Strategic Formulary Design in Medicare Part D Plans

The design of Medicare Part D causes most beneficiaries to receive fragmented health insurance, with drug and medical coverage separated. Fragmentation is potentially inefficient since separate insurers optimize over only one component of healthcare spending, despite complementarities and substitutabilities between healthcare types. Fragmentation of only some plans can also lead to market distortions due to differential adverse selection, as integrated plans may use drug formularies to induce enrollment by patients that are profitable in the medical insurance market. We study the design of insurance plans in Medicare Part D, and find that formularies reflects these two differences in incentives.

A growing share of all public expenditures on health insurance in the US is paid to private companies that deliver public health insurance benefits. Medicare Advantage (MA), the private alternative to traditional Medicare, and Medicare’s Part D prescription drug coverage together cause about one-third of all Medicare expenses to be delivered entirely by private companies. Similarly, Medicaid is highly dependent on delivery by private managed-care organizations in most states (KFF, 2014). Although a frequently cited goal of privatization is increased efficiency, in the case of Medicare Part D the split between public and private provision of benefits has also introduced several strategic opportunities for insurers that have the potential to reduce efficiency and increase total costs.

First, the decision to privately deliver Part D prescription drug benefits led to fragmentation of insurance for enrollees in traditional Medicare. These beneficiaries receive public hospital and physician insurance, but their private drug insurance is delivered by firms that have no incentive to consider the complementarity or sub-stitutability between different classes of medical treatment options when designing insurance plans. A minority (about 30 percent) of beneficiaries are enrolled in MA plans, and have integrated private insurance that bundles together hospital, physician, and prescription drug coverage, creating an incentive for insurers to internalize spillovers across these categories. Empirical evidence from other insurance settings—including Goldman and Philipson (2007); Chandra, Gruber, and McKnight (2010); and Fendrick et al. (2001)—suggests that spillovers between drug and medical spending exist, and may be quite large. In fact, internalizing such spillovers was a key stated reason for the creation of Medicare Part D. 1

Second, the fragmentation of some plans but not others can lead to market distortions due to differential adverse selection. In addition to any selection incentives faced uniformly by Part D plans, MA Part D (MAPD) plans are incentivized to design their formularies to induce enrollment by beneficiaries from whom they expect to earn profits in the Medicare hospital and physician insurance market (Parts A & B), while stand-alone Part D plans are not. This creates selection incentives that differ across plans competing in the same market. In contrast to the relatively coarse cost-sharing rules in medical insurance, cost-sharing decisions in drug plans are made at the drug level, 2 providing a mechanism to precisely alter the attractiveness of a plan for individuals with conditions that are associated with specific drugs.

In this paper we study how the decision to add Part D benefits into the existing Medicare benefit structure by offering beneficiaries a choice between integrated MA plans or fragmented insurance affected the ways in which private insurers designed Medicare plans to take advantage of these two relationships between drug and non-drug medical spending. We test whether integrated plans choose more generous formulary rules for drugs taken by patients that tend to be profitable in Medicare Parts A & B, empirically evaluating the type of selection-based insurance design distortions described by Rothschild and Stiglitz (1976). 3 In addition, we test whether integrated plans set cost-sharing rules in ways that internalize spillovers between drug and non-drug medical spending. For example, if the use of albuterol inhalers has the potential to reduce hospitalizations among asthma patients, MA plans should have stronger financial incentives to set low copayments to ensure that their enrollees have access to inhalers, since MA plans are liable for hospital costs whereas stand-alone Part D drug plans (SAPDs) are not. This difference in incentives between plan types has potential implications for the efficient allocation of medical spending across major classes of medical care. Considering these two incentive differences jointly could be important to the extent that there is tension between them.

Many papers have studied adverse selection in health insurance markets, including theoretical discussions such as Rothschild and Stiglitz (1976) and empirical studies such as Handel (2013) and Polyakova (2015). Until very recently, however, few studies had empirically tested how adverse selection affects the design of insurance plans. Lustig (2010) studies how adverse selection affects the generosity of coverage in Medicare Advantage. In contrast, we study the mechanism itself that is likely to lead to beneficiary selection: whether plans design insurance formularies differently in response to the potential for enhancing advantageous selection. Our study is related to Carey (2017), which finds evidence of plan design distortions in response to selection incentives within stand-alone Part D, as opposed to selection into MA.

Advantageous selection by MAPD plans is a well-known concern, and Medicare has several policies aimed at limiting selection. First, MA plans are required to be guaranteed issue, which prevents plans from overtly selecting beneficiaries by declining some applicants. However, guaranteed issue does not eliminate selection since plans can strategically design their benefits to induce non-random self-selection, or advertise to targeted audiences. In addition, Medicare uses risk-adjustment of payments, with the conceptual goal of equating the expected profit of each potential enrollee. In practice, however, risk-adjustment does not completely eliminate selection incentives either. For example, Brown et al. (2014) find that when Medicare changed the risk-adjustment formula to account for differences in the average costs of treating medical conditions, MA plans simply changed their strategy from selecting beneficiaries with the lowest total cost to selecting the lowest-cost beneficiaries conditional on medical diagnoses. As a result, they find that these efforts to improve the risk-adjustment formula had little effect on overpayments to MA plans. Carey (2017) also shows that the separate risk-adjustment formula used for Medicare Part D has systematic errors caused in part by technological change over time, so that risk-adjustment does not neutralize selection incentives in Part D.

