TECHNOLOGICAL CHANGE AND RISK ADJUSTMENT: BENEFIT DESIGN INCENTIVES IN MEDICARE PART D ∗ Colleen Carey Department of Policy Analysis and Management Cornell University Martha van Rensselaer Hall Ithaca, NY 14850 [email protected] January 22, 2016

Abstract Subsidized health insurance markets use diagnosis-based risk adjustment to induce insurers to offer an equitable benefit to individuals of varying expected cost. I demonstrate that technological change after risk adjustment calibration – new drug entry and the onset of generic competition – made certain diagnoses profitable or unprofitable in Medicare Part D. I then exploit variation in diagnoses’ profitability driven by technological change to show insurers designed more favorable benefits for drugs that treat profitable diagnoses as compared to unprofitable diagnoses. In the presence of technological change, risk adjustment may not fully neutralize insurers’ incentives to select through benefit designs. JEL codes: I11 (Analysis of Health Care Markets); I18 (Government Policy, Regulation, Public Health); L51 (Economics of Regulation)

∗ First draft: June 2011. This research received support from the Agency for Healthcare Research and Quality Grants for Health Services Research Dissertation Program as well as from the Robert Wood Johnson Foundation Scholars in Health Policy Research Program. This research was approved by the Institutional Review Board of the University of Michigan. A version of this paper circulated under the title “Government Payments and Insurer Benefit Design in Medicare Part D.” The author gratefully acknowledges the advising of Joseph Harrington, Elena Krasnokutskaya, Stephen Shore, Thomas Buchmueller, Richard Spady, and Tiemen Woutersen. Gerard Anderson, Jonathan Weiner, Brian Pinto, and staff at Thomson Reuters assisted in accessing the data. Two anonymous referees suggested substantial improvements. This paper benefited from the editorial assistance of Varanya Chaubey.

1

1

Introduction

Recent publicly-subsidized health insurance expansions have enshrined a requirement that health insurers accept all applicants, both the sick and the healthy, at the same premiums. Shortly after the passage of the Affordable Care Act, President Barack Obama celebrated that “discrimination against Americans with preexisting conditions will be banned for good.” In order to make insurers willing to enroll the sick at the same premiums as the healthy, the government risk adjusts subsidies so that insurers receive more for a sick enrollee than a healthy one. Risk adjustment underlies several health insurance markets: Medicare Part D, Medicare Advantage, Medicaid managed care, and the Affordable Care Act exchanges. In this paper I measure the success of risk adjustment in the presence of another common feature of health care markets – technological change. This paper studies the interaction of risk adjustment and technological change in Medicare Part D, a publicly-subsidized prescription drug insurance program. The risk adjustment system in Medicare Part D subsidizes insurers on the basis of enrollee diagnoses. The risk adjustment levels for each diagnosis were held fixed over time despite new drug entry and the onset of generic competition – i.e., technological change in drugs – changing the cost of treating each diagnosis. I demonstrate that treatment costs fell for diagnoses that experienced the onset of generic competition, leading such diagnoses to be profitable. Conversely, new drug entry led to higher average treatment costs relative to risk adjustment, making a diagnosis unprofitable. For example, several highly-effective new drugs for the diagnosis Multiple Sclerosis entered after the risk adjustment system was calibrated, while major branded statins for the high-cholesterol diagnosis began to face generic competition. The typical plan receives $334 per MS patient but spends $1185; meanwhile the plan receives $207 per enrollee with Hypertension, but spends only $147. Economic theory has long shown that, in a setting like Part D, insurers will design benefits to select profitable individuals and deter unprofitable individuals. I take advantage of differences in diagnosis-specific profitability driven by technological change to provide the first direct test of selection through benefit design. I find that insurers design more favorable benefits – i.e., lower out-of-pocket costs – for diagnoses that are made profitable by exposure to the onset of generic competition. In contrast, out-of-pocket costs are higher for unprofitable diagnoses, where new drug entry has raised treatment costs above risk adjustment levels. There are several features that make Part D an excellent setting for examining the interaction of risk adjustment, technological change, and benefit design incentives. Most importantly, the diagnosis-specific risk adjustment in Part D diverged from treatment costs because the government calibrated them based on data from the early 2000s and held them steady until 2011. About 40% of the diagnoses in risk adjustment are exposed to at least one new drug entrant after calibration, and nearly two-thirds have at least one previouslybranded drug that began to experience generic competition. In addition, prescription drug benefits represent a useful screening mechanism for selecting on the basis of diagnoses because drugs tend to treat a particular diagnosis. An insurer wishing to attract beneficiaries with profitable diagnoses can simply set lower out-ofpocket costs for drugs that treat that diagnosis. Such selection through benefit design is allowed in Part D; while insurers need to hit an overall actuarial value target, they have control of out-of-pocket costs for specific drugs. Moreover such selection through benefit design can be effective in Part D because Part D enrollees tend to know their diagnoses when they select a plan. In Section 2, I explain these aspects of Medicare Part D in more detail. I first examine the relationship between a diagnosis’s exposure to technological change and its profitability in Medicare Part D. Data on technological change in the form of new drug entry and the onset of generic competition is collected from the Food and Drug Administration. Using the claims of a random 5% sample 2

of beneficiaries, I measure the average treatment costs of each of the risk adjustment diagnoses. Diagnoses with high risk adjustment relative to treatment costs are defined as profitable, while diagnoses with low risk adjustment relative to treatment costs are unprofitable. I find that each new drug entrant introduced after risk adjustment’s calibration lowers profitability by about $58 per beneficiary per year; each new generic raises it by $14. I find similar results when I attempt to account for the importance of technological change by looking at expenditure on entrants or the market size (number of takers) of new generics. I then analyze how insurers design benefits in response to profitable or unprofitable diagnoses. When insurers cannot discriminate among applicants but know what services attract profitable and unprofitable enrollees the model of Frank et al. (2000) (and others reviewed in Section 3) shows they will set more favorable benefits for services (e.g., drugs) that attract profitable enrollees. I test this provision by predicting the benefit designs of the universe of Part D insurers using measured profitability (OLS) and profitability instrumented by technological change (2SLS). Consistent with the theory, insurers set lower out-of-pocket costs for diagnoses made profitable by technological change and higher out-of-pocket costs for unprofitable diagnoses. I also find that drugs that treat profitable diagnoses are on lower formulary tiers, consistent with insurers wishing to attract profitable individuals who use those drugs. The data do not permit me to analyze how the drug prices insurers negotiate with drug makers respond to the same profitability that affects benefit designs. However, I show that drugs that treat profitable diagnoses are more likely to use a copay rather than a coinsurance, which is a benefit design strategy that insurers may prefer if drug firms demand higher prices for drugs that treat profitable diagnoses. Lastly, I suggest an improvement to Part D regulation driven by the observation that diagnosis-based risk adjustment will lead insurers to have diagnosis-specific benefit design incentives. Insurers in Part D must cover at least two drugs in each therapeutic class, but need only meet an overall actuarial value target. I find no effects of profitability on rates of coverage. While therapeutic classes and diagnoses do not exactly correspond, the fact that coverage regulation is based on therapeutic classes appears to constrain insurers’ ability to act on the incentives provided by inaccurate diagnosis-specific risk adjustment. An out-of-pocket cost regulation that required targets within therapeutic classes could eliminate insurers’ latitude to react to such incentives. The economic contribution of this paper is two-fold. It is the first to empirically document the impact of technological change on a risk adjustment system. It contributes to an active literature analyzing the success of risk adjustment in neutralizing insurer incentives in theory and practice. Secondly, it provides clear empirical support for the theory of insurer benefit design. Research such as Brown et al. (2014) has shown that insurers can successfully select profitable individuals and deter unprofitable individuals; this paper documents the differences in benefit design that enable this selection. Econometrically, I propose and verify a method for linking drugs to the risk adjustment diagnoses they treat (Section 5.3). Finally, this paper demonstrates a policy challenge for risk-adjusted health insurance markets that serve more than 75 million Americans.1 In particular, the government generally announces risk adjustment before the market begins, and leaves the system in place for several years. With rapid technological change, even the most perfect risk adjustment system in year t will already be impaired by year t + 1, restoring insurers incentives to select through their benefit designs. 1 There are currently 16 million Medicare Advantage enrollees, 19 million enrollees in (free-standing, non-employer) Medicare Part D, and approximately 40 million individuals are enrolled in an individual or small group health insurance plan and therefore subject to the ACA risk adjustment mechanism. In addition, at least 32 million Americans are enrolled in a Medicaid riskbased managed care, although not every Medicaid managed care program necessarily uses risk adjustment (32 million reflects enrollment from 19 states reporting enrollment to the Kaiser Family Foundation).

3

2

Background on Risk Adjustment, Technological Change, and Benefit Design

In this section, I discuss several features of Medicare Part D that allow us to examine the interaction of risk adjustment, technological change, and benefit design.2 I first describe the risk adjustment system for Medicare Part D. I then discuss how technological change in pharmaceuticals would affect the functioning of risk adjustment. Next, I explain how insurers can use benefit design to attract individuals with profitable diagnoses.

2.1

Diagnosis-Based Risk Adjustment

The goal of the Medicare Part D risk adjustment system is to make insurers indifferent between all enrollees. Medicare uses risk adjustment in its managed competition markets because, as a public insurance program, Medicare aims to ensure that both sick and healthy beneficiaries are well-served by the program’s private insurers. To that end, Medicare provides a larger subsidy to an insurer for enrolling sick beneficiaries than for healthy beneficiaries. In theory, these subsidies make insurers indifferent between all enrollees by providing the expected cost of treating each type of beneficiary. In practice, risk adjustment is a system of subsidies conditioned on observables; in Medicare Part D, subsidies are conditioned on diagnoses and demographics. Risk adjustment must first be calibrated, meaning the subsidy level that induces indifference estimated for each diagnosis. Medicare Part D’s risk adjustment calibration was particularly challenging because no comparable insurance existed for elderly and disabled beneficiaries before Part D began. The risk adjustment designers, Robst et al. (2007), took advantage of medical claims and prescription drug spending from a sample of federal retirees in 2002 and Medicaid beneficiaries in 2000. Working from a list of diagnoses used in Medicare Advantage risk adjustment, they identified a set of 86 diagnoses that significantly raised total drug spending in their sample. They defined demographic categories of age × gender and gender × originally disabled.3 Since federal retirees and Medicaid beneficiaries pay much lower cost-sharing than Part D enrollees, they reduced each individual’s total drug spending by 19% (an estimate from the Medicare Office of the Actuary). Finally, they calculated the expenditure Ej of a standard Part D plan insuring person j – i.e., Ej represents what a Part D plan would contribute to these individuals’ total drug bill. The following regression calibrates the risk adjustment system: X X Ej = Djx Wxcal + Djg Wgcal + νj x

(1)

g

Djx and Djg are dummies for the 86 diagnoses (indexed by x) and the demographic categories (indexed by g). Wxcal and Wgcal represent the treatment costs associated with each diagnosis or demographic category in the calibration data. In addition, the risk adjustment system includes multiplicative factors that express the degree to which spending is higher for low-income individuals and those who are long-term institutionalized. CMS uses these coefficients to compute subsidies to Part D plans. First, they multiply the coefficients by two uniform factors to account for (1) overall prescription drug inflation and (2) trends in diagnosis prevalence in the claims data. To account for overall inflation, Medicare collects from each plan a bid that is its expected cost of providing benefits to a “typical beneficiary” given its benefit design. The Wxcal and Wgcal are uniformly inflated by N AB t /E, where N AB t is the enrollment-weighted national average bid in year t and E is the average expenditure in Equation 1. Trends in diagnosis prevalence may reflect true changes in prevalence or changes in a physician’s propensity to record the diagnosis in the claims. CMS established 2 For

a more comprehensive review of Part D’s market design, see Duggan et al. (2008). who are “originally disabled” are those over 65 who were entitled to Medicare prior to age 65 due to disability.

3 Individuals

4

an upcoding deflation factor, 1.085 in 2009, that helps ensure that a plan bidding the N AB enrolling an average-risk beneficiary is paid the amount specified by statute. In this paper, I use Wx and Wg to refer to the dollar payments a plan bidding the national average bid would receive in 2009. Next, CMS observes diagnoses in plan enrollees’ medical claims. Suppose a typical insurance plan in 2009 enrolled a 66-year-old man (never disabled, noninstitutionalized, and not low-income) whose 2008 medical claims show an infectious disease and hypertension. The demographic risk adjustment Wg for this individual is $331.10, and the diagnosis risk adjustment Wx , reported in Table IV, is ($68.09+$207.06=)$275.15, meaning the plan would receive $606.25 in total. Risk adjustment in Part D faces one more challenge: plans may differ in their overall generosity. A generous plan will spend more for every diagnosis than a less generous plan. Risk adjustment payments to each plan are therefore multiplied by the ratio of the plan’s bid to the national average bid, so that plans with high bids (low out-of-pocket costs) receive higher risk adjustment payments. If the man from the previous example enrolled in a plan with a bid of $1500, the risk adjustment his plan received would be equal to $898.63 (= $606.25 ∗ $1500/N AB 2009 = $606.25 ∗ $1500/$1011.96).

2.2

Technological Change in Medicare Part D

New drug entrants and new generics entering upon patent expiry can dramatically change the cost of treating a given diagnosis, but risk adjustment makes no provision for such changes. The calibration data for risk adjustment was from 2000 and 2002, but the levels for each diagnosis were held steady until 2011, when risk adjustment was reorganized and recalibrated (Kautter et al., 2012).4 Risk adjustment was insensitive to new drug entry or patent expiry between these years. In the year studied in this paper, 2010, at least eight years had passed since the expenditures used to calibrate diagnosis-specific risk adjustment were incurred. In addition, insurers were encountering this system for the fifth year, allowing plenty of time for insurers’ internal analysis to detect the difference between risk adjustment and average treatment costs. For example, the popular cholesterol drug Zocor (simvastatin) began to face generic competition in September 2006. Expenditures on Zocor by elderly Americans totaled $2.3 billion in 2002; presumably at least some of these purchases appear in the calibration data used by Robst et al. (2007). Inasmuch as generic simvastatin costs less than brand name Zocor, this patent expiration lowered the average treatment costs of High Cholesterol. The payment for High Cholesterol, however, was held steady. At the other end of the spectrum, the treatment of multiple sclerosis (as opposed to only management of its symptoms) became widespread in the first decade of the new millennium due to the introduction and expansion of several expensive immunological drugs (Miller, 2011). Recalibrating risk adjustment more often would allow it to incorporate technological change and improve its accuracy. However, in the classic trade-off identified by Newhouse (1996), doing so could introduce inefficiencies. In the extreme, risk adjustment can be a function of past-year utilization, so that risk adjustment automatically rises when treatment becomes more expensive (Hsu et al., 2009; Kautter et al., 2012). But as risk adjustment becomes more accurate, it comes to resemble a fee-for-service system. Insurers lose their incentive to select the profitable and deter the unprofitable, but they also lose their incentive to deter unnecessary utilization: an insurer who incurs high expenditure for a given diagnosis this year can count on higher risk adjustment for that diagnosis the next year (McAdams and Schwarz, 2007). To my knowledge no research addresses the optimal recalibration timing given background technological change and upstream providers with market power. 4I

am separately researching how plans and enrollees responded to the risk adjustment revision (Carey, 2015b).

