 10 years since 1900, that is, before as well as after Medicare was enacted, provide an alternative way to test for the effect of Medicare on longevity. They also provide strong support for the hypothesis that Medicare increased the survival rate of the elderly, by about 13 percent.

Effects of Medicare on Health Care Utilization and Outcomes

51

Notes 1. I calculated this by estimating the following regression: Aln(Privt) = 3.92 - 0.319Dln(Pubt) - .0020t (t = 4.24) (3.45) (4.22) Adjusted R2 = 0.327 Sample period: 1961-1998 Privf = real private health expenditure Pubt = real public health expenditure

Priv' = mean real private health expenditure Pub' = mean real public health expenditure 2. The monthly Social Security benefit is about 25 percent lower if one retires at age 62 than it is if one retires at age 65. As a general rule, early retirement will give one about the same total Social Security benefits over one's lifetime, but in smaller amounts to take into account the longer period during which they will be received. 3. NAMCS was not conducted in 1974,1982-1984, and 1986-1988. 4. A Medicare visit is defined as a visit in which Medicare is the expected principal source of payment. 5. In 1998, the elderly accounted for 23.8 percent of physician office visits. Medicare was the expected primary source of payment for 19.2 percent of physician office visits. 6. Proof of date of birth requires the submission of a public record of birth or a religious record of birth or baptism. Where no such document is available, the individual must submit another document or documents that may serve as the basis for a determination of his or her date of birth, provided that such evidence is corroborated by other evidence or by information in the records of the Social Security Administration. 7. The corresponding increase for women is only about one-third as large because women's death probabilities at given ages are significantly lower than are men's. 8. In principle, one could calculate the effect of the decline in mortality growth rate on life expectancy at age 65, which is perhaps the most interesting summary statistic. However, this requires predicting counterfactual mortality rates at advanced ages, a potentially speculative undertaking. 9. When it was introduced, 100 percent of Medicare beneficiaries were elderly; today about 14 percent of them are nonelderly disabled. 10. The survival rate is 1 - the death rate. Here, the survival rate is defined as the 5-year rate, for example, the probability of surviving from age 65 to 70.

52

Lichtenberg

References Anderson, R. N. (1999). United States Life Tables, 1997. National Vital Statistics Reports, Vol. 47, No. 28. Hyattsville, MD: National Center for Health Statistics. Health Care Financing Administration (2000). Medicare 2000:35 Years of Improving Americans' Health and Security, July. Wilkin, John C. (1981). "Recent Trends in the Mortality of the Aged," Transactions of the Society of Actuaries, Vol. XXXIII, 11-62.

3 Effects of Competition Under Prospective Payment on Hospital Costs Among High- and Low-Cost Admissions: Evidence from California, 1983 and 1993 David Meltzer, University of Chicago Jeanette Chung, University of Chicago

Executive Summary Competition and prospective payment systems have been widely used to attempt to control health care costs. Although much of the increase in medical costs over the past half-century has been concentrated among a few high-cost users of health care, prospective payment systems may provide incentives to reduce expenditures selectively on high-cost users relative to low-cost users, and this pressure may be increased by competition. We use data on hospital charges and cost-to-charge ratios from California in 1983 and 1993 to examine the effects of competition on costs for high- and low-cost admissions before and after the establishment of the Medicare Prospective Payment System (PPS). Comparing persons above and below age 65 before and after the establishment of PPS, we find that competition is associated with increased costs before PPS in both age groups, but decreased costs afterwards, especially among those above age 65 with the highest costs. We conclude that the combination of competition and prospective payment systems may result in incentives to reduce spending selectively among the most expensive patients. This conclusion raises important issues relevant to the use of competition and prospective payment to control costs. It also implies that, at minimum, there is a need to carefully monitor outcomes for the sickest patients under prospective payment systems in competitive environments. I. Introduction After a half-century of extraordinary growth in health care expenditures in the United States, there is now evidence that health care spending growth is slowing. Why this is occurring and how long it may last is not known, but a substantial body of literature suggests that two key elements of the efforts to contain costs may have played a role: the use of prospective payment systems (Russell and Manning 1989) and the encouragement of competition among providers (Melnick and Zwanziger 1988). Indeed, the combination of these two approaches seems to

54

Meltzer and Chung

be particularly important because competition in the absence of prospective payment systems has been suggested to increase costs (Robinson and Luft 1985), and prospective payment in the absence of competition provides no financial incentive to provide quality care. While most theoretical discussions of the effects of prospective payment hinge on the incentives to provide lower levels of care under fixed reimbursement and do not discuss the differential incentives to provide care to different types of patients,1 a few theoretical examinations of prospective payment have also incorporated the differential incentives for spending on profitable and unprofitable patients (for example, Allen and Gertler 1991, Ellis 1998). Meanwhile, several empirical studies have examined the differential effects of prospective payment systems on low- and high-cost patients. For example, Ellis and McGuire (1996) show how prospective payment for mental health services under Medicaid in New Hampshire resulted in reduced expenditures selectively among the sickest patients. In the context of the Medicare PPS, Newhouse (1989) finds that, while patients in unprofitable diagnosis related groups (DRGs) were not more likely than other patients to be transferred under PPS, they are more likely to be found in "hospitals of last resort," suggesting that there is selection according to profitability. Similarly, Meltzer and Chung (2000) show that hospital spending for the elderly in California under Medicare PPS was selectively reduced among the most expensive patients. Indeed, these reductions occurred despite an overall pattern among the young and elderly prior to the implementation of Medicare PPS for cost growth to be greatest among the most expensive patients, as reflected in the increasing concentration of health care expenditures over this century (Cutler and Meara 1998). Meltzer and Chung show that this same pattern of selective cost reduction for the most expensive patients is present within the twelve largest DRGs, the categories by which Medicare reimburses hospitals under PPS. The possibility that prospective payment systems may lead to a redistribution of resources from sick and costly persons within a payment category to healthier and more profitable ones cuts, in many ways, against a fundamental tenet behind prospective payment systems: namely, the subsidization of unprofitable patients by the profitable. Nevertheless, competitive pressures could lead to such an outcome as hospitals that attempt to support the care of unprofitable patients with revenue from profitable ones find the profitable patients wooed away by other hospitals that have chosen to invest resources in

Effects of Competition Under Prospective Payment on Hospital Costs

55

amenities that may appeal to patients and their doctors, but that are not necessarily directly associated with producing better outcomes for the most severely ill. In this paper, we use California data on patient charges and hospital cost-to-charge ratios from 1983 and 1993 to explore the effects of competition under prospective payment on hospital costs for low- and high-cost admissions within the twelve largest DRGs.2 Since Medicare PPS was implemented nationwide by states nearly simultaneously, we have to identify the effects of PPS on hospital costs mainly through temporal cross-sectional analyses rather than cross-state analyses. In an attempt to separate the effects of PPS from temporal changes in market competition, however, we contrast the effects of competition on costs for admissions of persons older than 65 versus costs for admissions of persons under 65. Complicating this analysis is the consideration of contemporaneous changes in the organization and financing of MediCal, California's Medicaid agency. In particular, the development of a selective provider contracting program and per-diem reimbursement system by Medi-Cal, in addition to the increasing use of managed care arrangements, all contributed to suppressing hospital growth among the young in California over the 1983-1993 period. Although we cannot prove that the patterns we observe are due to Medicare PPS, we find clear evidence that increased competition is associated with increased costs among the elderly before the implementation of PPS, but decreased costs afterward, with the reductions in costs clearly much greater among high-cost admissions than among low-cost admissions. This is consistent with the idea that the incentives created by Medicare PPS may have selectively reduced expenditures on the high-cost elderly. We begin in Section II with a short overview of the most important cost-containment efforts prevailing in California during this period: the Medicare Prospective Payment System, the California Medi-Cal selective provider contracting program, and the expansion of managed care. The description of PPS provides the institutional context for the effects of PPS we aim to investigate, while the discussion of the changes in reimbursement strategies among the young provides some insight into the use of the temporal changes in the effects of competition on costs for the young as a comparison. In Section III, we develop the theoretical motivation for our analyses using a model of provider response to fixed-rate, prospective reimbursement, in which quality can be varied for patients who differ in their underlying severity of illness and,

56

Meltzer and Chung

hence, profitability. In Section IV, we describe our data, and in Section V, we present the results of our analyses of the effects of competition on cost. Section VI concludes and examines the implications of our work for the design of reimbursement strategies, for quality assessment, and for outcomes research. II. Background on Cost-Containment in California, 1983-1993 Between 1983 and 1993, diverse cost-containment strategies were undertaken in California. They led to a widespread transition to prospective payment systems as well as intensified hospital market competition.3 Here, we discuss briefly major cost-containment strategies that were implemented over this period: the Medicare Prospective Payment System; selective provider contracting among California's Medicaid program, Medi-Cal, and health care providers; and the expansion of managed care arrangements. Medicare PPS Prospective payment systems certainly existed prior to the establishment of the Medicare PPS in 1983. Nevertheless, the scale and influence of Medicare made the shift from retrospective reimbursement on the basis of reasonable costs to PPS a change of fundamental importance for hospitals. With the establishment of PPS, reimbursement for nearly all hospitalizations under Medicare were made on the basis of prospectively fixed rates according to diagnosis-related groups. Each hospitalization is assigned a DRG based on principal diagnosis or the performance of a very limited number of particularly costly procedures (for example, coronary artery bypass graft surgery). Each DRG is assigned a fixed weight that reflects its relative cost of treatment with respect to a base rate. Because hospitals are paid a fixed amount per DRG based on the DRG weight, the classification system and aimed to create groups of patients as homogeneous as possible with respect to resource consumption. DRGs were also stratified with respect to age and the presence of complications. After a phase-in period of four years, during which hospital reimbursement reflected a mix of national, regional, and facility-specific rates (Smith and Fottler 1985), hospitals were reimbursed for each case according the national average cost of treating a base case (with adjustments to reflect location and local wages), multiplied by the DRG

Effects of Competition Under Prospective Payment on Hospital Costs

57

weight (Davis et al. 1990). Thus, reimbursement under PPS was fully prospective from the onset, but the persistence of differences in payment rates based on historical local costs meant that the competitive aspects of PPS increased progressively over its phase-in. Medi-Cal Selective Contracting The same year that Medicare PPS was implemented, California enacted legislation authorizing the state Medicaid program, Medi-Cal, to negotiate contracts with health service providers for the care of Medi-Cal beneficiaries. This was done with the intent to promote price competition in the Medicaid market. Under this legislation, eligible, short-term, acute-care general hospitals were offered the opportunity to negotiate service provision contracts with Medi-Cal on the basis of fixed per-diem rates (Johns 1985). Failure to secure a contract meant that hospitals would not be reimbursed for care given to Medi-Cal patients except in cases of emergency (Langa 1992). Although the per-diem reimbursement established under this legislation did not result in a fully prospective payment system for Medi-Cal patients, the resulting declines in Medi-Cal reimbursement also intensified the competitive pressure on California hospitals during this period. Expansion of Managed Care During the 1980s, managed care spread rapidly throughout the United States, but particularly in California. By 1988, California ranked first in the nation in terms of its HMO enrollment rate, with roughly 28.5 percent of the state population (7.68 million individuals) belonging to an HMO (Johns 1989). This was more than double the national rate in 1987, when only 12.1 percent of the U.S. population was enrolled in an HMO (Davis et al. 1990), and even well above the national rate of 19.7 percent in 1994 (Institute of Medicine 1997). Likewise, the number of Preferred Provider Organizations (PPOs) in California grew 94 percent, from 34 PPOs in 1984 to 72 in 1988 (Johns 1989). Some managed care payers adopted prospective payment systems for hospital care similar to Medicare PPS. However, the majority adopted other approaches to cost control, such as selective contracting, per diem reimbursement, and utilization review. These mechanisms did not necessarily provide any particular incentive to decrease expenditures for high-cost users relative to low-cost users (Gold et al. 1995).4

58

Meltzer and Chung

Nevertheless, many aspects of managed care served to further intensify competition in California during these years. Indeed, the empirical evidence suggests that Medicare PPS, Medi-Cal selective contracting, and managed care arrangements all contributed to curbing cost growth. From 1967 to 1984, Medicare hospital care expenditures had been growing at an average annual rate of 16.5 percent; in the seven years immediately following PPS, growth fell to 7.3 percent (Davis and Burner 1995). Based on projections of Medicare expenditures, Russell and Manning (1989) estimated savings of $12 to $18 billion for 1990 under PPS. Medi-Cal selective contracting in California also appears to have been largely successful in raising the level of competition in hospital markets while simultaneously suppressing cost growth (Johns 1989, Robinson and Phibbs 1989, Melnick et al. 1992). The growth of managed care organizations also contributed to lower cost growth, both by delivering health care at lower costs due to lower service intensity (Manning et al. 1984) and by increasing competition in hospital markets (Melnick and Zwanziger 1995). Thus, between 1983 and 1993, hospitals in California became increasingly subject to prospective payment systems as a result of Medicare PPS, and increased competition due to the effects of Medi-Cal selective contracting and the growth of managed care. In this context, economic theories of provider behavior under prospective reimbursement suggest incentives to decrease expenditures on high-cost patients while increasing expenditures on low-cost patients, as we explore below. III. Economic Theories of Provider Behavior Under Fixed-Rate Prospective Payment Systems Many cost-containment strategies rely on supply-side cost sharing to achieve cost-containment objectives. Whereas retrospective reimbursement systems largely insulate providers from increases in costs, providers under prospective payment systems are paid a fixed rate per unit of output that is defined in advance. If the patient population is taken as given, such payment schemes that hold providers financially responsible for the marginal costs of treatment can create incentives to reduce provision of unnecessary services to patients. This is reflected by the common view of managed care as reducing services. What is less appreciated, however, is that when providers have to compete for patients, prospective payment systems also create a new distinction among patients, namely, a distinction between profitable and unprofitable patients, depending on their ex-

Effects of Competition Under Prospective Payment on Hospital Costs

59

pected costs relative to the level of prospective reimbursement (Newhouse 1989). Thus, when profit-maximizing hospitals under fixed-rate prospective reimbursement face a patient population of variable illness within a reimbursement category (such as a DRG), they may have incentives to provide excessive levels of care for the less ill and/or to choose and advertise quality of care or amenities that differentially attract profitable patients while avoiding unprofitable ones (Hornbrook and Rafferty 1982, Ellis and McGuire 1986, Dranove 1987, Luft and Miller 1988, Newhouse 1989, Hodgkin and McGuire 1994, Ellis 1998).5 When intensified competition decreases overall profit levels and increases the price responsiveness of patient volume, such strategies may become matters of institutional survival. Thus, as Ellis (1998) has shown, incentives to engage in patient selection and discrimination in quality provision are exacerbated under increased competition, a condition that has been realized in many U.S. hospital markets in recent years due to greater market penetration by managed care organizations (Ellis 1998, Dranove and White 1994). The empirical implications of these theories is that, where providers are subject to fixed-rate prospective payment systems, declines in hospital cost growth will be concentrated at the top of the spending distribution. In other words, high-cost (unprofitable patients) will experience greater reductions in resource consumption relative to low-cost (profitable) patients. Furthermore, these effects will be magnified under competition. To illustrate this, we develop the following model of choice of quality of care for patients with differing severity of illness (s) given a prospective payment rate (P). Specifically, we model the choice of quality of care for patient of severity s at cost c(qs ). To capture the variation in costs of patients who differ in severity of illness, we allow the cost of providing basic care to depend also on severity [c(s)]. Thus, the total cost of caring for a patient of a given severity is c(s) + c(qs). The first component, c(s), is nondiscretionary, whereas the second component, c(qs), is subject to choice depending on the desire of a hospital to provide additional quality. In other words, we model the profit from caring for a patient of a given severity level (s) under prospective payment as:

We assume that cs > 0, css > 0 and that, with respect to the cost of delivering discretionary quality, c(0) = 0, cq > 0, and cqq > 0. To go from

60

Meltzer and Chung

this patient-level profit to the profit earned from caring for the class of patient of severity s, we allow the demand for care by patients of severity s when quality is qs to be D(s,qs) = D(qs), Dqs>0,Dqsqs< 0. The hospital chooses qs to maximize profit: Max P = D(qs)[P - c(s) - c(qs )] qs

(3.2)

Taking the derivative of equation (3.2) with respect to qs yields the first-order optimality condition that hospitals set the marginal revenue from additional quality equal to the marginal cost of providing that quality:

Equation (3.3) implies:

Totally differentiating and checking second-order conditions demonstrates:

as long as P — c(s) > 0, and qs = 0 otherwise. Thus, discretionary quality falls with severity for all profitable patients and is set to zero for all unprofitable patients. Comparison to Retrospective Reimbursement Since one of the empirical comparisons we will make is between prospective reimbursement and retrospective reimbursement, it is useful to contrast this result with the result that would transpire under a retrospective reimbursement system, according to our model. In particular, instead of a fixed price P that is independent of severity and quality, a retrospective reimbursement system may in general depend on both, for example, P(s, qs ). Under some circumstances, this makes the comparison between prospective and retrospective reimbursement easy; in others, it is more difficult. To illustrate this, assume P(s, qs) takes the general form P(s, qs ) = P0 + Psc(s) + Pgsc(qs ), where P0, Ps, and Pqs are the

Effects of Competition Under Prospective Payment on Hospital Costs

61

rates at which hospitals are reimbursed for, respectively, the basic admission (as in prospective payment), expenditures on severity-related costs, and expenditures on discretionary dimensions of quality (for example, amenities). In general, the latter two categories may be hard to distinguish in practice, but the distinction is worth making to reflect the idea that there may be some expenditures that may not be covered fully under a retrospective reimbursement system but are nevertheless desired by hospitals to attract patients.6 For our purposes, the most straightforward case is when reimbursement provides a fixed amount of profit per admission by providing a lump-sum profit (K) per admission and exactly reimburses severity-related costs while not covering quality-related costs. In that case, P0 = K > 0 , P s = 1, and Pqs = 0. Equation (3.4) becomes

and quality is independent of s. In this case, the shift to prospective payment would be expected to decrease spending for more expensive patients relative to less expensive ones. Perhaps even more relevant is the case in which retrospective reimbursement provides no fixed profit per admission but instead offers a markup over costs for severity-related costs, for example, P0 = 0, Ps > 1, and Pqs = 0. Here equation (3.4) becomes

Thus, with Ps> I, hospitals make more profit on more expensive patients and therefore will spend more on quality for the more expensive patients. Again, the switch to prospective payment will lead to a reduction in spending among the sicker patients. Finally, it is also worth considering a system in which no fixed profit per admission is given but in which all costs related to both severity and quality are reimbursed retrospectively with a markup. Some might consider this most like the retrospective reimbursement system as it was applied prior to prospective payment. In this case, P0 = 0, and Ps = Pq s > 1. Equation (3.4) now becomes:

and again

62

Meltzer and Chung

This seems to suggest that quality would rise with severity, but this conclusion would be misleading because full retrospective reimbursements for amenities provide no incentives for hospitals to limit expenditures on amenities. Hence, the second-order conditions actually imply that the optimal quality hospitals should provide is infinite quality for all patients. Thus, there must be some other constraint on the reimbursement of discretionary care, which seems most likely to be a combination of the possibility of doing harm to the patient (and associated risk of liability) and whatever limits are placed by the payer. Whichever the case, it is not possible to predict how prospective payment will affect discretionary expenditures on low- and high-cost patients. To summarize, except in the case where discretionary expenditures are not limited by economic incentives, there appears to be a fairly broad set of assumptions under which prospective payment would be expected to reduce expenditures selectively for the most expensive patients relative to retrospective reimbursement. Effects of Competition Equation (3.4) implies that the ratio of profit to cost for quality falls with increasing elasticity of demand with respect to quality so that, accordingly, quality rises with increasing elasticity of demand with respect to quality. Since the out-of-pocket cost of a hospitalization to a Medicare patient is independent of the hospital she or he chooses, it seems likely that competitive pressures will make this elasticity quite large, although such competitive forces will surely be limited by geographic factors in areas where there are few hospitals so that patient options are limited due to high search and transportation costs, and where changes in quality are more likely to be coordinated (Bain 1951, Stigler 1968, White 1972, Tirole 1988).7 Rearranging equation (3.4) and solving for c(qs ) yields:

Quality is set to a minimum for unprofitable patients, so equation (3.6) applies where patients are profitable (thus, the numerator is positive), and quality for profitable patients rises with the degree of competition. As indicated above, quality falls with increasing severity, and here the rate at which expenditures on quality fall with increasing severity rises with increasing elasticity of demand with respect to quality

Effects of Competition Under Prospective Payment on Hospital Costs

63

(for example, competition), so that the positive effect of competition on costs is dampened for more costly patients. Thus, an increasingly competitive environment under prospective payment has the effect of raising quality most for the least costly patients. Because competition under prospective payment may also increase efficiency, this may not result in an absolute increase in costs but should at least lessen cost decreases for the least expensive patients relative to the most costly patients, for whom the clear incentive is to reduce expenditures if possible because they are not profitable. Indeed, in the limit, as the elasticity of demand with respect to quality approaches infinity, expenditures on quality fall dollar for dollar with increasing severity of illness because all profits are competed away at each level of severity. IV. Data and Methods Data Description: California Hospital Cost and Financial Data We use the 1983 and 1993 hospital discharge and financial data released for public use by the California Office of Statewide Health Planning and Development (OSHPD). The financial data is described in detail below. The discharge data cover all inpatient discharges from every licensed, nonfederal hospital in California, as well as discharges from some specialized facilities such as psychiatric hospitals and rehabilitation and nursing facilities. Data elements available for each patient abstract in the public-use files include facility identifiers, patient age, zip code of patient residence, expected source of payment, total charges incurred by patients during their hospitalization episode, and patient DRG classification. Additional data for calculating per-capita spending and utilization rates comes from the U.S. Bureau of the Census Intercensal Population Estimates by Age, Sex, and Race (U.S. Department of Commerce, Bureau of the Census 1993,1998). We limit our analysis to all California state residents (identified by zip code) discharged from acute-care facilities for which data on total hospital charges are available. Certain institutions, many of which are managed-care facilities such as Kaiser hospitals, do not report total charges on their discharge abstracts because they are exempt from standard OSHPD accounting procedures. As a convention, patients discharged from these hospitals have zero charges recorded in their abstracts, although true costs of treatment were nonzero. Since total hospital charges for these patients cannot be ascertained, they are excluded from our analyses.8

64

Meltzer and Chung

To calculate costs, we begin with charge data that we convert to 1993 constant dollars using the general Consumer Price Index and then to costs using annual, institution-specific ratios of costs-to-charges (RCCs). These RCCs are calculated using the OSHPD Financial Disclosure Data, which report facility-level data on total operating expenses, gross patient revenue, and other nonoperating revenue. Because other nonoperating revenue consists of revenue from hospital enterprises such as the outpatient pharmacy and gift shop, we follow the approach recommended by the Office of Statewide Health Planning and Development (1993) in calculating RCCs:

RCCs are commonly used to estimate costs from charges, but OSHPD data do not permit disaggregation of inpatient charges into its component departments and services. Thus, institution-level RCCs must be used, which is an important limitation because they cannot reflect discrepancies between costs and charges that arise due to internal cross-subsidization across departments and services within a facility. Nevertheless, facility-level RCCs can adjust for certain discrepancies between costs and charges [for example, whether or not a facility treats a large proportion of charity cases (Finkler 1982)1 and have been found to perform somewhat better than charges as proxies for costs (Newhouse, Cretin, and Witsberger 1989; Schwartz, Young, and Siegrist 1995). While this suggests some justification for analyzing RCC-adjusted charges rather than raw charges, the most compelling reasons during the period we study is the growing inflation of charge rates to full-paying patients and the concomitant use of rebates for managed care contracts. The upshot of this is that charge growth based on charges may overstate real cost increases (Dranove, Shanley, and White 1991) if managed care rebates are not taken into consideration. The advantage of using RCCs in this case is that increase in gross patient charges that are offset by increases in rebates will result in a decrease in the RCC as calculated above. As a result, estimates of costs based on patient-level charges and RCCs are not inflated inappropriately by the use of rebates. In addition to the effects of discrepancies between costs and charges on aggregate charge growth, it is also important to consider the possibility that such discrepancies could have effects on costs across the spending distribution if they do not apply uniformly across it. Indeed,

Effects of Competition Under Prospective Payment on Hospital Costs

65

it is possible that the discrepancy between costs and charges could vary across the spending distribution. For example, if the markup on low-cost services and departments exceeds the markup on high-cost services and departments, then the actual distribution of costs across patients will be more concentrated than suggested by the distribution of charges. Although it is not clear that this is the case, it is possible that such markups may change over time, for example, if competition is particularly intense in high-cost services so that cost containment differentially reduces charges in these areas. If so, it is possible that an analysis of hospital charges may overstate costs at the bottom of the distribution in later years and understate costs at the top of the cost distribution. Although this would lead to patterns in hospital costs similar to those we find, we do not believe that internal cross-subsidization drives our results because we study a period in which all payers were tightening their reimbursement policies, thereby imposing a constraint on the extent to which hospitals could shift costs to other payers and departments. Indirect support for this comes from Dranove and White (1998), who studied the responses of California hospitals to Medicaid fee reductions between 1983 and 1992 and found significant reductions in levels of services provided to all patients, and Medicaid patients in particular, but no evidence of cost shifting. The ideal data to test this would allow us to assess whether rebates were more likely for sicker patients within a hospital, but the available data do not permit this disaggregated analysis because rebates are not made on a patient-level basis. As an alternative check, however, we examined whether hospitals in the OSHPD data that care for sicker patients (as measured by either greater average age, length of stay, or in-hospital mortality) were likely to give greater rebates to payers as a percentage of net revenue. Our results suggest no evidence of any significant relationship or change in relationship over time between rebates and patient age or length of stay, but they do suggest a positive relationship between rebates and mortality in the first six years we study and that this relationship is eliminated by the end of the period. While this latter result could suggest an artificial inflation of costs for the sickest patients initially that is later eliminated, the effect is not large. Thus, while there are possible reasons to be concerned that changes in the relationship between costs and charges across patients who differ in severity of illness could influence our results, we cannot find evidence of any changes in such relationships.