To test whether integrated MAPD formularies are designed to advantageously select beneficiaries with conditions that are profitable on the hospital and physician insurance segment, we use data from the universe of fee-for-service (FFS) Medicare beneficiary claims from 2008–2010. We first create a measure of average potential profits to MA plans from selection for each medical condition. Following the approach developed by Brown et al. (2014), we calculate the total expenditures of all beneficiaries enrolled in FFS Medicare in each year. We then use the full population of Medicare beneficiaries who subsequently switch to MA plans, calculate what the MA capitation payment would have been if the individual had been in an MA plan instead, and compare this to the actual FFS expenditures of that individual. Since the risk-adjustment formula is designed to equate expected profits across medical conditions, but is estimated using only FFS enrollees, any systematic variation in the difference between capitation payments and FFS expenditures by medical condition suggests that beneficiaries who switch into MA plans are not randomly selected from the pool of FFS beneficiaries.

Next, we use claims data from the population of FFS Medicare beneficiaries, including diagnosis codes and prescription drug purchases, to estimate the relationships between each medical condition and each prescription drug active ingredient. These joint probability distributions are used to calculate the expected risk-adjusted selection incentive of MA plans by drug active ingredient, 4 and test whether active ingredients taken by individuals with more profitable conditions are covered more generously by MAPDs than by SAPDs. Throughout the paper, we use the term ‘MA Switcher Surplus’ to refer to this expected risk-adjusted difference between counterfactual costs and revenue for each drug active ingredient.

Our second primary hypothesis is that MAPDs more generously cover drugs that can causally reduce medical spending. This hypothesis is based on the body of literature documenting substantial spillovers between drug and non-drug healthcare spending, especially among the elderly and individuals with chronic conditions. Chandra, Gruber, and McKnight (2010) find that 20 percent of the savings from increasing copayments for prescription drugs and physician visits are offset by increases in hospital costs, and 43 percent of the savings are offset among patients with chronic illnesses. The implication is that Part D formulary decisions could have substantial effects on both prescription drug spending and other medical spending.

To test whether drugs with medical spending offsets are covered more generously by MAPD plans, we first isolate the set of drugs where spillovers are most likely to occur. We use several alternative definitions that have been developed previously for similar purposes, each of which is based on information from medical experts. The first set of definitions comes from Chandra, Gruber, and McKnight (2010), who assembled a team of physicians and pharmacists to create drug groups for this purpose. They define spillover drugs as those that “if not taken, will increase the probability of an adverse health event within the year.” 5 We also use the list from Tamblyn et al. (2001), who define classes of “essential” drugs as “medications that prevent deterioration in health or prolong life and would not likely be prescribed in the absence of a definitive diagnosis.” We rely on medical expertise in constructing these lists because we are not aware of an alternative statistical approach for quantifying the causal effects of drug consumption on subsequent medical costs for every drug. One limitation with this approach is a loss of precision from using a binary measure of spillover incentives.

Although MAPD and SAPD plans have similar generosity levels on average, we show the average masks considerable differences in generosity across classes of drugs used to treat different medical conditions. For example, an MA beneficiary purchasing Ace Inhibitors, Beta Blockers, or Coronary Vasodilators would pay a 32–38 percent higher share of the total costs on average than a beneficiary in an SAPD would pay, while the same share would be about 11 percent lower for Antipsychotic and Antimanic drugs.

These coverage differences are not random. We find that a one-standard-deviation increase in a drug’s MA switcher surplus ($151 annually) is associated with beneficiaries paying on average 7.4 percent less out-of-pocket for the same product (NDC) in MAPDs relative to SAPDs during open enrollment, when the selection incentive is strongest. We also provide direct evidence that the effects are driven by a roughly equal combination of differences in formulary tier choices and differences in cost-sharing rules across tiers. Moreover, these differences remain about the same when comparing MAPD and SAPD plans owned by the same parent organization, suggesting the generosity differences are strategic, rather than due to information differences or firm-level design strategies.

Consistent with evidence of spillover effects between drugs and other medical care, we also estimate that MA plans more generously cover drugs that causally reduce medical spending. Out-of-pocket costs are about 6 percent lower in MAPD plans for CGM Spillover drugs, and about 8 percent lower for Tamblyn Essential drugs. Moreover, the selection and spillover incentives interact in ways that are consistent with intuition. We show that MAPD plans set higher cost-sharing during open enrollment for spillover drugs that have low switcher surplus, but then increase generosity after open enrollment when the selection risk subsides and the spillover incentive prevails. All of these effects persist when we compare generosity across plans and within drug class, or within exact NDCs. MA plans are also less likely to use non-price formulary hurdles like prior authorization requirements and step therapy restrictions on CGM Spillover drugs.

A related study by Starc and Town (2016) extends our work on spillover effects; using reimbursement rate discontinuities as an instrument for enrollment in MA plans, this study estimates that MA enrollment causes beneficiary spending on drugs to increase, and that the effects are largest for CGM drugs. The study’s primary goal is to estimate a structural model to evaluate alternative policy scenarios, such as forcing SAPD plans to internalize spillover effects, which they estimate would cause drug spending to increase by 13 percent in SAPDs. Our paper instead focuses on the supply-side responses to spillover incentives in formulary design and also incorporates selection incentives.