5

2.3

Insurer Strategies in Medicare Part D

The last section suggests that, due to technological change and a static risk adjustment system, insurers may have found certain diagnoses profitable and others unprofitable. Theoretical models described in the next section suggests that insurers should respond by using benefits as a screen. In this section, I describe three institutional features that make the theory very likely to hold in Medicare Part D. Firstly, a prescription drug benefit can be tailored quite finely to generosity for particular diagnoses. For example, we will later find that Hypertension is a profitable diagnosis in Medicare Part D, while Vascular Disease is not. If Part D were a medical benefit, insurers would face contradictory incentives in the generosity of cardiovascular services. But since different prescription drugs treat Hypertension and Vascular Disease, a prescription drug benefit can be made generous for one and not the other. Secondly, Part D regulations do not preclude designing benefits that are more favorable to profitable diagnoses. Insurer benefit design is characterized by two choices: coverage and out-of-pocket payments.5 Beneficiaries pay the full amount for drugs that are not “covered”, while insurers pay all but the out-ofpocket cost for drugs that are. Because coverage is an obvious way for an insurer to deter sick beneficiaries, Part D insurers have to comply with certain regulations. Firstly, insurers had to cover all the drugs in six “protected” therapeutic classes: antiretrovirals, antineoplastics (anti-cancer drugs), antidepressants, antipsychotics, anticonvulsants, and immune suppressants. In addition, plans were required to cover two drugs in each United States Pharmacopeia “therapeutic class”. In addition to choosing drug coverage, the insurer sets out-of-pocket costs for each drug, again constrained by program rules. The Part D legislation defined a “basic benefit”: plans pay 75% of drug costs in the initial coverage zone, 0% in the coverage gap, and 15% in the catastrophic zone (where government reinsurance pays 80%). Plans could choose to deviate from the 25% out-of-pocket costs in the initial coverage zone prescribed by the basic benefit, although they had to show that, for the type of Medicare beneficiaries they expect to attract, total out-of-pocket costs equaled 25% of total drug costs.6 In practice, this constraint means that plans can raise out-of-pocket costs for certain drugs as long as they lower them for others. Finally, a Part D insurer is likely to respond to diagnosis-specific profitability via its benefit design because enrollment in Part D is likely to respond to benefit design. Medicare beneficiaries know their diagnoses and will enroll in the Part D plan that offers them the most generous benefits. We know that beneficiaries know their diagnoses via a few types of evidence. Direct evidence shows that among Medicare beneficiaries, prescription drug utilization in one year is a very good predictor of prescription drug utilization in the next year (Soni, 2008, Hsu et al., 2009). Heiss et al. (2013) show in their analysis of Part D enrollment decision that individuals choosing a plan based purely on current year prescriptions choose the ex-post optimal plan more often than any of a number of rational expectations models they test, a result that follows from a high degree of persistence of exact drug purchases. In addition, the presence of substantial knowledge of one’s own drug utilization can be inferred from the adverse selection observed in the private prescription drug benefits that preceded Part D (Pauly and Zeng, 2004; Goldman et al., 2006).7 5 In this analysis I account for each plan’s premium, but since insurers have limited control over premium and certainly cannot set a diagnosis-specific premium, I do not study its response to diagnosis-specific risk adjustment. In Part D the premium for plan i is set to premi = (bidi − N AB) + γN AB where bidi and N AB are defined in Section 2.1 and γ is a fixed percentage (36% in 2009). Then plans with overall generous benefit designs (many covered drugs, low out-of-pocket costs) collect dollar-for-dollar higher premiums, while those with low costs can charge low premiums (subject to a zero lower bound). 6 To be specific, insurers first assumed the basic benefit and estimated resulting enrollment and drug utilization. Holding enrollment fixed, they then reestimate drug utilization under the alternative out-of-pocket costs. Insurers offering “enhanced” benefits reduce out-of-pocket costs by paying more than 75%, 0%, or 15% in each zone of coverage; enhanced plans collect the same government payments and finance the additional benefits through a supplemental premium. 7 Medicare Part D has been the setting of a substantial literature documenting deviations from rationality. Most relevant

6

3

Related Literature: Benefit Design and Risk Adjustment

Here, I describe the theoretical models of insurer benefit design that provide my hypothesis for how insurers should react to inaccurate risk adjustment. I then review literature on risk adjustment’s performance in theory and practice.

3.1

Insurer Incentives

A robust theoretical literature analyzes insurer incentives in a setting similar to Part D. These models, reviewed in Ellis (2008), assume that (1) individuals know the characteristics that determine their profitability (such as diagnoses), (2) insurers cannot deny enrollment or charge premiums based on those characteristics, and (3) beneficiaries vary (in a known way) in their preference for different medical services. Rather than an out-of-pocket cost paid in dollars, insurers in these models influence the utilization of medical services via a “shadow price” for each service that beneficiaries pay in time and hassle. The authors show that if the plan expects positive profits from a given type, the shadow price for services demanded by that type should be lowered. Jack (2006) places the model in a traditional Rothschild-Stiglitz framework, but the basic conclusions do not change. Insurers’ benefit design incentives can be neutralized in these models if an omniscient social planner pays each insurer the expected cost of each individual (Glazer and McGuire, 2002, 2000).

3.2

Risk Adjustment

A recently-revived applied literature assesses the assumptions required for successful risk adjustment. Geruso and McGuire (2015) provide a framework for analyzing tradeoffs contained in a risk adjustment system. Layton (2014) shows that risk adjustment’s success requires that those with high risk adjustment also demand more generous coverage, a requirement that standard risk adjustment assumes but does not assess. Risk adjustment under imperfect competition is explored in Mahoney and Weyl (2014), who show that adverse selection in a monopoly insurance market can reduce the monopolist’s quantity distortion; risk adjustment in such a case can actually reduce welfare under a range of plausible parameter values. Finally, Kifmann and Lorenz (2011) and Bijlsma et al. (2014) explore optimal risk adjustment under various assumptions about the nature of insurance or insurance demand. Medicare’s own risk adjustment system has produced a growing empirical literature studying its impacts. Exploiting the transition to diagnosis-specific risk adjustment in Medicare Advantage, Brown et al. (2014) show that Medicare Advantage plans successfully select for individuals who are inexpensive relative to the risk adjustment plans receive for them (i.e., individuals with milder forms of each diagnosis).8 This paper begins from the same premise – that inaccuracies in risk adjustment affect insurer behavior. In their case, risk to my paper are analyses of plan choice; early research suggested many enrollees were not choosing the best benefit design for their drug needs (Abaluck and Gruber, 2011). But newer research by Ketcham et al. (2015) suggests the original finding may be attributable to data issues, and that in any case Part D enrollees’ plan choices can be rationalized with brand preferences; in addition, choice behavior became more rational over time, suggesting such issues are less of a concern in my analysis of 2010 benefit designs. My analysis can accommodate errors in plan choice that are orthogonal to diagnoses’ profitability, since benefit design remains insurers’ best way to select among diagnoses. Another line of research documents the presence of switching costs (Ericson, 2014; Polyakova, 2015); in separate research I show how switching costs influence insurers’ response to risk adjustment recalibration (Carey, 2015b). Finally, several papers exploit enrollee behavior around the coverage gap to suggest Part D enrollees discount the future hyperbolically (Gowrisankaran et al., 2014; Einav et al., 2013); again, my benefit design hypotheses are not affected as long as diagnosis-specific profitability does not covary with discount rates. It is worth pointing out that cognitive diagnoses are not, as a group, distinctly profitable or unprofitable (although if individuals with cognitive diagnoses use health care proxies their decisionmaking would be unimpaired). 8 However, other research (McWilliams et al., 2012; Newhouse et al., 2013) finds weaker evidence that Medicare Advantage plans successfully selected those made profitable by risk adjustment. For example, Medicare Advantage insurers do not disproportionately enroll individuals with diagnoses made profitable by effective care coordination services.

7

adjustment is inaccurate because multiple “types” (i.e., mild and severe forms of a diagnosis) are associated with a single risk adjustment diagnosis. In my case, it is inaccurate due to the speed of technological change in prescription drugs. Brown et al. demonstrate that insurers differentially select for those who are inexpensive relative to risk adjustment. I demonstrate how insurers accomplish this selection: by setting more favorable benefits for the drugs taken by those with inaccurately high risk adjustment. Lavetti and Simon (2014) recognize that Medicare Advantage plans, who also pay for medical costs, have different incentives in the benefit design of their prescription drug benefit than free-standing Part D plans. They show that Medicare Advantage plans use their drug benefits to deter individuals with high medical spending; they also, however, cover drugs that can causally reduce medical costs more favorably. Finally, Decarolis (2013) and Miller (2016) find evidence that Part D plans are responding strategically to another form of government subsidy: the Low Income Subsidy (LIS). The LIS further subsidizes premiums and out-of-pocket costs for about 40% of Part D enrollees. Decarolis (2013) shows that Part D plans can manipulate the threshold at which plans can enroll LIS recipients and can also exploit the rules that automatically assign some LIS recipients to plans. Miller (2016) finds that plans with a large fraction of LIS recipients remit more profits to CMS via the Part D risk corridors; one potential explanation is that the uniform risk adjustment factor for LIS recipients is too high.

4

Conceptual Model

My hypothesis in this paper is that technological change makes risk adjustment inaccurate and that insurers design benefits to exploit those inaccuracies. In this section, I briefly discuss the two parts of this hypothesis and why it is likely to hold in Part D. My argument that technological change makes risk adjustment inaccurate is, to my knowledge, novel. The argument is straightforward for the onset of generic competition. The onset of generic competition generally reduces drug prices from monopoly levels to marginal cost (Berndt et al., 2011); since the same drugs can be purchased at lower costs, insurers’ expenditure will weakly fall,9 and (static) risk adjustment will exceed average treatment costs. I argue conversely that a new drug entrant will raise average treatment costs relative to risk adjustment. This argument holds as long as new entrants are more expensive than incumbent treatments (either brand or generic) and preferred by at least some patients. My argument that insurers will then design benefits in response to risk adjustment inaccuracies is derived from Frank et al. (2000). In a related theoretical paper (Carey, 2015a), I show how their argument analogizes to a setting like Part D. Beneficiaries are characterized by diagnoses that determine their valuation of different drugs. Beneficiaries evaluate plans’ benefits for the drugs that treat their diagnoses and enroll in the plan that allows them highest utility. Diagnosis-specific risk adjustment is unrelated to actual treatment costs. With a mild restriction on the derivatives of enrollment with respect to out-of-pocket costs, the model predicts that benefits are more favorable when diagnosis-specific risk adjustment is higher. Every prospective payment system will by its nature not fully compensate insurers for incurred liabilities; within diagnosis-based risk adjustment, there will be variance in spending even for individuals with the same diagnoses. The key requirement for the benefit design effects above is that profitability is (1) predictable to insurers and (2) correlated with demand for benefit design elements. Diagnosis-specific profitability is predictable to an insurer because it is driven by observable technological change or because it is observable in their own claims experience. Diagnosis-specific profitability is correlated with demand preferences because 9 Inasmuch as Huckfeldt and Knittel (2012)’s finding that firms successfully shift utilization to branded reformulations holds, the onset of generic competition could have no effect on average treatment costs.

8

drugs tend to treat a particular diagnosis, and benefits can be designed at the level of the drug under Part D rules.10 In Medicare Part D and other risk-adjusted markets, insurers purchase medical services from upstream providers with substantial market power. To my knowledge, little theoretical literature addresses how service prices should vary with diagnosis-specific profitability. If inaccurate risk adjustment is comparable to a tax, than it should partially pass-through to drug firms with market power (Weyl and Fabinger, 2013).11 This phenomenon could affect how an insurer structures benefit designs. If a drug firm with market power demands higher prices for drugs that treat profitable diagnoses, an insurer setting out-of-pocket costs as percent coinsurances may not be able to set low coinsurance amounts (the product of prices and coinsurance rates). The actuarial value regulation in Part D constrains plans using coinsurances to often set a coinsurance rate of 25%. Therefore, insurers facing powerful upstream drug firms may structure their benefits as flat copays, rather than coinsurances, for profitable drugs. Flat copays break the collinearity of prices and coinsurance amounts, which allows an insurer to simultaneously pay higher drug prices and set lower out-of-pocket costs. Frank et al. (2000) focus on a symmetric equilibrium in which all insurers set the same benefits for each service; in a symmetric equilibrium, all insurers receive equal shares of profitable and unprofitable diagnoses and obtain the same profits. If only some insurers respond to these incentives, than those insurers will obtain more individuals with profitable diagnoses and fewer with unprofitable diagnoses, leading to larger profits.12 Large insurers, for example, could more precisely observe diagnosis-specific profitability in their claims. Supposing my hypothesis holds, what are its welfare implications? The hypothesis states that, against an ideal of benefit designs set under perfectly accurate risk adjustment, individuals with profitable diagnoses will get more generous benefits than individuals with unprofitable diagnoses. If drug demand is inelastic, this results in a transfer from those with unprofitable diagnoses to those with profitable diagnoses. If drug demand is elastic, than those with profitable diagnoses will consume an inefficiently high quantity of drugs and those with unprofitable diagnoses will consume an inefficiently low quantity, both of which may have health consequences. However, these welfare impacts result from a laudable insurer effort to respond to technological change. That is to say, a finding that insurer benefit design did not respond to inaccuracies in risk adjustment would also be worrisome. Fundamentally I show that when generic competition begins and a diagnosis becomes more profitable, Part D enrollees benefit from lower out-of-pocket costs. In contrast, when new drugs enter and a diagnosis becomes less profitable, enrollees bear some of the costs of the advance in technology.

5

Measuring Technological Change, Risk Adjustment Accuracy, and Benefit Designs

In this section, I describe my measurement of technological change, risk adjustment, and benefit design strategies. I first demonstrate that diagnoses vary considerably in their exposure to technological change in the form of new drug entry and the onset of generic competition. Secondly, I compare risk adjustment for each diagnosis with actual treatment costs in Medicare claims data, showing that for many diagnoses risk adjustment greatly exceeded or was exceeded by average treatment costs. I then propose a method of 10 In fact, the prescription drug benefit of any insurance with diagnosis-based risk adjustment would be useful for screening for profitable or unprofitable diagnoses. Jacobs and Sommers (2015) suggest insurers in the ACA exchanges are using drug benefit design to deter AIDS patients. 11 I cannot examine drug prices directly because the price in the data is prior to a drug × plan rebate. 12 The Part D risk corridors (see Appendix) will reduce such insurers’ profits; if insurers are truly symmetric, the risk corridors will have no effect.