66

Meltzer and Chung

Limitations of Cost and Financial Data Several data and analytic limitations should be recognized at the outset. First, the 1983-1993 period was one during which hospital accounting and reimbursement systems were in flux. Hospitals are instructed by OSHPD to report the total charges incurred during a patient's hospitalization according to the facility's full-established rates prior to any prepayment deductions. At a minimum, hospitals are to include all charges associated with daily hospital services, ancillary services, and patient care services in calculations of total inpatient charges per discharge. Physician fees are omitted. Due to the volume of discharges processed, OSHPD does not conduct comprehensive accounting checks; hence, the reliability of reported data on charges is unknown. Nevertheless, the OSHPD charge data has been widely used by several researchers, for example, Robinson and Phibbs 1989, Stafford 1990, Langa 1992, Langa and Sussman 1993, Melnick and Zwanziger 1995, and Dranove and White 1998. Another issue is related to our lack of data concerning charges associated with outpatient care and forms of postdischarge care. Since the introduction of PPS and managed care, many have speculated that any decline in hospital spending may be offset by growth in other sectors such as ambulatory and long-term care. Since we are unable to account for cost shifting across sites of delivery, our finding that growth in hospital charges fell among high-cost admissions does not imply that the total cost of treatment among high-cost admissions also fell because these patients may be heavy consumers of postdischarge health care resources. However, we found no tendency for differential cost reduction among high-cost admissions with increasing competition in diagnoses with high or increasing levels of discharge to skilled nursing facilities. Even if such a pattern were found, it could be interpreted as providing insight into a mechanism by which quality discrimination was accomplished. A final point concerns the period over which we have data to analyze. The earliest data we have date back to 1983, the year in which Medicare's DRG-based Prospective Payment System was implemented and legislation authorizing selective contracting between Medi-Cal and service providers took effect. Also, throughout the 1983-1993 period, HMOs and various other managed care organizations emerged and proliferated. Because we do not have comparable data that antedate these major changes and because important changes were

Effects of Competition Under Prospective Payment on Hospital Costs

67

happening in the reimbursement strategies for younger persons at the same time, it is important to be cautious in drawing a causal connection between these specific policies and observed trends in charge growth. On the other hand, because Medicare PPS, Medi-Cal, and managed care all rely on different approaches to achieve cost containment, we can, to interpret our empirical findings, borrow insights from the theoretical models of provider responses to alternative reimbursement systems described above. Measures of Competition A large body of literature attempts to identify the appropriate measures of markets and competition in health care. Key debates in this literature concern methods for defining health care markets (for example, geopolitical boundaries, patient flow, or economic measures such as cross-price elasticities), the appropriate level of analysis (for example, facility level, medical service level, or patient level), and mathematical measures for computing market concentration (for example, the Hirschman-Herfindahl Index, spatial density of competitors, or entropy).9 Although these approaches may differ in their theoretical appeal, both in general and in individual applications, expediency has often been the operative criterion by which methodology has been chosen. By far, the most common approach has been to define markets on the basis of geopolitical boundaries (counties, Metropolitan Statistical Areas (MSAs), and/or Health Service Areas (HSAs) and to measure concentration using the Hirschman-Herfindahl Index (HHI) for total admissions at the county level.10 In the analyses reported here, we follow the same approach. Several studies have examined the robustness of the empirical findings in hospital markets compared to alternative methods of market delineation and concentration measurement. Some have found that results are not robust compared to methods of market definition (see, for example, Dranove, Shanley, and Simon 1992; Sohn 1996; Kessler and McClellan 1999). Consequently, future plans for our research include replicating our analyses using alternative measures of competition. Analytic Plan To analyze the effects of competition across the distribution of health care expenditures, we include measures of competition in quantile re-

68

Meltzer and Chung

gressions of cost for patients above and below age 65, and before the implementation of PPS in 1983 and after, in 1993. Our basic hypothesis is that competition under PPS will exert a downward pressure on costs among the most expensive elderly patients in 1993 relative to its effects among the less expensive elderly in 1993, relative to the expensive elderly in 1983, and relative to its effects among the young. To select the most appropriate comparison group among the young, we focus on persons age 55-64, though our results are not substantially different when we include persons age 5-54. Since our theory does not specify a specific measure of concentration, and since we have no reason to suspect a linear relationship between any particular measure of concentration and costs, we define a set of indicator variables to categorize counties in terms of their competitiveness based on the HHI ("less competitive" HHI > 0.20, "moderately competitive" 0.20 > HHI > 0.10, "competitive" 0.10 > HHI > 0.05, and "very competitive" HHI < 0.05). We also control for payer (Medicare, Medi-Cal, other nonprivate, and private), as well as various market-level and hospital characteristics. The market-level characteristics are log physicians per capita, log HMO enrollment rate, log county population, and log income per capita. Hospital-level characteristics in our model are ownership status (for-profit versus nonprofit), teaching status (teaching hospital versus nonteaching hospital), number of licensed beds, and total number of discharges per year. In our basic specification, we do not control for patient characteristics such as age or comorbidity because PPS does not base much, if any, of its reimbursement rate on those factors. As a result, selectively caring for patients who are younger or have fewer comorbidities may be a mechanism by which hospitals respond to PPS and limit costs. In other words, controlling for age and comorbidity could mask the effect we aim to identify. In alternative specifications, we also include patient age and the number of secondary diagnoses, but find little change in our overall results. We limit our analyses to the twelve largest DRGs by volume of discharges, more specifically, those DRGs with at least 10,000 discharges over the age of 4 in 1983 and 1994 combined. An important concern in this analysis relates to the incentives under Medicare PPS for hospitals to engage in "DRG creep," that is, the practice of progressively upcoding patients into DRGs with a higher reimbursement rate for a given condition (Carter, Newhouse, and Relies 1990). As a result of DRG creep, changes in charges within each stratified DRG may reflect trends in coding and classification rather than changes in service provi-

Effects of Competition Under Prospective Payment on Hospital Costs

69

sion. To address this concern, we aggregated DRGs for the same procedure and/or condition that are stratified for severity in calculating utilization rates and growth in charges.11 Adjustments for Changes in Discharge Rates To know how to interpret changes in the effects of competition at different points in the spending distribution over time, and to generate meaningful estimates of cost growth over time at different points in the spending distribution, it is important to consider the dramatic decline in admission rates in California over this period because a given position in the spending distribution may reflect a different degree of severity in different years. The California data show that per-capita hospital discharge rates declined steadily from 112 discharges per 1,000 total population in 1983 to 69 discharges per 1,000 total population by 1993. The decline in California's discharge rates is consistent with national utilization trends, which began slowing in the 1970s but declined even further since the 1980s. Much of the decline has been attributed to more widespread use of utilization control mechanisms by Medicare, state Medicaid programs, managed care, and other third-party payers. These controls include peer-review organizations, physician gatekeepers, and precertification requirements employed by Medicare and other third-party payers. In California especially, declining rates of discharges may also reflect the expansion of HMO enrollment and the shift of many services to outpatient settings. Assuming stable population morbidity from year to year, a falling admission rate implies that in each successive year, a smaller proportion of episodes of illness result in hospitalization. If one were to rank all admissions in order of increasing severity of illness, it would be reasonable to assume that, given the nature of utilization control measures, the distribution would tend to be truncated from the left, leaving the least severely ill patients denied hospital admission. Hence, not only does the proportion of the population experiencing hospitalization shrink over time, but the average severity level of the hospitalized population would also be expected to increase because there are fewer "healthy" admissions to dilute the spending distribution. Figure 3.1 illustrates how shifts in utilization rates can complicate intertemporal comparisons of expenditures at specific locations within the population spending distribution. The x axis plots the percentage of the population ranked in order of increasing severity of illness. The y

(a) Discharge distribution in year 0

(b) Discharge distribution in year 1 Figure 3.1 Intertemporal comparisons of percentile locations in the spending distribution, adjusting for changes in hospital utilization rates

Effects of Competition Under Prospective Payment on Hospital Costs

71

axis plots the frequency or number of individuals at each severity level. The curves in figure 3.1 (a) and 3.1(b) depict the distribution of illness in a given population, which we assume to be stable, at two time points: year 0 and year 1. H0 and H1 represent the discharge rates in year 0 and year 1, respectively. In this hypothetical population, the top 50 percent of the population ranked in terms of morbidity were hospitalized in year 0. In year 1, the admission rate fell to 40 percent. Suppose we wish to compare the effects on median hospital charges between year 0 and year 1. In year 0, the median discharge (M0) was the patient at the 75th percentile of the disease distribution. In year 1, the median discharge (M1) was at the 80th percentile of the disease distribution. Because of the falling discharge rate between year 1 and year 2, these two discharges are not directly comparable. This is seen in figure 3.2 by the dotted line that traces M0 down to the disease distribution in year 1, and by the dotted line that traces M1 above to the disease distribution in year 0. Thus, the median discharge in year 0 was less ill than the median discharge in year 1. Without taking into account falling discharge rates, a simple comparison between the median hospitalizations in year 0 and year 1 will compare patients that differ in their severity of illness. To address this concern due to falling admission rates, we also performed all our analyses based on adjusted percentiles in which we aim to compare persons with comparable levels of severity of illness. Therefore, we examine growth rates or the effects of competition at adjusted percentiles wherever discharge rates fell between two time points, according to the following formula:

where P0 is the adjusted percentile in year 0, P1 is the percentile in the later time period (year 1), and H0 and H1 are the discharge rates in the two corresponding years. For example, to compare costs at the median of the spending distribution of the hypothetical population, we should compare the median discharge in year 1 to the discharge at the 60th percentile of discharges in year 0 because:

We use this approach directly to calculate growth rates at different percentiles in the spending distribution. To analyze the effects of competi-

Figure 3.2 (a-1) The distribution of RCC-adjusted charges in DRG in 1993: age 65+

Effects of Competition Under Prospective Payment on Hospital Costs

73

tion, we implement this adjustment by performing our regression analyses using the same number of observations drawn from the top of the distribution of the 1983 data as we have in the 1993 data. Our method of adjustment exploits the fact that discharges fell over time and the fact that utilization control mechanisms typically raised the threshold of illness severity for hospital admission. This raises several potential problems. One is that in DRGs where discharge rates rise over time, it is not clear whether expanded services were extended to the less severely ill, or if improvements in technology and medical management enabled treatment of a greater number of the severely ill who would otherwise have remained untreated. Thus, in analyzing spending growth for the few DRGs where admission rates rose, we use all the observations from 1983 and analyze only unadjusted percentiles. Probably more important is the possibility that reductions in admission did not come uniformly from the left tail of the distribution during this period. In an extreme example, suppose that, although we assumed that the reduction of 43 admissions per 1,000 population between 1983 and 1993 came from the left of the distribution (the "healthy" side), the reductions actually came entirely from the left side of the distribution. This might happen, for example, if the 43 fewer admissions in 1993 were terminally ill individuals who had been shifted into hospices but who would have died in hospitals at high cost in 1983. The top of the 1993 distribution would then be expected to have a lower average severity of illness level compared to the top of the 1983 distribution—the opposite of our assumption. This implies that an unadjusted comparison would understate growth, and that our adjustment procedure would further exacerbate this. Fortunately, for the diagnoses we examine, we believe that most of the reductions in admissions are due to the movement of less severely ill patients to the outpatient setting. This is supported by the observation that the greatest declines in admission rates in our sample were among admissions for esophageal and gastrointestinal disorders, which likely results from a movement toward treatment of the least severely ill patients in an ambulatory setting. It is also supported by additional analyses we performed showing that the degree of comorbidity of patients in these DRGs increased over our sample period.12 Nevertheless, we also examined the robustness of our findings under the assumption that the reduction in admissions is distributed evenly across

74

Meltzer and Chung

the spending distribution by examining growth rates at the unadjusted percentiles. V. Results Distribution of RCC-Adjusted Charges by DRG: 1983 and 1993 Tables 3.1 and 3.2 show the number of cases and the distribution of costs in 1983 and 1993 for the twelve DRGs we examine. As figure 3.2 shows clearly, the distribution of costs in every DRG is highly skewed to the right, with about two-thirds of all admissions having costs below the mean (tables 3.1 and 3.2). This lays out the basic incentives implicit in PPS: that the majority of patients are profitable, while a minority are unprofitable but potentially responsible for large losses. Growth of RCC-Adjusted Charges by DRG: 1983 and 1993 Table 3.3 shows the growth of costs from 1983 to 1993 at unadjusted and adjusted percentiles for persons older than age 65. Although there are few exceptions, the vast majority of the unadjusted and adjusted growth rates clearly show falling growth with increasing position in the spending distribution, as predicted by the theoretical predictions of the effects of prospective payment on costs. Table 3.4 repeats these analyses for persons age 55-64. While the pattern is not as strong in several diagnoses as for those persons age 65 and older, there is still a clear trend for falling growth with increasing position in the spending distribution. This is not predicted by the theoretical model, and we will discuss possible reasons for this anomaly later. Effects of Competition on Hospital Expenditures: 1983 and 1993 For the sake of parsimony, we present the full results of quantile regression analyses examining the effects of competition on cost at selected points in the distribution for one DRG only—acute myocardial infarction (AMI). The rest of the results are summarized in a separate table. Table 3.5 reports the quantile regression results for AMI admissions among persons age 65 and older in 1983 and 1993. table 3.6 reports results of parallel analyses for persons in the 55-64 age group. Table 3.7 reports the coefficients on the competition variables from the quantile

Table 3.1 The distribution of RCC-adjusted charges by DRG in 1983 and 1993: age 65+ 1983 unadjusted distribution Unadjusted percentiles

Acute myocardial infarction Angina Arrythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

Standard deviation

25th

50th

75th

90th

95th

24,093 18,790 18,522 26,833 18,900 27,802 10,769 29,849 12,296 9,560 11,345 25,375

9,068 3,926 4,490 8,928 7,678 3,586 5,149 6,720 11,872 5,642 4,923 8,207

9,095 3,253 5,469 11,808 11,167 3,988 6,511 8,307 9,695 6,095 6,678 11,311

4,264 2,187 1,986 2,966 2,919 1,613 2,241 2,773 7,039 2,430 1,941 3,167

6,916 3,181 3,200 5,406 4,880 2,544 3,543 4,588 9,477 3,974 3,181 5,290

10,769 4,685 5,271 10,281 8,458 4,118 5,810 7,654 13,349 6,660 5,586 9,355

16,771 6,909 8,503 19,263 14,870 6,822 9,728 13,139 20,042 10,981 9,912 16,174

22,693 8,755 11,751 27,876 22,121 9,380 13,665 18,755 26,198 14,968 14,416 22,984

Table 3.1 (continued) 1983 adjusted distribution Adjusted percentiles

Acute myocardial infarction Angina Arrythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

Standard deviation

25th

50th

75th

90th

95th

22,368 18,790 18,522 26,833 18,900 17,169 10,769 29,849 12,296 9,560 11,345 25,375

9,651 3,926 4,490 8,928 7,678 4,950 5,149 6,720 11,872 5,642 4,923 8,207

9,184 3,253 5,469 11,808 11,167 4,538 6,511 8,307 9,695 6,095 6,678 11,311

4,837 2,187 1,986 2,966 2,919 2,709 2,241 2,773 7,039 2,430 1,941 3,167

7,344 3,181 3,200 5,406 4,880 3,617 3,543 4,588 9,477 3,974 3,181 5,290

11,188 4,685 5,271 10,281 8,458 5,422 5,810 7,654 13,349 6,660 5,586 9,355

17,352 6,909 8,503 19,263 14,870 8,545 9,728 13,139 20,042 10,981 9,912 16,174

23,610 8,755 11,751 27,876 22,121 11,710 13,665 18,755 26,198 14,968 14,416 22,984

Table 3.1 (continued) 1993 unadjusted distribution Unadjusted percentiles

Acute myocardial infarction Angina Arrythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

Standard deviation

25th

50th

75th

90th

95th

22,368 23,182 20,470 32,227 24,572 17,169 22,264 52,857 13,790 15,258 20,248 36,783

8,855 3,388 4,151 6,585 6,243 4,157 5,438 6,015 10,176 5,147 4,587 6,537

7,446 2,655 4,486 7,131 5,744 4,283 5,065 6,723 7,342 4,716 4,912 10,075

4,415 1,932 1,955 2,975 3,184 1,982 2,713 2,861 6,522 2,649 2,081 5,136

6,947 2,774 2,996 4,620 4,853 3,096 4,118 4,440 8,446 3,954 3,282 5,136

10,795 4,075 4,769 7,612 7,467 4,899 6,414 7,087 11,453 6,137 5,396 7,964

16,494 5,853 7,719 12,624 11,413 7,710 10,055 11,248 16,186 9,331 8,810 12,015

21,672 7,452 16,623 17,735 15,027 10,538 13,387 15,287 21,012 12,456 12,159 15,538

Table 3.1 (continued) 1993 adjusted distribution Adjusted percentiles

Acute myocardial infarction Angina Arrythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

Standard deviation

25th

50th

75th

90th

95th

22,368 23,182 20,470 32,227 24,572 17,169 22,264 52,857 13,790 15,258 20,248 36,783

8,855 3,388 4,151 6,585 6,243 4,157 5,438 6,015 10,176 5,147 4,587 6,537

7,446 2,655 4,486 7,131 5,744 4,283 5,065 6,723 7,342 4,716 4,912 10,075

4,415 1,932 1,955 2,975 3,184 1,982 2,713 2,861 6,522 2,649 2,081 5,136

6,947 2,774 2,996 4,620 4,853 3,096 4,118 4,440 8,446 3,954 3,282 5,136

10,795 4,075 4,769 7,612 7,467 4,899 6,414 7,087 11,453 6,137 5,396 7,964

16,494 5,853 7,719 12,624 11,413 7,710 10,055 11,248 16,186 9,331 8,810 12,015

21,672 7,452 10,623 17,735 15,027 10,538 13,387 15,287 21,012 12,456 12,159 15,538

Table 3.2 The distribution of RCC-adjusted charges by DRG in 1983 and 1993: ages 55-64 1983 unadjusted distribution Unadjusted percentiles

Acute myocardial infarction Angina Arrythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders

N

Mean

Standard deviation

25th

50th

75th

90th

95th

10,340 8,505 5,760 5,012 6,538 10,634 3,525 5,756 1,635 2,062 2,370

8,203 3,441 3,816 9,844 7,664 3,144 4,693 6,519 12,537 4,530 4,687

7,668 2,604 4,453 13,760 12,205 3,443 6,063 8,544 12,728 4,292 7,201

4,036 1,978 1,696 2,891 2,837 1,490 2,028 2,735 6,382 2,066 1,705

6,538 2,860 2,702 5,405 4,677 2,318 3,273 4,393 9,038 3,238 2,870

9,883 4,131 4,368 10,975 7,924 3,681 5,374 7,385 13,967 5,552 4,967

14,577 5,941 6,890 22,116 14,590 5,792 8,854 12,467 22,227 8,853 8,753

19,435 7,583 9,984 33,398 22,530 7,763 12,482 17,664 32,370 11,804 13,524

Table 3.2 (continued) 1983 adjusted distribution Adjusted percentiles

Acute myocardial infarction Angina Arrythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

Standard deviation

25th

50th

75th

90th

95th

6,004 7,066 3,988 4,921 6,269 5,360 3,525 5,756 1,074 2,062 2,370 5,402

11,543 3,883 4,939 10,015 7,951 4,793 4,693 6,519 16,304 4,530 4,687 7,230

8,568 2,644 4,947 13,829 12,383 4,218 6,063 8,544 14,295 4,292 7,201 10,658

7,207 2,410 2,615 3,015 3,058 2,878 2,028 2,735 9,155 2,066 1,705 2,714

9,143 3,205 3,571 11,207 4,873 3,668 3,273 4,393 12,057 3,238 2,870 4,508

12,654 4,469 5,315 22,423 8,130 5,191 5,374 7,385 17,133 5,552 4,967 7,864

18,343 6,385 8,214 33,618 14,873 7,741 8,854 12,467 22,122 8,853 8,753 13,792

24,625 8,030 12,085 69,655 23,048 9,925 12,482 17,664 40,432 11,804 13,524 21,123

Table 3.2 (continued) 1993 unadjusted distribution Unadjusted percentiles

Acute myocardial infarction Angina Arrythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

Standard deviation

25th

50th

75th

90th

95th

6,004 7,066 3,988 4,921 6,269 5,360 4,425 8,152 1,074 2,370 3,626 5,859

8,702 3,336 3,916 7,505 6,191 3,916 5,401 6,564 12,010 5,188 4,660 6,397

7,378 2,265 4,372 8,618 5,621 4,179 5,581 7,810 11,015 5,202 5,475 5,427

4,529 1,874 1,782 3,113 3,193 1,816 2,642 2,953 6,307 2,503 1,874 3,220

7,116 2,700 2,762 4,915 4,769 2,854 3,962 4,615 8,925 3,916 3,178 4,948

10,600 3,980 4,475 8,508 7,318 4,570 6,160 7,317 13,807 6,002 5,484 7,616

15,326 5,819 7,316 11,645 11,359 7,260 9,662 12,384 21,119 9,613 9,348 12,003

19,809 7,422 10,281 21,951 15,193 9,866 14,108 17,572 28,568 12,632 13,165 15,875

Table 3.2 (continued) 1993 adjusted distribution Adjusted percentiles

Acute myocardial infarction Angina Arrythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

Standard deviation

25th

50th

75th

90th

95th

6,004 7,066 3,988 4,921 6,269 5,360 4,425 8,152 1,074 2,370 3,626 5,859

8,702 3,336 3,916 7,505 6,191 3,916 5,401 6,564 12,010 5,188 4,660 6,397

7,378 2,265 4,372 8,618 5,621 4,179 5,581 7,810 11,015 5,202 5,475 5,427

4,529 1,874 1,782 3,113 3,193 1,816 2,642 2,953 6,307 2,503 1,874 3,220

7,116 2,700 2,762 4,915 4,769 2,854 3,962 4,615 8,925 3,916 3,178 4,948

10,600 3,980 4,475 8,508 7,318 4,570 6,160 7,317 13,807 6,002 5,484 7,616

15,326 5,819 7,316 11,645 11,359 7,260 9,662 12,384 21,119 9,613 9,348 12,003

19,809 7,422 10,281 21,951 15,193 9,866 14,108 17,572 28,568 12,632 13,165 15,875

Table 3.3 Ten-year annualized growth in RCC-adjusted charges at selected percentiles of the spending distribution: age 65+ 1983-1993 growth within unadjusted distribution Unadjusted percentiles

Acute myocardial infarction Angina Arrhythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

25th

50th

75th

90th

95th

-0.70 2.10 1.00 1.80 2.70 -4.70 7.50 5.90 1.20 4.80 6.00 3.80

-0.20 -1.50 -0.80 -3.00 -2.00 1.50 0.50 -1.10 -1.50 -0.90 -0.70 -2.20

0.30 -1.20 -0.20 0.00 0.90 2.10 1.90 0.30 -0.80 0.90 0.70 0.60

0.00 -1.40 -0.70 -1.60 -0.10 2.00 1.50 -0.30 -1.10 -0.10 0.30 -1.60

0.00 -1.40 -1.00 -3.00 -1.20 1.80 1.00 -0.80 -1.50 -0.80 -0.30 -1.60

-0.20 -1.60 -1.00 -4.10 -2.60 1.20 0.30 -1.50 -2.10 -1.60 -1.20 -2.90

-0.50 -1.60 -1.00 ^.40 -3.80 1.20 -0.20 -2.00 -2.20 -1.80 -1.70 -3.80

Table 3.3 (continued) 1983-1993 growth within adjusted distribution Adjusted percentiles

Acute myocardial infarction Angina Arrhythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

25th

50th

75th

90th

95th

0.00 2.10 1.00 1.80 2.70 0.00 7.50 5.90 1.20 4.80 6.00 3.80

-0.90 -1.50 -0.80 -3.00 -2.00 -1.70 0.50 -1.10 -1.50 -0.90 -0.70 -2.20

-0.90 -1.20 -0.20 0.00 0.90 -3.10 1.90 0.30 -0.80 0.90 0.70 0.60

-0.60 -1.40 -0.70 -1.60 -0.10 -1.50 1.50 0.30 -1.10 -0.10 0.30 -0.30

-0.40 -1.40 -1.00 -3.00 -1.20 -1.00 1.00 -0.80 -1.50 -0.80 -0.30 -1.60

-0.50 -1.60 -1.00 -4.10 -2.60 -1.00 0.30 -1.50 -2.10 -1.60 -1.20 -2.90

-0.90 -1.60 -1.00 -4.40 -3.80 -1.00 -0.20 -2.00 -2.20 -1.80 -1.70 -3.80

Table 3.4 Ten-year annualized growth in RCC-adjusted charges at selected percentiles of the spending distribution: ages 55-64 1983-1993 growth within unadjusted distribution Unadjusted percentiles

Acute myocardial infarction Angina Arrhythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

-5.30 -1.80 -3.60 -0.20 -0.40 -6.60 2.30 3.50 -4.10 1.40 4.30 0.80

0.60 -0.30 0.30 -2.70 -2.10 2.20 1.40 0.10 -0.40 1.40 -0.10 -1.20

25th

1.20 -0.50 0.50 0.70 1.20 2.00 2.70 0.80 -0.10 1.90 0.90 1.70

50th

0.90 -0.60 0.20 -0.90 0.20 2.10 1.90 0.50 -0.20 1.90 1.00 0.90

75th

90th

95th

0.70 -0.40 0.20 -2.50 -0.80 2.20 1.40 -0.10 -0.10 0.80 1.00 -0.30

0.50 -0.20 0.60 -4.00 -2.50 2.30 0.90 -0.10 -0.50 0.80 0.70 -1.40

0.20 -0.20 0.30 -4.10 -3.90 2.40 1.20 -0.10 -1.20 0.70 -0.30 -2.80

Table 3.4 (continued) 1983-1993 growth within adjusted distribution Adjusted percentiles