We study insurer behavior in responding to these Part D incentives when designing plans, and not on the resulting consumer behavior or net welfare effects. 6 The welfare effects of plan integration are ambiguous, because despite the selection effects we identify clearly increasing costs for Medicare, MA plans also internalize spillover incentives, potentially reducing total costs. In addition, since all MA plans have similar incentives, competition for consumers with profitable medical conditions may cause some of the potential rents from selection to be transferred back to consumers in the form of enhanced insurance benefits. 7

I. Medicare Part D Background

Several important institutional details and regulations governing plan design and reimbursement in Medicare Part D affect how incentives could possibly manifest in plan formularies. We provide a brief description of the relevant Medicare Part D rules, but more detailed explanations can be found in MedPAC (2005 in MedPAC (2006a), Hoadley and Simon (2010), and in the public law itself (U.S. Congress, 2003).

A. Plan Design

Beginning in 2006, Medicare Part D has provided prescription drug insurance to Medicare beneficiaries, 8 who face a choice between traditional Medicare with a private stand-alone drug plan or integrated private coverage of all medical and drug care through an MA plan. MAPDs and SAPDs receive a subsidy for each beneficiary to whom they provide drug coverage, and the subsidies are risk-adjusted based on the demographics and diagnosed illnesses of the beneficiary. Plans are forbidden from declining to insure anyone eligible. As of 2015, 15 million people received drug coverage through MAPD plans and 23.5 million through SAPDs (Hoadley et al. 2015).

Part D insurance is heavily subsidized, although beneficiaries pay some out-of-pocket costs in addition to monthly premiums. In the 2010 standard benefit structure set by Medicare, beneficiaries pay the first $310 of annual drug costs (the deductible), then a 25 percent coinsurance on the next $2,520 spent, then 100 percent of the cost for the next $3526 (the “doughnut hole”), and 5 percent of all costs beyond that (the “catastrophic zone”).

Although there was very active debate about the design of the standard benefit structure, 9 the law allows Part D plans a great deal of freedom in designing formularies, which was done to encourage private sector competition. By 2015, zero SAPDs and only 1 percent of MAPDs (enrollment weighted) offered the standard drug benefit (Hoadley et al. 2015), suggesting that plans have been very active in designing formularies. Most Part D plans have four coverage tiers, with the first, second, and third tiers having sequentially higher cost-sharing requirements, and a fourth specialty tier reserved for “very high cost and unique items” (CMS 2007). 10 Generally, plans can choose which drugs are listed on their formularies (i.e., covered at all), on which tiers drugs are listed, how out-of-pocket costs are assigned to each tier, and drug-specific non-price hurdles, such as whether prior authorization is required. Once a formulary has been offered for sale during the open enrollment period, between Oct. 15 and Dec. 7 of each year, insurers are generally not allowed to make the formulary more restrictive thereafter without written approval from CMS (CMS 2008, section 30.3.3.1). However, plans may increase the generosity of coverage without prior approval. In Section III.C we discuss empirical evidence on within-year changes in plan formularies.

Plans are allowed to deviate from the standard benefit design provided they follow certain regulations. First, the alternative cost-sharing structures must be actuarially equivalent to the standard plan, and must be “in accordance with standard industry practices” (CMS, 2007). Plan formularies must also include at least two drugs in each therapeutic category (see CMS 2007 for details of categories), and must include substantially all drugs in six key therapeutic classes, although there is no restriction on how generously each drug must be covered. Plans are also forbidden from designing formularies that discriminate against those with costly medical conditions (Hoadley, 2005), although it is not known how these requirements are audited. The rules governing formulary design apply equally to MAPDs and SAPDs, so the regulations themselves should not generate differences in formulary design. The existence of these rules simply limits the degree to which MA plans can respond to the differential economic incentives they face.

B. Selection Incentives

Concerns about adverse selection are present in nearly every insurance market. These concerns can be especially heightened when risks are systematically correlated, as they are in the context of prescription drugs, where demand is highly autocorrelated from one year to the next. Selection may be heightened in prescription drug insurance markets as individuals often have private information that allows them to better predict future demand for drugs than for other types of medical care. 11 In fact, Pauly and Zeng (2004) find that adverse selection problems may be so heightened in stand-alone prescription drug insurance that this market would not exist unless plans were subsidized, as they are under Part D, or bundled with other coverage to create a more comprehensive insurance product with less persistent spending, as is the case with MAPDs. Recent data show empirical evidence that insurers feared adverse selection when designing plans in the MA market prior to Part D. Lustig (2010) finds that insurers responded to adverse selection by constraining plan design, leading to a 14.5 percent reduction in the economic surpluses created by MA between 2000–2003.

Several factors reduce the incentive to strategically select beneficiaries. Starting in 2007 all reimbursements to MA plans became fully risk-adjusted, the culmination of a multi-year phase-in of a system known as the hierarchical condition category system (MedPAC 2006b), whereas prior rates were only adjusted for geographic and demographic factors. 12 However, there is evidence that even after condition-based risk-adjustment was fully introduced in 2007, MA plans are still able to profit by favorably selecting healthier patients conditional on risk scores. 13 MedPAC (2006b) explains that these risk-adjustment changes were intended in part to encourage participation by private insurers in traditionally under-served rural areas of the country. Whereas SAPD plans operate regionally, with 34 regions in the country, most MAPD plans operate at the county level, and are able to enter the market in urban counties without necessarily entering nearby rural counties.