9

determining what drugs treat each diagnosis, again using the Medicare claims data. Finally, I describe the benefit designs that serve as my dependent variable.

5.1

Measuring Technological Change

In my setting, technological change in the form of new drug entry and the onset of generic competition exogenously shifts the profitability of enrolling an individual with a given diagnosis. I gather data on new molecules and new generics from the Food and Drug Administration.13 I use technological change between 2003 (the year following the Federal retiree data that was used to calibrate risk adjustment) and 2008 (the year preceding the claims data I use to measure the accuracy of risk adjustment), inclusive. I create two measures that characterize the exposure of a given diagnosis to new drug entry: the number of ingredients treating the diagnosis that entered between 2003 and 2008, and the expenditure among a 5% random sample of Medicare Advantage enrollees14 in 2009 on ingredients treating the diagnosis that entered between 2003 and 2008. The measure based on expenditure recognizes that a new drug entrant that is inexpensive or rarely used will not greatly affect profitability for the diagnosis it treats. I create two analogous measures for a diagnosis’s exposure to generic competition: the number of ingredients treating the diagnosis that began to face generic competition between 2003 and 2008, and the number of people in Medicare Advantage taking such ingredients in 2009. The importance of a new generic is measured by number of takers because many new generics are very inexpensive compared to their patent-protected version; total expenditure can be modest even for very popular generics. Figures I.a and I.b illustrate the exposure of the diagnoses in the risk adjustment scheme to technological change, i.e., the stock of technological change incurred since calibration. Each dot in the figures represents one or more of the 69 diagnoses we use in the analysis in Section 7. In the top panel, the vertical axis counts the number of ingredients treating each diagnosis that began to face generic competition, while the horizontal axis displays the number of ingredients that entered since calibration. Of the 69 diagnoses matched to ingredients, 41 have zero new molecules, while the remainder range from one new molecule to eight. Twenty-six diagnoses have no new generics, while the remaining 43 range from one new generic to 23. Dots are scaled by the number of diagnoses represented; the dot at the origin represents thirteen diagnoses with neither new entrants nor new generics. The bottom panel depicts each diagnosis’s exposure to technological change as measured by the takers of its new generics and the expenditure of its entrants; similarly, the dot at the origin represents thirteen diagnoses with no technological change. In both panels, I’ve labeled the diagnoses most exposed to technological change. Congestive Heart Failure was exposed to the onset of generic competition for an important and popular therapeutic class, ACE Inhibitors. Disorders of Lipoid Metabolism is unusual in its exposure to both significant new generics and new entrants. Two 13 While FDA data on new molecules is straightforward, there are two sources of measurement error in my identification of new generics. Firstly, FDA approval of new generics is not always synonymous with market entry. Generic manufacturers frequently apply for FDA approval while relevant patents are still in effect. The patent-holding pharmaceutical firm then sues the generic manufacturer to stave off market entry. I counter this by dropping ingredient combinations that the FDA reports as “new generics” that do not appear as generics in the 2009 Part D claims data. The second source of measurement error is introduced by the fact that drugs with the same ingredients vary in strength. The FDA reports approval of a new strength as a “new generic” even if the ingredients have previously been available as generics in other strengths. But the most dramatic price decreases occur when the most popular strengths are introduced right after patent expiry. This means I probably classify some ingredients as “new generics” that were introduced in a new strength between 2003 and 2008 but were in fact available before 2003 as generics in the most common strengths. I compared my list of new generics to those investigated in Ellison and Ellison (2011) and found, reassuringly, no overlap with their list of new generics from the 1980s and 1990s. 14 Drug-specific expenditures in Medicare Advantage are a good proxy for drug-specific expenditures in Part D because the populations are similar in disease burden; however, Medicare Advantage drug expenditures are less affected by the strategic incentives I explore in this paper because Medicare Advantage insurers receive the majority of their risk adjustment through a medical risk adjustment system (i.e., approximately $9000 dollars for medical costs vs. $1300 for prescription drug costs). Essentially, I expect MA plans’ incentives to be mostly orthogonal to the inaccurate risk adjustment I show in Section 7.1.

10

entrants, rosuvastatin (Crestor) and ezetimibe (Zetia), represented more than 400 million of expenditure in Medicare Advantage in 2009. As discussed in Section 2.1, the very popular simvastatin, among other statins, began to face generic competition in the post-calibration period.

5.2

Measuring Average Treatment Costs

This paper argues that technological change makes risk adjustment inaccurate. In this subsection, I determine the accuracy of risk adjustment payments by comparing them to the average treatment costs actually incurred by Medicare Part D plans. I measure average treatment costs for each diagnosis using the medical and prescription drug claims of a 5% random sample of Medicare beneficiaries. The medical claims allow me to observe whether an individual has each of the diagnoses in the risk adjustment system; using the same algorithm as Medicare, I assign diagnoses to beneficiaries on the basis of medical codes (ICD-9 codes) in their Inpatient, Outpatient, or Carrier (Physician) claims. For each beneficiary, I directly observe in the Part D claims her plan’s contribution to her total drug costs – i.e., her plan’s expenditure. Conceptually, I measure average annual treatment costs by predicting plan expenditure using diagnosis dummies; essentially, I am redoing the work of Robst et al. (2007) with an accurate sample of Part D demand (instead of demand from 2002). In order to isolate the portion of plan expenditure that diagnosis-based risk adjustment is meant to offset, I make three modifications to raw plan expenditure that account for three other types of subsidies in Medicare Part D: demographic-based risk adjustment, reinsurance, and overall inflation by bids. In the Appendix, I describe in detail how Part D plans are paid and explain each of these three modifications. What’s left is the portion of plan expenditure that corresponds to diagnosis-specific risk adjustment. The equation that recovers average treatment costs for each of the risk adjustment diagnoses is given below: fj = E

X

ωx Djx + εj

(2)

x

where

fj E

is modified plan expenditure for beneficiary j in dollars

ωx

is average treatment costs for each individual in dollars

Djx

is 1 if beneficiary j has diagnosis x, 0 otherwise

I estimate the above equation using OLS. If risk adjustment is accurate, risk adjustment payment Wx and average treatment costs ωx will be approximately equivalent. I define profitability of a diagnosis as Wx − ωx . When risk adjustment exceeds average treatment costs, the diagnosis is profitable for insurers; when average treatment costs are higher than risk adjustment, the diagnosis is unprofitable. Table I about here The top panel of Table I reports the characteristics of the beneficiaries used in estimating Equation 2 (top panel). The estimation sample is 848,780 beneficiaries enrolled in traditional Medicare (not Medicare Advantage) in 2008 and Part D in 2009. The first rows of Table I report the unmodified expenditure and total out-of-pocket costs (including the low-income cost-sharing subsidy, if applicable) incurred by these beneficiaries in 2009. The next row shows the number of diagnoses appearing in each beneficiary’s medical claims in 2008. The next rows show the distribution of annual payments received by these beneficiaries’ 11

insurance plans. I compute total payments by summing diagnosis-specific risk adjustment, demographicspecific risk adjustment, and reinsurance, and then report the decomposition into the first component and the fj . For individuals with very low expenditure, other two.15 The next row shows plans’ modified expenditure E the removal of demographic-specific risk adjustment makes modified expenditure negative. In the next line, I describe the distribution of diagnosis-specific risk adjustment minus modified expenditure, which varies widely. When this difference is negative, this beneficiary was (ex post) unprofitable and conversely when this difference is positive this beneficiary was (ex post) profitable. One explanation for the large disparity in profitability among beneficiaries is that the diagnosis-specific risk adjustment was wrong.

5.3

Linking Drugs and Diagnoses

I predict that insurers will design less generous benefits for drugs that treat unprofitable diagnoses. In order to test this prediction, I must specify which drugs treat each diagnosis. Unfortunately, no reference work links drugs to the risk adjustment diagnoses. Instead, I use the empirical association between ingredient combinations and diagnoses found in three years of Medicare beneficiaries’ medical and prescription drug claims. For each ingredient combination c, I run a probit model: X P r(Yjc ) = γxc Djx + ηj

(3)

x

where Yjc is equal to 1 if beneficiary j takes ingredient combo c and Djx are flags for 84 risk adjustment diagnoses.16 Each coefficient γx gives the increase in the probability of taking c associated with having the diagnosis x. Ingredient combination c treats the diagnosis x that has the largest γxc . An advantage of linking drugs and diagnoses using actual Part D data (rather than a reference work) is that I account for diagnosis-specific undercoding. A diagnosis is undercoded if health professionals do not tend to report it in medical claims. If the diagnosis is not coded in medical claims, Part D plans do not receive risk adjustment for it, even if the beneficiary takes related drugs. In the extreme case, if a diagnosis is never recorded, for drugs that treat it should not be affected by the diagnosis’s risk adjustment. Research suggests that undercoding varies by diagnosis and is particularly common for mental diagnoses such as depression (Townsend et al., 2012). In other settings, insurers sometimes encourage physicians to upcode individuals by reporting diagnoses the individuals do not have in order to activate diagnosis-specific risk adjustment (Geruso and Layton, 2015). In those settings (e.g., Medicare Advantage), insurers contract with physicians to provide services to their enrollees; an advantage of Part D is that no such contracts exist, making it harder for insurers to influence physicians. I estimate Equation 3 for every ingredient combination taken by at least 200 beneficiaries in 2007, 2008, and 2009, or 732 ingredient combinations. The estimation sample is every individual enrolled in free-standing Part D and fee-for-service Medicare in those years, or approximately three million beneficiary-years. Basic facts about the estimation sample can be found in the bottom of Table I. On average, the largest coefficient (i.e., the one for the treated diagnosis) exceeds the second largest coefficient by a factor of six, suggesting that in general the connection between ingredient combinations and diagnoses is relatively strong. The 732 ingredients or combinations evaluated treat 69 of the risk adjustment diagnoses. I exclude from the analyses in Section 7 the 15 diagnoses that, according to my linkage, are not treated by any ingredients. 15 As described in the Appendix, the total payment reported in Table I is the sum of annual premium and annual government subsidies. That is, CMS applies risk adjustment and reinsurance formulas to calculate the total payment to which the insurance plan is entitled; Table I reports the distribution across beneficiaries of this amount. Some portion of total payments is collected from the beneficiary premiums, and CMS contributes the remainder. 16 Robst et al. (2007) discuss 87 diagnoses, but two are constrained in the estimation of Equation 1 to have the same coefficient, so I combine their ICD9 codes into a single diagnosis, leaving 86. I drop two diagnoses, Cystic Fibrosis and Cystic Fibrosis & Age<65, because fewer than 200 (less than 0.02%) of the beneficiaries have these diagnoses.

12

Unmatched diagnoses are often catch-all categories, such as Other Neurological Conditions/Injuries or Other Blood Diseases, or diagnoses such as Pelvic Fracture where drugs are used for pain and infection but do not treat the actual diagnosis. I check my ingredients-diagnosis linkage against the Prescription Drug Morbidity Groups (RxMGs) of the Johns Hopkins Adjusted Clinical Groups Case-Mix System, a commercial product used by insurers to model expected liabilities. The ACG Case-Mix System predicts the “morbidity groups” an individual is likely to have based on prescription drug claims. The “morbidity groups” do not correspond perfectly to risk adjustment diagnoses, but many are very similar. Comparing an ingredient’s “morbidity group” under the ACG Case-Mix system to the diagnosis it treats under my linkage suggests that my linkage performs fairly accurately. When problems arise, it is because the risk adjustment diagnoses are narrowly defined. For example, while pain is a common complaint for many diagnoses, pain drugs tend to be linked to only a few (mostly Disorders of the Spine and Migraine Headaches). Meanwhile, contraceptive hormones tend to be linked to Vaginal and Cervical Diseases. I include pain and hormonal drugs in my drug sample despite spurious matching; results are strengthened when they are excluded.

5.4

Part D Benefit Designs

I collect my benefit designs from the publicly-available Prescription Drug Plan Formulary files for 2010.17 For each of 1550 Part D plans, the files contain the list of drugs covered by the plan. If a drug is covered, the out-of-pocket cost is also available. I characterize the generosity and structure of benefit designs in 2010 using the following variables: • Coverage • Out-of-pocket costs for a thirty day supply in the initial coverage zone • Copays (i.e., out-of-pocket costs when equal to a flat dollar amount) • Coinsurance amounts (i.e.,out-of-pocket costs when equal to the drug price multiplied by the coinsurance rate) • Tier (i.e., plans assign every covered drug to a “tier” where the copay or coinsurance amounts rise with the tier) • Whether the out-of-pocket cost is a coinsurance I focus on out-of-pocket costs in the initial coverage zone because, due to regulation, out-of-pocket costs are usually collinear with drug prices in other zones of coverage. In 2009, nearly two-thirds of beneficiaries only purchased drugs in the initial coverage zone and copays for purchases in the initial coverage zone represented 54% of total copays. While my profitability measure is annual I measure benefit design generosity via the out-of-pocket costs for a thirty day supply. If every person with a diagnosis took a drug for the diagnosis, and all diagnoses were chronic (i.e., every drug is purchased for twelve months), it would be appropriate to use a 360 day supply to compare with annual profitability. Not every person with a diagnosis takes a drug that treats the diagnosis, with the share who do so varying by diagnosis (as low as 1.5% of those with Quadriplegia, as high as 92% of those with Congestive Heart Failure). In 2010, the days supply for the median beneficiary × drug is 90 days (when demand is aggregated to the beneficiary × drug × year level, half the observations represent 90 days or less of supply), and the mean is 135 days. I base my arguments off of the sign of the relationship between profitability and out-of-pocket cost, not its magnitude. 17 Appendix

Table AI reviews the time series of data used for each step of the empirical analysis.

13

Table II about here Table II summarizes the 3611 drugs studied in this analysis, averaged across all Part D plans. The table is structured to compare drugs within quintiles of list price. The first panel describes drugs in the lowest quintile of list price, reporting four characteristics averaged across all Part D plans: coverage, monthly outof-pocket cost, out-of-pocket cost as a percentage of drug price, and total annual expenditure in Medicare Advantage enrollees (i.e., expenditure aggregated to the drug level in my 5% sample of MAPD drug claims, inflated by a factor of 20). At the median, these inexpensive drugs are covered by virtually all plans at an out-of-pocket cost of $6 (averaged across plans). When expressed as a percentage of price, the median outof-pocket cost is a bit under half the total price. Among drugs in this quintile of list price, total expenditure in Medicare Advantage on the median drug was about $416,000; the massive skewness of total expenditures is apparent in this row. Two patterns emerge from this table. Firstly, out-of-pocket costs are commonly far above or below 25% of negotiated prices. Inaccurate risk adjustment is not the only potential explanation, but clearly insurers are taking advantage of benefit design latitude. Secondly, while drugs with higher list prices are covered less often and at higher out-of-pocket cost, benefit designs are surprisingly similar across quintiles, suggesting that list price is not the primary factor. Table III about here As discussed in Section 4, profitability may affect benefit structure as well as generosity. Table III describes the plan × drug observations where out-of-pocket cost is a flat copay (top panel) and a coinsurance (bottom panel). Copays are generally used for cheaper generic drugs, while coinsurances are used for more expensive branded drugs. This is consistent with the theoretical predictions in Berndt et al. (2011), who show that coinsurances are the optimal benefit design for an insurer facing an upstream monopoly.