Acute myocardial infarction Angina Arrhythmia Cerebrovascular accident Chronic obstructive pulmonary disease Esophageal and gastrointestinal miscellany Gastrointestinal hemorrhage Heart failure and shock Hip and femur procedures Kidney and urinary tract infection Nutritional and metabolic disorders Pneumonia

N

Mean

25th

50th

75th

90th

95th

0.00 0.00 0.00 0.00 0.00 0.00 2.30 3.50 0.00 1.40 4.30 0.80

-2.80 -1.50 -2.30 -2.80 -2.50 -2.00 1.40 0.10 -3.00 1.40 -0.10 -1.20

-4.50 -2.50 -3.80 0.30 0.40 -4.50 2.70 0.80 -3.70 1.90 0.90 1.70

-2.50 -1.70 -2.50 -7.90 -0.20 -2.50 1.90 0.50 -3.00 1.90 1.00 0.90

-1.80 -1.20 -1.70 -9.20 -1.00 -1.30 1.40 -0.10 -2.10 0.80 1.00 -0.30

-1.80 -0.90 -1.20 -8.00 -2.70 -0.60 0.90 -0.10 -2.80 0.80 0.70 -1.40

-2.20 -0.80 -1.60 -10.90 -1.10 -0.10 1.20 -0.10 -3.40 0.70 -0.30 -2.80

Table 3.5 Quantile regression parameter estimates: acute myocardial infarction, age 65+

1983

25th percentile Median

75th 90th 95th percentile percentile percentile

Patient level variables Payer (omit: private and HMO) 310 449a 827b Medicare 207 547a l,239a Medi-Cal -165 -197 509 Other nonprivate Market level variables 494C 867C 1,733C Log physicians per capita -21 42 61 Log HMO enrollment ratio -78 -273C -323C Log population 1,232C 1,568C 1,372C Log income per capita Level of competition [omit low]d 327a 612C 1,123C Moderate 418a 899C 1,034C Competitive 1,243C 2,327C 3,775C Very competitive Hospital level variables 374c -468C -566C Investor-owned (omit: NFP and other) 0 2C 2C Number of licensed beds 0 0 0 0 Total number of discharges (1983) Teaching hospital (omit: nontech) 607C 839C 1,042C -12,370 -13,067 -6,102 Constant c

1993 Patient level variables Payer (omit: private and HMO) Medicare Medi-Cal Other nonprivate Market level variables Log physicians per capita Log HMO enrollment ratio Log population Log income per capita Level of competition (omit low)d Moderate Competitive Very competitive Hospital level variables Investor-owned (omit: NFP and other) Number of licensed beds Total number of discharges (1983) Teaching hospital (omit: nontech) Constant a

c

c

2,237a 2/954a 804

1,350 2,683 1,634

3,064C 46 -523b 893

3,444a -278 -672 621

2,219a 2,807c 7,232C

2,879a 4,226a 11,730C

-690 4a

-797 llc

1,809C

2,498C

b

25th percentile Median

Qb

6,976

17,089

75th 90th 95th percentile percentile percentile

712C 1,125C -47

1,017C 1,857C 44

1,517C 2,726C 585

2,384C 4,778C 1,854

-358 15 356C 3,180C

-575a 32 628C 4,781C

-931a 67 1,037c 6,969C

-954

-642C -822C -1,454C

-1,048C -1,382C -2,423C

-1,623C -2,287c -3,544C

135 -lc 0c -28

260a -2C 0c 255a

693a -3C 0c 666C

2,977C 9,405C 970

0 1,737C 7,788C

-1,947 -14 2,892C 11,828c

-1,903C -3,729C -4,745C

-4,222C -6,394C -8,338C

1,718C -4C 0c 1,222C

2,998C -8a 0c l,244b

-34,452C -52,872C -79,199C -92,734C -149329c

p < 0.10 p < 0.05 c p < 0.01 d Low [1.00 > Herf > 0.20], moderate [0.20 > Herf > 0.10], competitive [0.10 > Herf > 0.05], very competitive [0.05 > Herf] b

Table 3.6 Quantile regression parameter estimates: acute myocardial infarction, ages 55-64

1983 Patient level variables Payer (omit: private and HMO) Medicare Medi-Cal Other nonprivate Market level variables Log physicians per capita Log HMO enrollment ratio Log population Log income per capita Level of competition [omit low]d Moderate Competitive Very competitive Hospital level variables Investor-owned (omit: NFP and other) Number of licensed beds Total number of discharges (1983) Teaching hospital (omit: nontech) Constant

1993 Patient level variables Payer (omit: private and HMO) Medicare Medi-Cal Other nonprivate Market level variables Log physicians per capita Log HMO enrollment ratio Log population Log income per capita Level of competition (omit low)d Moderate Competitive Very competitive Hospital level variables Investor-owned (omit: NFP and other) Number of licensed beds Total number of discharges (1983) Teaching hospital (omit: nontech) Constant a

25th percentile Median

126 114 -106

75th 90th 95th percentile percentile percentile

106 509C -186

438b 864C -65

1,820C 2,956c 326

2,665a 6,865c 990

232 16 -2 936C

290 -54 -74 1,804C .

642b -48b -119 1,922C

553 42 -26 2,158b

1,581 -172 -64 1,493

269 623a 839a

441a 987C 1,955C

1,240C 1,461C 3,268C

l,625b 1,264 4,275C

2,610b 2,469 7,826C

-226b -429C -588C -658 0 1 3C 3 0c 0c 0a 0 263a 646C 1,519C 2,708C -9,022a -18,760C -17,396 -18,392

-l,782b 2 0b 3,837C -3,923

25th percentile Median

75th 90th 95th percentile percentile percentile

331b 443a 242

583a 886C 527

946C 1,728C 749

3,222C 3,321C 924

4,274C 6,131C l,457b

-1,038C -87 453C 4,069C

-1,289C -167b 325b 5,803C

-919 -271b 508a 5,637C

-1,424 -177 l,141a 7,231a

-3,497b -651 l,173a 11,508C

-914C -1,049C -1,811C

-903 -989a -l,458a

- 326 - 271 -l,606a -2,724a -l,827a -3,643a

1,001 1,894 -1,814

391C 817C 1,577C 3,489c 4,243c - 1 0 0 0 4 0c 0c 0c 0c 0a a a C 148 385 547 1,295 2,548c c c c a -48,210 -63,669 -59,541 -83,656 -137,415a

p < 0.10 p < 0.05 c p < 0.01 d Low [1.00 > Herf > 0.20], moderate [0.20 > Herf > 0.10], competitive [0.10 > Herf > 0.05], very competitive [0.05 > Herf]

b

Table 3.7 Summary of quantile regression parameter estimates: effect of competition on RCC-adjusted costs within 12 largest DRGs in 1983 and 1993, by age group Age 55-64

Age 65+

Year 1. Acute myocardial infarction

1983

1993

2. Angina

1983

1993

3. Arrhythmia

1983

1993

a

p < .05 p<10 c p < .01 b

Median

75th percentile

90th percentile

95th percentile

441a 987 1,955 -903a -989a -l,458a

1,240 1,461 3,268 -326 -l,606a -l,827a

l,625b 1,265 4,275 -271 -2,724a -3,643a

2,610b 2,469 7,826C 1,001 -1,894 -1,814

151b 139 448C -87 -156b -208b

257 214b 721 -351 -356a -507a

555C 389a 1,1 78C -668 -542a -744a

623a -86 l,313a -977 -626 -78

701 2,627 1,186 -1,322C -l,448a -336

189b 443 775 -378 -442 -841

337a 541 956 -581 -693 -1,275

l,193b 1,048 2,887a -2,505 -3,020 -5,125

l,936b 3,039a 3,661b -1,495 -1,994 -3,369

Median

75th percentile

90th percentile

95th percentile

25th percentile

327a 418a 1,243C -642C -822C -1,454C

612C 899 2,327 -1,048 -1,382 -2,423

1,123C 1,045 3,775 -1,623 -2,287 -3,544

2,219 2,807 7,232 -1,903 -3,729 -4,745

2,879a 4,227a 11,730 -4,222 -6,394 -8,338

269 623 839 -914 -1,049 -1,812

Moderate Competitive Highly competitive Moderate Competitive Highly competitive

243C 252C 656C -118C -257C -433C

395 440 1,147 -338 -528 -817

529 614 1,665 -567 -854 -1,240

960 616 2,436 -951 -1,474 -1,939

l,202a 563 3,242 -1,271 -1,876 -2,255

Moderate Competitive Highly competitive Moderate Competitive Highly competitive

257C 698C 1,197C -453C -422C -747C

845 1,426 2,426 -729 -534 -865

1,427 2,386 3,984 -1,265 -1,385 -1,742

3,048 5,688 8,713 -1,187 -2,598 -3,120

2,085 5,527 11,652 -3,689 -4,872 -6,906

Level of competition (omit low)

25th percentile

Moderate Competitive Highly competitive Moderate Competitive Highly competitive

673 905 1,678 -1,322 -1,388 -2,428

Table 3.7 (continued) Age 55-64

Age 65+

Year 4. Cerebrovascular accident

6. Esophageal and gastrointestinal miscellaneous disorders

a

p < .05 p<10 c p < .01

b

25th percentile

Median

75th percentile

90th percentile

95th percentile

25th percentile

Median

75th percentile

90th percentile

95th percentile

Moderate Competitive Highly competitive Moderate Competitive Highly competitive

c

263 741C 1,100C -477c -456C -829C

602 1,341 1,874 -529 -468 -767

842 1,945 3,059 -769 -890 -1,222

l,130 4,635 6,205 -1,269 -l,241a -l,748a

1,714 6,635 10,159 -l,379b -1,596 -2,645b

318 935 1,263 -355b -476a -634b

625 1,395 l,688a -652a -654a -683

1,011 3,639 3,471a -l,403b -1,294 -1,770

8,104 16,727 18,941 -1,442 -952 -1,272

10,304a 27,287c 32,865C -1,958 240 -1,594

1983

Moderate Competitive Highly competitive

97 557C 1,084C

106 803 1,577

23 1,206 2,416

866 3,223 5,429

1,154 4,763a 9,139

80 389b 634a

-332 338 793

-808 l,091b 1,494

-1,329 2,956 2,824

-2,734 2,034 4,228

1993

Moderate Competitive Highly competitive

-160a -387C -330a

-159 -509 -473a

-497a -1,097 -970a

-578 -l,525a -l,541a

-510 -2,001a -2,005b

-362a -545b -423

-662C -568a -414

-930a -798b 254

-l,337b -l,720b 237

-2,455b -2,986b -504

1983

Moderate Competitive Highly competitive Moderate Competitive Highly competitive

213C 287C 519C -16 -67 -226a

333 428 824 -302 -329 -585

428 569 1,257 -555 -671 -1,160

598a 819a 1,660 -823a -964b -l,165b

919b l,534a 3,169 -l,644a -l,706a -1,492

132a 233 447 -181b -22 -298b

196a 336 579 -323a -185 -661a

415C 385a 616a -536 -346 -908

583b 528 815 -680 -130 -472

1983

1993

5. Chronic obstructive pulmonary disorder

Level of competition (omit low)

1993

a

b

654 1,223 1,022 -232 -1,473 -1,973

Table 3.7 (continued) Age 55-64

Age 65+

Year 7. Gastrointestinal hemorrhage

9. Hip and femur procedures

a

p < .05 p<10 c p<.01

Median b

75th percentile

90th percentile

95th percentile

25th percentile

b

Median

75th percentile

90th percentile

95th percentile

146 128 516c -225 -361 -705

221 226 933 -320 -495 -917

344 236 1,705 -653 -935 -1,737

l,297 800 2,730a -l,063a -1,712 -2,375

847 -96 1,647 -835 -l,950a -2,088

165 78 196 -62 35 -168

336b 358 957a -235 -190 -787a

543 378 1,338 -532 -490 -851

l,635a 1,265 4,275 -271 -2,724a -3,594a

-1,290 -432 1,503 -2,480 -3,590 -6,150b

1983

Moderate Competitive Highly competitive

225 418 1,054

318 544 1,593

546a 774 2,648

774b l,089b 4,542

l,618a 1,510 7,687C

89 392a 951

344b 770 1,704

69 926 2,197a

2,528a 4,587 8,463

6,176C 10,280C 17,259C

1993

Moderate Competitive Highly competitive

-333 -583 -857

-414 -782 -1,002

-717 -1,351 -1,655

-1,310 -2,336 -2,684

-2,654C -4,777C -5,081C

-267a -664 -793

-554a -1,128 -1,464

-649b -1,384 -l,154b

-929 -3,261a -1,970

-1,283 -7,014C -4,748

1983

Moderate Competitive Highly competitive Moderate Competitive Highly competitive

952 1,669 3,024 35 -242 -962

1,494 2,621 4,696 51 -217 -910

2,274 4,039 6,804 267 72 -131

3,263 4,736 9,188 -456 -l,874a -1,902

5,236a 8,736c 14,942 -1,568 -3,217 -2,940

424 166 1,111 l,059a 958 445

228 -2 1,730 236 347 -1,255

1,766 3,344b 7,425 1,397 1,888 2,281

-334 4,002 8,672 6,976b 5,160 6,872

276 3,631 11,293 2,633 -731 349

1993

b

25th percentile

Moderate Competitive Highly competitive Moderate Competitive Highly competitive

1983

1993

8. Heart failure and shock

Level of competition (omit low)

Table 3.7 (continued) Age 55-64

Age 65+

Year 10. Kidney and urinary tract infection

1983

1993

11. Nutritional and metabolic disorders

1983

1993

12. Pneumonia

1983

1993

a

p < .05 p <10 c p <.01

b

Level of competition (omit low) Moderate Competitive Highly competitive Moderate Competitive Highly competitive Moderate Competitive Highly competitive Moderate Competitive Highly competitive Moderate Competitive Highly competitive Moderate Competitive Highly competitive

25th percentile

Median

75th percentile

a

b

386 471

321

1,119 -267 -456 -524

1,551

31 104 329a 42 -174a -236a 186a

504 1,075 -153a -104 -230b

594 -573 -840 -884

504 954a 2,413 -786 -1,186 -1,269

185 398a 904 -40 -313 -491

360 797a 1,651 -237 -783 -826a

324a 844 1,884 -273 -250 -501

582b 1,476 3,553 -174 -422b -549b

90th percentile

95th percentile

25th percentile

Median

-174

-1,330 -61 624 -2,077a -2,270a -1,250

-53 -91 118

119 283 839

-197

-121

895 1,242 -1,793 -2,321 -2,103a

496 1,513 3,095a -722b -1,667 -l,713a

934 2,398a 6,314

1 -349 -450

785 3,407a 5,576a -1,031 -2,356a -2,984a 2,339b 3,607a 8,823C 272 -890 -156

-39 340

572 905b

5 144 428b 155

-63 225 714

-256 -143

225 445b 1,141

357 -491 -682

75th percentile

23 l,271b 2,010b -346 -49 501 -197

194

-331 -977 -940a

1,624 -649 -l,156b -296

576a 799a 2,220 -258 -309 -78

997b l,541a 3,596 -504 -221 209

90th percentile

819 1,875 3,520a

-1,256 1,569 1,253

241 938 3,620

95th percentile

580 2,418 2,040 -2,053 -2,551 -5,661

-303 -890 1,494

-506 -1,267 3,386 -705 -1,862 1,010

1,671 1,711 3,731 -704 -1,183 1,025

1,988 3,043 3,318 -1,339 -1,050 1,553

Effects of Competition Under Prospective Payment on Hospital Costs

93

regressions for all twelve DRGs for 1983 and 1993, and for both age groups. Beginning with the left panel of table 3.7, we find that in 1983, in every case, costs rise with increasing competition, and particularly for the most expensive admissions. This is consistent with the "medical arms race" literature, which suggests that under the retrospective reimbursement system in place at the beginning of the period we study, a more competitive hospital market will raise costs as hospitals compete to attract doctors and their patients by offering added services (Robinson and Luft, 1985). In contrast, in 1993, increasing competitiveness is associated with decreased costs in all twelve DRGs. This is consistent with previous findings such as those of Melnick and Zwanziger (1988), who found that costs fell by more than 11 percent for hospitals in the most competitive markets in California during this period while actually rising in the least competitive markets. Not addressed in their findings, however, is the strong pattern we observe for the reductions in expenditures with increasing competition to increase progressively along the spending distribution, as predicted by the incentive structure of PPS to induce selective reduction of expenditures among the most expensive patients. Among the elderly, for example, AMI admissions at the 95th percentile are associated with an estimated $8,338 reduction in hospital costs, compared to a reduction of $1,454 in costs at the 25th percentile. The columns of table 3.7 labeled age 55-64 repeats the above analyses for the younger age group. In general, we still find strong positive effects of competition on costs in 1983, but in 1993, we find much smaller negative effects of competition on costs that are statistically significant for only four of the twelve DRGs. Several DRGs also show a statistically insignificant trend toward lower costs with competition. This raises the question of whether some of the difference is due to reduced sample size. However, increasing the sample analyzed to include all persons below age 65, or adding additional years of data (for example, 1992), did not meaningfully alter these results. This suggests that whatever forces led to changes in the distribution of hospital expenditures in these diagnoses among the elderly between 1983 and 1993 may have also affected patients younger than age 65, although the effects do not appear to have been as powerful. Adjustments for Changes in Discharge Rates To address the concern that percentiles in one year may not be comparable to percentiles in another year due to changes in severity of ill-

94

Meltzer and Chung

ness, especially due to declines in admission rates as described previously, we also examined quantile regressions for DRGs in which admission rates fell from 1983 to 1993, and thus limited the number of observations in 1983 to the number in 1993, to compare "comparable patients" assuming no change in the underlying distribution of disease. These results were not substantively different from the regressions reported in tables 3.5 through 3.7. VI. Summary and Conclusion Using annual patient discharge data from all nonfederal, acute-care hospitals in the state of California from 1983 and 1993, we examined growth in hospital costs and the effects of competition on costs at various points in the spending distribution for persons above and below age 65 in the twelve largest DRGs. Our analyses of cost growth show cost growth falling with increasing position in the spending distribution in every DRG we studied, as predicted by the effects of Medicare PPS. However, a very similar pattern is also evident among admissions of patients age 55-64. Our analyses of the effects of competition show a strong trend for increasing competition to increase expenditures in all age groups in 1983, with increasing effects at higher locations in the spending distribution. Our analyses for those older than age 65 in 1993 show the opposite pattern, however, with increasing competition associated with decreased costs, and the effects far larger among the most expensive patients. This pattern is not as pronounced among those younger than age 65, suggesting that spending on persons older than age 65 during this period may have been subject to some forces different than those affecting spending on persons younger than age 65. These findings are broadly consistent with the model of provider behavior under alternative reimbursement schemes that we present. This predicts a tendency for hospitals to skimp on unprofitable patients and to milk profitable patients under fixed-rate prospective reimbursement. Although several studies have documented lower resource utilization associated with fixed-rate reimbursement systems, fewer have considered the possibility that such reductions might differentially affect profitable and unprofitable classes of patients, and none have demonstrated these patterns of increasing cost reductions among high-cost patients for Medicare PPS or shown that these reductions rise with increasing competition. Certainly, the establishment of PPS and its associated incentives to decrease costs among the most costly patients is a plausible explana-

Effects of Competition Under Prospective Payment on Hospital Costs

95

tion for the patterns we observe among the elderly, but several other possible explanations are worth considering. One explanation is that there were changes in particular medical technologies or in the underlying severity of illness among the elderly over this period, and that these changes somehow selectively reduced expenditures for the high-cost elderly relative to the low-cost elderly. The fact that we see a fairly similar pattern of growth among the young is somewhat suggestive of this alternative account. It is not clear why this is the case, but it is highly plausible that practice patterns are likely to be similar for older and younger patients, so that incentives implicit in PPS end up affecting practice patterns for patients below age 65 as well. If the changes in spending we observe are explained by some specific change in underlying severity of illness or medical technology, it is not clear why such changes should occur over such a broad range of diagnoses or why they should be associated with increased competition. Also, in additional analyses, we stratified the elderly according to age and controlled for measurable aspects of underlying comorbidity and found no changes in our results. Another possibility is that our results reflect changes in coding practices under Medicare, often referred to as "DRG creep." We have tried to address this concern in our analysis by combining related DRGs with and without complications, but it is possible that this does not capture all the changes that could have occurred. Indeed, one particular concern is that the development of the tracheostomy DRG in 1991 for patients requiring mechanical ventilation may have drawn some expensive patients out of the upper part of the distribution of costs from some of our DRGs. While this is important to consider, the fraction of all admissions coded into the tracheostomy DRG is only 0.1 percent to 0.2 percent. This seems to explain the broad changes we see across the spending distribution for such a broad range of diagnoses. In addition, some of the pattern we identify is clearly present by 1991, when the tracheostomy DRG was just being introduced. It is somewhat surprising that such receding would be present only in the most competitive markets, though that would also certainly be of interest if it were the case. It should also be noted that, to the extent that some of our diagnoses may be more highly reimbursed than other closely related diagnoses, they may also be the recipients of upcoding, in which case one would expect expenditures at the lower end of the distribution to decline as healthier patients are added to the distribution. Finally, it should be noted that our finding that cost reductions are greatest for the most expensive patients can also be interpreted as sim-

96

Meltzer and Chung

ply reflecting the idea that it is easier to save large amounts of money where more money is spent. We are sympathetic to this concern, but note that we find a similar pattern of reductions in DRGs that are more expensive as well as in DRGs that are less expensive. It is not always the case that it is easier to decrease spending where more money is spent. To illustrate, analysis of data from a natural experiment comparing the cost of hospital care provided by doctors who specialize in inpatient care to hospital care by doctors who spend only a small fraction of their time taking care of inpatients reveals no evidence that cost savings differed across the distribution of costs (Meltzer et al. 2000). With an understanding of the limitations of our analysis, it appears that increasing competition in the context of prospective payment is associated with selective reductions of expenditures for the most expensive patients. Whether this is desirable is impossible to determine without an analysis of the effects on outcomes. Nevertheless, our results suggest several clear lines for such analysis. First, the possibility that costs are selectively reduced for the most costly patients suggests that outcomes may be selectively affected. While more than a few studies have examined the effects of prospective payment on outcomes (for example, Rodgers et al. 1990 and accompanying articles, Cutler 1995), none has stratified outcomes according to patient cost. Our results suggest that such analyses might be very useful because it is possible that adverse effects among the most costly patients might be masked by their inclusion with less costly patients, whose outcomes may even improve if increased resources allocated to attracting them to a particular hospital have some positive (albeit small) effect on outcomes. The same conclusion applies for attempts to measure the effects of competition on outcomes (see, for example, Kessler and McClellan 1999). Our results also have important implications for measuring the quality of care under prospective payment systems, and especially in competitive environments, because they suggest that high-cost patients may be at particular risk in such contexts. Thus, it is important that quality measures reflect the concerns of that potentially vulnerable group. Even when a single measure of quality is used, our findings may have implications for how to measure quality of care. For example, our findings may provide a justification to prefer outcomes measures to process measures because process measures can suggest that quality is high over the whole population when the quality of care for certain parts of the population are actually poor. On the other hand,

Effects of Competition Under Prospective Payment on Hospital Costs

97

outcomes are often favorable for less severely ill (less costly) patients in any case, so expending greater resources on them is unlikely to improve outcomes. It is worth noting that this basic conclusion remains, regardless of whether one believes that cost reductions were largest among the most costly patients due to selective incentives within a prospective payment, or whether one simply believes that cost reductions are largest for the most costly patients simply because, as the old adage goes, "that's where the money is." A related issue is whether the effects of competition on costs should be interpreted as reducing quality or, rather, improving efficiency. Resolution of this question will be possible only with data that permits a comprehensive assessment of outcomes. It should be noted that the combination of prospective payment and competition studied here is not unique to Medicare PPS, but in fact, is the basic idea behind the increasing use of capitated managed-care arrangements and competition to control costs, including Medicare managed care. Indeed, such managed competition arrangements present similar incentives to expend resources to attract less costly participants while avoiding more costly ones. It is not difficult to imagine that these incentives result in substantial investments in wellness programs and preventive services, amenities that improve access for working persons, reductions in copayments, etc., that would attract relatively healthy participants. Even casual observation of the offerings of health maintenance organizations leaves little question that many of these benefits are indeed occurring, but whether such expenditures are an efficient use of health care resources and how they may affect the care received by the most severely ill are important questions for future work. This is especially true given evidence that quality of care in HMOs may be the worst for patients who are chronically ill (Miller and Luft 1997), and that HMOs may limit expenditures for severely ill persons in intensive care (Rapoport et al. 1992, Cher and Lenert 1997). Finally, it should be noted that various approaches can be tried to improve upon existing prospective payment systems, for example, risk adjustment and the use of blended payment schemes that include both prospective and retrospective components designed to mitigate incentives for patient selection or discrimination in care provision. Indeed, the Medicare PPS has always tied reimbursement to the amount of care provided to some extent, and thus never truly became fully prospective (McClellan 1997). Proposals have been seriously considered to expand this retrospective aspect of Medicare PPS as well as to improve

98

Meltzer and Chung

risk adjusters by developing a DRG system that allows a finer classification of admissions (Newhouse, Buntin, and Chapman 1997). Our work provides support for the value of further examination of both approaches. Notes 1. For standard textbook discussion, see Peter Zweifel and Friedrich Breyer (1997), Health Economics, NY: Oxford Press, and Charles Phelps (1997), Health Economics Second Ed., Reading, MA: Addison Wesley. 2. By annual volume of discharges within DRGs. 3. For a good summary of the cost-containment effort in California during the 1980s, see Langa 1992. 4. A detailed overview on forms of managed care can be found in Gold et al. 1995. 5. In practice, this may be implemented by increasing spending on infrastructural elements that may be most important to health patients (such as a pleasant cafeteria or waiting area), and decreasing spending on infrastructure that is most important to the sickest patients (such as expensive imaging machines or intensity of ICU care). It might also be implemented by reducing pressure on physicians to discharge relatively healthy patients quickly. 6. The classic example of this is the free car seat sometimes offered to expectant parents to induce them to deliver their child at a particular hospital. 7. Note, however, that the relationship between competition and quality in general may be much more complex than this in settings where both price and quantity may be varied because it will also depend on the complementarity between quantity and quality (Spence 1975, Saving 1982). Another alternative view is reflected in Satterthwaite (1979), in which an increasing number of sellers in a market effectively raises search costs by decreasing the value of information held by any individual about a particular seller. 8. Discharges from managed care facilities exempted from standard accounting requirements were identified in the data by a zero in the field for total charges, although actual charges were nonzero. In total, this involves omitting 8.8 percent of discharges. 9. A large amount of literature defines hospital markets for the purpose of measuring competition. Traditional measures have included market definitions based on geopolitical boundaries such as counties or metropolitan statistical areas (for example, Joskow 1980), distance (Robinson and Luft 1985), or patient flows (for example, Melnick and Zwanziger 1988). These measures have all been criticized for varying reasons, including the (ir)relevance of geopolitical boundaries or distance with respect to competition, and the endogeneity of patient flows. While some newer approaches (for example, Kessler and McClellan 1999) have tried to address these concerns, such approaches are substantially more difficult to implement, and their merits have not yet been demonstrated. While a comparison of multiple measures of competition would be of value, we defer it for future work. 10. See also Stigler (1968) and Cowling and Waterson (1976) for a theoretical rationale for the use of the HHI.

