Estimating the Impact of Medical Innovation: A Case Study of HIV Antiretroviral Treatments

Transcription

1 Estimating the Impact of Medical Innovation: A Case Study of HIV Antiretroviral Treatments Mark G. Duggan Department of Economics, Maryland Population Research Center, and NBER University of Maryland College Park, MD Vmail: (301) duggan@econ.umd.edu William N. Evans Department of Economics, Maryland Population Research Center, and NBER University of Maryland College Park, MD Vmail: (301) evans@econ.umd.edu February, 2006 Preliminary and Incomplete Comments Welcome JEL Classification: H51; I12; I18 Acknowledgements: This work was supported by grants from the Robert Wood Johnson Foundation (Duggan), the Alfred P. Sloan Foundation (Duggan), the National Science Foundation (Evans), and the Maryland Population Research Center. We thank Julian Crista, Scott Imberman, and Daniel Perlman for outstanding research assistance and Jim Klein from the California Department of Health Services for assistance with the Medicaid data. We are also grateful to Dana Goldman, Steven Levitt, Frank Lichtenberg, Justin McCrary, and seminar participants at the Brookings Institution, the Centers for Medicare and Medicaid Services, the University of Chicago, Columbia University, RAND, Rochester University, Stanford University, the World Bank, and the NBER Health Care meetings for helpful suggestions. The views expressed in this paper are those of the authors and not necessarily those of any of the individuals or institutions mentioned above. All errors are our own. 1

2 Abstract During the 2004 calendar year, health care spending in the U.S. accounted for 16 percent of GDP, more than double the corresponding share from One factor that has contributed to this growth is the introduction and diffusion of new and more expensive health care treatments. As budgetary pressures continue to grow, there is likely to be more scrutiny given by public and private insurers to the impact of new treatments. Estimating these effects with observational data is difficult because treatment is not randomly assigned and because the available data often does not contain the clinical information used by patients and their physicians when making treatment decisions. But unfortunately the random assignment clinical trials that are widely used in medicine have important limitations of their own. In the current study we aim to estimate the impact of one category of health care treatments HIV antiretroviral drugs (ARVs) using administrative data from California s Medicaid program. We use the detailed information on health care utilization in our data to construct proxies for individuals health status. It is essential to control for health status because ARVs were recommended only for the sickest HIV/AIDS patients. Our findings demonstrate that the increase in the use of four drugs approved by the FDA in late 1995 and early 1996 was largely responsible for the well-documented decline in AIDS mortality rates. Our estimates for mortality effects are in line with those from random assignment clinical trials but show substantial heterogeneity across patients with respect to both mortality and expenditures. Despite the approval of more than a dozen ARVs since these four, we find little further decline in mortality despite a continued increase in pharmaceutical spending. Our results for this one category of treatments suggest that the claims data that is widely available to both public and private insurers can be useful for estimating whether the benefit of new health care treatments are sufficient to justify their greater cost. This may be a more efficient instrument for reining in health care spending than other commonly proposed policies. 2

3 I. Introduction During the last several decades, health care expenditures in the U.S. have accounted for an ever increasing share of national output. In the 2004 calendar year, approximately 16 percent of GDP was devoted to health care, with this share more than twice as large as the corresponding value in The Centers for Medicare and Medicaid Services (CMS) project that real health care expenditures will grow at an annual rate of more than 7 percent between now and 2014, which is twice the rate that the Congressional Budget Office (CBO) predicts for real GDP growth over the same time period. Much of the past and projected future growth in health care spending has been fueled by the Medicaid and Medicare programs, which provided health insurance to nearly 90 million individuals in 2005 at a cost of more than $650 billion. According to the CBO, the share of all federal spending accounted for by these two programs will increase from 22 percent to 35 percent during the coming decade. Previous research has suggested that the driving force behind rising health care costs is the introduction and diffusion of new and more expensive treatments (Newhouse, 1992; Cutler, 2004). To date, most private and public insurers base coverage decisions for these advances solely on efficacy grounds. However, budgetary pressures are likely to force insurers to start requiring that coverage decisions also be based on cost effectiveness by considering whether the increment to health from a new treatment is sufficiently large to justify its cost (Garber, 2004). Recent evidence for the growing importance of cost-effectiveness can be found in the Medicare Prescription Drug and Modernization Act (MMA), which allows the private drug plans that provide Medicare recipients with coverage for prescription drug expenses to use cost-effectiveness as one criterion for excluding treatments from their list of covered prescription drugs. This is the first time that cost-effectiveness has been established as a criterion for coverage in the Medicare program. Unfortunately, research on the cost effectiveness of alternative treatments lags far behind what would be necessary for consumers, physicians, insurers, and the government to make well-informed decisions about new procedures and pharmaceutical treatments. Part of the problem stems from the fact that the gold standard for determining causal relationships in medicine is random assignment clinical 3