6

Empirical Strategy

In this section, I first explain my empirical model for relating technological change, profitability, and benefit designs. I then explain why instrumenting is necessary and why technological change is a good instrument.

6.1

Predicting Benefit Designs from Profitability

I predict benefit design outcomes for plans and drugs using the profitability implied by inaccurate risk adjustment. We transform profitability at the diagnosis level to a plan-drug measure Rd(x)i = (Wd(x) − i ωd(x) ) Nbid AB . I refer to Rd(x)i as the profitability in plan i of the diagnosis x that drug d treats, where the d(x)

subscript references the drug-diagnosis linkage. Rd(x)i is plan-specific because risk adjustment payments are linear in plan bids. A plan with a large bid who spends proportionally more on drugs would find a profitable diagnosis more profitable than a plan with a small bid. Rd(x)i is not plan i’s actual liabilities for diagnosis x; instead, Rd(x)i are the expected plan-specific profits accruing to i if an average individual with diagnosis x enrolls.18 I use least squares to test the empirical relationship of profitability Rd(x)i and benefit design outcomes Ydi . My main estimation equation is: Estimation Model: Yd(x)i = α

Rd(x)i | {z }

instrumented 18 Using

+δi +

+νdi

(4)

when Y=coverage

the same profitability for each plan Wd(x) − ωd(x) instead of scaling by

14

βT Cd | {z }

bidi , N AB

as in Table IX, does not change results.

A full set of plan fixed effects δi net out all plan-level patterns in benefit design such as overall generosity. Recall from Section 2.3 that plans must cover at least two drugs in each “therapeutic class.” When Ydi is coverage, I include dummies for each therapeutic class, T Cd , to control for the insurer’s choice set. νdi is clustered on drugs.19 Rd(x)i is a generated regressor (with known error reported in Table IV), but Pagan (1984) shows inference can proceed as normal under 2SLS estimation of Equation 4 or under OLS for the special case of testing the null α = 0.20 I weight each observation by the total expenditure for the drug among Medicare Advantage enrollees. Weighting is used because drugs vary in their total expenditure by a factor of hundreds of millions. If insurers and beneficiaries have limited resources, they will base their decision (benefit design or enrollment) on drugs accounting for more expenditure. In the framework of Solon et al. (2015), the use of expenditure weights recovers the average partial effect of profitability in the presence of unmodeled heterogeneity across total expenditure levels in how strongly Part D agents respond to incentives.21 I check for robustness to equal weights in Table VIII.

6.2

Instrumenting With Technological Change

This analysis uses technological change instruments to recover the relationship between profitability and benefit design. IV addresses three sources of bias: omitted drug quality, moral hazard, and measurement error due to of unobserved drug × plan rebates. The first two sources of bias are due to the fact that I am measuring profitability using claims data incurred under the same risk adjustment system I wish to assess. Drug quality affects both benefit design and profitability. Suppose that for a particular drug, perceived quality in Part D is higher than perceived quality in the calibration data (either because of differences in disease burden or updated clinical knowledge). Holding other variables fixed, insurers will set more favorable benefits – lower out-of-pocket costs – for higher-quality drugs to attract enrollees. Therefore, modified plan fj for individuals taking higher-quality drugs will be higher, raising treatment costs for their expenditure E diagnoses and lowering measured profitability. OLS will recover the true (negative) relationship between profitability and out-of-pocket costs plus the negative relationship between quality and profitability (Cameron and Trivedi, 2005). Instrumentation uses only the portion of profitability that is related to technological change, rather than the portion of profitability endogenously determined by benefit design. For out-of-pocket costs, IV estimates should be less negative than OLS estimates. Another source of bias arises from elastic drug demand. Suppose generic entry has made a particular diagnosis profitable. A plan that sets generous benefits to attract individuals with that diagnosis may find that, once enrolled, their utilization rises in response to the generous benefits. This response raises modified fj and the diagnosis’s average treatment costs, and lowers its measured profitability. plan expenditure E Conversely, reduced demand for drugs with high out-of-pocket costs for an unprofitable diagnosis would raise my measurement of profitability. The drug demand response can be interpreted as measurement error that negatively covaries with true profitability, such that measured profitability is biased towards zero. Under the assumption that measured profitability is at least half signal, OLS estimates can be biased in either direction (Bound and Krueger, 1991). As a matter of interpretation, the effect on benefit designs for a profitable diagnosis will balance the incentive to attract such individuals (lower OOPs) with the incentive 19 Clustering at the drug results in the most conservative (largest) standard errors among clustering at the drug, plan, or contract level. 20 A bootstrap analysis, available upon request, confirmed that accounting for the estimation error of R (x)i does not meand ingfully affect inference. 21 The authors demonstrate that, while ordinary least squares is always inconsistent in the presence of heterogeneous effects, weighted least squares only completely relieves the inconsistency when standard errors are homoscedastic. Note that my analyses cluster standard errors on drugs.

15

to moderate their demand response (higher OOPs). Each drug × plan can negotiate an unobserved rebate that I can not account for in modified plan fj . Therefore, profitability for each diagnosis is measured with error: the true profitability plus expenditure E an error term comprised of unobserved rebates. If rebates are classical measurement error, then two-stage least squares will remove the attenuation bias; for the case of out-of-pocket costs, IV estimates should be more negative than OLS estimates. If rebates are nonclassical and we assume that measured profitability is at least half signal, the effect of instrumentation again depends on whether rebates rise or fall with true profitability. Without research on the strategic role of rebates, it is hard to forecast their covariance with true profitability. Suppose a diagnosis has low true profitability, and the insurer wishes to set a high coinsurance amount for it while simultaneously paying a low post-rebate price. Since coinsurance amounts and pre-rebate price are linearly related, rebates make this combination of coinsurance amounts and post-rebate prices possible. Under that hypothesis, rebates fall with true profitability, but the direction of bias is unknown. Instrumenting ensures I am using only the portion of profitability attributable to technological change. Technological change is plausibly exogenous to risk adjustment. In a general sense, brand or generic drug firms making decisions about entrants or new generics would recognize that Part D represents only about 25% of US pharmaceutical demand. For the case of generic entry, most generic entry is pre-determined by the dates of long-ago patent filings, and so generic entry is unlikely to be determined by a diagnosis’s profitability. A challenge to the exogeneity of new entrants would arise if drug firms targeted drug development at profitable diagnoses (under the expectation that they could extract some profitability via drug price negotiations with Part D insurers). There are three reasons why such targeted development is unlikely to occur. Firstly, Blume-Kohout and Sood (2013) and Dranove et al. (2014) suggest an eight-year rule of thumb for the length of time between Phase I clinical trials and market entry; therefore, even if drug firms adjusted their investments immediately after risk adjustment levels were published in April 2005, it is unlikely that new entrants would be approved by 2008 (the last year of technological change data I use). Secondly, it could be difficult for drug firms in April 2005 to immediately recognize or anticipate that diagnoses would be profitable or unprofitable (delaying any investments and their results even further). Thirdly, drug firms would be unlikely to make investments in response to a risk adjustment system that they expect to be recalibrated relatively soon. To my knowledge, no research establishes general patterns of how benefit designs should respond to exposure to entrants or new generics, probably because drugs range from perfect substitutes to perfect complements for one another. One hypothesis is that entrants and new generics both increase competitive pressure on other drugs treating the diagnosis; if so, they should affect the benefit designs in the same direction. In Appendix Table AIII, I show the direct effect of the technological change instruments on benefit designs. Entrants and new generics generally have opposite signs, which is consistent with the theory on profitability.

7

Results

In this section, I first describe my results from estimating the profitability of each diagnosis. I then show how technological change relates to profitability, which is then the first stage for a 2SLS prediction of benefit designs from profitability. I investigate the relationship between benefit designs and profitablity across a number of subsamples and specifications.

16

7.1

Results: Measuring Diagnosis-Specific Profitability

I estimate Equation 2 on the 848,780 beneficiaries described in the top panel of Table I. The independent variables are flags for 84 risk adjustment diagnoses from 2008 and the dependent variable is modified plan expenditure from 2009. Table IV about here The results of estimating Equation 2 are clear: risk adjustment overpaid for certain diagnoses and underpaid for others. Table IV reports the results of Equation 2 for the 69 diagnoses that I link to drugs (i.e., the 15 unmatched diagnoses described in Section 5.3 are included in the regression but not reported in this Table IV). The second and third columns report the estimated treatment costs for each diagnosis: the coefficients (ωx ) of Equation 2 and their standard errors.22 Risk adjustment in the typical 2009 plan (Wx ) is reported in the next column, and then I subtract the treatment costs from the risk adjustment to compute the profitability of each diagnosis. For example, I find that beneficiaries with Multiple Sclerosis cost Part D plans $1185 on average, but they only received about one-third of that amount in diagnosis-specific risk adjustment. Therefore, Multiple Sclerosis is an unprofitable diagnosis. Lipoid Metabolism, discussed in Section 2.1, appears at the 70th percentile of profitability. Standard errors are small and the CI of treatment costs frequently exclude the payments. The average profitability across the 69 diagnoses is -$74 (SD $192), although a plan that enrolled a representative sample of beneficiaries (i.e., weighting each diagnosis by its prevalence) would approximately break even. Figures II.a and II.b about here The findings of Table IV are illustrated in Figures II.a and II.b. Figure II.a represents the risk adjustment Wx (solid circles) and estimated treatment costs ωx (open circles), of diagnoses reported in Table IV.23 Figure II.b depicts the profitability of each diagnosis (last column of Table IV). In each figure, diagnoses are sorted by risk adjustment level to show that diagnoses at a range of risk adjustment levels can be unprofitable.

7.2

Results: Technological Change and Profitability

In the next section we will instrument for profitability using two sets of technological change instruments. The bilateral relationship between each technological change instrument and profitability is illustrated in Figure III. Figure III about here Each marker represents a diagnosis, with the size of the marker showing the total expenditure on ingredients that treat the diagnosis in Medicare Advantage. The y-axis measures each diagnosis’s profitability (as reported in Table IV) and the x-axis measuring its exposure to a technological change instrument: a simple count of entrants or new generics in the upper figures, and the expenditure on new entrants and takers of new generics in the lower figures. The dotted line is the weighted least-squares line; the slope of this line will differ from the first stage reported in Table V for a number of factors: a single instrument vs. 22 Nothing constrains diagnosis-specific treatment costs to be positive, and negative estimates can arise through the cooccurrence of diagnoses. Most negative estimated treatment costs are statistically zero. 23 To preserve scale, the diagnosis HIV/AIDS is excluded from this figure; both payments and treatment costs for HIV/AIDS are much higher than for any other diagnoses.

17

two,24 a change in the unit of analysis from the diagnosis to drug × plan, and the presence of plan and class fixed effects. The positive relationship between new generics and profitability is apparent; the negative relationship between entrants and profitability is clear only for the number of entrants (top left). Therefore, an insurer setting benefit designs should expect higher profits from individuals with diagnoses exposed to many new generics, and conversely should expect lower profits from individuals with diagnoses exposed to many entrants.

7.3

Results: Profitability and Benefit Design Table V about here

Table V reports the results of estimating Equation 4. The first panel reports OLS estimates; the next panels report the results of instrumenting via two sets of technological change instruments. We find no significant effects on coverage (first column). The next column looks at the impact of profitability on overall out-of-pocket cost in dollars per thirty-day supply. We find the expected negative effect for OLS and with the expenditure and takers instruments (p value=0.06). In the next two columns, we separately test the response of flat copays and coinsurance amounts, both in dollars per thirty-day supply. Copays have a negative relationship with profitability in every specification, while the effect on coinsurances is negative but insignificant in both IV specifications. The next column shows that drugs that treat more profitable diagnoses are on lower tiers than drugs that treat less profitable diagnoses; an extra $500 in profitability reduces a drug’s tier by approximately one (i.e., from tier 4 to tier 3). Finally, we show that drugs that treat more profitable diagnoses are less likely to have a coinsurance (more likely to have a copay); an extra $100 in profitability reduces the probability that a drug has a coinsurance by approximately five percentage points. I find, therefore, different responses to profitability for the two types of out-of-pocket costs, as well as differences in the likelihood of an out-of-pocket cost being a coinsurance. One explanation of these results is that negotiated drug prices are also responding to profitability. Suppose firms that make drugs that treat profitable diagnoses demand higher prices; even if the plan still wants to attract individuals with these high-expected-profit diagnoses, coinsurances (the product of drug prices and coinsurance rates) may not fall. Plans may be more likely to charge copays for drugs that treat profitable diagnoses precisely because they have more control over copays than they do over coinsurances. In each IV panel I also report the results of the first stage regression. Columns have slightly different first stages because they differ in the number of observations (out-of-pocket cost is only observed for covered drugs, each out-of-pocket cost is either a copay or a coinsurance) In the full sample, each new entrant results in profitability falling by $57.64, while each new generic raises profitability by $13.82. Expenditure on new entrants always reduces profitability but the result is only different from zero in one first stage; each million takers of new generics that treat a given diagnosis raise its profitability by $13.87. The results in Table V show that (1) copays and out-of-pocket costs generally (although not in every specification) are lower for profitable diagnoses and (2) coverage has no statistical relationship with profitability. Together, these findings suggest that plans are reacting to profitability within the bounds of what Part D regulation allows. Recall that coverage regulations require plans to cover two drugs per therapeutic class, while the actuarial value regulations simply require that total out-of-pocket costs amount to 25% of expenditures. Plans wishing to set unfavorable benefits for drugs that treat unprofitable diagnoses still must 24 We use the instruments in pairs since diagnoses can be exposed to both types of technological change. In addition, Branstetter et al. (2014) find that generic entry may affect drug firm investments, leading to a co-determination of generic entry and lagged new entry. In Appendix Table AII, I instrument with each measure singly; the first stage is weaker, but results are consistent with Table V.