Effects of Competition Under Prospective Payment on Hospital Costs

99

11. Another possibility would be to analyze the effects of competition on cost growth within ICD-9 codes (International Classification of Disease, 9th ed.). However, we elected not to do this because the incentives created by Medicare PPS that may differentially affect high- and low-cost patients refer to high- and low-cost patients within DRG groups rather than within ICD-9 codes. 12. Specifically, we calculated the Charlson comorbidity index based on secondary diagnoses for the index admission, and found that the distribution of scores shifted upward in all our DRGs (Charlson et al. 1987, Deyo and Romano 1993, Romano et al. 1993).

References Allen, R., and P. Gertler (1991). "Regulation and the Provision of Quality to Heterogeneous Consumers: The Case of Prospective Pricing of Medical Services," Journal of Regulatory Economics, 3:361-375. Bain, J. (1951). "Relationship of Profit Rate to Industry Concentration: American Manufacturing, 1936-1940," Quarterly Journal of Economics, 65:293-324. Carter, G., J. Newhouse, and D. Relies (1990). "How Much Change in the Case Mix Index Is DRG Creep?" Journal of Health Economics, 9:411-128. Charlson, M., P. Pompei, K. Ales, and C. McKenzie (1987). "A New Method of Classifying Prognostic Comorbidity in Longitudinal Studies: Development and Validation," Journal of Chronic Disease, 40:373-383. Cher, D., and L. Lenert (1997). "Method of Medicare Reimbursement and the Rate of Potentially Ineffective Care of Critically I11 Patients," JAMA, 278(12):1001-1007. Cowling, K., and M. Waterson (1976). "Price-Cost Margins and Market Structure," Econometrica, 43(171):267-274. Cutler, D. (1995). "The Incidence of Adverse Medical Outcomes Under Prospective Payment," Economica, 63(1):29-50. Cutler and Meara (1998). "The Medical Costs of the Young and Old: A Forty-Year Perspective," in Frontiers in the Economics of Aging, David A. Wise, ed. (Chicago, IL: University of Chicago Press, pp. 215-242). Davis, K., et al. (1990). Health Care Cost Containment, Baltimore, MD: Johns Hopkins University Press. Davis, M., and S. Burner. (1995). "Three Decades of Medicare: What the Numbers Tell Us," Health Affairs, 14(4):231-243. Deyo, R., D. Cherkin, and M. Ciol. (1992). "Adapting a Clinical Comorbidity Index for Use with ICD-9-CM Administrative Databases." Journal of Clinical Epidemiology. 45(6):613-619. Dranove, D. (1987). "Rate-Setting by Diagnosis-Related Groups and Hospital Specialization," RAND Journal of Economics, 18(3):417-427. Dranove, D., M. Shanley, and C. Simon (1992). "Is Hospital Competition Wasteful?" RAND Journal of Economics, 23(2):247-262. Dranove, D., M. Shanley, and W. White. (1991). "How Fast are Hospital Prices Really Rising?" Medical Care. 29(8):690-696.

100

Meltzer and Chung

Dranove, D., and W. White (1994). "Recent Theory and Evidence on Competition in Hospital Markets," Journal of Economics and Management Strategy, 3(1):169-209. Dranove, D., and W. White (1998). "Medicaid-Dependent Hospitals and Their Patients: How Have They Fared?" Health Services Research, 32(2):163-185. Ellis, R. (1998). "Creaming, Skimping, and Dumping: Provider Competition on the Intensive and Extensive Margins," Journal of Health Economics, 17:537-555. Ellis, R., and T. McGuire (1986). "Provider Behavior Under Prospective Reimbursement," Journal of Health Economics, 5:129-151. Ellis, R., and T. McGuire (1996). "Hospital Response to Prospective Payment: Moral Hazard, Selection, and Practice-Style Effects," Journal of Health Economics, 15(3):257-277. Finkler, S. (1982). "The Distinction Between Costs and Charges," Annals of Internal Medicine, 96(1):102-109. Gold, M., et al. (1995). "Behind the Curve: A Critical Assessment of How Little Is Known About Arrangements Between Managed Care Plans and Physicians," Medical Care Research and Review, 52(3):304-341. Hodgkin, D., and T. McGuire (1994). "Payment Levels and Hospital Response to Prospective Payment," Journal of Health Economics, 13:1-29. Hornbrook, M., and J. Rafferty (1982). "The Economics of Hospital Reimbursement." Advances in Health Economics and Health Services Research, 3:79-115. Institute of Medicine (1997). Managing Managed Care: Quality Improvement in Behavioral Health, Washington, D.C.: National Academy Press. Johns, L. (1985). "Selective Contracting in California," Health Affairs, 4(3):32-48. Johns, L. (1989). "Selective Contracting in California: An Update," Inquiry, 26:345-353. Joskow, P. (1980). "The Effects of Competition and Regulation on Hospital Bed Supply and the Reservation Quality of the Hospital," Bell Journal of Economics, ll(2):421-447. Kessler, D., and M. McClellan (1999). "Is Hospital Competition Socially Wasteful?" NBER Working Paper 7266, NBER, Cambridge, MA. Langa, K. (1992). Medicaid Cost-Containment in the 1980s: Did It Encourage Interpayer Differences in Hospital Care? Unpublished dissertation. University of Chicago, Chicago, IL. Langa, K., and E. Sussman (1993). "The Effect of Cost-Containment Policies on Rates of Coronary Revascularization in California," The New England Journal of Medicine, 329(24):1784-1789. Luft, H., and R. Miller (1988). "Patient Selection in a Competitive Health Care Environment," Health Affairs, 7(3):97-119. Manning, W. et al. (1984). "A Controlled Trial of the Effect of a Prepaid Group Practice on Use of Services," The New England Journal of Medicine, 310(23):1305-1310. McClellan, M. (1997). "Hospital Reimbursement Incentives: An Empirical Analysis," Journal of Economics and Management Strategy, 6(1):91-128. Melnick, G., et al. (1992). "The Effects of Market Structure and Bargaining Position on Hospital Prices," Journal of Health Economics, 11:217-233.

Effects of Competition Under Prospective Payment on Hospital Costs

101

Melnick, G., and J. Zwanziger (1988). "The Effects of Hospital Competition and the Medicare PPS Program on Hospital Cost Behavior in California," Journal of Health Economics, 7(4):301-320. Melnick, G., and J. Zwanziger. (1995). "State Health Care Expenditures Under Competition and Regulation, 1980 Through 1991," American Journal of Public Health, 85(10):1391-1396. Meltzer, D., et al. (2000). "Effects of Medical Specialization on Costs and Outcomes in a General Medicine Service: Results of a Randomized Trial of Hospitalists." Unpublished manuscript. Meltzer, D., and J. Chung (2000). "Cost Growth Among High- and Low-Cost Hospital Admissions Under Cost-Containment: Evidence from California, 1983-1994." Unpublished manuscript. Miller, R., and H. Luft (1997). "Does Managed Care Lead to Better or Worse Quality of Care?" Health Affairs, 16(5):7-25. Newhouse, J. (1989). "Do Unprofitable Patients Face Access Problems?" Health Care Financing Review, ll(2):33-42. Newhouse, J., M. Buntin, and J. Chapman (1997). "Risk Adjustment and Medicare: Taking a Closer Look," Health Affairs, 16(5):26-3. Newhouse, J. P., S. Cretin, and C. J. Witsberger (1989). "Predicting Hospital Accounting Costs," Health Care Financing Review, ll(l):25-33. Office of Statewide Health Planning and Development (1993). Accounting and Reporting Manual for California Hospitals, Second Edition. Sacramento, CA: OSHPD. Rapoport, J., et al. (1992). "Resource Utilization Among Intensive Care Patients: Managed Care Versus Traditional Insurance," Archives of Internal Medicine, 152:2207-2212. Robinson, J., and H. Luft (1985). "The Impact of Hospital Market Structure on Patient Volume, Average Length of Stay, and the Cost of Care," Journal of Health Economics, 4:333-356. Robinson, J., and C. Phibbs (1989). "An Evaluation of Medicaid Selective Contracting in California," Journal of Health Economics, 8:437-455. Rodgers, W., et al. (1990). "Quality of Care Before and After Implementation of the DRG-Based Prospective Payment System: A Summary of Effects," JAMA, 264(15):1989-1994. Romano, P., L. Roos, and J. Jollis. (1993). "Adapting a Clinical Comorbidity Index for Use with ICD-CM Administrative Data: Differing Perspectives." Journal of Clinical Epidemiology. 46:1075-1079. Russell, L., and C. Manning (1989). "The Effect of Prospective Payment on Medicare Expenditures," The New England Journal of Medicine, 320(7):439-444. Satterthwaite, M. (1979). "Consumer Information, Equilibrium Industry Price, and the Number of Sellers," Bell Journal of Economics, 10(2):483-502. Saving, T. (1982). "Market Organization and Product Quality," Southern Economic Journal, 48(4):855-867. Schwartz, M., D. Young, and R. Siegrist (1995). "The Ratio of Costs to Charges: How Good a Basis for Estimating Costs?" Inquiry, 32:476-481.

102

Meltzer and Chung

Smith, H., and M. Pettier (1985). Prospective Payment: Managing for Operational Effectiveness. Rockville, MD: Aspen Systems Corporation. Sohn, M. (1996). From Regional to Local Markets: Network Study of Competition in California Hospital Markets. Unpublished dissertation. University of Chicago, Chicago, IL. Spence, A. (1975). "Monopoly, Quality, and Regulation," Bell Journal of Economics, 6:417-429. Stafford, R. (1990). "Cesarean Section Use and Source of Payment: An Analysis of California Hospital Discharge Abstracts," American Journal of Public Health, 80(3):313-315. Stigler (1968). The Organization of Industry. Homewood, IL: R. D. Irwin. Tirole, J. (1988). The Theory of Industrial Organization. Cambridge, MA: MIT Press. U.S. Department of Commerce, Bureau of the Census (1993). Revised Estimates of the Population of Counties by Age, Sex, and Race [United States]: 1890-1989 [Computer File]. Washington, D.C.: U.S. Department of Commerce, Bureau of the Census [Producer], 1992. Ann Arbor, MI: Inter-University Consortium for Political and Social Research [Distributor], 1993. U.S. Department of Commerce, Bureau of the Census (1998). Estimates of the Population of Counties by Age, Sex, Race, and Hispanic Origin [United States]: 1990-1996 [Computer File]. Washington, D.C.: U.S. Department of Commerce, Bureau of the Census [Producer], 1998. Ann Arbor, MI: Inter-University Consortium for Political and Social Research [Distributor], 1998. White, L. (1972). "Quality Variation When Prices Are Regulated," Bell Journal of Economics, 3:425-436.

4 Tax Credits, the Distribution of Subsidized Health Insurance Premiums, and the Uninsured Mark V. Pauly, Department of Health Care Systems, University of Pennsylvania Bradley Herring, Institution for Social and Policy Studies, Yale University David Song, Department of Health Care Systems, University of Pennsylvania

Executive Summary This paper investigates the impact of a $1,000 refundable tax credit for self-only coverage on net premiums and insurance purchases for a representative sample of potential buyers in the individual insurance market. Two methods are used to estimate the distribution of premiums: predicted premiums based on a sample of actual purchasers, and premium quotations drawn from an e-insurance web site. In most of the simulations, the net premiums for half or more of the prospective buyers are reduced to zero or low levels. The number of uninsured is reduced by 21 to 85 percent, depending on the size of the deductible in the benchmark plan. However, the results are sensitive to assumptions about insurer underwriting practices. I. Introduction One way to reduce the number of uninsured Americans is to help them pay private health insurance premiums. Proposals for refundable tax credits, such as those from the Bush administration or from members of Congress, would offer many people credits or vouchers that could cover part or all of premiums. For most of the currently uninsured, the most convenient and most likely place to obtain insurance is the individual market, and sometimes proposed credits would be limited to use in such markets. A key to understanding the possible impact of credits of different amounts on insurance purchases is estimation of the extent to which they reduce "net" premiums—the market premium minus the credit—to moderate levels. Previous work has indicated that reasonably generous credits, on the order of about 50 percent of an average premium, might reduce the number of uninsured by half or more.1

104

Pauly, Herring, and Song

However, the measures of premiums in those studies were based largely on estimates of individual insurance premiums generated by applying industry-average administrative loading factors to expected or average expenses, and the credit plans were hypothetical or sample plans. An alternative strategy is to develop direct measures or estimates of premiums that would be or have been charged in real-world markets, apply the actual credits that would be offered under a specific legislative proposal (or variation thereof), and calculate the net premium. One can then estimate the demand for insurance at that net premium. For people for whom the net premium is zero (credit exceeds market premium), one would expect high, almost universal demand; for people whose net premiums remain positive, estimates of the effect of subsidized premium levels on the demand for insurance can generate estimates of the number of people who would be willing to buy at least some insurance coverage at that price. II. What Is the Price of Individual Insurance? Only about 6 percent of Americans obtain private health insurance in individual insurance markets. There are two commonly remarked characteristics of such insurance. First, industry-level data suggest that premiums are high relative to money benefits received. Since the primary benefit from insurance is the payment for medical services, these data suggest that nongroup insurance is (relative to group insurance) generally expensive for what one gets. It is widely believed that this high "price" for benefits partially explains why only about one-quarter of those who are not already covered by group or public insurance choose to buy nongroup coverage. Second, it is commonly believed that persons with higher expected expenses pay higher premiums in the individual market. Pauly and Herring showed that this is only partially true: premiums paid are higher for higher risk people, but they increase significantly less than proportionally with premiums. They increase with individual age and the costliness of the local health care market, but (given age and location) they are not significantly higher for people with chronic conditions or other high risk characteristics.2 Pauly and Herring attribute some of this behavior to the widespread prevalence of guaranteed renewability provisions in individual insurance policies. They do find, however, that premiums paid for similar policies vary substantially for reasons unrelated to observable risk. These observations suggest that it would be desirable to relate credits to the premiums people pay or would pay in the individual market,

Tax Credits, Subsidized Health Insurance Premiums, and the Uninsured

105

as well as estimating the likelihood of purchase (or the proportion purchasing) based on those net premiums and a model of the demand for insurance. This is the task we perform in this paper. Specifically, we will explore the impact on the net premium for self-only coverage of flat dollar credits at a level of $1,000. This credit can be claimed for either individual or group coverage. In the latter case, any tax subsidy would be subtracted from the credit to be offered. We focus on the use of the credit in the individual market. The conventional estimate of the nongroup premium for comprehensive self-only coverage is about $2,500, so such a credit would reflect an average subsidy of only 40 percent. Because actual premiums vary considerably across persons, however, the actual net premium can likewise vary. For the take-up rates presented in this paper, we examine long-run behavior so that any short-run "friction" in take-up rates is not considered. Also, while most proposals envision credits that phase out as income increases and that are different for family coverage, we assume for simplicity here that credits equal $1,000 for all individuals. Two populations are of interest here. The most obvious group to study is the population of people already purchasing individual coverage; for them, we know what insurance costs, and no estimation is involved. For this population, we use evidence from a survey of a representative sample of the U.S. population without group insurance to describe premiums paid and some characteristics of insurance purchased at those premiums. Second, the set of people who did not buy insurance is also important—precisely because this is the set of people whose insurance-purchasing behavior tax credits are supposed to affect. This population may have sought insurance but may have discontinued searching after confronting premiums higher than the premium they would have been willing to pay (their "reservation price"), or they may not have searched at all. We develop several methods to estimate the price (or distribution of prices) they would face. Our work reported here supplements previous work by ourselves and others.3 Our earlier work based measures of gross premiums not on actual premium data but rather on estimates of premiums constructed using information about expected benefits and average administrative loading in the nongroup insurance market. However, the actual premiums people could pay or do pay may differ from these estimates because actual nongroup insurance premiums have a very wide dispersion about the average.4 The other approach to estimating individual insurance premiums is to use an average or median of premiums that insurers quote. In 2001,

106

Pauly, Herring, and Song

this approach yields an estimate of average annual premiums for self-only coverage of about $2,400 per year. However, no rational buyer who has obtained price quotes would choose to pay the average or median price. Rather, the buyer would pay the lowest price (or something close to it). In short, the most relevant measure for the analysis of a tax credit is the premium actually paid by those who obtained insurance and the lowest premium the person would have found for those who do not purchase insurance. Either actual or potential transaction prices may differ substantially from the average or typical premium posted by insurers. Our work reported in this paper generates estimates of net premiums paid or payable for a representative sample of those who currently do not obtain employment-based insurance coverage. This sample is obviously more relevant than two self-selected samples recently discussed in the literature. One is the sample of people who chose to purchase insurance on a large web site.5 Those who actually purchased may have been the lucky few who were able to find low premiums; others who visited the site may have seen premiums they regarded as high and therefore decided to remain uninsured. That is, the population of all purchasers and the premiums they paid may both differ from those for users of the site, and many current nonpurchasers would not have visited the site at all. In short, the sample of web site transactions, while instructive, cannot be assumed to be representative of potential nongroup insurance purchasers. The other sample is the "selected" sample in a small number of cities or in high-risk categories for which only family premiums have been investigated, and for which high premiums have been found.6 A more recent similar study presented premiums and coverages for a small number of hypothetical buyers (in a small number of cities) who list previous or current medical conditions on their insurance applications.7 Our sample is also superior to these two studies as a description of the effect of credits on the overall population of potential nongroup purchasers. Another way in which this work differs from some common policy analyses is the treatment of tax credits in the case of people who face positive net premiums. A typical approach implicitly or explicitly judges the effectiveness of such a program by comparing the resulting net premiums to the person's income and judges insurance to be less "affordable," and therefore less likely to be purchased, if the net premium remains relatively high compared to income.8 However, there is more involved in the purchase of insurance than only this kind of

Tax Credits, Subsidized Health Insurance Premiums, and the Uninsured

107

affordability. Specifically, compared to people facing positive but lower premiums, a person would be expected to be more likely (than someone charged a lower premium) to purchase insurance even at a relatively high premium if the person expected substantially larger benefits from the insurance. Since higher premiums sometimes (although not always) reflect higher risk, we might expect that, at least up to a point, higher risk people would be likely to buy insurance at moderately high premiums if the insurance provides them with high levels of benefit payments and/or shields them from high levels of outof-pocket payment. Consistent with this argument, the data suggest that, even in the relatively low income (but not impoverished) population, the proportion of people obtaining insurance is actually relatively high among middle-aged people.9 Thus, translating net premiums into probabilities of purchase will be an important adjunct to the estimation of the determination of net premiums. Therefore, we use several alternative estimates of the net price of individual insurance that a person might face. The most direct measures use actual premiums paid by a random sample of nongroup insurance purchasers. This sample illustrates the actual variation in premiums paid, variation that is related both to pricing/search behavior and to characteristics of insurance. For those who did not buy insurance, we use two approaches to estimate the premium they would have paid. In one, we use data on the characteristics of the uninsured to generate premium quotations from an online web site. We specify the level of coverage to be held constant, and we offer alternative simulations of purchasing behavior—selecting either the lowest decile of premiums or the premium at the lowest quartile. The second approach develops a premium prediction regression from the data on actual purchases and then uses this regression to predict premiums for the currently uninsured. (All of these approaches may be optimistic because we do not know the actual rejected premiums, if any, of those who did not purchase.) It will be of interest to compare the net effect of various tax credits in states with community rating laws and states without such laws. If community rating laws reduce the extent to which premiums vary with risk (their intended purpose), they will reduce the proportion of a population with relatively high and relatively low net premiums. Determining the magnitude of the difference that laws actually make, and the implications of these differences for insurance purchasing, will be of value.

108

Pauly, Herring, and Song

If we generate a pattern of estimated net premiums for the uninsured, how can we determine whether they would be willing to buy coverage at those premiums? Just observing the distribution of net premiums can tell us a great deal. Those who face zero or nominal net premiums would be expected to take up the free or nearly free coverage. But what will the other people who face positive but subsidized premiums do? The main approach we pursue here is to convert one of the distributions of net premiums into estimates of the distribution of insurance purchases. We assume that people whose credit equals or exceeds the premium for a given insurance policy, for whom the net price is zero, would prefer to obtain that policy rather than remain uninsured. For those with positive net premiums, we use the models developed by Pauly and Herring to estimate the probability or proportion of insurance purchasing.10 III. Net Premiums Among Insureds We examine a sample of persons purchasing nongroup insurance provided by the 1996-1997 Community Tracking Survey (CTS), a large random household sample selected from a nationally representative set of communities. Within the set of all individuals, 1,050 nonelderly adult respondents reported purchasing nongroup insurance on a self-only basis. Of this set, 908 (86 percent) reported the premiums paid. All of these individuals are classified as "insured" in the analysis of the data. The 1996 premium data are "inflated" to 2001 insurance price levels by using an annual premium growth rate of 7 percent. If we assume that each of these individuals would be eligible for a $1,000 tax credit, the distribution of net premiums would be as indicated in table 4.1. Approximately 20 percent of these purchasers would pay a zero net premium. At the median net premium, the dollar amount is $809 per year, which is 45 percent of the total premium. About 60 percent of all purchasers would have had their premium halved by a $1,000 credit. We also split the sample at the median income and found essentially similar results for each subsample. IV. Online Premiums for Nonelderly Uninsured Adults There were 6,083 nonelderly adults in the CTS sample who indicate that they currently had neither public nor private health insurance. We

Tax Credits, Subsidized Health Insurance Premiums, and the Uninsured

109

Table 4.1 Distribution of net premiums for nongroup self-only policies under fixed $1,000 tax credit plan

Mean 10th percentile 25th percentile 50th percentile 75th percentile 90th percentile

Actual premium

Net premium (NP)

NP as a percentage of original premium

$1,989 $708 $1,088 $1,809 $2,753 $3,932

$1,122 $0 $88 $809 $1,753 $2,932

36 0 8 45 64 75

Representative sample of 821 nongroup policyholders in the CTS household survey (1996-1997).

identified a comprehensive (indemnity or PPO) medical-surgical plan with annual deductibles less than or equal to $1,000, or an HMO with similar or lower deductibles. We then examined the publicly available online premium quotations available from the e-healthinsurance web site (www.ehealthinsurance.com) that would be available to the survey respondents given their age, gender, smoking behavior, and zip code location. The lowest premium plan that meets these criteria is certainly a possible choice. If its premium were less than or close to the credit, it surely would be preferable to the no-insurance option. In many cases, however, the absolute lowest priced plan that meets the deductible specification also has other exclusions; for instance, it may exclude coverage of nonsurgical routine outpatient care entirely. Accordingly, we show the net premium at the 10th and 25th percentile of the distribution of premiums. We were able to match 72 percent (N = 4,383) of the CTS sample with web premiums. Failure to match usually was due to missing CTS data (for example, smoking behavior), or to the absence of a web insurance option in the sample person's zip code. The geographic patterns themselves are of interest. Web premiums were more likely to be missing for people who lived in community rating states. The form of community rating varies across the states that require it, with only New York and New Jersey using pure community rating. Of the nine states identified as community rating states with cities in the CTS sample (New York, New Jersey, Washington, Massachusetts, Kentucky, North Dakota, New Hampshire, Maine, and Vermont), there were no web site

110

Pauly, Herring, and Song

premiums available in six. In New York, the highest deductible listed was $250. In the remaining community rating states (Washington and New Jersey), median premiums were about three times higher than in other states, presumably reflecting both higher premiums for given coverage and less availability of high-deductible options. For example, deleting the community rating states from table 4.1 would cut the net premium at the 50th percentile in half. Results of the analysis on the complete matched data set are shown in tables 4.2 and 4.3. For the population of potential insurance purchasers, the 10th percentile net premiums for insurance obtained from the web are zero or close to zero for 50 to 75 percent of all such persons. As in table 4.1, the top 10 percent of net premiums are quite large, reflecting both the absence of a high-deductible option and high overall health insurance premiums in some locations. If we move up the frequency distribution of premiums to the 25th percentile, there is still a sizeable proportion of the uninsured who can obtain insurance for free and must pay net premiums that are a small fraction of the total. The net premium at the 50th percentile is $252. V. Estimating Premiums for the Uninsured Based on Transactions Data The third approach to generating a distribution of net premiums assumes that uninsured persons would have available to them the same premiums as those with similar characteristics who actually purchased insurance. This approach also assumes that the uninsured would have the same preferences for plan types and the same search behavior as those who actually purchased. In reality, the uninsured probably would search less, but they might also might seek less coverage than those who purchased. To generate a distribution of premiums, we first regress the premiums paid by purchasers on purchaser characteristics. Table 4.4 shows the result of such a regression. The most important predictive variables turn out to be age, race, and community location. Given the distribution of characteristics of the uninsured, we then generate a distribution of predicted or average premiums. Table 4.5 shows the distribution of net premiums based on this approach. The distribution (perhaps not surprisingly) is fairly close to the distribution of paid premiums, but with a somewhat larger proportion of the population facing low net premiums (since more of the uninsured are young).11

Tax Credits, Subsidized Health Insurance Premiums, and the Uninsured

111

Table 4.2 Distribution of net 10th percentile online premiums for uninsured individuals, under fixed $1,000 tax credit plan

Mean 10th percentile 25th percentile 50th percentile 75th percentile 90th percentile

Online premium

Net premium (NP)

of original premium

$1,326 $478 $640 $984 $1,791 $2,884

$508 $0 $0 $0 $791 $1,884

27 0 0 0 44 65

Representative sample of 4,383 uninsured individuals in the CTS household survey (1996-1997). For each individual, the online premium reflects the 10th percentile premium for a menu of individual health insurance plans (whose deductibles are no greater than $1,000) in the corresponding locale. Premium quotes were obtained from www.ehealthinsurance.com. Table 4.3 Distribution of net 25th percentile online premiums for uninsured individuals, under fixed $1,000 tax credit plan

Mean 10th percentile 25th percentile 50th percentile 75th percentile 90th percentile

Online premium

Net premium (NP)

NP as a percentage of original premium

$1,631 $683 $873 $1,252 $1,995 $2,952

$710 $0 $0 $252 $995 $1,952

27 0 0 20 50 66

Representative sample of uninsured individuals in the CTS household survey (1996-1997). For each individual, the online premium reflects the 25th percentile premium for a menu of individual health plans (whose deductibles are no greater than $1,000) in the corresponding locale. Premium quotes were obtained from www. ehealthinsurance.com.