4 trials (RACT). In many cases, clinical trials are not feasible or are simply too expensive for determining cost effectiveness. RACTs are also typically done on small and potentially unrepresentative populations for just a short period of time. Moreover, the results from a trial may not apply to a real-world setting, where adherence to a treatment regimen may be quite different. And finally, RACTs tend to focus on average effects, which could obscure important differences in the effect across patients. It is therefore necessary that researchers develop alternative methods to analyze questions of costeffectiveness for real world populations. A promising strategy is to use observational data such as the data on health care utilization available to both public and private insurers. This type of data has been used for just this purpose by a number of authors including McClellan, McNeil and Newhouse (1994) and Cutler (1995). Claims data sets often have very large sample sizes and are inexpensive to analyze, yet they have important limitations. Treatment is assigned in a non-random manner and most claims data bases have limited clinical information that could be used to measure baseline health status. The two studies outlined above use quasi-experimental variation to alleviate the problem of non-random treatments. However, not all problems have this type of variation and timely questions such as the recent Vioxx controversy may require alternative strategies to investigate the impact of new treatments. In this paper, we investigate whether claims data can be useful for estimating the effect of health care treatments, even in the absence of quasi-experimental variation, by focusing on one category of pharmaceutical treatments - HIV antiretrovirals (ARVs). We focus particularly on four prescription drugs that were approved by the U.S. Food and Drug Administration during a four-month period from November of 1995 to March of 1996: Epivir (a nucleoside reverse transcriptase inhibitor, of which there were already four at the time), and three drugs from a new class known as protease inhibitors. Early clinical trials of these new treatments demonstrated their ability to reduce the concentration of the virus in patients, increase the presence of important blood cells that fight opportunistic infection, slow the progression from HIV to AIDS, and reduce mortality. Use of these new treatments spread rapidly and mortality rates in the U.S. among AIDS patients fall by 70 percent between 1995 and 1998 (CDC, 2001). To estimate the effect of these treatments on health outcomes and health care expenditures, we 4

5 use individual-level claims and eligibility data for a large and random sample of Medicaid patients from the state of California with one or more months of eligibility for the program between 1993 and Because almost half of individuals with HIV/AIDS in the U.S. are on Medicaid, our focus on recipients of this program is not as limiting as it otherwise might be. Our data includes patient demographic characteristics along with detailed information on each individual s health care utilization (while they are eligible for the program) for the eleven-year period from 1993 to Additionally, the data has been linked to individual-level mortality data maintained by the state s Center for Health Statistics. The full data set contains information for more than four million residents of the state of California. Estimating the impact of new antiretroviral treatments (ARVs) among HIV/AIDS patients with a claims data base presents a number of difficult challenges. First, as we outline below, the current protocol suggests that patients abstain from antiretroviral drugs until health falls below a certain threshold. Second, initiation of ARV therapy is determined by biological markers such as viral loads and cell counts that are not available in claims data bases. Despite these challenges, the case-study of ARVs is useful for a number of reasons. Previous studies with more detailed clinical data and random assignment but smaller samples have demonstrated that ARVs reduce mortality. Therefore, our first job is to replicate these basic findings with decidedly inferior data to see how far we can get with a simple claims data base. The starting point for our analysis is a sample of more than 13,000 California Medicaid recipients with two or more claims with an HIV/AIDS diagnosis between 1993 and No prior study of ARVs has examined as large a sample of patients for as long a period as the one studied here. The annual number of patients in our sample and their mortality rates over time closely track the numbers for all AIDS patients in California as reported by the U.S. Centers for Disease Control and Prevention. We estimate that the fraction of California residents living with HIV/AIDS who are on Medicaid is close to 50 percent, which is consistent with estimates for the U.S. (Bhattacharya, Goldman, and Sood 2003). Our first set of analyses use aggregate data for our sample to investigate the effect of Epivir and three protease inhibitors on mortality and Medicaid expenditures. One year after the approval of these treatments in the fourth quarter of 1996, more than fifty percent of the individuals in our sample are 5

6 taking Epivir or a protease inhibitor. Our estimates suggest that the average effect of these treatments for those taking them was to reduce quarterly mortality rates by 7.9 percentage points. Though this exceeds the average mortality rate in our sample in late 1995, it is of a plausible magnitude given that sicker patients with higher baseline mortality rates are the ones most likely to initiate treatment. Our findings also demonstrate that the increase in the utilization of these treatments is not significantly related with changes in average Medicaid spending, as the increase in spending on prescription drugs is offset by a reduction in spending on other types of care. Probing more carefully on the expenditure effects, we do find significantly negative effects for the most expensive patients and significantly positive effects for patients at or below the seventieth percentile in the expenditure distribution. We next use individual-level data to investigate the determinants of treatment decisions. As outlined above, pharmaceutical treatments are not randomly allocated with patients with more advanced cases more likely to use the treatments. We exploit the panel nature of our data and use pre-treatment medical claims to construct a patient-level severity index. Consistent with suggested protocols, we find that sicker patients, as measured by our severity proxy, are more likely to initiate treatment following the approval of Epivir and protease inhibitors. This partially explains the variation by gender in the takeup of the treatments, as women had mortality rates that were 45 percent lower than men s in the latter half of 1995 and were 35 percent less likely to use the treatments following their approval. Estimating the effect of Epivir and protease on patient outcomes is difficult given the endogeneity of the treatment decisions and the absence of detailed clinical information in our data. Simple OLS specifications that fail to account for this endogeneity substantially understate the benefit of the treatments. We demonstrate that utilizing our claims-based severity index can substantially lower this bias. Moreover, by interacting our treatment indicators with these measures of health status, we uncover substantial heterogeneity in the benefit of the treatment. In all, we estimate that these four new treatments reduced patient mortality probabilities by 70 percent. This corresponds quite closely with the decline in overall mortality rates for AIDS patients observed during the 1995 to 1998 period suggesting these treatments were the driving force behind the drop in mortality for AIDS patients over this period. 6