18

meet the two-drug coverage requirement for each therapeutic class (diagnoses and therapeutic classes are related but do not correspond perfectly). But the insurer is free to set high out-of-pocket costs for these drugs as long as out-of-pocket costs for other drugs (e.g., those that treat profitable diagnoses) are lowered. Given the success of the therapeutic class coverage regulation at limiting plans’ ability to act on benefit design incentives, policymakers should consider revising out-of-pocket cost regulations to require the 25% standard to be met within therapeutic classes. When risk adjustment is diagnosis-specific, regulation should acknowledge plans’ diagnosis-specific benefit design incentives. 7.3.1

Profitability’s Effect By Drug and Plan Subsamples

We now repeat the analysis of Table V for three drug subsamples, described in Table VI. The first subsample is 480 branded drugs not in one of the six “protected” therapeutic classes covered in Section 2.3, which represent approximately half of Part D spending. The next two panels describe 2587 generic and 544 drugs, representing 30% and 20% of spending, respectively. I test this analysis because we may expect somewhat different reactions to benefit design incentives for each subsample. Benefit designs for generic drugs may represent attempts to steer enrollees away from branded substitutes. Protected drugs are subject to strict coverage regulation, although plans still control their out-of-pocket costs.25 Tables VI and VII about here The relationship between profitability and benefit design varies across branded, generic, and protected drugs, as shown in Table VII. In each panel, I predict outcomes for all drugs of that type across all plans; profitability and technological change for each diagnosis are the same as what is used in Table V. Thus, the first panel assesses how plans design benefits for the branded drugs that treat profitable diagnoses. We find nearly no significant effects, except that they are slightly more likely to be covered (expenditure/takers IV). The results for generic and protected drugs are similar to the patterns we saw in the overall sample: for drugs that treat profitable diagnoses, the copays and tiers are lower, and the cost-sharing is less likely to be a coinsurance. Note that the 544 protected drugs treat (i.e., are linked to in Section 5.3) only 20 diagnoses; the limited variation in the first-stage dependent variable may explain why some F tests are dangerously low for the protected subsample. The finding that the benefit design for brands does not respond is, again, what we might expect if profitability affects drug prices. Branded drugs are supplied uncompetitively (monopolistically or oligopolistically for substitutes such as statins) and therefore their makers have the opportunity to extract profitability via higher drug prices. Generic drugs are supplied competitively, so the insurer can retain the profitability. The first stage coefficients for each drug subsample are reported in Appendix Table AIV. Online Appendix Table AV assesses heterogeneity across plans in response to benefit design incentives. This table repeats the analysis of Table V across high (≥25,000), medium, and low(<500) enrollment plans. Due to the concentration of Part D enrollment, 163 high-enrollment plans represent nearly two-thirds of Part D beneficiaries. We find no differences in the 2SLS coefficients across plan sizes. 7.3.2

Robustness Tables VIII and IX about here

Tables VIII and IX test the robustness of the results in Table V to bootstrapped standard errors, equallyweighted observations, a profitability measure uniform across plans, and a control for the co-occurrence of 25 Approximately

70% of protected drugs are covered by every plan; the remainder are largely branded drugs with exact

generic equivalents.

19

diagnoses. The bootstrap panel first demonstrates that the baseline model is robust to the exclusion of plan dummies, which are suppressed in the bootstrap due to computational demands. Although all models cluster the standard errors at the drug level, bootstrapping the dispersion of estimates under the null of α = 0 allows arbitrary error correlation. The bootstrap algorithm predicts benefit design outcomes for drugs using independent variables – profitability and its instruments – for other ingredients. For example, benefit design outcomes for cholesterol drug simvastatin are predicted from the profitability for reflux drug omeprazole, or the predicted profitability for omeprazole given its technological change. Significance stars are assigned when 10, 5, 1, or 0.1 percent of coefficients exceed the true coefficient in absolute value. The bootstrap standard errors are larger than the clustered standard errors, although the significance stars tend to persist. The next panel weights each plan × drug observation equally rather than weighting drugs by their expenditure in Medicare Advantage. Magnitudes are smaller, but the results persist. The next robustness check, second panel of Table IX simply predicts benefit design outcomes using a profitability measure uniform across plans instead of the profitability adjusted for the plan’s bid (see Equation 4). The final check, at the bottom of Table IX acknowledges that beneficiaries have multiple diagnoses. Empirically, unprofitable diagnoses have a slight tendency to co-occur in Part D, meaning a dollar of unprofitability as measured may result in more than a dollar of unprofitability as experienced by plans.26 P For each diagnosis x, I compute an “expected other profitability” measure: E(Prof)x = y6=x (Wy − ωy ) ∗ P (having y|having x). In words, for each other diagnosis y 6= x, we multiply the profitability of that diagnosis and the posterior probability of having that diagnosis conditional on having x. Controlling for “expected other profitability”, however, does not appear to substantially change my results.

8

Conclusion

Risk adjustment that conditions subsidies on enrollee type is essential when insurer discrimination is prohibited. In this paper, I first demonstrated that technological change after recalibration made Part D’s diagnosis-specific risk adjustment inaccurate. Diagnoses exposed to new drug entry were unprofitable for Part D insurers, while diagnoses exposed to the onset of generic competition were made profitable by static risk adjustment. Next, I provide empirical support for theoretical models of benefit design in a setting like Part D. I show that, consistent with the theory, insurers designed more favorable benefits, in particular lower out-of-pocket costs, for more profitable diagnoses. Insurers receiving risk adjustment now cover at least 75 million Americans, and yet the first diagnosisspecific risk adjustment system only began in 2004. Current risk adjustment systems were designed and calibrated without the benefit of much theoretical research on insurer incentives or empirical research on how this new example of “administered pricing” works in practice. Risk adjustment payments do not, for example, minimize insurers’ incentives to select nor maximize the welfare of enrollees; instead, the payments minimize the sum of squared errors in Equation 1. Among a number of open questions that remain for future research is how risk adjustment affects negotiations between insurers and upstream medical providers. For example, suppose risk adjustment for a given diagnosis is exogenously high relative to treatment costs. This high risk adjustment can affect both the threat points and the payoffs of a Nash bargain over prices for services that treat the diagnosis. How is an extra dollar in risk adjustment divided between service providers (who may get higher prices), enrollees (who may get improved benefits), and insurers (who are the residual 26 A

later iteration of Frank et al. (2000)’s model incorporates this possibility explicitly (Ellis and McGuire, 2007).

20

claimant)? How is does this answer depend on the market power of the service provider? Further research on risk adjustment could help improve risk adjustment policies and develop our understanding of market power in health care.

21

Appendix A

Subsidization Scheme for Medicare Part D Plans

A Part D plan i who enrolls beneficiary j will receive four types of payments for that beneficiary: Paymentsij = (DPij + GPij − premi ) +premi + RIij | {z } Direct Subsidy

where

DPj

is the diagnosis-specific risk adjustment for beneficiary j in plan i

GPij

is the demographic-specific risk adjustment for beneficiary j in plan i

premi RIij

is chosen plan i’s premium paid by each beneficiary is the government reinsurance payment for beneficiary j in chosen plan i

The Direct Subsidy contains the risk adjustment that the plan receives for beneficiary j. The diagnosisspecific portion is the sum of risk adjustment over the individual’s diagnoses, scaled up or down by the plan’s bid. DPij =

bidi X Wx Djx N AB x

Wx are the risk adjustment amounts (in dollars) described in Section 2.127 and Djx is 1 if beneficiary j has diagnosis x. These weights are inflated by the ratio of the plan’s bid to the national average bid for the year. Demographic-specific risk adjustment GPij are computed in the same way for the demographic categories described in Section 2.1. Insurance plans also receive premiums and reinsurance payments. Plan i’s premium premi is subtracted from the Direct Subsidy and then collected directly from beneficiaries. Reinsurance payments reimburse plans directly for 80% of plan liabilities in the catastrophic zone. The government makes two other payments to plans that I ignore in my empirical strategy. Firstly, because low-income beneficiaries pay reduced copays, the Low-Income Cost-Sharing Subsidy reimburses plans directly for the difference between the reduced copay and the plan’s stated copay. I can ignore this payment because plan expenditure is the same regardless of a beneficiary’s low-income status. Secondly, a Risk Corridor payment partially offsets the losses of any plan whose total expenditure exceeds its total receipts by five percent. If instead an plan’s total receipts exceed its expenditure by 5%, the insurer remits part of its profits to the government under the rules of the Risk Corridor. Since risk corridor payments apply at the plan level rather than the beneficiary level, I cannot incorporate them into my adjustments.

B

Adjustments to Raw Plan Expenditure

In the Medicare prescription drug claims, I observe the plan’s raw expenditure on each beneficiary. I make four modifications to this expenditure in order to isolate the portion of expenditure that corresponds to diagnosis-specific risk adjustment. 1. I subtract reinsurance payments. Reinsurance payments are simply calculated as 80% of the plan’s expenditure above the catastrophic threshold. 2. I subtract demographic risk adjustment. 3. I subtract supplemental premiums collected by plans offering enhanced benefits. The supplemental premium is paid by beneficiaries but not subtracted from the Direct Subsidy, therefore it does not 27 To

reiterate, Wx in this paper are the weights in Robst et al. (2007) times the 2009 national average bid and divided by the 2009 upcoding normalization factor.

22

drop out of total payments. 4. Finally, I divide by

bidi N AB .

This rescales plan expenditure by overall plan generosity to make expenditure

incurred by plans of varying generosity comparable. fj is the dependent variable in Equation 2. Modified expenditure E

C

Use of Datasources Over Time

The table below describes the timeline of data used to for risk adjustment and this paper’s measurement and estimation. The first column shows that the Part D risk adjustment calibration data comes from 2000 (disabled Medicaid beneficiaries used to represent disabled Medicare beneficiaries) and 2002 (Federal retirees used to represent elderly Medicare beneficiaries). I collect the cumulative technological change (entrants and new generics) between 2003 and 2008, inclusive. Medical and prescription drug claims for a 5% sample of Medicare beneficiaries are used to link drugs to diagnoses (Section 5.3). Risk adjustment payments are based on diagnoses from the previous calendar year, so in estimation of Equation 2, I use diagnoses from 2008 and modified expenditure from 2009. We call the resulting coefficient treatment costs for 2009. The difference between risk adjustment in 2009 and treatment costs in 2009 is profitability for 2009, which is the independent variable in the estimation of Equation 4. I use 2009 profitability to explain benefit design outcomes in 2010; finally, this equation is weighted by expenditure on drugs in Medicare Advantage in 2009. Table AI: Datasources Over Time

2000 2001 2002 2003 2004 2005 2006

Risk adjustment calibration data Medicaid

Stock of Tech. Change from FDA

Measuring Profitability

RHS: diagnoses LHS: drug choices RHS: diagnoses LHS: drug choices RHS: diagnoses LHS: drug choices

RHS: diagnoses

Effect of Profitability on Outcomes

Federal Retiree X X X X

2007

X

2008

X

2009

Linking Drugs to Diagnoses

2010

23

LHS: spending

RHS: profitability weights: MA spending LHS: outcomes

References Abaluck, Jason T. and Jonathan Gruber, “Choice Inconsistencies Among the Elderly: Evidence from Plan Choice in the Medicare Part D Program,” American Economic Review, 2011, 101 (4), 1180–1210. Berndt, Ernst R., Thomas McGuire, and Joseph P. Newhouse, “A Primer on the Economics of Prescription Pharmaceutical Pricing in Health Insurance Markets,” Forum for Health Economics and Policy, 2011, 14 (2). Bijlsma, Michiel, Jan Boone, and Gijsbert Zwart, “Competition Leverage: How the Demand Side Affects Optimal Risk Adjustment,” RAND Journal of Economics, 2014, 45 (4), 792–815. Blume-Kohout, Margaret E. and Neeraj Sood, “Market Size and Innovation: Effects of Medicare Part D on Pharmaceutical Research and Development,” Journal of Public Economics, 2013, 97, 327–336. Bound, John and Alan B Krueger, “The Extent of Measurement Error in Longitudinal Earnings Data: Do Two Wrongs Make a Right?,” Journal of Labor Economics, 1991, 9 (1). Branstetter, Lee, Chirantan Chatterjee, and Matthew J. Higgins, “Starving (or Fattening) the Golden Goose?: Generic Entry and the Incentives for Early-Stage Pharmaceutical Innovation,” September 2014. NBER Working Paper 20532. Brown, Jason, Mark Duggan, Ilyana Kuziemko, and William Woolston, “How Does Risk Selection Respond to Risk Adjustment? New Evidence from the Medicare Advantage Program,” American Economic Review, 2014, 104 (10), 3335–64. Cameron, A. Colin and Pravin K. Trivedi, Microeconometrics: Methods and Applications, Cambridge University Press, 2005. Carey, Colleen, “Pass-Through of Public Subsidies in Health Insurance Markets,” 2015. Available at http://www-personal.umich.edu/~careycm/Carey%20Subsidy%20Pass-Through%2020130710.pdf. , “A Time to Harvest: Evidence on Consumer Choice Frictions from a Payment Revision in Medicare Part D,” January 2015. Decarolis, Francesco, “Medicare Part D: Are Insurers Gaming the Low Income Subsidy Design,” American Economic Review, 2013, 105 (4), 1547–80. Dranove, David, Craig Garthwaite, and Manuel Hermosilla, “Pharmaceutical Profits and the Social Value of Innovation,” June 2014. Duggan, Mark and Fiona Scott-Morton, “The Effect of Medicare Part D on Pharmaceutical Prices and Utilization,” American Economic Review, 2010, 100 (1). , Patrick Healy, and Fiona Scott Morton, “Providing Prescription Drug Coverage to the Elderly: Americas Experiment with Medicare Part D,” Journal of Economic Perspectives, 2008, 22 (4), 69–92. Einav, Liran, Amy Finkelstein, and Paul Schrimpf, “The Response of Drug Expenditures to NonLinear Contract Design: Evidence from Medicare Part D,” August 2013. NBER Working Paper 19393.