Table 4.4 Regression analysis of the determinants of health insurance premiums, population ages 18-64 with nongroup coverage3 Variable Male Age 18-24 25-34 35-44 45-54 55-64 Female 18-24 25-34 35-44 45-54 Squared age Smoker White African-American Hispanic Family income Family education High school grad Some college College grad Graduate school Metropolitan area New England Mid-Atlantic East S. Central West N. Central West S. Central South Atlantic Mountain Pacific Constant a

N = 740, adjusted R-squared = 0.19. Significant at 5%. Significant at 1% or less.

Coefficient -622.3 -715.0 -74.8 38.18 404.2b -645.7 -600.3 -560.3 -187.0 0.619 -179.3 -149.4 -921 .3C -22.5 0.0042C 207.4 162.4 262.8 63.25 210.3 101.1 51.66 -62.94 -281.8 -171.7 -82.00 -154.7 -252.1 1550b

Tax Credits, Subsidized Health Insurance Premiums, and the Uninsured

113

Table 4.5 Distribution of net predicted CTS premiums for uninsured individuals, under fixed $1,000 tax credit plan

Mean 10th percentile 25th percentile 50th percentile 75th percentile 90th percentile

Predicted premium

Net premium (NP)

NP as a percentage of original premium

$1,558 $735 $1,166 $1,519 $1,864 $2,475

$619 $0 $166 $519 $864 $1,475

31 0 14 34 46 60

Representative sample of 6,083 uninsured individuals in the CTS household survey (1996-1997). Premiums are predicted from estimated coefficients in a regression of nongroup policyholder premiums on characteristics (see table 4.4).

VI.

Simulating Underwriting

The estimates using the web data show the premiums that individuals would see if they searched the relevant part of the web site, but the process of obtaining coverage requires more than just agreeing to buy at one of those proposed premiums. (The estimates based on the CTS "purchasing" sample already include any higher premiums based on underwriting.) For new purchases of insurance (but not for renewals), individuals must apply for coverage and will be asked questions about health status and use of medical services in the recent past. If an individual provides answers that suggest high risk, the insurer may decline to insure or may propose a higher premium. While some high-risk individuals obtain individual insurance coverage at premiums that do not differ from the average, some do not.12 To simulate the operation of an underwriting process, we assume that uninsured individuals in the CTS survey who report their health on the survey to be "poor" or "fair" (whatever they might tell an insurer), or who report that they were ever denied coverage, will be faced with the premium in the 90th percentile of the distribution of premiums for persons with their characteristics. Of the 4,352 individuals in the sample, 895 (18 percent) report that they were in poor or fair health, and 202 report that they were ever denied coverage. Because of overlap, the net proportion of the sample thus classified as "high risk" is 21 percent. Table 4.6 modifies table 4.3 based on these assumptions. It is still true that more than one-quarter of all uninsured would face zero net

114

Pauly, Herring, and Song

Table 4.6 Distribution of net premiums after underwriting, under fixed $1,000 tax credit plan

Mean 10th percentile 25th percentile 50th percentile 75th percentile 90th percentile

Online premium

Net premium (NP)

NP as a percentage of original premium

$1,849 $663 $831 $1,390 $2,436 $3,962

$936 $0 $0 $390 $1,436 $2,962

31 0 0 28 59 75

Representative sample of uninsured individuals in the CTS household survey (1996-1997). For each individual who reported no denial of coverage and good to excellent health, the online premium reflects the 25th percentile premium for a menu of individual health insurance plans in the corresponding locale. Those individuals (approximately 20 percent of the sample) who reported fair or poor health or denial of coverage were matched to the 90th percentile premium. Premium quotes were obtained from www.ehealthinsurance.com.

premiums, but the net premium at the 50th percentile rises a moderate amount, to $390. The proportion who would be classified as high risk by these assumptions is larger than the proportion of applicants insurers estimate they would classify in this way.13 But probably many high risks do not apply for insurance, both because they would not expect to be quoted premiums they would be willing to pay and because high risk especially characterizes low-income uninsured, who would often not pay even average-risk premiums. VII. Predicting Purchase We now wish to determine whether uninsured individuals facing the distribution of net premiums described above would be willing to buy coverage. To produce estimates of such take-up rates, we use two simulation techniques described in Pauly and Herring.14 One technique constructs the distribution of the reservation prices directly and assumes an individual will obtain coverage if the individual's reservation price exceeds the net premium he or she faces. The other technique estimates an individual's probability of obtaining insurance as a function of the net premium; this model is derived from the observed relationship between coverage and net loading in the employment-based setting. For our first approach, we construct a "synthetic" estimate of a reservation price for insurance for an uninsured individual based on his or

Tax Credits, Subsidized Health Insurance Premiums, and the Uninsured

115

her change in expected expenses for being insured relative to being uninsured. Specifically, one's willingness to pay for insurance is specified as the sum of the expected decrease in out-of-pocket expense, the Arrow-Pratt risk premium paid for the decrease in the variation in out-of-pocket expense, the increase in consumer surplus from consuming more medical care, and the reduction in disutility felt from free care when uninsured. We set the Arrow-Pratt absolute risk aversion coefficient equal to 0.00095 and assume that the increase in consumer surplus equals half of the difference in total expenditures. Further, we assume that the disutility associated with being a charity care patient equals 20 cents per dollar of free care received; for more detail regarding results generated from varying this assumption, see Pauly and Herring (forthcoming). Since the CTS lacks detailed data for medical expenses, we first developed a distribution of reservation prices for a sample of individuals in the 1996 Medical Expenditure Panel Survey (MEPS). After inflating this data to 2001 dollars, we then assigned that distribution of reservation prices to the uninsured CTS sample by randomly selecting a reservation price from an MEPS subsample of individuals with similar age, gender, and self-reported health status. Table 4.7 shows the distributions of reservation prices of the uninsured for three separate plans that were the most common of those found online. The first is a PPO plan with a $1,000 deductible, 20 percent coinsurance, and a $2,000 upper limit on out-of-pocket spending. The second is a similar plan with a $500 deductible, and the third has a $250 deductible. The assumptions we make for estimating an individual's reservation price for a particular plan are those that we presented in the midrange case presented in Pauly and Herring.15 As seen in the table, the median uninsured individual would be willing to pay $592 for a $1,000 deductible PPO plan, $707 for a $500 deductible plan, and $787 for a $250 deductible plan; these distributions are considerably skewed. To examine the level of subsidy required to induce the uninsured to purchase insurance, we show the distribution of the difference between an individual's reservation price and the premium he or she faces. Table 4.8 shows the distribution of subsidies needed when we assume that the premium faced by an uninsured individual is at the 10th percentile of premiums quoted in our automated online search. The last three lines of this table show the percentile at which the difference between our estimated reservation price and the premium they face falls below $1,000. For example, we find that 85 percent of the unin-

116

Pauly, Herring, and Song

Table 4.7 Distribution of reservation prices for insurance, using an expected utility frameworka

Mean 10th percentile 25th percentile 50th percentile 75th percentile 90th percentile

$1,000 deductible

$500 deductible

$250 deductible

$1,106 $176 $312 $592 $1,332 $2,612

$1,245 $240 $395 $707 $1,513 $2,900

$1,353 $290 $469 $787 $1,647 $3,126

a

Details of these simulations are provided in the text.

Table 4.8 Distribution of subsidies required for the purchase of insurance (assuming premiums obtained are at the 10th percentile)a

Mean 10th percentile 25th percentile 50th percentile 75th percentile 90th percentile 85th percentile 78th percentile 34th percentile

$1,000 deductible

$500 deductible

$250 deductible

$534 $0 $0 $168 $608 $1,480 $1,000 NA NA

$1,176 $0 $0 $316 $880 $1,966 NA $1,000 NA

$1,615 $0 $487 $1,464 $2,312 $3,524 NA NA $1,000

a

Details of these simulations are provided in the text.

sured sample requires a subsidy of under $1,000 for the purchase of a $1,000 deductible PPO plan, while only 34 percent of the uninsured would respond to such a subsidy for purchase of a $250 deductible plan. If we assume that all of those for whom the net premium falls below their reservation price were to obtain coverage, the take-up rate of a $1,000 credit for a $1,000 deductible PPO would be 85 percent. Table 4.9 shows the distribution of subsidies needed when we assume instead that individuals are able to obtain premiums only at the 25th percentile of those quoted for their specific age, gender, and location. The subsidies reported here are larger. Our expected utility model generates reservation prices for insurance for the minority that fall below the unsubsidized premium (we assume that) they face currently. We therefore present take-up rates of a

Tax Credits, Subsidized Health Insurance Premiums, and the Uninsured

117

Table 4.9 Distribution of subsidies required for the purchase of insurance (assuming premiums obtained are at the 25th percentile)

Mean 10th percentile 25th percentile 50th percentile 75th percentile 90th percentile 77th percentile 57th percentile 32nd percentile

$1,000 deductible

$500 deductible

$250 deductible

$741 $0 $0 $390 $951 $2,044 $1,000 NA NA

$1,580 $0 $156 $866 $1,500 $2,798 NA $1,000 NA

$1,678 $0 $638 $1,518 $2,379 $3,583 NA NA $1,000

Details of these simulations are provided in the text. Samples differ slightly due to the unavailablity of certain plans in a few markets.

$1,000 credit using this "implied" take-up rate of unsubsidized insurance as our baseline. For instance, we observe that the 10th per- centile "net" premium (i.e., the "absolute" premium minus the $1,000 credit) for the $1,000 deductible PPO plan is less than our simulated reservation price for 85 percent of our sample of currently-uninsured individuals. However, the 10th percentile of the "absolute" premium is actually less than the reservation price we produced for 23 percent of this sample of uninsured. We therefore present overall take-up rates using as a baseline this 23 percent of uninsured that our model predicts would have already purchased unsubsidized insurance. That is, the overall take-up rate we present is equal to 62 percent, i.e., about 85 percent minus 23 percent. Table 4.10 shows more detail on the take-up rates that we estimate for various assumptions about the plans facing the uninsured. If we assume instead that the uninsured face the 25th percentile of premiums for a $1,000 deductible PPO, we estimate a lower take-up rate of 56 percent. Assuming that the $1,000 credit is made available only to a more generous $250 deductible PPO, we estimate a reduction in the uninsured of only 21 percent when we also assume that an individual faces the premium at the 25th percentile of online quotes. If we use instead the individual's predicted premium from the CTS nongroup sample of purchases, we estimate a take-up rate of 43 percent. For this final case, the distribution of reservation prices we apply is that for the typical nongroup plan seen in the CTS followback data.

118

Pauly, Herring, and Song

Table 4.10 Take-up rates of private insurance given a $1,000 refundable tax credita Premium assumption Internet premiums, 10th percentile (of $1,000 deductible plans) Internet premiums, 25th percentile (of $1,000 deductible plans) Internet premiums, 25th percentile (of $250 deductible plans) Predicted CTS nongroup premiums

Reservation price approach

Net loading approach

62%

85%

56%

77%

21% 43%

32% 61%

a

Details of the simulations are provided in the text.

The second column of table 4.10 shows the results generated from applying a second estimation technique described in Pauly and Herring.16 This second technique instead estimates a reduced-form version of the demand for insurance to produce an individual's likelihood of obtaining coverage as a function of the net price she or he faces. Specifically, using the MEPS data for workers and their dependents, we estimated a probit model for the probability of purchasing insurance as a function of one's age, gender, income, education, race, region, and the net price one faces. Defining this "net price" in the traditional sense as the administrative loading and as a percentage of expected benefits, we constructed values for net loading determined by household marginal tax rates and the administrative loading that coincides with the median firm size in the worker's industry. Thus, the coefficients from estimating this model for the probability of being insured as a function of these various controls and the net price of insurance allows us to simulate a new predicted probability of purchasing insurance given that uninsured individual's demographic controls and the "new" subsidized net price. Here too, however, the take-up rate for insurance must be specified as the change in predicted probabilities from the unsubsidized nongroup price to a subsidized one. If we consider the 25th percentile of online quotes for a $250 deductible PPO plan and assume nongroup administrative loading equal to 40 percent of premiums, we estimate a 32 percent take-up rate of a $1,000 credit for that insurance. If instead we consider a $1,000 deductible PPO plan, the fixed-dollar credit lowers the net loading considerably; here, we estimate a take-up rate of 77 percent. If we assume that the premium faced by an uninsured individ-

Tax Credits, Subsidized Health Insurance Premiums, and the Uninsured

119

ual is the one we generate from the average CTS nongroup premium, we estimate a reduction in the uninsured equal to 61 percent. Overall, the take-up rates we estimate from our net loading approach are somewhat higher than those we estimate from our reservation price approach. As we argued before, however, this uncertainty—whether in the form of specifying a model or in the form of assuming what premiums the uninsured face—should be front and center in the evaluation of tax credit schemes because we, as analysts, have minimal experience with large subsidies directed at low-income individuals.17 VIII. Discussion The results in this paper yield estimates of the effectiveness of a modest tax credit in reducing the number of a representative sample of uninsured that are consistent with, but somewhat more optimistic than, our earlier conjectures, largely because the premiums we estimate or use in this paper are generally lower than those used earlier. If the 10th percentile web site premiums represent genuine offers to sell insurance, the results are even more optimistic. While many of the low-risk uninsured face low or zero premiums, however, the minority of people who are high risk will still pay high net premiums. Our simulations indicate that even many of these older or higher risks would be willing to pay higher premiums because the alternative is to pay large amounts out of pocket or put up with less attractive charity care. The main conclusion, as we have noted before, is that fixed dollar premiums are less effective than proportional or risk-adjusted premiums at getting a smaller number of higher risks covered, but they are more effective for the larger number of lower risks. Indeed, without additional risk-adjusted credits or a high-risk pool, it is unlikely that insurers will insure truly high risks who seek coverage for explicitly acknowledged active medical conditions. Why do our results in this paper differ somewhat from our earlier findings? It is possible that web premiums tend to be lower than actual transactions premiums for people who did not use the web. Maybe the benchmark loading estimates, which are based on aggregate premiums and benefits for a set of large commercial insurers, are overestimates. The Blue plans, which are active in the nongroup market, were left out of some of these benchmark measures, as were some smaller commer-

120

Pauly, Herring, and Song

cial insurers, and accident insurance was included with health insurance. The key remaining unknown for our study and all others, however, is how underwriting affects the premiums people actually pay (not what insurers quote). Our previous work indicated that many buyers with a history of prior chronic conditions were able to avoid paying unusually high premiums and were still able to obtain coverage. But the nature of the interplay between sellers trying to charge high premiums to high (and low) risks, and buyers searching for reasonable options, is unknown. Current nongroup health insurance premiums are not set in expectation of purchases by a large number of formerly uninsured persons armed with tax credits. Such a surge in demand would cause a profound transformation of this small and sleepy market. What effects might it have on premiums and coverage? We think it is likely that it would lead to lower administrative costs and less severe problems of adverse selection than nongroup insurers currently face, both because of sheer volume and because most of these new purchasers would tend to be average risks who are strongly motivated to seek coverage with little selling effort needed. If these conjectures are right, the final outcome could be even more optimistic than the estimates presented here. It is also possible, however, that cautious insurers might respond by raising premiums in the face of what is perceived as a more risky (or at least different) market. Tax credits would put a heavy obligation on nongroup insurers to offer attractive policies at affordable net premiums. Notes Research was supported by a grant from The Leon Lowenstein Foundation. 1. Pauly and Herring 2001. 2. Pauly and Herring 1999. 3. Pauly and Herring 2001; Prague 2001; eHealthInsurance.com 2001a, eHealthInsurance.com 2001b, eHealthInsurance.com 2001c, eHealthInsurance.com 2001d. 4. Pauly and Herring 1999. 5. Frogue 2001. 6. Lav and Friedman 2001. 7. Pollitz, Sorian, and Thomas 2001.

Tax Credits, Subsidized Health Insurance Premiums, and the Uninsured

121

8. Trude and Ginsburg 2001. 9. Shactman and Altaian 2000. 10. Pauly and Herring, forthcoming. 11. We also used standard errors of prediction to construct 95 percent confidence intervals for expected premiums, conditional on individual characteristics. Because the errors tend to cancel out, we would not expect that the prediction error matters much to the distribution of net premiums. We varied the distribution of predicted premiums through random draws of these intervals or through the application of the intervals' lower and upper bounds. We conclude that the intervals are small because they do not affect the overall interpretation of the original distribution of predicted premiums. 12. Pauly and Herring 1999. 13. ChoUet and Kirk 1998. 14. Pauly and Herring, forthcoming. 15. Pauly and Herring 2001. 16. Pauly and Herring, forthcoming. 17. Pauly and Herring 2001.

References Chollet, D., and A. Kirk (1998). Understanding Individual Health Insurance Markets: Structure, Practices, and Products in Ten States. Memo Park, CA: Henry J. Kaiser Family Foundation. eHealthInsurance.com (2001a). "Letter to Congress: eHealthlnsurance Endorsement of the 'Reach' Act," March 14,2001. http:/ /www.ehealthinsurance.com/ehealthinsurance/ Press.html (June 12,2001). eHealthInsurance.com (2001b). "Press Release: eHealthlnsurance Supports U.S. Senators on Bipartisan 'Reach' Act to Help Uninsured Americans," March 14,2001. http://www. ehealthinsurance.com/ehealthinsurance/Press.html (Tune 12,2001). eHealthInsurance.com (2001c). "Letter to Congress: eHealthlnsurance Endorsement of the Fair Care for the Uninsured Act," April 3, 2001. http: / /www.ehealthinsurance. com/ehealthinsurance/Press.html(Tune 12,2001). eHealthInsurance.com (2001d). "Press Release: eHealthlnsurance Supports Congress' Bipartisan Tax Credit Provisions in the Fair Care for the Uninsured Act," April 3, 2001. http:/ /www.ehealthinsurance.com/ehealthinsurance/Press.html (Tune 12,2001). Frogue, J. (2001). "Recent Survey Points to Affordable Individual Health Insurance." The Heritage Foundation Executive Memorandum, No. 740 (April 17). Lav, I., and J. Friedman (2001). "Tax Credits for Individuals to Buy Health Insurance Won't Help Many Uninsured Families." Center on Budget and Policy Priorities Report (February 15). Pauly M., and B. Herring (1999). Pooling Health Insurance Risks, Washington, D.C.: AEI Press.

122

Pauly, Herring, and Song

Pauly, M., and B. Herring (2001). "Expanding Insurance Coverage Through Tax Credits: Tradeoffs and Options." Health Affairs 20(1):1-18. Pauly, M., and B. Herring (forthcoming). Options and Effects of Tax Credits for Health Insurance. Washington, D.C.: AEI Press. Pollitz, K., R. Sorian, and K. Thomas (2001). How Accessible Is Individual Health Insurance for Consumers in Less-Than Perfect Health? Menlo Park, CA: Henry J. Kaiser Family Foundation, June. Shactman, D., and S. Altaian (2000). "The United States Confronts the Policy Dilemmas of an Aging Society." Health Affairs 19(3):252-258. Trude, S., and P. Ginsburg (2001). "Tax Credits and Purchasing Pools: Will This Marriage Work?" Center for Studying Health System Change, Issue Brief No. 36 (April).

5 Hospital Ownership Conversions: Defining the Appropriate Public Oversight Role Frank A. Sloan, Duke University andNBER

Executive Summary This paper reviews recent empirical evidence on the effects of hospital ownership conversions on quality of care and provision of public goods, such as uncompensated care, and presents new results on these topics based on hospital discharge data from the Healthcare Cost and Utilization Project's (HCUP) Nationwide Inpatient Sample. My analysis of these data reveals that conversion from government or private nonprofit to for-profit ownership has no effect on in-hospital mortality, but rates of pneumonia complications increased following conversion to for-profit status. Other research, discussed in the paper, found increased mortality rates following discharge from the hospital for patients admitted to hospitals that had converted to for-profit ownership. There was no effect of such conversions on the propensity to admit uninsured or Medicaid patients. Clearly, there is considerable heterogeneity in outcomes attributable to conversions. Overall, the evidence suggests a role for public scrutiny of hospital ownership conversions. I.

Introduction and Policy Context

The hospital industry attracts much public scrutiny, given its important role in providing personal health services, its size—about 3 percent of gross domestic product, the importance of hospitals as employers in the communities they serve, and the high share of hospital revenue from public sources. During the last two decades, the industry has experienced a dramatic downsizing at the same time, and for that reason, the methods of paying for hospital care have changed. The market for hospital care in the year 2000 was much more competitive than it was

in 1980. Downsizing has taken various forms. First, there has been a reduction in the number of hospitals. In 1980, there were 5,830 "community" hospitals (nonfederal short-term general hospitals located outside

124

Sloan

institutions). By 1999, the latest year for which data are publicly available, the number had fallen to 4,956 (figure 5.1). The number of community hospitals peaked during the 1970s. Second, existing hospitals have reduced bed capacity from a peak of slightly over 1.0 million beds in 1983 to 830,000 beds in 1999 (figure 5.2). Third, hospitals have diversified, integrating both vertically and horizontally, and they have sought new ownership and management. As an extreme measure, many hospitals have closed. For-profit (F) hospitals run counter to the national trend in number and bed capacity (figures 5.1 and 5.2). The number of such hospitals has remained relatively constant in terms of numbers of hospitals and has risen in terms of number of beds under such hospital ownership. Although the number of private nonprofit hospitals (N) peaked in 1984, as did the number of beds, the subsequent decline in both has been small relative to the decline in number of public (G) hospitals and beds in such hospitals. In fact, most of the decline in community hospital capacity has occurred in state and local government community hospitals. Thus, the trend has been to a privatized hospital system with some relative increase in the share of private hospitals under for-profit ownership. These changes reflect hospital closings, mergers, as well as ownership changes among existing hospitals. This study focuses on hospital ownership changes and the effects of such changes. Although government or private nonprofit ownership to for-profit ownership receives the most publicity, in fact, changes have occurred in all directions (Needleman et al. 1997). Relatively more attention has been paid to conversions from G and N to F ownership for several reasons. There is a concern that hospitals with a profit-seeking mission as an explicit motive are less likely to accept unprofitable cases for treatment, such as persons who lack health insurance or who are underinsured. Given the goal of maximizing profit, such hospitals may be more likely to exploit loopholes in the reimbursement rules; they may be more willing to reduce quality of care, especially quality attributes that are difficult for patients, and perhaps even their physicians, to monitor ("noncontractable quality"; see, for example, Hart et al. 1997). Allegations of adverse effects associated with ownership conversion are easily made. But obtaining rigorous empirical support for such allegations is a much more difficult matter. Any in-depth study of the effects of ownership change on access to and quality of care and on business practices should account for the following factors.

Hospital Ownership Conversions

125

Figure 5.1 Number of hospitals in the United States Source: American Hospital Association (1976, 1979, 2001). The data before 1979 are for nonfederal, short-term general and other special hospitals, which include community hospitals plus hospitals in institutions.