7 Additionally, our estimates correspond closely with those from RACTs for these same treatments. Our findings for Medicaid expenditures reveal substantial heterogeneity, with the treatments leading to large reductions in spending for the sickest patients but modest increases for the healthiest patients. The mechanism for this appears to be that the treatments improved health and therefore reduced the need for hospitalizations and other types of medical care. We find little evidence of offset in spending for the one-third of our patients also eligible for Medicare. This is because the Medicare program would have realized the savings from this offset because it is the primary expected payer for dual eligibles. In the final section of our paper, we investigate the effect of the new treatments on long-term Medicaid spending, life expectancy, and the corresponding cost per life-year saved. The new treatments reduced quarterly Medicaid spending by an average of 26 percent for those not on Medicare but increased life expectancy by more than a factor of three. Combining these two effects leads us to estimate the cost per life year saved of the four treatments introduced in late 1995 and early 1996 at close to $16,000, which is well within the range of what is considered to be cost effective. But the 12 drugs approved since 1996 appear to have done little to reduce mortality rates still further, raising the question of whether the further increases in pharmaceutical spending have been cost effective. Taken together, our findings suggest that one can use claims data from the time both before and immediately following the release of new health care treatments to reliably estimate their effects on health outcomes and health care spending. This is true despite the endogeneity of treatment decisions and the limited clinical information in claims data. While studies that use observational data have limitations and should therefore not be considered a substitute for RACTs, they can be an important complement as patients, health care providers, insurers, and policymakers strive to allocate resources optimally in an environment of rapidly increasing health care spending and an ever expanding set of available health care treatments. II. Background and Significance A. Previous Research Using Claims Data and Replicating RACTs 7

8 Insurance claims data bases have large sample sizes, are inexpensive to construct, and contain a wealth of information about patients treatments in real-world situations. Previous research has used samples from claims data bases to analyze the effectiveness of various treatment options. McClellan, O Neil and Newhouse (1994) use a sample of Medicaid patients to analyze the effectiveness of various heart attack treatments. Cutler (1995) uses data on hospital discharges for Medicare patients to examine whether movement to the Prospective Payment System altered patient outcomes. Duggan (2005) used data from California Medicaid to examine whether new pharmaceutical treatments for schizophrenia generated better medical outcomes and lower costs than existing options. To date however, very little research has examined whether analyses using a claims data base can replicate a stylized result generated from a more controlled research environment such as a random assignment clinical trial (RACT). These types of replication exercises have been important in economics in shaping research agendas and fostering the use of new statistical techniques. Lalonde s (1986) results which demonstrated that econometric evaluations of job training programs could not replicate results from random assignment trials led to a long line of research evaluating the statistical methods used in such studies. In contrast, the results in Dehejia and Wahba (1999) which showed that propensity score matching techniques could replicate the results from these same job training experiments, helped launch the use this technique in economics. In a more recent example, Todd and Wolpin (2003) helped encourage the use of structural models by showing that a dynamic structural model of school choice and fertility was able to replicate some results from a famous anti-poverty experiment in Mexico. In this paper, we aim to follow the lead of these previous authors and examine whether an econometric technique applied to a claims data base with non-random assignment can replicate the results of a previous studies. To do this we use administrative data from California s Medicaid program, which according to recent research insures approximately half of individuals with HIV/AIDS in the state (Bhattacharya et al, 2003). The context within which we are conducting this exercise is interesting in its own right. The Centers for Disease Control report that through 2004, approximately one million people in the U.S. had been diagnosed with AIDS and half of these people had died by the end of that same year. 8