Within Part D there are several features that attenuate selection incentives for both MA and SA plans. First, Medicare pays Part D insurers on a risk-adjusted basis using a formula that accounts for the average difference in drug spending among FFS Medicare beneficiaries with different medical conditions. Carey (2017) finds that there are substantial imperfections in this risk-adjustment formula. For MA plans, Part D payments (and the risk-adjustment formula) are separate from medical insurance payments.

Second, Part D has risk-corridors to prevent plans from earning excessive profits or losses, and a reinsurance feature that subsidizes plans if any enrollee’s costs exceed a certain threshold. During our study period, the impact of risk-corridors was small. Since 2008, Part D plans have received risk-corridor payments that subsidize 50 percent of losses whenever drug spending exceeds plan bids by 5 to 10 percent, and cover 80 percent of losses that exceed bids by more than 10 percent. The risk-corridors are symmetric, imposing a large effective tax rate on profits that exceed 5 percent of bids. Nonetheless, these corridors leave sufficient opportunity for typical profit margins that are observed in markets without risk-corridors. For example, a 2013 report by CMS estimated that, on average, private insurance companies in the US earned profit margins of about 5.3 percent in small group markets and 3.8 percent in large group markets. 14 Consistent with this pattern, MedPAC (2015) reports that in 2013, insurers with profits greater than 5 percent paid CMS a total of $737 million as a result of this tax, but 70 percent of these payments came from just two parent organizations, suggesting that most plans are within the normal industry profit range. 15 Reinsurance payments, however, affect many insurers. Although these payments were fairly stable during the time period we study, in subsequent years reinsurance payments have rapidly increased, and represented 37 percent of all Part D expenditures by 2015. 16

Conceptually, both risk corridors and reinsurance attenuate the responsiveness of profits to selection, although they do not eliminate the incentive for plans to maximize profits. Thus, these risk-protection features may make the selection incentive we study more important to MA insurers relative to selection within Part D alone. They may also lead MA insurers to be more aggressive in pursuing spillover incentives, knowing that they are protected against large losses caused by high drug spending.

The ability to induce self-selection through strategic formulary design also depends on how sensitive beneficiaries are to differences in generosity when choosing insurance plans. The evidence from the literature on choice inconsistencies, including Abaluck and Gruber (2011), suggests that consumers were far less responsive to out-of-pocket costs than they were to plan premiums in the first year of Medicare Part D. However, Ketcham et al. (2012) find that overspending on out-of-pocket costs fell by 55 percent in subsequent years, suggesting that consumer learning may have increased responsiveness of plan choice to generosity. They also find direct evidence that the potential savings associated with switching plans greatly increased the probability of switching.

If consumers are sufficiently responsive to cost-sharing rules when choosing insurance plans, they presumably must also respond to costs when making drug purchase choices. There is a large body of evidence consistent with this, although there is variation in estimates. Duggan and Scott Morton (2010) estimate the price elasticity of demand for prescription drugs under Medicare Part D to be −0.38, Lichtenberg and Sun (2007) estimate it to be about −0.7, and Einav, Finkelstein, and Polyakova (2016) estimate the average elasticity across drugs in Part D to be −0.24, with substantial drug-level heterogeneity. Non-price formulary hurdles, such as prior authorization requirements or quantity limits, have also been shown to be important predictors of drug use and spending. 17

II. Conceptual Framework

To help clarify our empirical objects of interest, we start with a basic theoretical description of the selection and management incentives. For SAPD plans, profit maximization entails choosing a premium bid, which affects the calculation of federal premium subsidies, and choosing the coinsurance rates of each drug. Although Part D formulary coverage schedules tend to be nonlinear (in that coinsurance rates generally change as a function of total spending), we abstract by considering the average share of a drug’s total cost that is covered by the insurer, r_d. The SAPD profit function is:

where P (r_d) is the monthly plan premium paid by beneficiaries, which depends on the generosity of the plan’s coverage of the set of d drugs, S(r_d) is the monthly federal subsidy payment, which depends on P, c(r_d) is the cost of insuring a beneficiary, and Q is the number of enrollees, which may depend on the plan premium, federal subsidies, and plan generosity.

Consider the decision over the generosity of a single drug with index d = 1. The SAPD plan’s FOC is:

[ ∂ P ( r d ) ∂ r 1 + ∂ S ( r d ) ∂ r 1 − ∂ c ( r d ) ∂ r 1 ] Q + [ P ( r d ) + S ( r d ) − c ( r d ) ] ∂ Q ∂ r 1 = 0

In contrast, the profit function for an MA plan includes both drug and medical components. Consider the MA plan’s problem that takes into account interactions with medical profits.

max P , r d [ P ( r d ) + S ( r d ) − c ( r d ) + MAR ( r d ) − MAc ( r d ) ] Q ( P , S , r d , MAR , MAc )

where MAR(r_d) is the average risk-adjusted revenue that an MA plan receives for Part A and B coverage, which could depend on the drug formulary generosity insofar as formulary design affects the composition of enrollees, and MAc(r_d) is the average non-drug medical cost of enrollees that choose the plan. The difference between these terms is the selection incentive that MAPDs face, but SAPDs do not.