First, hospitals do not change ownership in a vacuum. Hospitals that convert may have specific attributes that distinguish them from hospitals that do not convert. These attributes may be a characteristic of the hospital and/or of the market in which the hospital operates. If a hospital did not change ownership in the particular way proposed, what was the alternative course of action? The hospital industry is a mature industry, in a sense, more like steel than e-commerce. In an industry that is downsizing, there are rarely many attractive alternatives. Thus, even if the outcomes are worse than before, such outcomes could have been even worse if the choice to change ownership had not been made. The alternative to conversion may have been closure. Under such circumstances, all persons in the locality may have experienced a decrease in access to hospital care, and the loss of jobs to the community may have been far greater than occurred as a consequence of "efficiency measures" implemented by the acquirer. It is essential to ask the "what if" question, both in policy and in empirical analysis of effects of ownership changes. Second, some of the observed changes reflect change in ownership per se, rather than the effects of change in type of ownership. This is

126

Sloan

Figure 5.2 Number of hospital beds in the United States Source: American Hospital Association (1976, 1979, 2001). The data before 1979 are for nonfederal, short-term general and other special hospitals, which include community

particularly true during the first few years following a conversion, as hospitals adjust to new management and strategies. A third point pertains to policy adoption based on empirical evidence of undesirable outcomes following specific types of hospital ownership conversions. If changes in mission and in behavior are observed, are there more direct and efficient policy instruments for ensuring desirable behavior than simply blocking a certain type of conversion? For example, if fraud and abuse in hospital reporting of patient information for purposes of reimbursement is a widespread practice, a more direct approach would involve direct public oversight and enforcement rather than indirectly affecting such behavior by influencing the mix of hospitals according to their propensities to engage in undesirable behaviors. As discussed below, some researchers have found that profit-seeking behavior is contagious. That is, when for-profit hospitals engage in certain kinds of behavior, their hospital competitors with different ownership forms may emulate it. There is a lot of empirical literature on the relationship between hospital ownership and various performance measures (see, for example, Sloan 2000). In general, the literature reveals that private hospitals,

Hospital Ownership Conversions

127

whether they are for-profit or nonprofit, are more alike than different. The vast number of studies, however, have assessed the effects of ownership type on hospital behavior rather than ownership conversions. The latter may be particularly insightful because they discuss location and characteristics that potentially influence behavior associated with location constant and examine changes that occurred post- versus preconversion. Norton and Staiger (1994) found differences in hospital behavior by ownership status, but the differences were attributable to where for-profit versus other types of hospitals decided to locate. For example, a profit-seeking hospital may decide to locate in a relatively affluent suburb, where well-insured persons are located, rather than in an inner city, where people are much more likely to be uninsured or to be enrolled in Medicaid (one of the less generous third-party payers). The dearth of studies on hospital ownership conversions reflects the difficulty of obtaining accurate data on conversion dates and conversion types. This problem has been remedied in the present study and in some studies reviewed below. After comparing ownership codes from two independent sources, the Annual Survey of Hospitals (conducted by the American Hospital Association) and Medicare Cost Reports, my students called hospitals for which the two sources differed to determine whether the conversion occurred, when it occurred, and the ownership types before and after conversions. In total, about 300 telephone calls were made. Compared to the number of quantitative studies, there have been very few qualitative studies of ownership and ownership conversions. Qualitative studies may be called "soft" but at the same time, they can reveal differences among and changes in decision-making processes that otherwise can only be inferred very indirectly from outcome changes. By peering inside hospital decision making, especially decision making in the presence of major stress and organizational change, we can enhance our understanding of these changes. Such analysis also provides an important cross-check on findings from the quantitative analysis. Qualitative analysis is not done mostly because it is so difficult to do. Decision makers undergoing change and/or associated with choices that did not lead to a successful outcome are not likely to want to have their decisions and the consequences scrutinized. Also, there is no explicit hypothesis testing, as in quantitative research. In the end, one gains an impression without the sharply defined results of a rejected null hypothesis.

128

.

Sloan

In this paper, I review and evaluate very recent research (mostly published in 2000 or later) on the relationship between hospital ownership and behavior, and I present new empirical results on ownership conversions based on empirical analysis of hospital discharge data for the years 1988 through 1996, the Healthcare Cost and Utilization Project's (HCUP) Nationwide Inpatient Sample (MS). HCUP is an intramural program of the U.S. Agency for Healthcare Research and Quality. Using qualitative as well as quantitative evidence, the public policy questions addressed in this paper are those posed above. Does ownership conversion, especially to the for-profit ownership form, lead to increased care barriers, diminished quality of care, and profit-seeking billing practices designed to maximize reimbursement? Do acquirers of hospitals pay prices for the facilities that are commensurate with earning a fair rate of return on their investments? Or do acquirers pay too little for community assets in the form of hospitals? If there is evidence of market failure in this context, what are the appropriate public remedies? Section II of this paper reviews empirical evidence on the effects of ownership and ownership conversions on behavior. Section III describes a new investigation of the effects of ownership conversions from G and N to F ownership and the reverse, from F to G or N. To the extent that hospital decision making becomes more profit-oriented when hospitals convert from G or N status, one should observe the opposite when hospitals convert from F to G or N status. Thus, by examining the effects of conversions in both directions—toward and away from for-profit status—the empirical analysis makes it possible to distinguish the effects of change of ownership type from the effects of conversion per se. Section IV presents results from a sample of Medicare beneficiaries and Section V results from a sample of patients not yet age-eligible for Medicare. Section VI further evaluates the findings and explores the implications of the results for public policies as they relate to hospital ownership conversions. II. New Empirical Evidence on Ownership Effects and on Hospital Conversions Why Hospitals Convert Using qualitative as well as quantitative methods, several recent studies have analyzed why hospitals change ownership status. The major

Hospital Ownership Conversions

129

determinants of ownership change are financial distress prior to the change, mostly due to a reduction in aggregate demand for hospital inpatient care and an increase in competition among hospitals, which necessitates implementation of new competitive strategies that were more difficult to introduce under the prior regime. A study of five public hospitals that converted from public to nonprofit ownership gave two reasons for the changes: (1) a reluctance of public facilities to expand beyond their political jurisdiction and (2) governance and management restrictions that made it more difficult to compete with other hospitals (Legnini et al. 2000). Burns et al. (2001) performed case studies of sixteen hospitals that changed ownership between 1993 and 1996. Among the sixteen hospitals were conversions from private nonprofit to for-profit, for-profit to nonprofit, public to for-profit, and public to nonprofit ownership. The case studies revealed two primary motivations for conversion regardless of ownership type, even though the two both come down to money. First, with one exception, the hospitals reported current or anticipated financial stress as a motivation. By joining a chain or reducing managerial constraints in the case of public hospitals, the hospitals hoped to improve financial performance. A motivator for the hospitals converting from nonprofit to for-profit governance was access to equity capital. In one case, the hospital was not considered to be "strategic enough" to justify further investment by its nonprofit owner, but the hospital fit the strategy of a for-profit purchaser. In the case of for-profit to nonprofit conversions, the hospital had not met the former owners' required financial return. Reasons for the public to private conversions included poor hospital financial management under the previous ownership. In the extreme, the alternative for the hospital was closure. Second, nearly all the hospitals said that they were faced with an inability to compete for managed-care contracts. By joining a larger, more dominant hospital in the local market, the hospital was in a stronger position when dealing with managed-care organizations. A desire to change its mission was a motive only rarely for an ownership conversion. Transitions to for-profit ownership were motivated by both "pull" and "push" factors. The former included an attractive financial offer; the latter included siphoning off hospital cash flow by the former owner. Transitions from for-profit ownership to the other forms tended to be motivated by push factors, including failure to realize a required return. In a review of conversion financing, Robinson (2000) reported that most conversions from private nonprofit to for-profit status are

130

Sloan

initiated by the nonprofit trustees and thus resemble friendly leveraged buyouts. At the same time, many nonprofit hospitals have created for-profit subsidiaries and continue to function as nonprofit organizations. Conover et al. (2001) provide a recent study of the determinants of hospital closings, mergers, and ownership changes. Hospitals that converted experienced on average a major decline in financial performance in the years immediately preceding conversion. In this sense, converting hospitals were not typically successful businesses prior to conversion. In the absence of a change in ownership, some other outcome, including closing, was often inevitable. What Is Being Bought and Sold? The financial transaction involves a purchase or lease price, and much more. The transaction may be parsed into two elements, the price and provisions of the transaction other than price.1 On price, the issue is whether or not acquirers pay a fair price for the facility, given discounted cash flows that are likely to accrue from the transaction. In the vast majority of transactions in the general economy, the underlying assumption is that buyers and sellers are sufficiently informed to allow the market to set the transaction price. In this context, there is a concern that sellers may not be sufficiently well informed and may obtain too little in return for relinquishing a major community asset. On the nonprice dimensions, there is a concern that acquirers will be driven by profit considerations, and provision of public goods by hospitals, such as provision of care to uninsured persons, public health programs, and education and research other than that sponsored by an outside public or private organization, will be reduced as a consequence. Much of the evidence on these two questions comes from case studies of ownership conversions. To date, the literature has provided no clear answers to the price question. As far as the nonprice question is concerned, the evidence is somewhat reassuring but preliminary. The price question has been addressed by two recent empirical studies. Conover and Sloan (2001) assessed rates of return on a sample of hospital purchases starting in the early 1980s. Wide variations in rates of returns were observed, with conversions to for-profit status exhibiting higher rates of return than other types of changes in hospital ownership. On average, rates of return were well above the cost of capital when hospitals converted to for-profit status, but somewhat closer to the cost of capital otherwise.

Hospital Ownership Conversions

131

In an earlier study, rates of return tended to be closer to the cost of capital, but that study was limited to ownership conversions that occurred in North and South Carolina and had a shorter postconversion follow-up period, meaning that more of the measured return was based on projected rather than actual cash flows (Sloan et al. 2000). Methodologies used in the two studies were similar. A weakness of both studies is that it was possible to make only crude adjustments for nonprice concessions granted by buyers to sellers. On the second question, it is necessary to glean bits of evidence from qualitative studies. Blumenthal and Weissman (2000) provided case studies of the sales of three teaching hospitals to investor-owned hospital chains, focusing in particular on the effects of the sales on the organizations' medical education missions. In all three, there were no adverse effects. The authors attributed the lack of an effect to three factors. First, the for-profit purchasers considered preservation of some unprofitable activities a cost of doing business at these institutions. The contracts of sale stipulated that specific resources be devoted to teaching, resource, and charity care. Second, private subsidization of medical education may not be that burdensome to the new owners, given external subsidies such as for graduate medical education. Third, at the time the study was written, all three institutions were doing well financially. If faced with financial stress, their missions could change. Cutler and Horwitz (2000) studied conversions of the Wesley Medical Center in Wichita, Kansas, and HealthOne in Denver, Colorado. Both were large teaching facilities. The first was purchased by Hospital Corporation of America in 1985. Two foundations were funded with proceeds from the sale. In the HealthOne transaction, the hospital entered into a joint venture with Columbia/HCA in 1995. After the transaction, the new HealthOne organization concentrated on graduate medical education, paying faculty and residents and administering medical education at its facilities. Cutler and Horwitz found improvements in financial performance after the conversion and the mission previous to ownership conversion was maintained. Improved financial performance came in part from the skill of the for-profits at increasing public sector reimbursements, not solely from efficiency gains. Outcomes following conversion are not uniformly favorable. Burns et al. (2000) described "the Allegheny system debacle." This is a case study involving the Allegheny Health, Education and Research Foundation, which consisted of a major teaching facility in western Pennsylvania and several affiliate hospitals throughout the state. The authors

132

Sloan

attributed failure in part to failure of external oversight mechanisms, including lack of performance of the organization's board, accountants, and auditors; bond-rating agencies; and state government. There was ambiguity regarding the powers of the state attorney general, state politics, and jurisdictional issues with federal bankruptcy court. Pennsylvania law is ambiguous regarding the attorney general's power over transactions with nonprofits. The sixteen case studies reported in Burns et al. (2001) revealed that, in all but one conversion, the financial status of the hospital improved after the conversion. There were funds from the sale or lease of the facility. New hospital owners invested in hospital plant and equipment, particularly in hospitals converting from private nonprofit to for-profit status, although in some cases, the investment after the conversion was not as great as hospital management had anticipated. Improvements in margins were achieved by cutting staff,2 improving purchasing practices, and consolidating services lines in a network approach. Changes in decision-making style depended mostly on whether the new organization was a multihospital system or a freestanding facility. In fact, the transition to being part of a larger system was on balance more important than was the specific change in ownership form. Those hospitals that became part of a chain lost some local autonomy in decision making. But there were differences in the strategic decisionmaking process among chains and even in treatment of individual facilities within a particular chain. Overall, the organization's general mission remained intact, as specified, for example, in the sales contract, while the methods for achieving its objectives changed. Anderson et al. (2001) studied changes in internal decision making after conversion in ownership status at the same set of converting hospitals as in Burns et al. (2001), but they also included a comparison with twenty-two nonconverting hospitals matched on location, ownership (prior to conversion), and bed size. They found that, relative to nonconverted hospitals, converted hospitals had greater levels of physician and nurse participation in hospital decision making; in the converted group, these health professionals had greater influence over final choices made. These agents' interconnections and interactions may have intensified as the organizations tried to cope with the changes brought about by the conversion. Alternatively, granting more influence to professionals may have been a compromise struck to overcome professionals' resistance to change. The study assessed participation at three to six years following ownership conversion.

Hospital Ownership Conversions

133

Thus, the change in participation was not only transitory. However, increased participation by health professionals is more closely related logically to provision of care to individual patients than to policies affecting the community as a whole, such as provision of care to the uninsured. Effects of Hospital Ownership Changes on Quality of Care Two recent studies have assessed differences in quality of care by hospital ownership status. One is a comparison of differences in quality by ownership. The other explicitly examined the effect of change in ownership on quality of care. Both studies relied on Medicare administrative records to gauge outcomes of care. At least to some extent, both studies account for methodological complexities of discerning effects of conversions discussed in the previous section. Both studies imply that conversion to for-profit status may lead to some reduction in quality of care as measured by mortality rates at various times after discharge from the hospital. McClellan and Staiger (2000) examined all Medicare hospital discharges for acute myocardial infarction for 1984-1994 and for ischemic heart disease for 1984-1991. The outcome measures were death within ninety days of admission and cardiac complications leading to readmission. Many of the details are beyond the scope of this review. One purpose of that study was to develop a hospital-specific measure of quality with a high signal-to-noise ratio. However, the findings on hospital quality measured in terms of patient survival, are highly pertinent here. First, when county-level fixed effects were included, the estimated mortality difference between N and F hospitals fell by roughly half, implying that for-profit hospitals tend to locate in geographic areas where hospital quality is not as high in general. This may be partly because for-profits often acquire facilities that are not doing well financially. This still left some amount of lower quality attributable to being a for-profit hospital. Second, in the three markets they considered in detail, for-profit hospitals did not have higher mortality rates. In one of these markets, a for-profit firm acquired a low-quality hospital, gauged in terms of mortality rates, and quality at that hospital subsequently improved. Overall, these results lend support to the conclusion that some of the results on quality differences reflect differential patterns by hospital

134

Sloan

ownership in the location of their facilities rather than in fundamental differences in hospital behavior. The authors concluded that: [O]n average, the performance of not-for-profit hospitals in treating elderly patients with heart disease appears to be slightly better than that of for-profit hospitals, even after accounting for systematic differences in hospital size, teaching status, urbanization, and patient demographic characteristics. This average difference appears to be increasing over time. However, this small average difference masks an enormous amount of variation in hospital quality within the for-profit and the private nonprofit categories. Our case-study results also suggest that for-profits may provide the impetus for quality improvements where, for various reasons, relatively poor quality of care is the norm. (p. Ill)

In a comment on the McClellan-Staiger study, Wolfram (2000) stressed that survival to ninety days after the admission date represents only one dimension of quality. While no informed person would seek admission at a hospital with a markedly higher mortality rate, other attributes such as the time that providers spend with patients, are also plausibly important. She also noted that McClellan and Staiger may not have adequately controlled for patient selection. For example, if more severely ill patients (in ways that the researcher cannot measure) seek care at private nonprofit hospitals, the true differential in quality between these institutions and for-profits will be higher than the difference the researchers measured. Several strategies deal with this issue, none of which is perfect. One is to include more explanatory variables, such as Wolfram's suggestion to include a binary variable for the presence of a trauma center.3 If private nonprofits are more likely to have trauma centers, they may attract the most vulnerable heart attack victims.4 Other approaches involve instrumental variables and difference-in-difference. In the end, more progress is likely to be made by supplementing statistical approaches with case studies that identify changes in processes of care that occur in hospitals with different ownership form. Picone et al. (forthcoming) also assessed health outcomes using Medicare data for the years 1984-1995. They studied death at thirty days, six months, and one year. The underlying hypothesis was that hospitals converting to for-profit ownership raise profitability after acquiring the facilities in part by reducing dimensions of quality not readily observed by patients or their physician agents and by raising prices. Hospital-specific mortality rates after discharge from the hospital are one important dimension of such hard-to-observe quality. The authors found that one to two years subsequent to conversion to

Hospital Ownership Conversions

135

for-profit status, the mortality rate of patients at one year following admission rose markedly. At the same time, hospital staffing decreased. A similar decline in quality was not observed after hospitals switched from for-profit status to government or to private nonprofit ownership status. At three or more years following conversion to for-profit ownership, the decline in quality was much lower (relative to the quality that prevailed at the hospital at five years before the conversion occurred). A plausible interpretation consistent with the data is that the new for-profit owners (or their managers) experienced adverse effects after the staffing cuts and reversed course at three years or so following the date of acquisition. By adding staff, some improvements in quality were then realized. Effects of Hospital Ownership Changes on Provision of Public Goods A major policy concern is that ownership conversions, especially to the for-profit form, will result in decreased provision of public goods. Most frequently mentioned among such public goods is provision of care to the uninsured, sometimes cast in terms of provision of uncompensated care. The empirical evidence on this score is mixed. Comparing provision of uncompensated care across ownership types, one is struck by the similarity between shares of dollar amounts of uncompensated care relative to hospital revenue provided by N and F hospitals. Government hospital uncompensated shares, not surprisingly, tend to be higher than for N and F hospitals (Sloan 2000). A recent study, which focused on ownership conversions, calls the conclusion that the main distinction in provision of uncompensated care is between private and public hospitals into question. Using unpublished and confidential data on revenue from the American Hospital Association, Thorpe et al. (2000) studied the effects of conversion from N to F status on hospital provision of uncompensated care. They measured uncompensated care as bad debt and charity care charges deflated by each hospital's cost-to-charge ratio for 1991-1997. To account for variations in hospital capacity and inflation, they divided this amount by total expenses for the hospital. They included several explanatory variables in their analysis, most notably hospital fixed effects. They found little effect of conversion to for-profit status on Medicaid patient loads. However, uncompensated care fell after conversion from N to F status, falling from 5.3 percent to 4.7 percent of hospital revenue on average. The 0.6 percentage-point reduction in

136

Sloan

uncompensated care amounted to about $400,000 less being spent per converted hospital on uncompensated care. For public hospitals converting to for-profit status, the decrease was greater, from 5.2 to 2.7 percent, or $800,000, per hospital. The authors expressed the amounts in terms of admissions lost even though they provided no indication whether the decrease occurred on the inpatient or outpatient sides of the hospital or hi some combination thereof. Thorpe et al. concluded that: Of concern, however, is our findings that the provision of uncompensated care is reduced when hospitals convert to for-profit status. Of particular concern is the large reduction in uncompensated care observed among public to for-profit hospital conversions. Because the bulk of these conversions occurred among smaller, rural public hospitals, such conversions could limit access to hospital care among the uninsured, (p. 192)

The implication is that conversions cause reductions in uncompensated care and that these reductions could not be accounted for by inward shifts in demand largely exogenous to the hospitals involved in the conversions. The authors noted that their result for nonprofit to for-profit conversions had not been found in previous studies (see, for example, Desai et al. 2000). I shall return to this issue in the next section when I discuss my own empirical analysis. Several recent studies have focused not on the effect of ownership change on the provision of public goods and more generally on product mix, but more specifically on the effects of competition from for-profit hospitals on the behavior of nonprofit and public hospitals.5 The Disproportionate Share (DSH) Program was implemented nationally to provide a greater financial incentive for hospitals to deliver care to the poor. Subject to this federal law, each state could design its own DSH program. Duggan (2000a, 2000b) studied the change in financial incentives created by California's variant of the DSH program on the propensity of hospitals to treat Medicaid recipients. This DSH program provided an explicit financial incentive for hospitals to admit Medicaid patients. It increased revenues to hospitals for which low-income patients constituted more than 25 percent of their patients. Hospitals above this threshold experienced a revenue increase, and those below this threshold had an incentive to increase their low-patient shares to this level. Duggan (2000b) found an appreciable difference in response between public and private hospitals, regardless of whether they were N

Hospital Ownership Conversions

137

or F. The public hospitals faced a "soft budget" constraint; that is, as their revenues from DSH increased, their sponsors lowered their subsidies accordingly. By contrast, both N- and F -owned facilities could accumulate wealth from this new revenue source. He concluded that the greatest difference in response to the DSH incentive was between public and private rather than between N and F hospitals. In the second study, Duggan (2000a) found that DSH resulted in a shift of Medicaid patients from public to private hospitals. The magnitude of the shift was directly related to the market share of for-profit hospitals in the county. In particular, the response of N hospitals to the DSH incentive was greater when they faced more competition from the for-profit sector. The implication is that N hospitals behave more like profit-maximizers when faced with the market discipline of the for-profits. Various tests that Duggan performed rejected alternative explanations for his finding (for example, quality of care in the public facilities). An examination of the effects of competition from F hospitals on the board composition of N hospitals revealed that N hospital boards had larger shares of physicians when they faced competition from the F hospitals. Duggan reported that F boards contained large numbers of physician members.6 In this sense, N boards in F-influenced areas may put physicians on the boards as their mission becomes more profit oriented. However, other interpretations seem at least as likely. For one, placing physicians on the board may be a competitive response by N hospitals to retain their medical staffs. Silverman and Skinner (2001) assessed the effects of hospital competition on mission, but from a different perspective. Since implementation of the Medicare Prospective Payment System (PPS), hospitals have known that the pattern of reporting of diagnoses and procedures can affect the diagnosis-related group (DRG) assigned to the case and hence the amount of revenue received from Medicare. Upcoding involves rearranging reports of diagnoses and/or procedures, with the result that the patient falls within a higher-priced DRG. Some upcoding may be perfectly legitimate. In some cases, it may improve accuracy of reporting. But given the financial incentive to upcode, there is a large gray area. More profit-oriented hospitals may be more willing to take advantage of the incentive to increase revenue from Medicare. Even for those hospitals that are not fully profit oriented, pressures from competition may force them to act in this way. The authors limited their analysis to hospital admissions involving pneumonia and respiratory infections. Between 1989 and 1996, the number of the most

138

Sloan

expensive DRG in this diagnosis family rose by 10 percentage points among stable N hospitals, 23 percent among stable F hospitals, but 37 percentage points among N hospitals that converted to F status. Silverman and Skinner obtained two major findings. First, holding many other factors constant, for-profit hospitals were more likely to upcode in this diagnostic category than most nonprofit and government hospitals. Second, the authors found evidence that upcoding among N hospitals was much more likely when they faced greater competition from F hospitals. By contrast, the upcoding behavior of the for-profits was not affected by the presence of nonprofits.7 During the latter part of the 1990s, Medicare became more aggressive in monitoring hospital billing practices, with the consequence that some hospitals made large payments to compensate for shortcomings in past billing practices.8 III. Evidence from a National Sample of Hospital Discharges Objectives of the Analysis To assess further the effect of hospital ownership conversions on quality, patient mix, especially willingness to treat publicly insured and uninsured persons, and upcoding, I assessed hospital discharge data from the Healthcare Cost and Utilization Project (HCUP) Nationwide Inpatient Sample (NIS). Results from this analysis are reported in this study for the first time. This data set offers many advantages. There is a very large number of observations, and the data come from many states and are available for several years. The hospital, but of course, not a patient identifier, is provided. The main disadvantage of the data is that no information is available on patients after they were discharged from the hospital. Also, there is no information on hospital outpatient care. Data and Methods The NIS is a compilation of data from state hospital discharge datacollection systems made available through the U.S. Agency for Healthcare Research and Quality. I limited the analysis to hospital admissions occurring during the years 1988-1996. Data from the following states were included for some or all of the observational period: Arizona, California, Colorado, Connecticut, Florida, Hawaii, Illinois, Iowa, Kansas,

Hospital Ownership Conversions

139

Maryland, Massachusetts, Missouri, New Jersey, New York, Oregon, Pennsylvania, South Carolina, Tennessee, Utah, Washington, and Wisconsin. Each year, the NIS provides discharge abstract data from about 5 to 7.1 hospital stays from over 900 hospitals per year. The NIS is designed to approximate a 20-percent sample of U.S. community hospitals. The NIS is not designed to be representative of community hospitals in terms of ownership. Of the states included in the NIS, California and Florida had the most admissions, by far, to for-profit hospitals. Tennessee has a high for-profit market share but is a much smaller state and was included in the NIS only for part of the observational period. Texas, another state with a high for-profit market share, was not included in the NIS. Separate analyses were conducted on admissions of (1) persons over the age of 65 at admission—hereafter called the "Medicare sample" and (2) births and admissions of persons who were between the ages of 1 and 64 on the admission date—the non-Medicare sample (although a minority of persons in the under-age-65 group also had Medicare as the primary payer). For the Medicare analysis, the sample was limited to persons who were admitted for one of five primary diagnoses: stroke, hip fracture, coronary heart disease, congestive heart failure, and pneumonia. To facilitate analysis, the NIS 10 percent sample was used with one exception. In the analysis of discharges of persons who were between the ages of 1 and 64 on the admission date, I used a 25 percent sample of NIS's 10 percent patient sample. The Medicare and non-Medicare samples were each divided into two subsamples. The first consisted of admissions to hospitals that were under government or private nonprofit ownership prior to conversion and changed to for-profit ownership or remained under government or private nonprofit ownership during the observational period, 1988-1996 (the GN sample). The second included admissions to hospitals that were under for-profit ownership but converted to GN status during the observational period, as well as those that remained for-profit (the F sample). Six hypotheses about effects of ownership conversions from GN to F and F to GN status were tested. Conversions from F to GN ownership were hypothesized to have the opposite effect of conversions from GN to F. By employing a two-sided test, I was able to distinguish between the effects of conversion and those attributable to a change in ownership. In combining the G and N categories, I focused on changes in be-