24

Ellis, Randall P., “Risk Adjustment in Health Care Markets: Concepts and Applications,” in Mingshan Lu and Egon Jonsson, eds., Financing Health Care: New Ideas for a Changing Society, Wiley-VCH Verlag, 2008, pp. 177–222. and Thomas G. McGuire, “Predictability and Predictiveness in Health Care Spending,” Journal of Health Economics, 2007, 26, 25–48. Ellison, Glenn and Sara Fisher Ellison, “Strategic Entry Deterrence and the Behavior of Pharmaceutical Incumbents Prior to Patent Expiration,” American Economic Journal: Microeconomics, 2011, 3, 1–36. Ericson, Keither M. Marzilli, “Consumer Inertia and Firm Pricing in the Medicare Part D Prescription Drug Insurance Exchange,” American Economic Journal: Economic Policy, 2014, 6 (1), 38–64. Frank, Richard G., Jacob Glazer, and Thomas G. McGuire, “Measuring Adverse Selection in Managed Health Care,” Journal of Health Economics, 2000, 19, 829–854. Geruso, Michael and Thomas G. McGuire, “Tradeoffs in Design of Health Plan Payment Systems: Fit, Power, and Balance,” February 2015. NBER Working Paper No. 20359. and Timothy Layton, “Upcoding: Evidence from Medicare on Squishy Risk Adjustment,” May 2015. NBER Working Paper No. 21222. Glazer, Jacob and Thomas G. McGuire, “Optimal Risk Adjustment in Markets with Adverse Selection: An Application to Managed Care,” American Economic Review, 2000, 90 (4), 1055–1071. and

, “Setting Health Plan Premiums to Ensure Efficient Quality in Health Care: Minimum Variance

Optimal Risk Adjustment,” Journal of Public Economics, 2002, 4, 1055–1071. Goldman, Dana E., Geoffrey Joyce, Pinar Karaca-Mandic, and Neeraj Sood, “Adverse Selection in Retiree Prescription Drug Plans,” Forum for Health Economics and Policy, 2006, 9 (1). Gowrisankaran, Gautam, Christina Marsh, and Robert Town, “Myopia and Complex Dynamic Incentives: Evidence from Medicare Part D,” April 2014. NBER Working Paper 21104. Heiss, Florian, Adam Leive, Daniel McFadden, and Joachim Winter, “Plan Selection in Medicare Part D: Evidence from Administrative Data,” Journal of Health Economics, December 2013, 32, 1325– 1344. Hsu, John, Jie Huang, Vicki Fung, Mary Price, Richard Brand, Rita Hui, Bruce Fireman, William Dow, John Bertko, and Joseph P. Newhouse, “Distributing $800 Billion: An Early Assessment of Medicare Part D Risk Adjustment,” Health Affairs, 2009, 28 (1), 215–225. Huckfeldt, Peter J. and Christopher R. Knittel, “Pharmaceutical Use Following Generic Entry: Paying Less and Buying Less,” July 2012. NBER Working Paper 17046. Jack, William, “Optimal Risk Adjustment with Adverse Selection and Spatial Competition,” Journal of Health Economics, 2006, 25, 908–926. Jacobs, Douglas B. and Benjamin D. Sommers, “Using Drugs to Discriminate – Adverse Selection in the Insurance Marketplace,” New England Journal of Medicine, 2015, 372. 25

Kautter, John, Melvin Ingber, Gregory C. Pope, and Sara Freeman, “Improvements in Medicare Part D Risk Adjustment: Beneficiary Access and Payment Accuracy,” Medical Care, 2012, 50 (12), 1102– 1108. Ketcham, Jonathan, Nicolai Kuminoff, and Christopher Powers, “Choice Inconsistencies among the Elderly: Evidence from Plan Choice in the Medicare Part D Program: Comment,” 2015. Forthcoming, American Economic Review. Kifmann, Mathias and Normann Lorenz, “Optimal Cost Reimbursement of Health Insurers to Reduce Risk Selection,” Health Economics, 2011, 20, 532552. Lavetti, Kurt and Kosali Simon, “Strategic Formulary Design in Stand-Alone and Integrated Medicare Advantage Prescription Drug Plans,” June 2014. Layton, Timothy, “Imperfect Risk Adjustment, Risk Preferences, and Sorting in Competitive Health Insurance Markets,” February 2014. Mahoney, Neale and E. Glen Weyl, “Imperfect Competition in Selection Markets,” November 2014. NBER Working Paper No. 20411. McAdams, David and Michael Schwarz, “Perverse Incentives in the Medicare Prescription Drug Benefit,” Inquiry, 2007, 44 (2). McWilliams, J. Michael, John Hsu, and Joseph P. Newhouse, “New Risk-Adjustment System Was Associated With Reduced Favorable Selection In Medicare Advantage,” Health Affairs, 2012, 31 (12), 2630–2640. Miller, Aaron E., “Multiple Sclerosis: Where Will We Be in 2020?,” Mount Sinai Journal of Medicine, 2011, 78 (2), 268–279. Miller, Daniel P., “Risk Adjustment and Low Income Subsidy Distortions in Medicare Part D,” January 2016. Newhouse, Joseph P., “Reimbursing Health Plans and Health Providers: Efficiency in Production Versus Selection,” Journal of Economic Literature, 1996, 34 (3), 1236–1263. , J. Michael McWilliams, Mary Price, Jie Huang, Bruce Fireman, and John Hsu, “Do Medicare Advantage Plans Select Enrollees in Higher Margin Clinical Categories?,” Journal of Health Economics, 2013, 32, 1278–1288. Pagan, Adrian, “Econometric Issues in the Analysis of Regressions with Generated Regressors,” International Economic Review, 1984, 25 (1), 221–247. Pauly, Mark V. and Yuhui Zeng, “Adverse Selection and the Challenges to Stand-Alone Prescription Drug Insurance,” in David M. Cutler and Alan M. Garber, eds., Frontiers in Health Policy Research, Vol. 7, NBER Books, 2004, pp. 55–74. Polyakova, Maria, “Regulation of Insurance With Adverse Selection and Switching Costs,” American Economic Journal: Applied Economics, Forthcoming 2015.

26

Robst, John, Jesse Levy, and Melvin Ingber, “Diagnosis-Based Risk Adjustment for Medicare Prescription Drug Plan Payments,” Health Care Financing Review, 2007, 28 (4), 15–30. Solon, Gary, Steven J. Haider, and Jeffrey Wooldridge, “What Are We Weighting For?,” Journal of Human Resources, 2015, 50 (2), 301–316. Soni, Anita, “The Top Five Therapeutic Classes of Outpatient Prescription Drugs by Total Expense for the Medicare Population Age 65 and Older in the U.S. Civilian Noninstitutionalized Population, 2005,” Technical Report, Agency for Healthcare Research and Quality, Rockville, MD 2008. Statistical Brief #199. Townsend, Lisa, James T. Walkup, Stephen Crystal, and Mark Olfson, “A Systematic Review of Validated Methods for Identifying Depression Using Administrative Data,” Pharmacoepidemiology and Drug Safety, 2012, 21 (S1), 163–173. Weyl, E. Glen and Michal Fabinger, “Pass-Through as an Economic Tool: Principles of Incidence under Imperfect Competition,” Journal of Political Economy, 2013, 121 (3).

27

Table I: Summary Statistics on Medicare Beneficiaries Measuring Average Treatment Costs: Estimation Sample for Equation 2 Percentile of Distribution 25th 50th 75th 581.75 1808.86 3637.265 255.38 622.54 1726.395 3 5 8

95th 10152.08 4592.43 12

Payments Received by Beneficiary’s Insurance Plan Total Payments 406.23 Diag. Risk Adjustment 0.00 Other Payments 289.64

($) 781.69 379.95 353.73

1,072.24 646.76 409.78

1,435.01 934.01 482.86

4,503.59 1,441.22 3,451.31

Expenditures by Beneficiary’s Insurance Plan ($) Modified Expenditure -492.45 Risk Adjustment – Mod. Expenditure -1,347.03

-188.73 -470.06

619.06 140.31

1,280.40 602.16

2,080.98 1,123.90

N=848,780 beneficiaries Unmodified Expenditure Out-of-Pocket Costs # of Diagnoses

5th 0 0 0

Linking Drugs and Diagnoses: Estimation Sample for Section 5.3 Percentile of Distribution N=2,908,936 beneficiaries × years 5th 25th 50th 75th # of Diagnoses 0 3 5 8 # of Ingredient Combos Taken 0 2 7 11

95th 13 20

The top panel reports the distribution of measured and computed outcomes across 848,780 beneficiaries enrolled in traditional Medicare (not Medicare Advantage) in 2008 and Part D in 2009. The first rows show the percentiles of raw drug expenditure and out-of-pocket costs (copay or coinsurance, including low-income cost-sharing subsidy) for the sampled beneficiaries in 2009. The third row shows the number of diagnoses (of 84 possible) appearing on the beneficiaries’ medical claims. The next rows show the total payments received by beneficiaries’ Part D plans, decomposed into diagnosis-specific risk adjustment and other payments (demographic-specific risk adjustment and reinsurance). Within each individual diagnosis-specific risk adjustment and other payments sum to total payments, but because individuals can appear at different points in each distribution the median diagnosis-specific risk adjustment plus the median other payment does not necessarily equal the median total payment. The next rows show the modified expenditures incurred by beneficiaries’ Part D plans – i.e., total plan expenditures less other payments (see text) and the difference between diagnosis-specific risk adjustment and modified expenditure. Positive values in this row indicate that diagnosis-specific risk adjustment exceeded modified expenditure. The bottom panel reports on the distribution of diagnoses and the number of drug ingredients taken for the beneficiary×years used to link drugs and diagnoses. Beneficiaries in this sample were enrolled in traditional Medicare (not Medicare Advantage) and free-standing Part D in 2007, 2008, or 2009.

28

Table II: Part D Benefit Design and Expenditures Percentile of Distribution 5th 25th 50th 75th 95th Lowest Quintile of List Price % of Plans Covering 19 89 100 100 100 OOP as % of Price 17 29 45 79 215 Out-of-Pocket Cost ($) 4.86 5.26 5.69 7.47 38.65 MA Expenditure ($1000) 2 57 416 2,000 13,536

N=3611 drugs

% of Plans Covering OOP as % of Price Out-of-Pocket Cost ($) MA Expenditure ($1000)

11 12 5.39 3

66 20 6.41 84

96 31 10.98 472

100 66 38.25 2,714

100 143 77.28 17,726

% of Plans Covering OOP as % of Price Out-of-Pocket Cost ($) MA Expenditure ($1000)

8 10 5.63 4

17 24 14.82 39

70 39 45.57 340

96 61 72.86 2,127

100 169 82.03 16,062

% of Plans Covering OOP as % of Price Out-of-Pocket Cost ($) MA Expenditure ($1000)

8 10 5.85 5

15 21 34.19 64

59 34 61.08 395

95 58 77.44 3,085

100 122 86.27 37,306

% of Plans Covering OOP as % of Price Out-of-Pocket Cost ($) MA Expenditure ($1000)

8 11 7.53 11

17 19 52.22 164

67 27 80.00 1,029

98 43 99.84 6,217

100 106 647.87 62,725

This table summarizes the distribution of benefit design and expenditure share for 3611 drugs, averaged across 1550 Part D plans operating in 2010. Drugs are grouped into quintiles by “list price” (Wholesaler Acquisition Cost). The first row in each group reports the distribution of the percent of plans covering a given drug. The second and third rows first average OOP (for a 30 day supply) and OOP as a percentage of plan-specific price across plans, and then report the distribution across drugs in the list price quintile. The next row reports the dollar value (in thousands) of total expenditures on the drug in Medicare Advantage enrollees in 2009.

29

Table III: Part D Benefit Design Structure: Outcomes When Out-of-Pocket Cost is a Copay or Coinsurance OOP is copay: 2,950,041 plan × drug observations % branded 11

% generic 73

1 0.58

2 0.24

3 0.14

copay ($) price ($)

p5 2.00 5.93

p25 5.00 17.61

Share in Tier 4 0.03 p50 7.00 44.66

% protected 16

5 0.002

6 0.0000

p75 36.00 121.24

p95 85.00 421.80

OOP is coinsurance: 862,083 plan × drug observations % branded 22

% generic 56

1 0.32

2 0.25

3 0.23

coins. amount ($) price ($)

p5 2.17 9.22

p25 9.69 38.60

Share in Tier 4 0.15 p50 33.68 115.47

% protected 22

5 0.043

6 0.0031

p75 108.20 353.49

p95 790.54 2718.88

This table describes benefit design outcomes in 2010 for 3611 drugs and 1550 plans for covered drugs where out-of-pocket costs are a flat copay (top panel) or coinsurance (bottom panel; coinsurance amount=coinsurance rate × price). The first row describes the percentage of each drugs that are brands, generics, or protected (either brand or generic). Next, we show the distribution across six tiers. Finally, we show the percentiles of the distributions of out-of-pocket costs and negotiated drug prices.

30

Table IV: Diagnosis-Specific Payments and Treatment Costs Diagnosis Multiple Sclerosis Leukemia HIV/AIDS Psoriatic Arthropathy Major Organ Transplant Metastatic Acute Cancers Age<65 & Schizophrenia Schizophrenia Huntington’s Ds Dementia w/ Depression Age<65 & Other Major Psych. Dsrs Seizure Dsr & Convulsions Diabetes w/ Comps Inflamm. Bowel Ds Motor Neuron Ds/Atrophy Other Endocrine Psoriasis Chronic Renal Failure Lung Cancer Parkinson’s Ds Hepatitis Asthma and COPD Migraines Connective Tissue Dsr Urinary Obstruction Vascular Disease Severe Hematological Dsr Polyneuropathy exc. Diabetic Incontinence Infectious Ds Salivary Gland Ds Cellulitis & Skin Ds Larynx/Vocal Ds Open-angle Glaucoma Myocardial Infarction/Unstable Angina Fecal Incontinence Mononeuropathy/Abnormal Movement Other Psych. Kidney Transplant Bronchitis & Congenital Lung Dsr Bullous Dermatoses Rheumatoid Arthritis Vascular Retinopathy exc. Diabetic Polycythemia Vera Lipoid Metabolism Other Organ Transplant Other Major Psych. Dsr Other Upper Respiratory Ds Glaucoma and Keratoconus Esophageal Ds Chronic Skin Ulcer exc. Decubitus Quadriplegia Dsr of Spine ADD Other Spec. Endocrine Pulmonary Embolism & Thrombosis Pancreatic Ds Cerebral Hemorrhage/Stroke Ulcer & Gastro Hemorrhage Bone Infections Vaginal & Cervical Ds Congestive Heart Failure Osteoporosis Heart Arrhythmias Hypertension Empyema, Abscess, & Lung Ds Polymyalgia Rheumatica Muscular Dystrophy Opportunistic Infections

Treatment Coeff.($) 1184.96 1001.21 2530.55 710.83 587.05 527.61 705.65 467.88 275.36 395.90 336.14 249.51 366.44 281.31 252.95 183.72 176.22 171.88 138.00 379.41 165.28 217.27 154.80 114.46 72.88 53.15 125.74 91.75 110.96 80.52 58.60 56.28 32.27 159.12 139.26 52.69 73.22 121.45 203.50 42.41 46.97 185.29 52.17 73.59 139.36 60.80 132.21 58.82 43.96 143.38 22.13 19.80 104.00 208.93 17.60 -4.31 11.27 24.54 -11.92 -21.86 -16.76 185.28 48.73 27.34 147.07 -32.57 -49.57 -80.42 78.40