A similar decision over the generosity of coverage of an arbitrary drug with index d = 1 is determined by the FOC:

[ ∂ P ( r d ) ∂ r 1 + ∂ S ( r d ) ∂ r 1 − ∂ c ( r d ) ∂ r 1 + ∂ MAR ( r d ) ∂ r 1 − ∂ MAc ( r d ) ∂ r 1 ] Q + [ P ( r d ) + S ( r d ) − c ( r d ) + MAR ( r d ) − MAc ( r d ) ] ∂ Q ∂ r 1 = 0

There are two sets of terms that cause an MAPD plan’s decision over r₁ to potentially differ from an SAPD’s decision. The first is ∂ MAc ( r d ) ∂ r 1 Q , which captures the spillover effect between drugs and the cost of medical treatment. For example, if choosing to generously cover asthma inhalers decreases the probability that an enrollee will have an adverse event leading to hospitalization, then this spillover term would be positive, and likely much larger in magnitude than ∂ c ( r d ) ∂ r 1 since the cost of inhalers is very low relative to emergency care. In our empirical application we do not observe the spillover derivative separately for each drug, so we rely on the knowledge of the medical experts that created the CGM Spillover and Tamblyn Essential drug lists to determine which drugs have the most positive derivatives. In theory there could also be drugs with negative average derivatives, for example if medical care like physician checkups are complementary to drug purchases, as may be the case in the treatment of depression.

The second set of terms is:

The first component could be nonzero if the choice of r₁ affects the composition of enrollees in the plan in a way that alters average risk scores. In this case the revenue effect is Q times the change in average medical revenue per enrollee. Second, the choice of r₁ could affect enrollment decisions of beneficiaries, differentially increasing the profits of an MA plan by the average difference between medical revenue and medical cost times the responsiveness of demand to the generosity of insurance coverage of the drug, r₁. This term differs from zero if the risk-adjustment formula does not fully eliminate the difference between revenue and cost for enrollees in MA plans. Although we do not have evidence on each of these terms, one possibility is that as plans grow larger the derivative of enrollment with respect to cost-sharing, ∂Q ∂r₁, becomes smaller relative to enrollment levels, Q, which could cause the spillover effect to increase in importance relative to the selection effect. We return to this point in the empirical analyses, where we test for differential responses to the two incentives by plan market share.

Of course, our conceptual framework contains simplified versions of much more complicated profit functions faced by plans. A potential interpretation concern could arise if there are other factors that cause elements of the FOCs to differ between MAPD and SAPDs. Two parameters in the model deserve particular attention in this regard. The first is the moral hazard parameter, ∂ c ( r d ) ∂ r 1 . If MAPD and SAPD enrollees have different price elasticities of demand, this term could differ and potentially confound the interpretation of our parameters of interest. However, there are a few pieces of empirical evidence that alleviate this concern. Einav, Finkelstein, and Polyakova (2016) estimate this price elasticity of demand for drugs in Part D, and find that for an average drug it is –0.24, but the standard deviation of the drug-level distribution of elasticities is 0.59, reflecting substantial across-drug heterogeneity. In our empirical specifications, we compare estimates from models that include drug NDC fixed effects, which absorb all of this across-drug variation in elasticities of demand, to models that do not, and find that our estimates are not sensitive to whether we condition on drug classifications or NDCs. If differences in average elasticity of demand of enrollees were an important factor, one would expect that the substantial differences in elasticity across drugs should also matter, but these differences do not appear to be strongly correlated with the selection and spillover incentive variables.

Geruso, Layton, and Prinz (2017) also directly estimate whether risk-adjustment errors in ACA exchange plans, which use a similar risk-adjustment formula developed for Medicare, are correlated with the drug-level elasticity of demand estimates from Einav, Finkelstein, and Polyakova (2016). Although their empirical approach is different than ours, it is based on similar intuition. They find no clear empirical relationship between selection incentives and drug elasticities of demand.

A second potential concern may be that ∂ Q ∂ r 1 , the derivative of enrollment with respect to coverage generosity for a particular drug, could differ between plan types. For example, since MA plans consist of a broader bundle of products, including medical and drug insurance, beneficiaries may be less responsive to drug formulary changes with respect to enrollment choices in MA plans than in SA plans. This term affects the selection incentive in our conceptual framework, and could attenuate the scope for inducing selection into MA plans using drug formularies, which would make it more difficult to observe significant formulary differences in response to selection incentives. However, as we show, there are significant formulary differences across plan types that are related to the selection incentive. Han and Lavetti (2017) also provide direct evidence that beneficiaries responded to our measure of MA Switcher Surplus, and estimate that the drug formulary mechanism we discuss increased the probability of an average Medicare beneficiary enrolling in an MA plan by 7.1 percent after Part D was introduced.