140

Sloan

havior to and from the for-profit ownership form rather than on the private versus public ownership distinction. The six hypotheses are listed below: 1. Quality of care falls following conversion from GN to F ownership. This may occur because profit-seeking hospitals seek to achieve higher profitability by cutting hospital inputs, such as personnel. 2. Quality may be increased, but only when it is profitable to do so, that is, when higher quality results in an additional payment sufficient to cover the additional marginal cost of the higher quality level. 3. Following conversion from GN to F status (when subject to a fixed payment per case, as under the Medicare Prospective Payment System), the hospital becomes more aggressive in reducing length of stay. Again, the motive for reducing stays is to increase profitability. 4. Increasing transfers to nursing homes postdischarge is one manifestation of a strategy for reducing the length of stay. 5. To increase payment from Medicare, the hospital upcodes diagnoses more frequently following a conversion from GN to F status. This is an extension of Silverman and Skinner's (2001) research, but for a different set of diagnoses. 6. Following conversion from GN to F ownership, the hospital becomes less likely to admit publicly insured and uninsured patients. This analysis used only the non-Medicare sample and sought to replicate the finding by Thorpe et al. (2000) that conversion to for-profit ownership led to reductions in hospital provision of uncompensated care. For the analysis of the admissions of persons aged 65 and over, the dependent variables were inpatient mortality; extended length of stay; length of stay; pneumonia complication; destination at discharge— home (omitted reference group), nursing home, other hospital, or death; and upcoding of diagnoses. Extended length of stay was defined as a stay two standard deviations above the mean stay for the primary diagnosis and the year in which the admission occurred. The pneumonia complication was taken from the list of secondary diagnoses provided on the hospital discharge abstract. In this analysis, elderly persons admitted with a primary diagnosis of pneumonia were excluded. Upcoding was tested by two analyses. In the first, the dependent variable was 1 if the primary diagnosis was listed as a transient

Hospital Ownership Conversions

141

ischemic attack (TIA) and 0 if the primary diagnosis was listed as a stroke. Medicare pays more for strokes than TIAs. There is some discretion in classifying patients between these two diagnoses. The second was the DRG weight assigned to the case, limiting the sample to the five primary diagnoses listed above. A sample of births was analyzed to test the sixth and second hypotheses. The dependent variables were patient did not have private insurance, patient stayed in the hospital for less than two days, and patient had a vaginal birth as opposed to a cesarean section—the underlying presumption being that C-sections are more profitable on average than are vaginal births. With the sample of births, the dependent variable was payment source other than private insurance versus private insurance. The other category included Medicaid, Medicare or other government insurance (such as Veterans Administration, Champus), self-pay or no charge (suggesting no health insurance coverage), and private insurance (the omitted reference group). With the sample of persons with any diagnosis or principal procedure who were between the ages of 1 and 64 at the admission date, the dependent variables were each of the payer categories for public payers and self-pay/no charge with privately insured individuals, the omitted reference group. With the exceptions noted below, four alternative specifications were employed. The methodology for the Medicare analysis is explained in detail here. Specifications for the non-Medicare were similar. Each successive specification added a set of explanatory variables, retaining the explanatory variables from the previous specification. The first specification included explanatory variables identifying admissions after ownership conversion occurred, patient characteristics, market characteristics, and binary variables for the year of admission. The patient characteristics were age; race; gender; source of admission—emergency room, nursing home, other hospital versus home; binary variables for the five primary diagnoses; DRG weight; and the DxCG score. The DxCG is a case mix measure that accounts for the patient's secondary diagnoses (see www.dxcg.com). The market characteristics were a Herfindahl index based on bed shares, the fraction of the population enrolled in health maintenance organizations, population density (population per square mile), hospital beds per 1,000 population, and per-capita income. For hospitals located in standard metropolitan statistical areas (SMSAs), the market area was assumed to be the SMSA. For hospitals located outside SMSAs, the market area was the county.

142

Sloan

In the second specification, I added a conversion fixed effect binary variable. The conversion fixed effects identified admissions to hospitals involved in a conversion from G or N to F, or F to G or N during the observational period. In the third specification, I added explanatory variables for hospital characteristics: bedsize; the number of resident physicians per bed, a measure of commitment to medical education; the hospital's operating margin; debt-to-asset ratio; and occupancy rate—all defined for the year in which the admission occurred. For hospitals with a negative operating margin, the operating margin was set equal to 0 and an additional binary variable, "no profit," was set to 1. Likewise, if the debt-to-asset ratio exceeded 1, the ratio was set to 1 and a binary variable was included to identify such cases. Although the hospital characteristics are possibly endogenous, there is an argument for including them. In their absence, the ownership and ownership change variables may represent other hospital characteristics, including financial distress that may be the true causal determinants of the dependent variables. By considering alternative specifications, it is possible to gauge the sensitivity of findings to inclusion/exclusion of these explanatory variables. Finally, the fourth specification added area fixed effects. These were binary variables for the SMSA in which the hospital was located and for non-SMSA hospitals, a binary variable measuring the hospital's state. When the dependent variable was more than one mutually exclusive alternative, I used multinomial logit analysis and did not include the fourth specification. With area fixed effects, there were too many parameters to estimate. IV. Results: Medicare Analysis Sample Characteristics The main GN sample consisted of 419,000 hospital discharges from 1,215 hospitals (table 5.1). Of these, over 16,000 hospital discharges were from forty-nine hospitals that changed from G or N to F ownership status. Slightly over 6,000 discharges were observed from thirtyfive hospitals after the ownership conversion occurred. The main F sample consisted of 56,000 discharges from 165 hospitals. Among these, thirty-two hospitals experienced a conversion from F to N or G status during the observational period. Data were available from

Hospital Ownership Conversions

143

Table 5.1 Samples3 Sample screen

Patients

Hospitals

GN sample GN to F subsample GN to F after

418,831 16,354 6,050

1,215 49 35

56,231 6,500 2,331

165 32 20

F sample F to GN sample F to GN after a

Does not include observations drawn for analysis of coding of stroke versus transient ischemic attack (TIA).

twenty hospitals converting from F to N or G ownership for the period after the conversion occurred. I constructed a counterfactual sample. This sample of hospital discharges was matched to the conversion sample with respect to base ownership (GN or F). An artificial conversion year was randomly assigned to the admission. The frequency distribution of conversion years in the counterfactual sample matched the frequency distribution of actual conversions. Given the large number of hospital discharges, there were many statistically significant differences between hospitals that did not convert and those that did convert, as well as for the pre versus postconversion comparisons (table 5.2). The differences, however, were often small. For the GN sample, these are the most noteworthy differences. Hospitals converting from GN to F status experienced increases in the proportion of nonwhites that they admitted. Admissions through the emergency room fell, as did the receipt of patients from other hospitals. For converting hospitals, the mean DRG weight fell after conversion. A drop of 0.2 in the mean DRG weight for the five primary diagnoses is substantial. By contrast, in the counterfactual comparison group, the DRG weight rose in the after group in contrast to the before group. The mean DxCG score rose from 2.1 to 2.2 before versus after conversion to F status. For the comparison group, the mean score rose from 2.0 to 2.2. Medicare payment is based on the DRG weight but not the DxCG score. This comparison suggests something less than a massive change in upcoding for purposes of obtaining higher Medicare payments postconversion to F ownership.

Table 5.2 Mean values of explanatory variables: converting and nonconverting hospitals: Medicare samples

Whole sample

All All nonconverting converting

Conversion sample

Counterfactual sample p

Nonconverting: before

Nonconverting: after

Converting: before

Converting: after

10,304

6,050

77.616 0.879 0.461 0.605 0.027 0.041 0.009 0.016 0.090 0.120 0.413 0.224 0.153 1.567 2.099

78.245 0.746 0.440 0.539 0.037 0.018 0.005 0.010 0.084 0.097 0.313 0.275 0.232 1.367 2.235

<.0001 29680.220 0.524 302.886 <.001 0.014 <.001 0.015 <.001 0.695 <.001 58.879

24253.720 127.584 0.030 0.085 0.555 51.803

p

P

GN sample Number of observations Patient Age White Male Admitted from ER Admitted from nursing home Admitted from other hospital Diagnosis of Alzheimer's Diagnosis of other dementia Diagnosis of hip fracture Diagnosis of stroke Diagnosis of coronary heart disease Diagnosis of congestive heart disease Diagnosis of pneumonia DRG weight DxCG score Hospital Income in zip code area ('000) Bed size Residents/beds Operating margin Debt-capital ratio Occupancy rate (%)

402,477

16,354

77.356 0.886 0.460 0.572 0.031 0.053 0.005 0.014 0.088 0.103 0.392 0.244 0.172 1.558 2.136

77.849 0.817 0.453 0.580 0.031 0.032 0.008 0.014 0.088 0.112 0.376 0.243 0.182 1.493 2.149

30444.360 337.657 0.061 0.022 0.461 66.274

27672.740 237.889 0.020 0.041 0.643 55.002

216,277

186,371

<.0001 <.001 0.118 0.042 0.956 <.001 0.000 0.373 0.863 0.001 <.001 0.645 0.001 <.001 0.333

77.250 0.893 0.458 0.560 0.030 0.047 0.007 0.020 0.088 0.122 0.388 0.237 0.164 1.512 2.043

77.479 0.880 0.462 0.587 0.032 0.059 0.003 0.008 0.089 0.082 0.397 0.252 0.180 1.610 2.245

<.0001 <.001 <.001 <.001 <.001 <.001

30191.710 337.425 0.053 0.022 0.524 67.110

30737.820 337.919 0.070 0.023 0.387 65.303

<.001 <.001 0.011 <.001 <.001 <.001 <.001 <.001 0.075 <.001 <.001 <.001 <.001 <.001 <.001

<.001 <.001 0.011 <.001 <.001 <.001 0.001 0.002 0.167 <.001 <.001 <.001 <.001 <.001 <.001

<.0001 <.001 <.001 <.001 <.001 <.001

Table 5.2 (continued) Whole sample

All All nonconverting converting Number of observations Market Herfindahl index HMO share Population density ('000/sq m) Hospital beds/COOO) pop. Income in zip code area ('000)

402,477 0.230 0.187 0.844 0.390 30.444

P

16,354 0.204 0.183 0.684 0.378 27.673

Conversion sample

Counterfactual sample Nonconverting: Nonconverting: after before 216,277

<.001 0.005 <.001 <.001 <.001

0.235 0.168 0.824 0.400 30.192

P

10,304

186,371 0.225 0.209 0.867 0.377 30.738

Converting: Converting: after before

<.001 <-001 <.001 <.001 <.001

0.177 0.162 0.670 0.387 29.680

P

6,050 0.252 0.218 0.709 0.363 24.252

<.001 <.001 <.001 <.001 <.001

F sample Number of observations Patient Age White Male Admitted from ER Admitted from nursing home Admitted from other hospital Diagnosis of Alzheimer's Diagnosis of other dementia Diagnosis of hip fracture Diagnosis of stroke Diagnosis of coronary heart disease Diagnosis of congestive heart disease Diagnosis of pneumonia DRG weight DxCG score

49,731 77.581 0.880 0.468 0.654 0.022 0.044 0.014 0.036 0.089 0.100 0.393 0.249 0.169 1.474 2.325

29,218

6,500 78.008 0.872 0.440 0.635 0.020 0.036 0.004 0.010 0.104 0.110 0.331 0.256 0.200 1.270 2.161

<.001 0.184 <.001 0.003 0.268 0.001 0.466 0.424 0.001 0.016 <.001 0.219 <.001 <.001 <.001

77.297 0.884 0.470 0.654 0.021 0.036 0.007 0.016 0.086 0.121 0.398 0.236 0.159 1.440 2.200

4,169

20,513 77.985 0.877 0.466 0.654 0.023 0.055 0.002 0.004 0.094 0.070 0.385 0.268 0.183 1.522 2.502

<.001 0.090 0.337 0.881 0.281 <.001 <.001 <.001 0.002 <.001 0.004 <.001 <001 <.001 <.001

77.861 0.863 0.451 0.614 0.021 0.034 0.006 0.013 0.101 0.120 0.342 0.243 0.193 1.284 2.106

2,331 78.272 0.884 0.421 0.673 0.018 0.038 0.000 0.004 0.107 0.092 0.310 0.279 0.211 1.246 2.260

0.045 0.048 0.020 <.001 0.317 0.397 <.001 <.001 0.466 0.001 0.007 0.002 0.082 0.111 <.001

Table 5.2 (continued) Whole sample All All nonconverting converting

Hospital Income in zip code area ('000) Bed size Residents /beds Operating margin Debt-capital ratio Occupancy rate (%) Market Herfindahl index HMO share Population density ('000/sq m) Hospital area('000) pop. Income in zip code area ('000)

27741.850 217.581 0.002 0.080 0.521 53.105

32848.850 155.585 0.002 0.039 0.659 45.105

0.180 0.155 0.823 0.414 27.742

0.155 0.141 1.192 0.351 32.849

Conversion sample

Counterfactual sample P

<.0001 <.001 0.004 <.001 <.001 <.001 <.001 <.001 <.001 <.001 <.001

Nonconverting: Nonconverting: after before

P

Converting: Converting: after before

27521.830 221.990 0.001 0.064 0.607 53.678

28054.770 211.460 0.003 0.102 0.399 52.290

<.0001 <.001 <.001 <.001 <.001 <.001

32228.350 170.079 0.002 0.037 0.721 44.431

33958.600 128.925 0.004 0.041 0.549 48.743

0.182 0.124 0.776 0.427 27.522

0.178 0.199 0.890 0.396 28.055

0.056 <.001 <.001 <.001 <.001

0.129 0.116 0.890 0.353 32.228

0.200 0.187 1.733 0.347 33.959

P

<.0001 <.001 <.001 0.059 <.001 <.001 <.001 <.001 <.001 0.007 <.001

Hospital Ownership Conversions

147

In the counterfactual group, mean bed size remained about constant. Among converting hospitals, however, mean bed size fell by more than 50 percent after conversion. The number of residents per bed rose among converting hospitals, but this change was fully attributable to the reduction in beds. Operating margins in converting hospitals rose dramatically. Before conversion, the mean margin per discharge in converting hospitals was 0.015. After conversion, the mean was 0.085. By contrast, in the counterfactual group, operating margins remained about constant. The debt-to-assets ratio fell for both converting and nonconverting hospitals, but much more for those hospitals that did not convert. The mean occupancy rate fell from 57 to 52 percent for converting hospitals, in spite of the substantial reduction in bed size. For nonconverting hospitals, however, mean occupancy decreased only from 67 to 65 percent. Clearly, low occupancy rates were far more characteristic of hospitals converting from GN to F than among hospitals that did not convert. For converting hospitals, the Herfindahl index rose, suggesting a decline in competition. By contrast, for hospitals that did not convert, the Herfindahl index was unchanged. Personal per-capita income fell on average in the zip code areas in which the converting hospitals were located. But there was no change in income among nonconverting hospitals. Many of the patterns were similar in the F sample. For example, the mean DRG weight declined slightly after conversion to GN status. By contrast, for hospitals in the counterfactual group, the mean DRG rose after the conversion. The major difference was in the operating margin. Margins increased much more among nonconverting than among converting hospitals. As seen above, for GN hospitals, margins increased appreciably for hospitals converting to F ownership. The mean in-hospital mortality rate in the GN sample hospitalized in nonconverting hospitals was 8.1 percent (table 5.3). For those hospitalized in converting hospitals, the mean mortality rate was 7.9 percent. Even with this large sample, this difference was not statistically significant at conventional levels. Although mortality fell from 8.2 percent before to 7.6 percent after conversion in converting hospitals, this difference was not statistically significant at conventional levels (p = 0.19). In the counterfactual sample, the decline was apparently larger, from 8.7 percent in the counterfactual before group to 7.5 percent in the after group (p < 0.0001). The decrease reflects the secular decrease in inpa-

Table 5.3 Mean values of dependent variables: converting and nonconverting hospitals Counterfactual sample

Whole sample Nonconverting

Converting

p

Before

After

Conversion sample

p

Before

After

P

GN sample Number of observations Mortality Extended stay Length of stay

402,477 0.081 0.020 8.124

16,354 0.079 0.015 7.982

0.375 <.0001 0.089

216,106 0.087 0.019 8.650

186,371 0.075 0.020 7.511

<.0001 0.001 <.0001

10,304 0.082 0.017 8.538

6,050 0.076 0.012 7.036

0.187 0.009 <.0001

Alive at discharge Number of observations Extended stay

368,827 0.024

15,053 0.021

0.004

198,805 0.025

172,022 0.023

<.0001

9,463 0.025

5,590 0.014

<.001

Non-pneumonia primary diagnosis patients Number of observations Pneumonia complications

333,401 0.042

13,378 0.042

0.847

180,531 0.042

152,870 0.042

0.081

8,731 0.038

4,647 0.050

0.001

Non-pneumonia primary diagnosis patients: alive at discharge Number of observations Pneumonia complications

307,866 0.034

12,438 0.034

0.803

165,727 0.034

142,139 0.035

0.075

8,089 0.030

4,349 0.043

0.001

Table 5.3 (continued) Counterfactual sample

Whole sample Nonconverting

Converting

p

Before

After

Conversion sample p

Before

After

P

F sample Number of observations Mortality Extended stay Length of stay

49,731 0.075 0.034 7.018

6,500 0.083 0.035 7,372

0.026 0.701 0.256

29,218 0.080 0.034 7.489

20,513 0.068 0.034 6.346

<.0001 0.992 <.0001

4,169 0.087 0.026 6.982

2,331 0.076 0.051 8.071

0.102 <.001 0.198

Alive at discharge Number of observations Extended stay

45,988 0.037

5,953 0.034

0.156

26,851 0.038

19,137 0.036

0.134

3,799 0.026

2,154 0.047

<.001

Non-pneumonia primary diagnosis patients Number of observations Pneumonia complications

41,313 0.037

5,201 0.044

0.011

24,542 0.0.35

16,771 0.040

0.006

3,363 0.045

1,838 0.044

0.829

Non-pneumonia primary diagnosis patients: alive at discharge Number of observations Pneumonia complications

38,530 0.030

4,808 0.035

0.064

22,770 0.028

15,760 0.034

0.001

3,092 0.034

1,716 0.038

0.485

150

Sloan

tient mortality, which apparently was not as favorable among hospitals that changed from GN to F ownership status. The rate of pneumonia complications rose in hospitals converting from GN to F after conversion. The increase was from 3.8 to 5.0 percent (p = 0.001). In the counterfactual sample, there was no change. Rates of extended stays, as did mean length of stay, declined much more in hospitals that were involved in ownership conversions than in those not involved in a conversion, implying that hospitals converting from GN to F made a substantial effort to reduce the very long lengths of stay. Mean length of stay declined in both converting and nonconverting hospitals, but more so in the former. For the F sample, mortality rates declined by the same absolute amount for those hospitals that converted versus the counterfactual sample, but those that converted had a higher mortality rate in the before period. Pneumonia complication rates rose among those hospitals that remained for-profit, but these rates did not rise among those that converted from for-profit to either government or private nonprofit status. The rate of extended stays rose dramatically among hospitals in the converting group, but remained constant among those that did not convert, a very different pattern from hospitals in the GN sample. Thus, overall, judging on the basis of mean values alone, a couple of negative features emerge in the GN-to-F case. The in-hospital mortality experience was not as good among hospitals that converted from GN to F status. Furthermore, pneumonia complication rates rose among those hospitals that converted from GN to F and among those for-profit hospitals that did not convert. Length of stay was related to ownership, with for-profit hospitals achieving lower lengths of stay overall. Effects of Ownership Type Conversions on Inpatient Mortality, Pneumonia Complications, and Extended Length of Stay The multivariate analysis of ownership type conversion effects begins with effects on inpatient mortality, and pneumonia complications during the stay occurring to patients admitted for stroke, hip fracture, coronary heart disease, and congestive heart failure (table 5.4). In the first specification in the GN sample, inpatient mortality fell after conversion to F ownership. The decline was 0.8 percent on average, or about 10 percent relative to mean mortality. (The numbers in the table in brackets are marginal effects—changes in the probability of an outcome for a

Table 5.4 Effects of ownership conversions on mortality, pneumonia complications, and extended length of stay3 GN sample Mortality

Specification 1 After conversion

Specification 2 After conversion

Pneumonia complication

F sample Extended stay

Full sample

Alive

Full sample

Alive

-0.155b (0.054) [-0.008]

0.126d (0.075) [0.003]

0.191C (0.082) [0.005]

-0.521b (0.141) [-0.004]

-0.472b (0.132) [-0.005]

-0.014 (0.066) [0.001]

0.260b 0.095) [0.007]

0.308b (0.106) [0.008]

-0.420b (0.163) [-0.003]

-0.144b (0.039) [-0.007]

-0.138C (0.060) [-0.003]

-0.120d (0.069) [-0.002]

-0.018 (0.040 [-0.001] -0.126b (0.040) -[-0.007]

0.211C (0.097) [-0.006] -0.188b (0.061) [-0.004]

0.259C (0.108) [0.006] -0.175C (0.070) [-0.003]

Mortality

Pneumonia complication

Extended stay

Full sample

Alive

Full sample

Alive

0.222b (0.090) [0.012]

0.148 (0.138) [0.003]

0.145 (0.151) [0.003]

0.342C (0.135) [0.007]

0.244d (0.140) [0.006]

-0.434b (0.148) [-0.005]

0.061 (0.108) [0.003]

-0.081 (0.164 [-0.002]

0.005 (0.184) [0.001]

0.443b (0.168) [0.010]

0.423C (0.173) [0.011]

-0.104 (0.083) [-0.001]

-0.039 -(0.039) [0.001]

0.171b (0.063) [0.009]

0.244C (0.097) [0.005]

0.148 (0.113) [0.003]

-0.109 (0.111) [-0.001]

-0.191d (0.113) [-0.004]

-0.329C (0.166) [-0.003] -0.025 (0.086) [-0.0002]

-0.347C (0.152) [-0.004] 0.037 (0.073) [0.001]

0.064 (0.110) [0.003] 0.164C (0.066) [0.009]

-0.077 (0.150) [-0.002] -0.113 (0.167) [-0.003]

0.047 (0.169) [0.0002] -0.023 (0.187) [-0.0004]

0.440b (0.167) [0.010] -0.026 (0.118) [-0.001]

0.409C (0.173) [0.010] -0.114 (0.119) [-0.002]

Fixed effect

Specification 3 After conversion

Fixed effect

Table 5.4 (continued) GN sample Mortality

Specification 4 After conversion

Pneumonia complication

F sample Extended stay

Full sample

Alive

Full sample

Alive

0.029 (0.070) [0.002]

0.196d (0.101) [0.005]

0.264C (0.112) [0.006]

-0.310d (0.172) [-0.002]

-0.337C (0.157) [-0.003]

-0.109C (0.045) [-0.006] 417,850

-0.096 (0.068) [-0.002] 346,765

0.107 (0.078) [-0.002] 320,292

0.119 (0.098) [0.001] 418,773

0.151d (0.083) [0.002] 383,864

Mortality

Pneumonia complication

Extended stay

Full sample

Alive

Full sample

Alive

-0.031 (0.127) [-0.002]

-0.115 (0.195 [-0.002]

0.079 (0.217) [0.001]

0.533b (0.199) [0.012]

0.367d (0.206) [0.009]

0.170C (0.079) [0.009] 56,864

0.126 (0.123) [0.003] 46,514

-0.070 (0.145) [-0.001] 43,338

0.041 (0.143) [0.007] 56,217

-0.024 (0.144) [-0.001] 51,941

Fixed effect

N

Specification 1 includes patient's characteristics and market characteristics and year dummies. Specification 2 includes specification 1 plus hospital conversion fixed effect (shown). Specification 3 includes specification 2 plus hospital characteristics. Specification 4 includes specification 3 plus area fixed effects. Standard errors in parentheses and marginal effects in brackets. Standard errors corrected for heteroscedasticity. b Significant at 1% level (two-tail test). c Significant at 5% level (two-tail test). d Significant at 10% level (two-tail test).