9 Prior to the introduction of Epivir and protease inhibitors in late 1995, annual mortality rates of AIDS patients stood at 30 percent but these rates declined by 70 percent during the next three years. The focus on examining the effectiveness of a pharmaceutical intervention in Medicaid is also of interest given the important role that new drugs have played in raising Medicaid costs. Between 1995 and 2004, the fraction of Medicaid costs devoted to prescription drugs nearly doubled, from 7.4 to 13.7 percent (Table 1). Much of this increase has been driven by the introduction of new and more expensive treatments, though some has also been explained by an increase in enrollment and in the average number of prescriptions per enrollee. In 2004, 65 percent of prescription drug costs for branded drugs in the Medicaid program were for treatments introduced in or after 1997, while only 6 percent were attributable to drugs introduced in 1990 or earlier. B. Background on HIV/AIDS and Antiretroviral Treatments AIDS is a chronic disease that damages, and ultimately destroys, an individual s immune system. AIDS is caused by HIV, an infection that kills the body=s "CD4 cells" (also called T-helper cells), a type of white blood cell that helps the body fight off fungal, viral and parasitic infections. When the HIV/AIDS epidemic first appeared, providers could only treat opportunistic illnesses encouraged by a weakened immune system rather than attack the virus itself. This changed with the entry of Retrovir (AZT) to the market in This drug was the first one approved by the FDA in the therapeutic class known as NRTIs (nucleoside reverse transcriptase inhibitors) and more than 500,000 prescriptions were filled for it in the U.S. in Despite the entry of three additional NRTIs from 1991 to 1994, use of these drugs among AIDS patients actually declined from 1992 through This trend reversed following the approval of Epivir (another NRTI) and three drugs from a new class known as protease inhibitors (PIs) in late 1995 and early The first NNRTI (non-nucleoside reverse transcriptase inhibitor) was approved in June of Twelve additional drugs from these three classes were approved 9

10 in the seven years from 1997 to Table 1 contains a complete list of drugs approved to treat HIV/AIDS by the end of The release of Epivir and protease inhibitors spawned the use of highly active antiretroviral therapy (HAART), which is the simultaneous use of two or more ARVs to treat HIV. The optimal time to initiate HAART depends both on the strength of the patient s immune system (measured by CD4 cell counts) and on the concentration of HIV in the patient s blood (also known as the viral load). 2 Current guidelines recommend HAART for all patients with less than 200 CD4 cells per cubic millimeter of blood and suggest that all patients with CD4 cell counts between 200 and 350 should be offered treatment (NIH, 2004; Yeni et al., 2002). The guidelines outline the costs and benefits of an early start of HAART therapy. Starting treatment relatively early might prevent both the degradation of the immune system and the elevation of viral loads. On the other hand, HAART therapy can reduce the quality of life because of severe side effects. 3 Patients may also develop drug resistance, thereby reducing drug options in the future. In general, however, the above guidelines suggest that those HIV-positive individuals who take the drugs will tend to be sicker than their counterparts who do not. In a short period after the approval of Epivir/PI, HAART became the standard treatment for those infected with HIV. Bozzette et al., (1998 and 2001) found that by the end of 1996, nearly 60 percent of HIV infected patients were using protease inhibitors while estimates for some urban clinics calculated HAART use rates in excess of 80 percent by the late 1990s (Palella et al., 1998; Sackoff, McFarland, and Shin, 2000; Ghani, Donnelly, and Anderson, 2003). 4 The sharp increase in the use of the drugs coincided with a substantial decline in the mortality rate among AIDS patients. Between 1995 and 1998 the mortality rate for individuals with AIDS fell by 70 percent. A large number of later studies, some using randomized research designs (e.g., Hammer et al., 1 A fourth group of drugs - known as fusion inhibitors - was approved by the FDA in These drugs were however introduced near the end of our sample period and will not be a part of our main analyses. 2 The goal of treatment is complete viral suppression, defined to be less than 50 copies of HIV RNA per milliliter. 3 Side effects that range from more minor medical conditions such as fatigue, fever, nausea, and headaches, to severe conditions such as liver damage, diabetes, high cholesterol, fat maldistribution, heart attacks and stroke. 4 Medicaid patients typically had lower use rates for the new drugs (Bozzette et al., 2001; Shapiro et al., 1999). 10

11 1997; Delta Coordinating Committee, 2001; Floridia et al., 2002) and others using observational data on patients (Palella et al., 1998; Detels et al., 1998; Schwarcz et al., 2000; Lewden et al., 2001; Egger et al., 2002; and the CASCADE Collaboration, 2003; Messeri et al., 2003) investigated the life saving benefits of new ARVs. All of these studies found that the new drug therapies generated statistically significant reductions in mortality among HIV/AIDS patients. For example, in a random assignment trial examining the effectiveness of Indinavir (a protease inhibitor) used in combinations with two NRTIs, Hammer et al. (1997) found 48-week mortality rates were 55 percent lower in the treated sample. These results have been replicated in larger studies with only observational data but with detailed data on clinical markers. Schwartz et al. (2000) examined the impact of protease inhibitor use on survival of 15,271 HIV/AIDS patients from San Francisco. Controlling for a variety of socioeconomic factors and CD4 count levels prior to treatment, the authors found that PI use reduced mortality by 64 for those who start treatment after progression to AIDS and by 75 percent for those who start treatment prior to progressing to AIDS. Each of these had detailed clinical information about patients, such as viral load or CD4 counts, prior to the initiation of treatment. Demonstrating that we could replicate these results within a claims data base offers a potential leap forward in the types of questions that can potentially be addressed with non-clinical data. Studies in which treatment is randomly assigned contain carefully selected patients, raising the question of whether the drugs would be as effective when used by a more diverse set of patients. A Medicaid sample such as the one we use above has a large and diverse population, allowing us to examine the extent to which the effects vary across groups rather than simply focusing on mean impacts. Cost considerations and other factors may further limit the types if questions analyzed with a clinical trial. Raffi et al. (2001) notes that because HAART has suppressed mortality rates to such low levels among AIDS patients, clinical trials of new HAART therapies would have to be impractically large to use mortality as an outcome. Finally, few have analyzed the clinical and financial consequences of ARV treatment within the same sample. None of the studies listed above examined the impact of ARVs on medical care costs and only a few studies (Gebo et al., 1999; Gebo, Deiner-West, and Moore, 2001) have found that the high cost of ARVs offset 11