III. Data and Empirical Methods

There are several steps required to construct the key variables used in our analyses, and to link them to Part D plan formulary data. We begin by describing how we calculate the “MA Switcher Surplus” variable for each drug, our measure of MA plans’ selection incentives. This calculation requires first estimating the difference between risk-adjusted MA revenue and expenditures by medical condition, and then mapping these average differences to the drug level using data on the joint distribution of drug purchases and medical diagnoses. This provides a drug-level estimate of the MA Switcher Surplus, which we merge at the NDC level to the plan formulary data. Next, we describe the Medicare Part D formulary data that we use, and provide summary statistics on the data. We then discuss the lists of drugs that we use from the literature to identify spillover incentives, and the steps required to link these lists to formulary data. The third set of explanatory variables that we link to the formulary data are Part D-specific estimates of risk-adjustment errors from Carey (2017), which come from a different risk-adjustment formula than the one we study. Before presenting the main empirical specification and identification strategy, we provide an illustrative example of the sources of variation used in our analyses.

A. Medicare Advantage Risk-Adjustment and Selection

The selection incentive we study arises because risk-adjustment in Medicare Parts A & B does not eliminate the potential for plans to profit through selection, as shown by Brown et al. (2014). The first step of our analyses is to quantify the magnitude of the selection incentives that MA plans face for each medical condition in the risk-adjustment formula. To do this, we use claims data from the universe of fee-for-service Medicare beneficiaries from 2008–2010. Using Medicare’s internal calculations of patient risk-scores, for each beneficiary that was in FFS Medicare for the full prior calendar year we calculate the difference between the actual observed FFS spending and the counterfactual capitation payment that an MA plan would have received if that beneficiary enrolled. If the entire FFS population were to switch into MA plans at the same time this difference would approximately equal zero due to the risk-adjustment formula. However, since the switchers from FFS to MA are non-randomly selected, the average annual spending of the population of FFS beneficiaries who subsequently choose to switch to MA plans (the estimated value of parameter β ^ in Equation 1) is $902 [SE $28] less than the spending of those who remain in FFS, conditional on medical diagnoses and other characteristics included in the risk-adjustment formula. This finding is consistent with advantageous selection of beneficiaries into MA plans as in Brown et al. (2014), and is very close to the similar statistic estimated by Batata (2004) of $1030 using data from the early 1990s.

Using this difference, counterfactual MA capitation payments minus FFS expenditures, as a dependent variable, we estimate the average surplus associated with each of the 70 medical conditions in the risk-adjustment formula using the the following fixed effects regression:

M A Switcher Sur p i t = α + β M A Switc h i t + ∑ k = 1 70 θ k 1 [ H C C i t − 1 = k ] + ∑ k = 1 70 γ k M A Switc h i t ∗ 1 [ H C C i t − 1 = k ] + π X i t + ψ c ( i t ) + ε i t

where MA Switch_it is an indicator that equals one if person i switched from FFS into an MA plan in year t, 1 [HCC_it−1 = k] is an indicator that equals one if person i had a diagnosis associated with Medicare Hierarchical Condition Code (HCC) k in the prior year, and X_it is a vector of control variables that includes year effects, age effects, race effects, a gender effect, interactions between race effects and an indicator that equals 1 if the individual originally entered Medicare due to a disability, and ψ_c(it) is a set of county effects that equals one for the county in which beneficiary i lived in year t.

The key parameters of interest from this model are the 70 estimated values of γ_k, which capture the average deviation from the mean MA Switcher Surplus for each of the 70 medical conditions, k. If there were no heterogeneity in patterns of switching into MA plans by HCC, these values would all be zero. Figure 1 shows the distribution of these estimates, along with the 95 percent confidence intervals. 18 The estimated MA switcher surplus is statistically significantly different from zero for 48 out of 70 HCCs. For 46 of these conditions the estimate is significant and positive, suggesting that lower cost beneficiaries within these HCCs are more likely to switch into MA plans.

MA Switcher Surplus by Hierarchical Condition Code

Note: Notes: This figure plots the estimated values of γ_k from Equation 1 corresponding to each of the 70 Medicare hierarchical condition codes, along with the 95 percent confidence interval for each estimate. HCC 130 (end-stage renal disease with dialysis) is dropped because Medicare rules restrict beneficiaries with this condition from switching into MA plans.

This evidence is consistent with the findings of Brown et al. (2014), who point out that because the variance of medical expenditures tends to increase with the expected mean, if plans are able to avoid beneficiaries in the upper tail of the expenditure distribution, they can potentially obtain large overpayments from enrolling beneficiaries with medical conditions that are on average more costly. They show that after HCC-based risk-adjustment was introduced, the difference in baseline medical expenditures between MA switchers and FFS stayers was larger for beneficiaries with higher risk-scores. This implies overpayments to MA plans are larger for beneficiaries with medical conditions (HCCs) that tend to be associated with higher spending. Similarly, we find substantial variation in overpayments across HCCs, which is the variation captured by the γ_ks and depicted in Figure 1 .

To be clear, these estimates do not necessarily equal MA plan profit for several reasons. The general approach in the literature of studying selection based on switching behavior (Morgan et al. 1997; Batata, 2004; Brown et al., 2014), which we follow conceptually in estimating Equation 1, is limited to the extent that the flow of switchers may be different than the stock of enrollees in any given plan. This is difficult to account for, because data on utilization and expenditures of individual MA enrollees are not generally available to researchers. Somewhat reassuringly, however, Brown et al. (2014) estimate that 75 percent of MA enrollees were switchers from FFS at some point, whereas only 25 percent joined an MA plan at the point of initial eligibility. They also find, using administrative risk scores linked to MCBS data, that the average risk score of the stock of MA enrollees increased over time at a rate that closely corresponded to the patterns observed for switchers. Although both of these findings suggest that studying switching choices may be informative about selection in MA plans more broadly, it is possible that some of the difference in average expenditures between switchers and stayers at the time of a switch may not persist.