Costs (SE) (16.0) (57.3) (16.8) (30.8) (28.0) (10.3) (8.4) (15.1) (39.9) (10.4) (4.8) (6.2) (4.3) (12.6) (46.3) (5.8) (12.2) (11.2) (5.0) (10.1) (14.6) (3.3) (8.9) (10.3) (5.6) (3.6) (21.6) (5.9) (6.0) (10.2) (17.5) (4.2) (26.3) (4.9) (3.2) (21.4) (5.4) (13.7) (21.4) (4.7) (3.8) (6.8) (6.7) (26.6) (2.7) (21.1) (3.9) (3.7) (12.6) (3.2) (7.1) (10.5) (3.3) (20.1) (2.8) (7.3) (10.7) (3.7) (5.4) (11.7) (7.1) (4.7) (3.6) (4.0) (2.8) (35.4) (14.1) (46.6) (19.7)

Risk Adjustment ($)

Profitability ($)

333.90 273.28 1904.54 139.90 73.68 162.29 349.76 233.17 51.30 206.12 153.89 118.45 240.63 169.75 141.77 72.75 71.82 69.02 46.63 298.46 85.81 152.03 98.86 61.56 44.77 32.64 105.39 71.82 95.13 68.09 46.63 44.77 22.38 150.16 130.58 44.77 66.22 118.45 200.53 40.11 44.77 184.67 52.23 85.81 152.03 73.68 147.36 77.41 63.42 164.15 44.77 44.77 131.51 236.90 45.70 25.18 44.77 58.76 30.78 21.45 30.78 234.10 107.26 86.74 207.06 40.11 40.11 77.41 239.70

-851.06 -727.94 -626.02 -570.92 -513.37 -365.33 -355.90 -234.71 -224.06 -189.77 -182.24 -131.06 -125.80 -111.56 -111.19 -110.97 -104.41 -102.86 -91.36 -80.95 -79.47 -65.25 -55.93 -52.90 -28.11 -20.51 -20.34 -19.93 -15.83 -12.43 -11.96 -11.51 -9.89 -8.96 -8.69 -7.92 -7.00 -3.00 -2.98 -2.31 -2.20 -0.62 0.06 12.22 12.67 12.88 15.16 18.59 19.47 20.77 22.64 24.97 27.51 27.97 28.10 29.49 33.50 34.22 42.70 43.32 47.54 48.82 58.52 59.40 59.98 72.68 89.67 157.84 161.30

This table reports the results of the estimation of Equation 2 on 848,780 Medicare Part D enrollees in 2009. The dependent variable is annual plan expenditure modified for other Part D payments, and the independent variables are dummies that are one if the beneficiary had the diagnosis in 2008. Risk adjustment levels are for a plan bidding the national average in 2009. Profitability is the difference between risk adjustment for a typical plan and treatment costs. Only the 69 diagnoses used in later analyses are reported.

31

Table V: Effect of Profitability on Benefit Design in Medicare Part D covered (p.p)

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OOP is coinsurance? (p.p.)

OLS Profitability

0.0032 (0.0038)

-0.7183 (0.1428)***

-0.0227 (0.0042)***

-0.5637 (0.1601)***

-0.0017 (0.0002)***

-0.0519 (0.0064)***

-0.0320 (0.4509)

-0.0023 (0.0004)***

-0.0465 (0.0108)***

-35.47 (10.94)** 16.60 (1.75)*** 50

-33.39 (6.45)*** 13.79 (1.12)*** 84

-33.39 (6.45)*** 13.79 (1.12)*** 84

IV: Entrants and New Generics Profitability

Entrants New Generics F

-0.0046 (0.0113)

-0.3004 (0.2227)

-57.64 (11.91)*** 13.82 (2.08)*** 31

-33.39 (6.45)*** 13.79 (1.12)*** 84

-0.0509 (0.0082)*** First stage results -31.89 (3.77)*** 10.80 (0.68)*** 131

IV: Expenditure on Entrants and Takers of New Generics Profitability

0.0351 (0.0214)

-0.5245 (0.2876)+

$ on Entrants Takers of New Generics F

-0.01 (0.16) 13.87 (2.53)*** 15

-0.16 (0.10) 22.16 (1.91)*** 117

N

5,597,050

3,812,124

-0.0708 (0.0101)*** First stage results -0.23 (0.07)*** 17.25 (1.17)*** 131 2,950,041

-0.3845 (0.6093)

-0.0029 (0.0005)***

-0.0654 (0.0129)***

-0.17 (0.17) 26.37 (2.74)*** 72

-0.16 (0.10) 22.16 (1.91)*** 117

-0.16 (0.10) 22.16 (1.91)*** 117

862,083

3,812,124

3,812,124

This table reports the results of estimation of Equation 4 on each of 3611 drugs in 1550 Part D plans in 2010. The first panel reports OLS and the remaining report 2SLS with the indicated instrument. Analyses are weighted by the expenditure on the drug in Medicare Advantage. Plan dummies are always included. When the outcome is coverage, controls for therapeutic class are included. Standard errors (in parentheses) are clustered on drugs. †, *, **, and *** represent significance at 10, 5, 1, and 0.1 percent.

Table VI: Summary Statistics on Drug Samples Percentile of Distribution 5th 25th 50th 75th 480 Branded Drugs (Not Protected) % of Plans Covering 15.03 50.19 80.00 91.68 OOP as % of Price 16.26 25.59 32.08 51.89 Out-of-Pocket Cost ($) 9.82 38.91 54.20 72.56 MA Expenditure ($1000) 13 254 2,046 11,714

% of Plans Covering OOP as % of Price Out-of-Pocket Cost ($) MA Expenditure ($1000)

% of Plans Covering OOP as % of Price Out-of-Pocket Cost ($) MA Expenditure ($1000)

2587 Generic Drugs (Not Protected) 7.61 21.81 84.32 11.15 21.47 36.17 5.07 6.19 21.44 3 50 369 544 Protected Drugs 7.74 59.74 9.67 19.62 5.22 8.51 11 137

100.00 28.58 49.68 587

95th 100.00 94.01 375.36 83,299

99.94 67.83 71.95 1,960

100.00 165.91 85.41 14,762

100.00 43.75 81.31 4,682

100.00 89.07 250.24 39,872

This table summarizes the distribution of benefit design and expenditure, averaged across 1550 Part D plans operating in 2010, for various drug samples used in the estimation of Equation 4. The first row in each group reports the distribution of the percent of plans covering a given drug. The second and third rows first average OOP (for a 30 day supply) and OOP as a percentage of plan-specific price across plans, and then report the distribution across drugs. The next row reports the dollar value (in thousands) of total expenditures on the drug in Medicare Advantage enrollees in 2009. The next row reports the dollar value (in thousands) of total expenditures on the drug in Medicare Advantage.

Table VII: Effect of Profitability on Benefit Design in Medicare Part D by Drug Type

Brands covered (p.p) Profitability

0.0101 (0.0078)

Profitability

0.0024 (0.0232) 14

first stage F Profitability first stage F N

0.0624 (0.0225)** 22 744,000

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.5850 -0.0013 -0.3060 -0.0011 (0.1743)*** (0.0037) (0.2242) (0.0003)*** IV: Entrants and New Generics 0.4669 -0.0104 0.7351 -0.0001 (1.0518) (0.0068) (1.0412) (0.0009) 21 25 12 21 IV: Expenditure on Entrants and Takers of New Generics -0.1864 0.0108 0.3305 -0.0011 (0.6590) (0.0120) (0.7908) (0.0008) 34 35 16 34 523,890 330,201 193,689 523,890

OOP is coinsurance? (p.p.) -0.0429 (0.0118)*** -0.0105 (0.0358) 21 -0.0455 (0.0316) 34 523,890

Generics covered (p.p) Profitability

0.0001 (0.0055)

Profitability

-0.0095 (0.0069) 27

first stage F Profitability first stage F N

-0.0093 (0.0203) 17 4,009,850

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.0179 -0.0187 0.0157 -0.0008 (0.0151) (0.0081)* (0.0453) (0.0004)* IV: Entrants and New Generics -0.0442 -0.0403 0.0918 -0.0022 (0.0267)+ (0.0203)* (0.0728) (0.0009)* 33 37 17 33 IV: Expenditure on Entrants and Takers of New Generics -0.1433 -0.0972 -0.0509 -0.0044 (0.0301)*** (0.0246)*** (0.0580) (0.0010)*** 22 28 15 22 2,636,717 2,156,415 480,302 2,636,717

OOP is coinsurance? (p.p.) -0.0130 (0.0068)+ -0.0365 (0.0163)* 33 -0.0791 (0.0168)*** 22 2,636,717

Protected covered (p.p) Profitability

0.0068 (0.0030)*

Profitability

0.0101 (0.0123) 18

first stage F Profitability first stage F N

0.0269 (0.0232) 2 843,200

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.9898 -0.0283 -0.6534 -0.0018 (0.2775)*** (0.0075)*** (0.3600)+ (0.0006)** IV: Entrants and New Generics -0.1827 -0.0944 0.6332 -0.0026 (0.2676) (0.0228)*** (0.6016) (0.0006)*** 19 12 23 19 IV: Expenditure on Entrants and Takers of New Generics -4.7510 -0.1160 -5.5597 -0.0052 (2.4520)+ (0.0451)* (2.8315)+ (0.0016)*** 11 12 3 11 651,517 463,425 188,092 651,517

OOP is coinsurance? (p.p.) -0.0579 (0.0204)** -0.0634 (0.0214)** 19 -0.1365 (0.0479)** 11 651,517

This table reports the results of estimation of Equation 4 across three types of drugs in 1550 Part D plans in 2010. The first panel reports OLS and the remaining report 2SLS with the indicated instrument. Analyses are weighted by the expenditure on the drug in Medicare Advantage. Plan dummies are always included. When the outcome is coverage, controls for therapeutic class are included. Standard errors (in parentheses) are clustered on drugs. †, *, **, and *** represent significance at 10, 5, 1, and 0.1 percent.

34

Table VIII: Effect of Profitability on Benefit Design in Medicare Part D: Robustness to Specification

Bootstrap covered (p.p) Profitability BS SE BS p value Profitability BS SE BS p value Profitability BS SE BS p value

0.00014 (0.0039) [0.0138] 0.443 -0.00519 (0.0135) [0.0122] 0.580 0.03382* (0.0224) [0.0147] 0.030

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.81599** -0.02983*** -1.09177* -0.00188*** (0.159) (0.005) (0.214) (0.0002) [0.179] [0.007] [0.468] [0.0003] 0.003 0 0.033 0 IV: Entrants and New Generics -0.34513 -0.06467*** -0.4052 -0.00258*** (0.251) (0.010) (0.648) (0.0004) [0.360] [0.014] [1.008] [0.0005] 0.293 0 0.627 0 IV: Expenditure on Entrants and Takers of New Generics -0.53387 -0.08257*** -0.77871 -0.00290*** (0.291) (0.012) (0.664) (0.0005) [0.381] [0.016] [1.066] [0.0006] 0.170 0 0.423 0

OOP is coins.? (p.p.) -0.05500*** (0.007) [0.006] 0 -0.05742*** (0.013) [0.014] 0 -0.06918*** (0.014) [0.015] 0

No Weights covered (p.p) Profitability

Profitability 1st stage F Profitability 1st stage F

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.1489 -0.0031 -0.2582 -0.0005 (0.0322)*** (0.0016)+ (0.0712)*** (0.0001)*** IV: Entrants and New Generics -0.0118 -0.1192 -0.0122 -0.1309 -0.0009 (0.0108) (0.0355)*** (0.0036)*** (0.1004) (0.0002)*** 146 514 487 363 514 IV: Expenditure on Entrants and Takers of New Generics -0.0308 -0.185 -0.0143 -0.4561 -0.0009 (0.0111)** (0.0549)*** (0.0034)*** (0.1716)** (0.0001)*** 152 907 905 551 907 -0.0069 (0.0040)+

OOP is coins.? (p.p.) -0.0156 (0.0024)*** -0.0197 (0.0033)*** 514 -0.0213 (0.0032)*** 907

This table reports the results of estimation of Equation 4 on each of 3611 drugs in 1550 Part D plans in 2010. The data and specification are the same as Table V with the indicated change in specification. The first panel bootstraps the standard errors by sequentially predicting the benefit design outcomes using the dependent variables – profitability and its instruments – of a randomly selected ingredient. The second panel excludes the expenditure weights. Analyses are weighted by the expenditure on the drug in Medicare Advantage in the first and second panels. Plan dummies are included in the first and third panels. When the outcome is coverage, controls for therapeutic class are included. Standard errors (in parentheses) are clustered on drugs. †, *, **, and *** represent significance at 10, 5, 1, and 0.1 percent.

35

Table IX: Effect of Profitability on Benefit Design in Medicare Part D: Robustness to Variable Definitions

National Average Bid instead of bidi covered (p.p) Prof. if NAB

-0.0015 (0.0046)

Prof. if NAB

-0.0050 (0.0122) 31

F Prof. if NAB F

0.0381 (0.0232) 15

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.8002 -0.0265 -0.6241 -0.0019 (0.1590)*** (0.0048)*** (0.1712)*** (0.0002)*** IV: Entrants and New Generics -0.3267 -0.0565 -0.0337 -0.0025 (0.2422) (0.0091)*** (0.4757) (0.0004)*** 83 131 55 83 IV: Expenditure on Entrants and Takers of New Generics -0.5708 -0.0785 -0.4019 -0.0031 (0.3127)+ (0.0111)*** (0.6396) (0.0005)*** 117 132 81 117

OOP is coins.? (p.p.) -0.0574 (0.0070)*** -0.0505 (0.0118)*** 83 -0.0712 (0.0140)*** 117

Including “Expected Other Profitability” covered (p.p) Profitability E(Prof)

Profitability E(Prof) F Profitability E(Prof) F

0.0030 (0.0049) 0.0017 (0.0136) -0.0058 (0.0141) 0.0134 (0.0250) 40 -0.0058 (0.0174) 0.0134 (0.0294) 27

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.7668 -0.0287 -0.6786 -0.0019 (0.1673)*** (0.0038)*** (0.2179)** (0.0002)*** 0.6368 0.0657 1.659 0.0026 (0.4417) (0.0135)*** (1.0964) (0.0006)*** IV: Entrants and New Generics -0.3367 -0.0504 0.0556 -0.0024 (0.2131) (0.0082)*** (0.4513) (0.0004)*** 0.1289 0.0821 0.4384 0.0031 (0.4874) (0.0152)*** (1.3720) (0.0008)*** 102 144 68 102 IV: Expenditure on Entrants and Takers of New Generics -0.4618 -0.0527 -0.3695 -0.0024 (0.2155)* (0.0076)*** (0.4627) (0.0003)*** 0.2766 0.0838 1.1452 0.0031 (0.4926) (0.0149)*** (1.4231) (0.0008)*** 118 163 74 118

OOP is coins.? (p.p.) -0.0554 (0.0066)*** 0.0459 (0.0172)** -0.0526 (0.0107)*** 0.0426 (0.0218)+ 102 -0.0548 (0.0097)*** 0.0452 (0.0211)* 118

This table reports the results of estimation of Equation 4 on each of 3611 drugs in 1550 Part D plans in 2010. The data and specification are the same as Table V with the indicated change in variable definitions. The first panel uses the national average bid rather than the plan’s actual bid in calculating profitability. The second panel creates an “expected other profitability” to account for the correlation among diagnoses. Analyses are weighted by the expenditure on the drug in Medicare Advantage. Plan dummies are always included. When the outcome is coverage, controls for therapeutic class are included. Standard errors (in parentheses) are clustered on drugs. †, *, **, and *** represent significance at 10, 5, 1, and 0.1 percent.