12 spending on hospital care and on other health care services. III. Constructing the Analysis Files A. The California Medicaid Claims and Eligibility Data We utilize claims and eligibility data for a random 24 percent sample of Medicaid recipients from the state of California to estimate the effect of HIV antiretroviral drugs. In our data there are 4.03 million people who were eligible for the program in at least one month between January of 1993 and December of The eligibility files contain demographic information for program participants including gender, month and year of birth, race, ethnicity, and zip code of residence. Additionally, there are two variables that allow us to determine whether an individual is dually eligible for health insurance through Medicare in each month. Finally, the eligibility file indicates whether the Medicaid recipient is enrolled in a Medicaid managed care plan in each month and if so, lists the plan number. The claims data includes all fee-for-service payments made from January of 1993 until June of 2004, though because there is often a lag in processing the claims, we focus on the eleven-year period ending in December of There are three types of claims in our data. Inpatient claims are for admissions to hospitals and long-term care facilities and include information about the patient s primary and secondary diagnosis, the dates of service, the amount paid by Medicaid, and the procedures performed. Outpatient claims have similarly-detailed data about payments to physicians, clinics, hospital outpatient facilities, laboratories, and other health care providers. Finally, prescription drug claims provide data on payments made to pharmacies for drugs covered by Medicaid. Each pharmacy claim includes an eleven-digit National Drug Code (NDC), the number of units of the prescription, and the date that the prescription was filled. The NDC is unique for each drug and dosage amount, which allows us to determine the drug treatment(s) that each patient is consuming at any point in time. Every claim includes the patient=s Medicaid identifier, which can then be matched to the eligibility files. Finally, we reached an agreement with the California Center for Health Statistics and the Medical Care Statistics Section that allowed us to merge death records for the 1993 through 2001 period to the Medicaid data. These records 12

13 identify date and cause of death for all residents of the state of California. 5 B. Defining the HIV/AIDS Sample A number of previous researchers have used Medicaid claims data to construct samples of HIV/AIDS patients (Eichner and Kahn, 2001; Kahn et al., 2002; Morin et al., 2002). Following this research, we use diagnosis codes on the Medicaid inpatient and outpatient claims to determine whether individuals are diagnosed with this illness. 6 Patients enter our sample on the date of their first HIV claim although they may have been enrolled in Medicaid for some time before that time. This algorithm yields a sample of 14,745 individuals who have one or more HIV/AIDS claims, are eligible for Medicaid at some point during our eleven-year study period, have a valid social security number, and have consistent age and gender information across years in the eligibility files. Previous research using Medicaid claims data has found that this algorithm captures the vast majority of recipients diagnosed with HIV/AIDS. 7 Our selection algorithm will however still produce false positives and negatives in our sample. There will be false positives if providers incorrectly code claims. To reduce this possibility, we restrict the sample to include patients with two or more non-prescription HIV/AIDS claims, which reduces the sample to 12,932 patients. False negatives are a potentially more important concern. There are three primary avenues through which we would not identify HIV/AIDS patients on Medicaid. First, an infected person may have no claims with an HIV/AIDS diagnosis because he/she is healthy and thus has limited contact with the health care system. Second, a patient may choose not to use medical care despite their illness. Third, a person may be treated for the illness but have no inpatient or outpatient claims with a primary or secondary diagnosis of HIV/AIDS in our data. 8 5 This match could only be done for the 92 percent of our sample with a valid social security number. See Duggan (2005) for a more detailed description of the California Medicaid claims and eligibility data. 6 California s Medicaid program uses the ICD-9 system of classifying diagnoses, and thus we code a claim as an HIV/AIDS claim if the first three characters are 042, 043, or Rosenblum et al., (1993) matched Medicaid claims data to medical records of patients known to be infected with HIV. These authors found that Medicaid claims successfully identified 91 percent of the patients. 8 For example, there are 1,944 individuals with one or more claims for an HIV antiretroviral drug but with no inpatient or outpatient HIV/AIDS claims during our study period. While we could include these individuals in our 13