The costs of treating beneficiaries in MA plans may also differ from the costs under FFS, regardless of risk scores, and this cost difference is not included in our calculation since we have no data on beneficiary utilization in MA plans. In addition, as discussed by Geruso and Layton (2016), to the extent that enrollment in MA plans has a causal effect on risk scores through upcoding, our estimated capitation revenue may understate the revenue that an MA plan would actually receive. The calculation that we use, which is similar to that used by Brown et al. (2014), captures only the selection component of profits.

B. Mapping Medical Conditions to Drug Purchases

The final step of the switcher surplus calculation, and the step in which we diverge from previous research studying switchers into MA plans, is to use the estimated values γ_k from Equation 1 to calculate the predicted MA switcher surplus by drug active ingredient, rather than by HCC code. The goal is to be able to link each drug on every plan formulary to a measure of the predicted MA surplus that a plan would earn if an average enrollee who takes a drug with that active ingredient were to enroll. In order to connect this selection incentive to formulary design, we use Medicare claims to construct a complete mapping of all medical conditions used in risk-adjustment calculations to each drug covered by Part D plans. To account for the fact that drugs with the same active ingredient are used to treat the same condition(s), and so they should conceptually have the same selection effect, we link each drug NDC to its active ingredient using the NDC product database. 19 Beginning with the universe of Medicare Part D claims data from 2008–2010, we link each NDC in the claims data to its active ingredient, and for each beneficiary-year we construct a set of binary variables indicating whether the beneficiary filled a prescription with that active ingredient in that year. We then link this file to a database of all of the HCC conditions for that beneficiary, which is derived from the patient’s diagnoses. This provides an individual-year level database of every active ingredient purchased and every medical diagnosis for the population of FFS beneficiaries.

We estimate a separate probit model for each active ingredient in the data. In each model the dependent variable equals one if the beneficiary purchased a drug with the given active ingredient in a given year, and zero otherwise. The independent variables are simply a set of binary variables for each of the 70 HCC condition codes used in the Medicare Advantage risk-adjustment formula.

This set of equations gives a d × h matrix of coefficients, ν, where d is the top 431 most frequently purchased active ingredients, 20 and h is 70, corresponding to the number of HCC codes. The dh th element in this matrix equals the marginal effect of HCC h on the probability of purchasing a drug with active ingredient d.

For each estimated v h ^ we calculate the predicted marginal effect of having HCC_h on the probability of taking a drug with the given active ingredient. We then calculate the predicted MA switcher surplus at the active ingredient level as:

where γ h ^ are the estimated coefficients from Equation 1 and v d h ^ is the row vector from the coefficient matrix from Equation 2 corresponding to the same active ingredient d. Equation 3 gives the probability that a beneficiary takes c drug d given their HCC codes times the MA switcher surplus for each of those HCC codes, which equals the expected MA switcher surplus at the active ingredient level, taking into account the full joint distribution of HCC codes and drug active ingredient consumption in the population. Intuitively, we can describe the two dimensions that contribute to variation in MA Switcher Surplus in Equation 3 as follows. First, drugs whose use is more strongly predicted by a medical condition that is associated with over-payments to MA plans increase drug-level MA Switcher Surplus because they have higher values of v d h ^ . Loosely, one can think of this term as capturing the strength of the information signal that connects an HCC to demand for a drug. Second, holding v d h ^ fixed, if the magnitude of HCC-level overpayments γ h ^ increases, so too does the magnitude of the expected drug-level MA Switcher Surplus for all associated drugs.

Figure 2 shows the distribution of drug-level MA switcher surpluses, which is the key variable we will use to test the selection hypothesis. The mean of the distribution is $55, and the standard deviation is $152.

Distribution of MA Switcher Surplus by Drug Active Ingredient

Note: Notes: This figure plots the distribution of values of ∑ h = 1 70 v d h ^ ∗ γ h ^ from Equation 3 corresponding to each of the 431 drug active ingredients in the data, weighted by their frequency of coverage on Part D formularies. The distribution is truncated at $400 for ease of presentation, although the full distribution has a longer upper tail with a maximum value of $1615.

C. Part D Formularies and MA Medical Spillover Incentives

We use the CMS Formulary Files and Quarterly Pricing Files to test for differences in Part D formulary designs in SAPD and MAPD plans. The Formulary Files contain monthly data on the universe of MAPD and SAPD plans offered in the country since Part D implementation at the beginning of 2006. In particular, they provide the formulary status of drugs, tier location, whether prior authorization, step therapy and quantity limits are imposed, and limited information about cost sharing policies corresponding to tiers. The Quarterly Pricing Files began being released in 2009, and include data on the average reimbursement prices negotiated between each plan and the in-area retail pharmacies for that plan at the NDC-level. Together, these two files provide complete information about the generosity of the universe of Part D plans. Prior to 2009 the Formulary files contained cost-sharing rules, but the cost of the drug to the insurance company was unknown, making it difficult to calculate the percentage of the total cost that the consumer must pay for drugs with fixed copayments. We use data on prices from the Quarterly Pricing Files between the first quarter of 2009 and the third quarter of 2011. Table 1 shows summary statistics of our measures of plan generosity from these data sources. In total, the sample contains 2,588 unique drug NDCs and 5,530 unique plans.