36

25

Figure I.a

Number of New Generics since Calibration 5 10 15 20

Congestive Heart Failure

0

Age<65 & Other Major Psychiatric Disorders

0

2 4 6 Number of Entrants since Calibration

8

Figure I.b Takers of New Generics since Calibration (millions) 0 5 10

Congestive Heart Failure

Lipoid Metabolism

0

100 200 300 Expenditure on Entrants since Calibration (millions)

400

These figures illustrate the exposure of diagnoses in the Part D risk adjustment to technological change, i.e., the stock of technological change since calibration. Each marker represents one or more of the 69 diagnoses we use in later analyses; markers are scaled to the number of diagnoses represented. In the top panel, the vertical axis counts the number of ingredients treating the diagnosis that began to face generic competition between 2003 and 2008, while the horizontal axis counts the number of ingredients treating the diagnosis that entered the market between 2003 and 2008. In the bottom panel, the vertical axis measures the number of individuals, in thousands, taking a new generic that treats the diagnosis in Medicare Advantage in 2009. The horizontal axis measures the expenditure, in millions, on market entrants in Medicare Advantage in 2009.

37

Figure II.a: Diagnosis-Specific Payments and Treatment Costs (2009)

Payment

0

Dollars 500

1000

Treatment Costs

Diagnoses (Increasing Payment Level −−>)

Profitability (Payment − Treatment Costs) Dollars −800 −600 −400 −200 0

200

Figure II.b: Diagnosis-Specific Profitability: Payments Minus Treatment Costs (2009)

Diagnoses (Increasing Payment Level −−>)

These figures illustrate the results reported in Table IV. In the top panel, the solid dots represent the annual diagnosis-specific risk adjustment for the typical 2009 plan in dollars for each diagnosis (69, excluding HIV/AIDS to preserve scale) treated by a sample drug (see text). The open dots represent the estimated annual treatment costs in 2009 for each diagnosis. In the bottom panel, bars represent the difference between diagnosis-specific payments and treatment costs in dollars. Positive values indicate that payments exceed treatment costs. Diagnoses are sorted by level of payment. 38

200 −800

Profitability (Payment − Treatment Costs) Dollars −600 −400 −200 0

200 Profitability (Payment − Treatment Costs) Dollars −600 −400 −200 0 −800

8

0

100 200 300 Expenditure on Entrants since Calibration (millions)

400

0

5

10 15 20 Number of New Generics since Calibration

25

Profitability (Payment − Treatment Costs) Dollars −600 −400 −200 0 −800

−800

Profitability (Payment − Treatment Costs) Dollars −600 −400 −200 0

200

2 4 6 Number of Entrants since Calibration

200

0

0

5 10 Takers of New Generics since Calibration (millions)

15

Figure III: Profitability and Technological Change: Bilateral Relationships These figures show the bilateral relationship between profitability and each of the four measures of technological change. Each marker is a diagnosis, and the size of the marker is the total expenditure on ingredients treating the diagnosis in Medicare Advantage. The y-axis is the diagnosis’s profitability (as reported in Table IV). The x-axis shows technological change between 2003 (after calibration) and 2008 (before the claims data used to measure profitability). The dashed line shows the weighted least squares regression line.

39

Table AII: Effect of Profitability on Benefit Design: Each Technological Change Instrument Singly

IV: Entrants covered (p.p) Profitability

Entrants F

-0.0103 (0.0179) -34.12 (10.08)*** 11

out-of-pocket cost ($) 0.2031 (0.5599)

copay ($) 0.0079 (0.0315) First Stage -14.58 (3.71)*** 15

-12.96 (6.64)+ 4

coinsurance ($) -0.5388 (0.9029) -15.61 (11.73) 2

0.0016 (0.0020)

OOP is coins? (p.p.) 0.0584 (0.0583)

-12.96 (6.64)+ 4

-12.96 (6.64)+ 4

tier

IV: New Generics

Profitability

0.0102 (0.0330)

out-of-pocket cost ($) -0.4471 (0.3302)

New Generics

4.91 (0.81)*** 36

9.45 (1.41)*** 45

covered (p.p)

F

copay ($) -0.0808 (0.0173)*** First Stage 6.74 (0.83)*** 66

coinsurance ($) 0.1404 (0.6581) 11.59 (2.47)*** 22

tier -0.0034 (0.0007)*** 9.45 (1.41)*** 45

OOP is coins? (p.p.) -0.077 (0.0179)*** 9.45 (1.41)*** 45

IV: Expenditure on Entrants covered (p.p) Profitability

$ on Entrants F

0.2179 (0.0768)** 0.16 (0.17) 1

out-of-pocket cost ($) -3.7528 (19.1437)

copay ($) -0.1765 (0.1445) First Stage -0.09 (0.06) 2

0.01 (0.10) 0

coinsurance ($) 10.1602 (34.9727) 0.04 (0.18) 0

tier 0.0116 (0.1023) 0.01 (0.10) 0

OOP is coins? (p.p.) -0.6785 (3.5075) 0.01 (0.10) 0

IV: Takers of New Generics

Profitability

0.0371 (0.0216)+

out-of-pocket cost ($) -0.5379 (0.2854)+

Takers of New Generics F

13.71 (3.31)*** 17

20.8 (2.58)*** 65

covered (p.p)

copay ($) -0.0624 (0.0123)*** First Stage 15.43 (1.59)*** 94

coinsurance ($) -0.2528 (0.5790) 24.83 (3.95)*** 40

tier -0.0028 (0.0005)*** 20.8 (2.58)*** 65

OOP is coins? (p.p.) -0.0679 (0.0143)*** 20.8 (2.58)*** 65

This table reports the results of estimation of Equation 4 on each of 3611 drugs in 1550 Part D plans in 2010. Each panel uses a different instrument in 2SLS estimation. Analyses are weighted by the expenditure on the drug in Medicare Advantage. Plan dummies are always included. When the outcome is coverage, controls for therapeutic class are included. Standard errors (in parentheses) are clustered on drugs. †, *, **, and *** represent significance at 10, 5, 1, and 0.1 percent.

40

Table AIII: “Reduced Form” of IV: Direct Regression of Technological Change Instruments on Benefit Design out-of-pocket coinsurance OOP is covered (p.p) copay ($) tier cost ($) ($) coinsurance? (p.p.) Entrants 0.3607 4.4930 0.9523 7.7807 0.0333 0.3975 (0.7096) (7.7405) (0.3939)* (16.3698) (0.0200)+ (0.6208) New Generics -0.0057 -4.8078 -0.6654 0.5273 -0.0362 -0.7792 (0.1877) (3.6456) (0.1044)*** (8.3765) (0.0063)*** (0.1777)***

Takers of New Generics

0.0348 (0.0111)** 0.0664 (0.3018)

out-of-pocket cost ($) 0.0500 (0.1380) -11.60 (6.7929)+

N

5,597,050

3,812,124

covered (p.p) $ on Entrants

copay ($) 0.0254 (0.0079)** -1.1655 (0.1771)*** 2,950,041

coinsurance ($) 0.5101 (0.3613) -11.029 (16.4003) 862,083

0.0006 (0.0003)+ -0.0641 (0.0105)***

OOP is coinsurance? (p.p.) 0.0039 (0.0084) -1.4449 (0.2976)***

3,812,124

3,812,124

tier

This table uses the same empirical model as Table V but reports the reduced form of instruments on the benefit design outcomes directly. Analyses are weighted by the expenditure on the drug in Medicare Advantage. Plan dummies are always included. When the outcome is coverage, controls for therapeutic class are included. Standard errors (in parentheses) are clustered on drugs. †, *, **, and *** represent significance at 10, 5, 1, and 0.1 percent.

41

Table AIV: First Stage for Table VI

First Stage for Brands out-of-pocket coinsurance copay ($) tier cost ($) ($) IV: Entrants and New Generics -85.38 -4.48 -28.58 4.93 -4.48 (40.59)* (15.42) (8.18)*** (17.03) (15.42) 19.2 10.63 8.85 14.61 10.63 (6.56)** (1.76)*** (1.27)*** (3.15)*** (1.76)*** 14 21 25 12 21 IV: Expenditure on Entrants and Takers of New Generics 1.64 0.23 -0.21 0.35 0.23 (0.92)+ (0.27) (0.13) (0.29) (0.27) -12.04 18.87 13.29 26.14 18.87 (11.85) (3.70)*** (2.07)*** (5.76)*** (3.70)*** 22 34 35 16 34

covered (p.p) Entrants New Generics F $ on Entrants Takers of New Generics F

OOP is coins? (p.p.) -4.48 (15.42) 10.63 (1.76)*** 21 0.23 (0.27) 18.87 (3.70)*** 34

First Stage for Generics out-of-pocket coinsurance copay ($) tier cost ($) ($) IV: Entrants and New Generics -89.65 -30.02 -31.72 -23.69 -30.02 (16.73)*** (8.78)*** (8.48)*** (10.89)* (8.78)*** 17.18 8.76 8.96 7.54 8.76 (2.62)*** (1.17)*** (1.17)*** (1.44)*** (1.17)*** 27 33 37 17 33 IV: Expenditure on Entrants and Takers of New Generics -0.12 -0.18 -0.18 -0.19 -0.18 (0.38) (0.07)* (0.07)** (0.09)* (0.07)* 11.41 12.85 12.75 13.46 12.85 (2.79)*** (1.92)*** (1.72)*** (2.43)*** (1.92)*** 17 22 28 15 22

covered (p.p) Entrants New Generics

$ on Entrants Takers of New Generics F

OOP is coins? (p.p.) -30.02 (8.78)*** 8.76 (1.17)*** 33 -0.18 (0.07)* 12.85 (1.92)*** 22

First Stage for Protected out-of-pocket coinsurance copay ($) tier cost ($) ($) IV: Entrants and New Generics -30.51 -51.81 -26.32 -76.15 -51.81 (7.55)*** (9.14)*** (5.54)*** (13.34)*** (9.14)*** 20.58 57.51 29.56 73.29 57.51 (7.12)** (12.58)*** (7.31)*** (17.29)*** (12.58)*** 18 19 12 23 19 IV: Expenditure on Entrants and Takers of New Generics 0.24 -0.71 -0.35 -1.05 -0.71 (0.29) (0.27)** (0.17)* (0.49)* (0.27)** -51.92 103.48 82.86 38.71 103.48 (28.85)+ (28.91)*** (16.70)*** -49.85 (28.91)*** 2 11 12 3 11

covered (p.p) Entrants New Generics F $ on Entrants Takers of New Generics F

OOP is coins? (p.p.) -51.81 (9.14)*** 57.51 (12.58)*** 19 -0.71 (0.27)** 103.48 (28.91)*** 11

This table reports the first stages that correspond to the IV analyses in Table VI. †, *, **, and *** represent significance at 10, 5, 1, and 0.1 percent.

42

Table AV: Effect of Profitability on Benefit Design in Medicare Part D By Plan Enrollment

High Enrollment Plans covered (p.p) Profitability

0.0028 (0.0033)

Profitability

-0.0048 (0.0078) 31

first stage F Profitability first stage F N

0.0202 (0.0212) 15 588,593

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.8100 -0.0254 -0.6463 -0.0017 (0.1631)*** (0.0048)*** (0.1893)*** (0.0002)*** IV: Entrants and New Generics -0.3299 -0.0478 -0.0747 -0.0021 (0.2500) (0.0082)*** (0.5123) (0.0004)*** 85 133 51 85 IV: Expenditure on Entrants and Takers of New Generics -0.5795 -0.0718 -0.4637 -0.0027 (0.3240)+ (0.0105)*** (0.6893) (0.0005)*** 116 130 70 116 440,285 346,088 94,197 440,285

OOP is coins? (p.p.) -0.0564 (0.0067)*** -0.0489 (0.0114)*** 85 -0.0685 (0.0137)*** 116 440,285

Medium Enrollment Plans covered (p.p) Profitability

0.0039 (0.0039)

Profitability

-0.0056 (0.0113) 31

first stage F Profitability first stage F N

0.0333 (0.0217) 15 3,809,605

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.7239 -0.0232 -0.5680 -0.0017 (0.1434)*** (0.0043)*** (0.1594)*** (0.0002)*** IV: Entrants and New Generics -0.3048 -0.0535 -0.0362 -0.0023 (0.2237) (0.0086)*** (0.4485) (0.0004)*** 84 131 51 84 IV: Expenditure on Entrants and Takers of New Generics -0.5291 -0.0728 -0.3852 -0.0029 (0.2890)+ (0.0104)*** (0.6053) (0.0005)*** 116 130 72 116 2,599,966 2,013,276 586,690 2,599,966

OOP is coins? (p.p.) -0.0525 (0.0063)*** -0.0478 (0.0111)*** 84 -0.0666 (0.0132)*** 116 2,599,966

Low Enrollment Plans covered (p.p) Profitability

0.0045 (0.0046)

Profitability

-0.0018 (0.0132) 31

first stage F Profitability first stage F N

0.0469 (0.0213)* 15 1,198,852

out-of-pocket cost ($)

copay ($)

coinsurance ($)

tier

OLS -0.6654 -0.0203 -0.5167 -0.0017 (0.1332)*** (0.0037)*** (0.1511)*** (0.0002)*** IV: Entrants and New Generics -0.2736 -0.0446 0.0021 -0.0023 (0.2071) (0.0074)*** (0.4298) (0.0004)*** 82 128 49 82 IV: Expenditure on Entrants and Takers of New Generics -0.4855 -0.0644 -0.3448 -0.0030 (0.2664)+ (0.0093)*** (0.5823) (0.0005)*** 119 132 72 119 771,873 590,677 181,196 771,873

OOP is coins? (p.p.) -0.0487 (0.0066)*** -0.0414 (0.0103)*** 82 -0.0599 (0.0121)*** 119 771,873

This table reports the results of estimation of Equations 4 across three samples of plans: 163 plans with more than 25,000 enrollees, 1055 plans with between 500 and 25,000 enrollees, and 332 plans with fewer than 500 enrollees. In each panel, the dependent variables are a binary coverage measure or, if covered, the copay or copay as a percentage of list price for each drug in 1550 Part D plans in 2010. The first results are OLS and the remaining are 2SLS with the indicated instruments. Plan dummies are always included. When the outcome is coverage, controls for therapeutic class are included. Standard errors (in parentheses) are clustered on drugs. †, *, **, and *** represent significance at 10, 5, 1, and 0.1 percent.

43