14 We can use external data to gauge the potential importance of false negatives for our sample. Our analysis of California death records indicates that between January 1, 1993 and December 31, 2001, there were approximately 31,000 deaths with a primary cause of death listed as HIV/AIDS. Of these, a total of 7,459 deaths have Social Security numbers that would have placed them in our 24-percent Medicaid sample if they were enrolled in Medicaid at some point during our study period. A match of the eligibility files to this death data indicate that 4,371 of these 7,459 (58.6 percent) individuals who died with a primary cause of HIV/AIDS were on Medicaid at some point between 1993 and Of this group, our algorithm captures 3,617 of the individuals who died (almost 83 percent). 9 We could identify only an additional 106 patients by using claims for antiretroviral treatments as well to identify patients. Although our claims data contain a rich set of information, it does have some important limitations. First, our data is for just one state, but, California has the second highest number of people living with AIDS. Second, we lose patients who temporarily or permanently exit because they become ineligible for Medicaid, but less than 2 percent of the sample exits the sample per quarter and this number does not change substantially during our period of analysis (Appendix Figure 1). Third, we do not know when patients were first diagnosed with HIV or AIDS but instead only the date of their first Medicaid HIV/AIDS claim during our study period. Fourth, claims data do not contain important diagnostic information about patients such as CD4 cell counts or HIV viral loads. 10 This information is important because it indicates who is recommended to receive ARVs. As we demonstrate below, we can to some extent control for the severity of the patient=s condition by using claims data about the patient=s prior medical care use. Fifth, we do not have Medicare expenditure data for people also eligible for that sample, we elect not to given that we would then be constructing the sample based on patients choice of treatment rather than on provider diagnoses. The fact that the fraction taking an ARV changes substantially over time provides an additional reason to exclude this group as it would increase the risk of composition bias. 9 The patients in our full sample who died of AIDS but were not identified by our claims algorithm look very different from the people we captured. For example, the false negative group has one half as many eligible months (11.6 versus 22.6) and a much larger fraction of eligible months in managed care (49.1 percent versus 9.7 percent). The first difference suggests that we do not capture some patients because they die early in the sample period before they have any claims. The second difference occurs because individuals in Medicaid managed care will not have fee-for-service claims (Duggan, 2004). 10 Other authors have matched claims data sets to clinical files with this information (e.g., papers from the HIV Cost and Services Utilization Survey, Gebo et al., (1999), Gebo, Diener-West, and Moore (2001), Sambamoorthi et al., (2001)) but we do not have this capacity in this instance. 14

15 program. Medicare will typically cover most of the hospitalization costs of dual eligibles. Thus while we can accurately measure utilization, we will understate total expenditures by the government on both inpatient and outpatient care for this group. 11 Finally, we have incomplete utilization data for patients who are enrolled in a Medicaid managed care plan and thus exclude them from our analyses. Even with these limitations, our data has a number of important benefits over the data sets used in all previous research. First, it is the largest sample of HIV/AIDS patients generated from one consistent source. 12 Second, our period of analysis covers an important time including three full years before and nearly eight years after the introduction of Epivir/PI, allowing us to investigate the effect of the treatments on both short and long-term health and spending. Third, because of the rich set of information in our claims data we can proxy for individual s pre-treatment health status and thus try to account for endogenous treatment decisions. Finally, we can estimate not just the average effect of antiretroviral treatments but also the extent to which this varies across individuals. C. Sample Characteristics On the left-hand axis in Figure 2, we plot the number of Medicaid recipients in our sample who were alive at the beginning of half-year periods starting in January of The patients in each halfyear cell had their first HIV/AIDS claim by the end of that period although they may have been enrolled in Medicaid for some time before that date. Roughly one-fourth of the sample appears in the first halfyear of the time period and the sample grows steadily after that date. On the right-hand axis of the figure, we graph the total number of people living with AIDS in California 13 at the end of each six month period 11 Medicare did not cover prescription drugs for dual eligibles during our study period. 12 The HIV Cost and Services Utilization Consortium (Shapiro et al., 1999) followed a cohort of 2,864 HIV patients for January of 1996 through January of The CASCADE Collaboration (2003) pooled data from 22 cohorts of HIV patients from European, Canadian and Australian studies to construct a sample of 7740 HIV patients covering the pre-1997 through 2001 period. The ART Cohort Collaboration (Egger at al, 2002) pooled data from 13 cohort studies of patients starting HAART from Europe and North American to generate a sample of 12,574 patients. 13 We should note that our sample includes not only patients with AIDS but also some who are just HIV-positive. Unfortunately, in most years California only reported to the CDC the number of people living with AIDS, not the number with HIV. Thus in one respect it is plausible that the patients in our sample would be healthier than the typical AIDS patient in California. However, most of the individuals in our sample qualify for Medicaid through the means-tested Supplemental Security Income (SSI) program. Thus they must be in relatively poor health to meet 15