Hospital Ownership Conversions

153

change in binary or a unit change in explanatory variables for continuous variables.9) In the remaining specifications, however, there was no change in inpatient mortality after conversion to F ownership. The results for the conversion fixed effect in specifications 2 to 4 imply that hospitals that converted from G or N to F ownership tended to have lower mortality rates. That is, if anything, the new owners selected GN hospitals with relatively good mortality records. In the F sample, the first specification implies a mortality increase after conversion to G or N status with an associated marginal effect of 0.012 (p — 0.01). However, the after conversion parameter becomes statistically insignificant in the other, more complete specifications. In general, the results on the key parameters of interest—the binary variables for after conversion and the conversion fixed effect—are quite sensitive to changes in equation specification. Thus, again, ownership conversion had no effect on inpatient mortality. The proportion of patients with very long stays declined in the G and N to F case and rose for conversions in the opposite direction, supporting the descriptive results presented above in table 5.3. Pneumonia complications rose after conversion among patients admitted to hospitals that converted to F ownership. The reverse occurred among patients admitted to hospitals that converted from F ownership. This conclusion holds for the full sample of Medicare patients in the four primary diagnosis categories (excluding those admitted with a primary diagnosis of pneumonia) and for a sample that excludes such patients who died during their stays. To conserve space and permit focus on the key parameter estimates and associated marginal effects, table 5.4 does not include most of the parameter estimates in the model. Table 5.5 shows complete specifications from the GN sample for four dependent variables shown previously, but the table does not show area fixed effects. Almost all the parameter estimates on the patient variables are statistically significant at conventional levels and have plausible signs. Results for the hospital and market variables are somewhat more mixed. The measures of financial distress show no consistent effects on outcomes. For example, hospitals with an operating loss (no profit) experienced higher rates of pneumonia complications and a higher mean length of stay, but there were no effects on either inpatient mortality or the rate of extended stays. Death rates were higher in larger hospitals but lower in major teaching hospitals. The result for hospital size plausibly reflects case mix not otherwise measured. The coefficient on ex-

Table 5.5 GN sample: mortality, pneumonia complication, extended stay, and length of stay: specification 4a Mortality Coefficient

Standard errorb

Conversion GN to F after GN to F fixed effect

-0.109 0.029

(0.045)d (0.070)

Patient Age White Male Admitted from ER Admitted from nursing home Admitted from other hospital Diagnosis of Alzheimer's Diagnosis of other dementia Diagnosis of stroke Diagnosis of coronary heart disease Diagnosis of congestive heart disease Diagnosis of pneumonia DRG weight DxCG score

0.047 0.106 0.072 0.342 0.443 0.285 0.348 -0.137 1.935 0.827 1.305 1.402 0.083 0.424

(0.001)c (0.027)c (0.013)c (0.016)c (0.034)e (0.032)c (0.052)c (0.034)c (0.031 )c (0.030)c (0.030)c (0.031)c (0.004)c (0.003)c

Pneumonia complication

Extended stay

Length of stay

Standard errorb

Coefficient

Standard errorb

Coefficient

Standard errorb

0.096 0.196

(0.068) (0.101)e

0.019 -0.310

(0.098) (0.172)e

0.088 -0.127

(0.099) (0.124)

0.036 -0.015 0.307 0.268 0.535 0.147 0.182 0.184 1.087 -0.428 0.913 — 0.063 0.533

(0.001)c (0.039) (0.019)c (0.024)c (0.051)c (0.047)c (0.081 )d (0.046)c (0.034)c (0.034)c (0.032)c — (0.005)c (0.005)c

0.016 -0.205 -0.170 0.040 -0.039 0.542 0.017 0.274 -0.145 -0.215 0.052 0.241 0.337 0.401

(0.002)c (0.049)c (0.026)c (0.032) (0.072) (0.047)c (0.120) (0.066)c (0.063)d (0.049)c (0.052) (0.051)c (0.005)c (0.006)c

0.057 -0.325 -0.618 0.267 -0.064 2.132 -0.202 0.461 1.288 -3.822 -0.416 0.119 1.743 1.285

(0.003)c (0.110)c (0.043)c (0.057)c (0.124) (0.275)c (0.313) (0.146)c (0.144)c (0.081)c (0.098)c (0.086) (0.026)c (0.022)c

Coefficient

Table 5.5 (continued) Mortality

Pneumonia complication

Extended stay

Length of stay

Coefficient

Standard errorb

Coefficient

Standard errorb

Coefficient

Standard errorb

Hospital Bed size ('000) Residents /beds Operating margin Debt-capital ratio Occupancy rate No profit Debt-capital ratio > 1

0.079 -0.169 0.290 0.004 0.112 0.012 0.038

(0.038)d (0.056)c (0.139)d (0.040) (0.517)d (0.015) (0.041)

-0.277 0.194 0.140 0.070 -0.103 0.113 0.106

(0.060)c (0.075)c (0.233) (0.059) (0.730) (0.023)c (0.061)e

0.228 -0.056 -0.136 0.195 0.494 -0.017 0.058

(0.067)c (0.094) (0.312) (0.089)d (1.231)c (0.033) (0.087)

0.003 -1.113 0.162 0.201 1.360 0.151 0.106

(0.137) (0.152)c (0.387) (0.125) (2.197)c (0.055)c (0.133)

Market Herfindahl index HMO share Population density ('000/sq m) Hospital beds/('000) pop. Income in zip code area (mil $) Constant

0.035 -0.025 -0.014 -0.140 0.033 -8.743

(0.038) (0.053) (0.008) (0.064)d (0.592) (0.132)c

0.292 0.078 0.004 -0.021 2.140 -8.216

(0.054)c (0.071) (0.012) (0.087) (0.867) (0.184)c

-0.220 0.175 0.012 0.526 -4.380 -9.121

(0.091 )d (0.108) (0.012) (0.160)c (1.170) (0.342)c

-0.108 0.020 0.177 0.607 -16.700 -0.735

(0.111) (0.182) (0.034)c (0.150)c (3.110) (0.385)e

N R2 Pseudo R2

417,833 — 0.140

a

Year, area, and indicators of missing value binary variables not shown. Standard errors in parentheses. Significant at 1% level (two-tail test). Significant at 5% level (two-tail test). Significant at 10% level (two-tail test). b

346,688 — 0.171

416,428 — 0.248

Coefficient

412,462 0.091 —

Standard errorb

156

Sloan

tended stay is positive and statistically significant at conventional levels. Finally, to the extent that hospitals converting from GN to F status achieved shorter lengths of stays after ownership conversion, was this achieved by increased rates of transfers to nursing homes or to other hospitals? To determine the answer to this question, multinomial logit analysis of the place to which the person was discharged after leaving the hospital was conducted (table 5.6). The four destinations were home (reference group), death, nursing home, and other hospital. The analysis controlled, among other factors, for source of admission (nursing home, other hospital, emergency room, home) because patients were more likely to return to the place from which they came to the hospital. But given the number of parameter estimates in a multinomial format, area fixed effects were not included. As in all of the other analyses, year binary variables were included. Results shown in the table 5.6 are based on specification 3. In the GN sample, 8.1 percent of persons died in the hospital, 15.3 percent were transferred to a nursing home, 10.1 percent were transferred to another hospital, and the remaining two-thirds (65.5 percent) returned home. The pattern of discharge destinations was similar for the F sample, with a slightly lower percentage of patients returning home and a somewhat higher percentage being transferred to other hospitals—possibly reflecting their smaller bed size. Holding other factors constant, the rate of discharges to other hospitals increased after conversion for GN to F hospitals (relative to discharge to home). The associated marginal effect of 0.039 is substantial relative to the sample mean transfer rate (about 40 percent of the sample mean). In view of the dramatic drop in bed size among GN-to-F converting hospitals noted earlier, an increase in the transfer rate to other hospitals is not surprising. In the F sample, discharges to nursing home increased after conversion, but the associated marginal effects were very small. Upcoding of Diagnoses For the measures studied, there is no evidence that ownership conversion from GN to F ownership status increased the rate of upcoding (table 5.7). In the analysis of mean DRG weight of the five primary diagnoses included in the Medicare analysis, the DRG weight fell after conversion from GN to F. In this specification, many hospital character-

Table 5.6 Multinomial logit analysis: destination of discharge3 GN sample Frequency After conversion Conversion fixed effect Pseudo R2 Number of observations

Death 8.12% 0.012 (0.068) [0.002] -0.228 (0.416) [-0.010]b 0.191 418,773

Nursing home 15.27% -0.086 (0.054) [-0.010] -0.181 (0.034) [-0.011]b

Other hospital 10.12% 0.373 (0.061) [0.039]b -0.505 (0.042) [-0.035]b

F sample

Frequency After conversion Conversion fixed effect Pseudo R2 Number of observations a b

Death

Nursing home

Other hospital

7.57% 0.095 (0.110) [0.005] 0.184 (0.068) [0.010]b 0.191 56,217

14.77% 0.153 (0.087) [0.006]b 0.070 (0.056) [0.013]

10.96% -0.051 (0.089) [-0.003] -0.012 (0.061) [-0.007]

Based on specification 3. Omitted group is destination to home. Standard errors in parentheses and marginal effects in brackets. Significant at 1% level (two-tail test).

158

Sloan

istics, including bed size, were held constant. Thus, the results for ownership conversion do not reflect just downsizing and loss of some sophisticated product lines. Of course, a decrease in the mean DRG may have reflected subtle changes in case mix. To the extent that this occurred, it more than obscures any change in upcoding for the five primary diagnoses included in this analysis that may have occurred. For coding of transient ischemic attack (TIA) versus stroke, there was an increase in the proportion of TIAs (significant at the 10 percent level) following conversion from GN to F status. By contrast, hospitals that converted from F to GN did have higher DRG weights after conversion. Similarly, the proportion of cases coded as stroke rather than TIA rose. V. Results: Patients Under Age 65 Payer Mix after Ownership Conversion As discussed in the previous section, there is widespread concern that ownership conversions work to the disadvantage of underserved populations. However, empirical evidence on actual effects has been mixed. Table 5.8 presents key results of a multinomial analysis of anticipated source of payment. Since the payment categories are confined to the primary source of payment, they are mutually exclusive. For patients who were between the ages of 1 and 64 at the admission date and were admitted to a facility in the GN sample, the probability of having Medicare, other government insurance, or Medicaid, or being classified as a self-pay/no charge patient increased after conversion to F status. G or N hospitals acquired by F owners tended to be those that had relatively low proportions of Medicaid and self-pay/no charge patients, or Medicare (disabled, end stage kidney disease) patients. However, the tendency to eschew the poor and the disabled was not sustained after conversion. Results for hospitals converting from F to GN ownership are mixed. The proportion of self-pay/no charge patients fell after conversion to GN status. For the sample of births, holding other factors constant, hospitals converting from GN to F were more, not less, likely to admit nonprivately insured patients after the conversion occurred (table 5.9). This finding is very sensitive to changes in equation specification. The conversion fixed effect is negative and the associated marginal effect is substantial. The implication is that for-profit organizations were more likely to acquire hospitals that were relatively oriented to privately in-

Table 5.7 Upcoding of diagnosis3 DRG weight

TIA versus stroke GN sample Conversion GN to F after GN to F fixed effect F to GN after F to GN fixed effect Patient Age White Male Admitted from ER Admitted from nursing home Admitted from other hospital Diagnosis of Alzheimer's Diagnosis of other dementia Diagnosis of stroke Diagnosis of coronary heart disease Diagnosis of congestive heart disease Diagnosis of pneumonia

F sample

0.183 (0.098)d -0.163 (0.058)b

GN sample -0.173 (0.023)b 0.176 (0.016)b

-0.505 (0.162)b

-0.003 0.142 -0.107 -0.136 -0.524 -1.489 — — — — — —

(0.001 )c (0.037)b (0.019)b (0.022)b (0.066)b (0.080)b — — — — — —

F sample

0.316

(0.107)b

-0.011 -0.099 -0.053 -0.279 -0.802 -1.220 — — — — — —

(0.003)b (0.099) (0.051) (0.059)b (0.197)b (0.188)b — — — — — —

0.102 (0.034)b

-0.013 0.072 0.181 -0.391 -0.047 0.671 -0.091 -0.114 -0.443 -0.068 -0.714 -0.517

(0.000)b (0.011 )b (0.005)b (0.006)b (0.014)b (0.018)b (0.010)b (0.007)b (0.007)b (0.006)b (0.005)b (0.006)b

-0.125

(0.023)b

-0.008 0.045 0.173 -0.302 -0.054 0.501 -0.036 -0.058 -0.505 -0.369 -0.741 -0.404

(0.001 )b (0.021)c (0.011)b (0.014)b (0.038) (0.045)b (0.032) (0.021)b (0.016)b (0.014)b (0.013)b (0.016)b

Table 5.7 (continued) TIA versus stroke GN sample Hospital Bed size ('000) Residents/beds Operating margin Debt-capital ratio Occupancy rate (%) No profit Debt-capital ratio larger than 1 Market Herfindahl index HMO share Population density ('000/sq m) Hospital beds/ ( '000) population Income in zip code area (mil $) Constant Number of observations R2 a

F sample

GN sample

F sample

-0.507 -0.410 -0.210 0.247 0.005 0.077 0.005

(0.055)b (0.088)b (0.203) (0.053)b (0.001 )b (0.021)b (0.058)

-0.381 -0.695 -0.597 -0.165 0.001 -0.116 0.157

(0.259) (1.488) (0.393) (0.104) (0.002) (0.071) (0.092)d

0.824 0.395 0.076 -0.107 0.003 -0.065 -0.034

(0.016)b (0.026)b (0.049) (0.014)b (0.000)b (0.006)b (0.013)b

1.009 1.397 0.677 -0.175 0.003 0.018 0.085

(0.065)b (0.586)c (0.088)b (0.025)b (0.000)b (0.015) (0.020)b

-0.092 -0.226 0.441 0.563 3.240

(0.041)c (0.054)b (0.084)b (0.065)b (0.802)

-0.004 0.093 0.843 0.037 -1.300

(0.125) (0.144) (0.239)b (0.194) (2.640)

-0.080 0.175 0.076 -0.072 0.562

(0.012)b (0.019)b (0.039)b (0.019)b (0.231)

-0.083 -0.062 -0.719 -0.023 2.150

(0.033)c (0.045) (0.125)b (0.041) (0.667)

2.941 (0.133)b 61,853 0.033

Year, area, and indicators of missing value binary variables not shown. Standard errors in parentheses and marginal effects in brackets. c Significant at 1% level (two-tail test). d Significant at 5% level (two-tail test).

b

DRG weight

2.383 (0.042)b 8,015 0.062

0.625 (0.340)d 418,773 0.141

-1.457 (0.125)b 56,217 0.113

Table 5.8 Multinomial logit analysis: payer type of patients aged 1 to 64a GN sample Medicare or other government insurance Frequency After conversion Conversion fixed effect Pseudo R2 Number of observations

7.94% 0.244 (0.052) [0.008]b -0.066 (0.027) [0.007]c 0.105 512,133

Medicaid

Self -pay or no charge

17.67% 0.582 (0.057) [0.067]b -0.412 (0.034) [-0.039]b

19.27% 0.579 (0.081) [0.033]b -0.601 (0.051) [-0.030]b

F Sample Medicare or other government insurance Frequency After conversion Conversion fixed effect Pseudo R2 Number of observations a

6.29% 0.068 (0.086) [0.002] -0.343 (0.059) [-0.049]b 0.103 40,405

Medicaid 12.90% 0.454 (0.096) [0.048]b -0.326 (0.077) [0.022]b

Self-pay or no charge 21.36% -0.266 (0.121) [0.011]c 0.210 (0.090) [0.013]c

Based on specification 3. Sample: 25% of one of the subsamples provided by HCUP data, excluding women giving birth. Omitted group is private insurance. Standard errors in parentheses and marginal effects in brackets. b Significant at 1% level (two-tail test). c Significant at 5% level (two-tail test).

162

Sloan

Table 5.9 Logit analysis of probability of payment other than private insurance for labor delivery, less than two-day stay, and vaginal birth3 GN sample

F sample

Probability of payment other than private insurance After conversion Conversion fixed effect Number of observations Pseudo R2

[0.224]b 0.916 (0.046) -0.417 (0.026) [-0.098]b 558,268

0.188 Probability of less than two-day stay After conversion -0.391 (0.056) [-0.025]b Conversion fixed effect -0.347 (0.028) [-0.023]b Number of observations 566,927 Pseudo R2 0.015

Probability of vaginal birth After conversion -0.340 (0.049) [-0.025]b [0.025]b Conversion fixed effect 0.252 (0.024) Number of observations 567,605 Pseudo R2 0.293

0.884 (0.107) [0.218]b -0.187 (0.083) [-0.044]c 32,527 0.213

-0.206 (0.117) [-0.017]d -0.085 (0.092) [-0.007] 32,635 0.024

0.123 (0.103) 0.299 (0.075) 32,635 0.315

[0.024] [0.061]b

a

Based on specification 3. Standard errors corrected for heteroscedasticity. Standard errors in parentheses and marginal effects in brackets. b Significant at 1% level (two-tail test). c Significant at 5% level (two-tail test). d Significant at 10% level (two-tail test).

sured patients, but after conversion, there was a major increase in nonprivately insured patients. The marginal effect is 0.224. Since time fixed effects were included, the marginal effect for the after-conversion dummy does not reflect a secular growth in persons not covered by private insurance. Patterns for conversions from F to GN status were quite similar to those from GN to F. Changes in Length of Stays: Birth Sample There is widespread concern that the length of hospital stays for labor and delivery have declined to unsafe levels. Such reductions are said to be motivated mostly by a desire to increase profit. Some states have implemented laws to restrict the ability of health care providers to limit stays for labor and delivery to less than two days. Table 5.9 shows that the probability of a less-than-two-day stay declined after ownership conversion, both for hospitals that converted from GN to F status and those that converted in the opposite direction.

Hospital Ownership Conversions

163

Changes in the Probability of Vaginal Births The proportion of vaginal births declined after hospitals converted from government or not-for-profit to for-profit status (table 5.9). For hospitals that converted from F to GN status, there was an increase in the probability of vaginal births, but this increase was not statistically significant at conventional levels. VI. Discussion, Conclusions, and Implications The main question addressed by this paper is whether or not the market for hospitals is fundamentally broke. In my review of the literature of cross-sectional evidence on differences in hospital behavior by ownership type, I conclude that hospitals with various types of ownership were more alike than different (Sloan 2000). With longitudinal data on the same set of hospitals (panel data), the research focus has been on changes in behavior for the same hospital following a change in ownership. With a panel, isolating the effect of the change in ownership can be done more precisely. In many respects, the empirical evidence from hospital conversions is reassuring. In particular, in this study, I did not find that in-hospital mortality changed as a result of changes in ownership type. Payer mix, including the proportion of persons admitted without health insurance, was not affected. On the whole, hospitals' missions appear to be preserved postconversion. In large part, constancy of mission has been safeguarded by contract provisions that are made part of the agreement between the buying and selling hospitals at the time of the transaction. The analysis presented in the previous section did not reveal systematic upcoding by hospitals that converted to for-profit ownership status. Thus, the upcoding for pneumonia and respiratory infections reported in an earlier study does not appear to generalize to other diagnoses. However, rates of pneumonia complications rose following conversions from government or private nonprofit to for-profit ownership. In another recent study cited in this paper, mortality rates following discharge rose appreciably following conversion at hospitals that converted from government or private nonprofit to for-profit ownership, especially during the first two years after the conversion occurred. One reason that there would be no change in in-hospital mortality, but an increase in mortality observed at thirty days, six months, and one year following admission to the hospital, may be that hospitals

164

Sloan

converting to for-profit ownership reduced lengths of hospital stay disproportionately. The evidence pointing to some adverse effects of conversions is a reason for concern and should affect public policy regarding hospital ownership conversions. There is a clear role for public oversight. Success is not guaranteed—the "Allegany system debacle/' where effective public oversight was lacking, is a case in point. Communities that place their hospital assets for sale would do well to exercise due diligence. This may take the form of oversight by the state attorney general or the state certificate of need agency, as well as local elected officials. Practices such as upcoding should be monitored by hospital compliance programs. Some public oversight of hospital staffing during the years immediately following conversion seems warranted based on the empirical evidence. Is there a chance that this study has failed to document important changes in quality of care and/or public goods provision associated with changes in hospital ownership status? Almost certainly yes. The issues involved are complex, and there are so many kinds of outcomes. This consideration adds further force to the conclusion that scrutiny should be case-specific. Rather than conclude that all conversions of a given type are "bad" or "good," public policy should recognize the heterogeneity and multiplicity of outcomes. Notes I acknowledge the capable research assistance of Jingshu Wang, Department of Economics and Center for Health Policy, Law, and Management, Duke University, and the grant, Hospital Ownership Conversions, from the Robert Wood Johnson Foundation and administered by the Academy for Health Services Research and Health Policy. 1. Of course, the two aspects are related. If there are many concessions on nonprice dimensions of the transaction, the price should reflect this. 2. This result is consistent with the finding by Thorpe et al. (2000) that margins increased after conversion to for-profit status, mainly because of expense reductions. 3. In some unpublished work, colleagues and I followed the suggestion to include a binary for trauma center in analysis of survival following inpatient stays. The variable typically did not affect outcomes. One might argue that a trauma center is endogenous to ownership. 4. Their technique includes the equivalent of hospital fixed effects. 5. A more complete discussion of the effects of hospital competition on quality of care is beyond the scope of this paper. On this subject, see, for example, Kessler and McClellan (2000).

Hospital Ownership Conversions

165

6. This has been documented by others as well. See, for example, Eldenburg et al. (2001). 7. This result is consistent with a view advanced by Lakdawalla and Philipson (1999). They argued that nonprofit behavior should not affect for-profit behavior because the latter, by virtue of the lack of tax advantages and charitable endowments, are the marginal firms and hence influence market outcomes (price, quality, etc.). 8. Medicare compliance, coding, self-referral, and joint ventures have become major issues to hospitals. See, for example, a newsletter for hospital administrators explicitly devoted to these issues: Eli Research, "Hospital Compliance Alert." 9. In table 5.4, all explanatory variables shown are binaries.

References American Hospital Association (1976). Hospital Statistics. Chicago, IL: AHA, 1976. American Hospital Association (1979). Hospital Statistics. Chicago, IL: AHA, 1979. American Hospital Association (2001). Hospital Statistics. Chicago, IL: AHA, 2001. Anderson, Ruth, Charlene Allerd, and Frank A. Sloan (2001). "Effects of Hospital Conversions on Organizational Decision Making and Service Coordination," unpublished. Blumenthal, David, and Joel S. Weissman (2000). "Selling Teaching Hospitals to Investor-Owned Hospital Chains: Three Case Studies," Health Affairs 19(2): 158-166. Burns, Lawton R., John Cacciamani, James Clement, and Welman Aquino (2000). "The Fall of the House of AHERF: The Allegany Bankruptcy," Health Affairs 19(1):7-41. Burns, Lawton R., Amit Shah, and Frank A. Sloan (2001). "The Impact of Governance Changes on Strategy: Results from a Study of Hospital Ownership Conversions," University of Pennsylvania. Conover, Christopher J., Michael E. Ezell, and Frank A. Sloan (2001). "Understanding Hospital Conversions," unpublished. Duke University. Conover, Christopher J., and Frank A. Sloan (2001). "Rates of Return from Hospital Conversions," unpublished. Duke University. Cutler, David M., and Jill R. Horwitz (2000). "Converting Hospitals from Not-for-Profit to For-Profit Status," in David M. Cutler, ed., The Changing Hospital Industry: Comparing Not-for-Profit and For-Profit Institutions. Chicago, IL: University of Chicago Press, 45-79. Desai, Kamal R., Carol VanDeusen Lukas, and Gary J. Young (2000). "Public Hospitals: Privatization and Uncompensated Care," Health Affairs 19(2):167-172. Duggan, Mark G. (2000a). "Hospital Market Structure and the Behavior of Not-for-Profit Hospitals: Evidence from Responses to California's Disproportionate Share Program," NBER Working Paper 7966. "Duggan, Mark G. (2000b). "Hospital Ownership and Public Medical Spending," Quarrly Journal of Economics, 115(4):1343-1373. ?nburg, Leslie, Benjamin E. Hermalin, Michael S. Weisbach, and Marta Wosinska U '.). "Hospital Governance, Performance Objectives, and Organizational Form," NBE Working Paper 8201.

166

Sloan

Hart, Oliver, Andrei Shleifer, and Robert W. Vishny (1997). "The Proper Scope of Government: Theory and an Application to Prisons," Quarterly Journal of Economics 112(4):1127-1162. Kessler, Daniel P., and Mark B. McClellan (2000). "Is Competition Socially Wasteful?" Quarterly Journal of Economics, 115(2):577-615. Lakdawalla, Darius, and Tomas Philipson (1999). "The Nonprofit Sector and Industry Performance," mimeo, University of Chicago. Legnini, M. W., L. E. Rosenberg, M. J. Perry, and N. J. Robertson (2000). "Where Does Performance Measurement Go from Here?" Health Affairs 19(3): 173-177. McClellan, Mark, and Douglas Staiger (2000). "Comparing Hospital Quality at For-Profit and Not-for-Profit Hospitals," in David M. Cutler, ed., The Changing Hospital Industry: Comparing Not-for-Profit and For-Profit Institutions. Chicago, IL: University of Chicago Press, 93-112. Needleman, J., D. J. Chollet, and J. Lamphere (1997). "Hospital Conversion Trends," Health Affairs 16(2):187-195. Norton, Edward C, and Douglas O. Staiger (1994). "How Hospital Ownership Affects Access to Care for the Uninsured," Rand Journal of Economics 25(1):171-185. Picone, Gabriel, Shin-Yi Chou, and Frank Sloan (forthcoming). "Are For-Profit Conversions Harmful to Patients and to Medicare?" Rand Journal of Economics. Robinson, James G. (2000). "Capital Finance and Ownership Conversions in Health Care," Health Affairs (19)1:56-71. Silverman, Elaine, and Jonathan Skinner (2001). "Are For-Profit Hospitals Really Different? Medicare Upcoding and Market Structure," NBER Working Paper 8133. Sloan, Frank A. (2000). "Not-for-Profit Ownership and Hospital Behavior," in A. J. Culyer and J. P. Newhouse, eds., Handbook of Health Economics, vol. 1B. Amsterdam: Elsevier, 1141-1174. Sloan, Frank A., Donald H. Taylor, and Christopher J. Conover (2000). "Hospital Conversions: Is the Purchase Price Too Low?" in David M. Cutler, ed., The Changing Hospital Industry: Comparing Not-for-Profit and For-Profit Institutions. Chicago, IL: University of Chicago Press, 13-44. Thorpe, Kenneth E., Curtis S. Florence, and Eric E. Seiber (2000). "Hospital Conversions, Margins, and the Provision of Uncompensated Care," Health Affairs 19(6):187-195. Wolfram, Catherine (2000). "Comment on Chapters 3 and 4: McClellan and Staiger," in David M. Cutler, ed., The Changing Hospital Industry: Comparing Not-for-Profit and For-Profit Institutions. Chicago, IL: University of Chicago Press, 155-158.