16 as reported by the CDC in their publication HIV/AIDS Surveillance Report. These two surveys track one another quite closely (ρ=0.98). Our numbers suggest that roughly 52 percent of people living with AIDS in California are on Medicaid, 14 a number close to the national average (Bhattacharya et al, 2003). Similarly the number of individuals in our sample grows at an almost identical rate to the statewide total (58 percent for both from 1994 to 2001). Given the possible limitations with using claims data outlined above, our algorithm for identifying Medicaid recipients with HIV/AIDS appears to work quite well. 15 In Figure 3, we graph half-year mortality rates for the Medicaid recipients in our sample of 12,932 patients during the period. On the second vertical axis of the table, we graph the halfyear mortality rate among all California AIDS patients. Death rates in our sample are percentage points higher than deaths rates for all AIDS patients in California, indicating that our sample is substantially sicker than the typical AIDS patient. The timing and magnitude of the change in mortality in our sample is similar to the changes found for California AIDS patients. Between the first half of 1995 and the second half of 1997, six month mortality rates fell by 69 percent in our sample and 79 percent among all California AIDS patients. In Table 3, we report descriptive information about our Medicaid HIV/AIDS sample at four points in time: 1994, 1997, 2000 and In constructing this sample we drop the 1,063 individuals who live in one of the eight counties that moved its Medicaid recipients into a county-organized health system during our study period because our claims data would often be incomplete for them. We also drop the 1,802 individuals with one or more months in a Medicaid managed care plan during our eleven- SSI s medical eligibility criteria. As we document below, the death rates for our sample are substantially higher than for non-medicaid AIDS patients in California. Therefore, comparing trends in the number of HIV/AIDS patients on Medicaid to overall trends of AIDS patients seems a reasonable compromise given the available data. 14 Consider the first half of 1994 when there are 3237 individuals in our sample. To estimate the number on Medicaid with HIV/AIDS one must multiply this by (1/.24). Additionally we must multiply by to account for the exclusion of those with an invalid SSN. This yields 14,270, which is 52.0% of the statewide total of 27,454. The actual fraction may be even higher given that we exclude individuals with just one HIV/AIDS claim or with prescription drug claims only during the study period. On the other hand, some of the individuals in our sample have not yet progressed to AIDS and thus our estimate will to some extent overstate the Medicaid fraction. 15 One possible concern with focusing just on Medicaid recipients is that the incentive to enroll in the program will change after new treatments become available (Goldman et al, 2001), raising the possibility of composition bias. The fact that our series tracks closely with the total number in the state suggests this is not too problematic. 16

17 year study period. 16 This leaves us with a final sample of 10,067 HIV/AIDS patients. As the table shows, the annual mortality rate in the sample fell from 23.0 percent in 1994 to 5.2 percent in 2000, contributing to a large increase in the average age of the sample. The fraction of the population under 40 fell from 50 percent in 1994 to 28 percent nine years later. During our study period the fraction of the sample that is black and female increased by 47 and 19 percent, respectively. In the bottom half of the table, we report some basic information about health care utilization in our sample. Patients have high medical care use but some measures are improving over time. Almost 48 percent of patients in the sample had an inpatient stay in 1994 and this number fell to 28 percent during the next nine years. Annual inpatient spending fell by an even larger percentage from $7125 to $3510. In contrast, annual outpatient spending increased slightly while spending on prescription drugs tripled, driven primarily by the increased use of antiretroviral drugs and their high cost. Although average annual spending on prescription drugs increased by $8,000 over the period, total spending increased by just $4,800. The fraction of HIV/AIDS patients who are also eligible for Medicare increased from 28 to 45 percent, with this change likely contributing to the fall in average Medicaid spending on inpatient care. IV. The Impact of HIV Antiretroviral Treatments: Time-Series Evidence The FDA s approval of Epivir in November of 1995 and of three protease inhibitors during the next four months coincided with a sharp decline in the mortality rate among the Medicaid recipients in our sample. As Figure 3 demonstrates, from the latter half of 1995 to the same period in 1997, the sixmonth mortality rate among California Medicaid recipients diagnosed with HIV/AIDS fell by 70 percent, from 11.3 to 3.4 percent. This decline was quite similar to the corresponding decline reported by the CDC for all California residents diagnosed with AIDS, though the figure demonstrates that the 16 The lack of data for patients in managed care could be problematic if it leads to changes in the composition of our sample over time. The fraction of all Medicaid recipients in a managed care plan does increase substantially during our study period, though as Duggan (2004) notes these changes differentially affected AFDC/TANF recipients who account for a small share of our sample. SSI recipients, who account for nearly 70% of the patients in our sample, were not required to enroll in managed care in any counties except those moving to a county-organized health system and it is for this reason that we drop COHS counties. The fact that the number of individuals in our sample tracks the number statewide with AIDS quite closely (Figure 1) suggests that this issue is not too problematic. 17

18 individuals in our sample appear to be in worse health than the average California AIDS patient. During the next four years the mortality rate in our sample declined gradually and was equal to 2.8 percent in the second half of In this section we investigate the contribution of the new pharmaceutical treatments to this decline in mortality and also examine their effect on Medicaid expenditures. Suggestive evidence that Epivir and protease inhibitors had a large impact can be seen in Figure 4, which depicts the fraction of individuals in the sample filling at least one prescription for an ARV in the quarter. From the third quarter of 1995 to the second quarter of 1997, this fraction more than doubled, increasing from 28.7 to 59.0 percent. This growth was entirely driven by an increase in the use of Epivir and protease inhibitors, with no one taking these drugs in the third quarter of 1995 and 56.0 percent taking one or more of these treatments in the second quarter of 1997 (Figure 5). There were no significant changes in utilization for other ARVs, as shown in Appendix Tables 1 and 2. A. Modeling the Effect of ARVs Taken together, the series depicted in Figures 3, 4, and 5 strongly suggest that Epivir and protease inhibitors were the primary cause of the significant decline in mortality rates observed during our study period. To probe more formally on this issue, consider the relationship between a mortality indicator M and a treatment indicator Z for individual j in period t: (1) M j,t+1 = α + β jt * Z jt + μ * H jt + Θ * X jt + ε jt In this equation, β jt captures the effect of the treatment, which is allowed to vary both across individuals and within an individual over time. The variables H jt and X jt control for health status and other characteristics (e.g. age and gender) that are plausibly related to an individual s baseline mortality probability. In considering whether to take the treatment, a forward-looking individual might be interested not only in the effect on mortality in the current period but also the effect in future periods. Put simply, an individual could rationally decide to delay treatment even if the magnitude of β jt was high because by 18

19 doing so he/she would increase the likely effect of the treatment in the future. But even considering this dynamic effect, one would expect individuals who take the treatment to have significantly higher values of β jt in (1) above. Thus the magnitude of the average effect on the treated (ATT) would likely exceed the average treatment effect (ATE) for all individuals with HIV/AIDS. Estimating the distribution of β jt in this context is difficult for the reasons outlined above. Sicker patients are the ones most likely to initiate treatment because they tend to have higher (in magnitude) values of β jt. Additionally, the information on health status that is available in the claims data is not complete. But given the rapid change in treatment utilization depicted in Figures 4 and 5, it seems reasonable to assume that there were no similarly large changes in other factors that would substantially bias an estimate of the relationship between changes in overall mortality rates and in the fraction using certain ARVs. We therefore aggregate over all individuals in our sample to estimate a first-difference version of equation (1) to estimate the average effect of the treatments on those who take them as follows: ( 2) Δ M t +1 = λ + β ΔZ + θ ΔX + ξ This specification relates the average change in mortality rates to the change in the fraction of Medicaid HIV/AIDS patients using particular pharmaceutical treatments. The variable λ is included to control for a constant trend over time in the mortality rate. We also control for changes in the fraction of patients who are black or in the fraction female in ΔX t. Because treatments are likely to influence mortality through an effect on health status H, we do not include a control for ΔH t, as we would otherwise be likely to understate the contribution of treatments to mortality rate reductions. If properly estimated, the OLS estimate for β captures the average effect of ARV utilization among those who take the treatment on the quarterly mortality rate. t t t B. Time-Series Estimates of the Effect of ARVs on Mortality To estimate the impact of HIV antiretroviral treatments, we calculate the fraction of individuals in the sample alive at the end of the quarter and who fill one or more prescriptions for an ARV during the 19

20 quarter. Individuals enter our sample in the quarter they first have an inpatient or outpatient claim with a primary or secondary diagnosis of HIV/AIDS. The person remains in the sample until they die or exit the Medicaid program. We restrict attention to the 10,067 patient sample defined in Section 3. In each quarter we calculate several different measures of treatment utilization including the fraction taking any ARV, the fraction taking Epivir, and the fraction taking a protease inhibitor. This time series data set has 32 quarterly observations from the fourth quarter of 1993 through the third quarter of 2001 and the key outcome is the quarterly mortality rate in quarter t, which is defined to be the fraction of individuals alive and in the sample at the end of quarter t-1 who die during quarter t. The potential explanatory power of Epivir/PI for the rapid decline in mortality among AIDS patients is depicted in Figure 6. On the left vertical axis, we report the fraction of patients that are using either Epivir or protease inhibitors and on the right vertical axis, we report the patient quarterly mortality rate. There are three things to highlight in this graph. First, notice that prior to the first quarter of 1996, quarterly mortality rates had been falling (though this trend had leveled off early in 1995). This suggests that there might have been some decline in AIDS mortality rates from late 1995 to late 1997 even if Epivir/PI had not been introduced. Second, as Epivir and protease inhibitor use increased from zero to 56 percent between the fourth quarter of 1995 and the second quarter of 1997, quarterly mortality rates fell by 72 percent, from 6.7 percent to just under 2 percent. As Epivir and protease inhibitor use stabilized in mid-1997, so did mortality rates. Between mid 1997 and the end of our sample, quarterly mortality rates were originally 1.8 percent, fell as low as 1.4 percent, and ended up at 1.6 percent. In Table 4, we report the results from first-difference specifications outlined in equation (2) that examine the relationship between changing ARV use and HIV/AIDS patient outcomes. In each column, the dependent variable is the change in the quarterly mortality rate in our sample. We explore the relationship of this variable with changes in the fraction of the sample with one or more prescriptions for 20