Table 1

Summary Statistics on Plan Generosity

Note: Summary statistics from formulary pricing files between the first quarter of 2009 and third quarter of 2011. ‘Initial Coverage Cost-Sharing’ is the out-of-pocket cost divided by the total average cost of the drug. ‘Number of Plan-Drug Pairs’ is the number of unique NDC and plan formulary pairs, and refers only to drugs that are on-formulary. ‘Percent of Drugs on Formulary’ is calculated as the share of NDCs that each plan covers among the full set of NDCs that were covered by at least one plan in the same quarter.

One caveat associated with studying formulary design is that many Medicare enrollees are dually eligible for Medicaid, and receive additional low-income subsidies that reduce both premiums and cost-sharing (see DeCarolis, 2015). For beneficiaries with incomes below 135 percent of the federal poverty line, subsidies reduce cost-sharing to nearly zero in all plans, 21 which effectively reduces or eliminates coverage generosity differences between MAPD and SAPD plans. This could potentially attenuate the differences in plans’ incentives somewhat, although the majority of Part D enrollees (70 percent) do not receive these subsidies (Hoadley et al., 2015). The cost-sharing rules included in the formulary files, which we focus on in our analyses, are those relevant to non-dual-eligible beneficiaries.

We find that there are negligible differences in negotiated drug prices for MAPD and SAPD plans. For the same NDC, the average price paid by MAPD plans is within 1 percent of the average price paid by SAPDs. There also does not appear to be any meaningful correlation between differences in negotiated prices and MA Switcher Surplus. A $100 increase in MA Switcher Surplus is associated with about a 0.06 percent difference between MAPD and SAPD prices.

As described in Section I.A, there are asymmetric restrictions on changes in plan design within years, permitting changes that increase generosity but not vice versa. One hypothesis we test is whether plans strategically alter their formularies within plan-years, since the selection incentive is theoretically strongest during the open enrollment period. Before directly testing this hypothesis, we look for summary evidence of changes to formulary designs within plan-years. To do this, we construct a balanced panel for each plan-year, which contains the set of all NDCs that the plan covers at any point during the plan year. On the extensive margin, we find that of the drugs covered by a plan at some point during the year, on average 11.5 percent of these drugs are excluded from the formulary during open enrollment, compared to only 5.2 percent, 3.4 percent, and 2.5 percent of drugs excluded during the other three quarters of the year. This pattern appears to be similar for MAPD and SAPD plans. On the intensive margin, however, we find evidence of relative differences in formulary changes for MA plans. To avoid the impacts of potential changes in negotiated prices, which could alter out-of-pocket costs for drugs with variable coinsurance, we calculate the fraction of drugs in the balanced plan-year panel for which the cost-sharing rule itself changes within the year. Consistent with regulatory restrictions, we find that only 0.4 percent of plan-drug-quarter observations that remain on a formulary have changes in cost-sharing rules that reduce generosity. However, in the other direction we calculate that 6.4 percent of plan-drug-quarter observations have changes in cost-sharing rules within the plan-year that increase plan generosity, and MAPD plans are about 23 percent more likely than SAPDs to make such changes.

We connect the formulary data by drug to two lists of spillover drugs to test the spillover hypothesis. The first list is the categorization of acute and chronic spillover drugs developed by Chandra, Gruber, and McKnight (2010). Since their categories were developed using a drug classification system from 1995, it cannot be linked to some more recent NDCs. We developed a mapping that links their categories to the current classification system used by CMS (the United States Pharmacopeia Classification System). This mapping was largely straightforward, but there were some classes that were not uniquely matchable to the USP system, so our versions of the lists are subsets of the original lists containing 73 percent of the Acute and 72 percent of the Chronic classes. 22

The second list we use was developed by Tamblyn et al. (2001), which similarly relied upon clinical experts to classify drugs according to whether they are ‘essential,’ which was defined as: “medications that prevent deterioration in health or prolong life and would not likely be prescribed in the absence of a definitive diagnosis.” Since the Tamblyn list is based on a different classification system than the CGM list, they are not directly overlapping, and both indicator variables can be included in regressions without causing substantial collinearity.

D. Part D Risk-Adjustment

In addition to the selection incentive described above that MA plans face, all Part D plans face separate selection incentives caused by differences in Part D profitability across medical conditions. Carey (2017) shows that this incentive arose primarily because the risk-adjustment formula did not quickly adjust following new drug entry or the onset of generic competition, which changed treatment costs. Although these selection incentives are uniform across all plans that we study, it is possible that by chance they may be correlated with our key variables of interest that measure differential selection and spillover incentives among MAPDs. To remove this source of potential omitted variable bias, we control for the Part D selection incentive using the expected profit or loss associated with each drug in the CMS plan formularies, which was calculated by Carey (2017). Carey constructs these profitability measures in two steps using data from a 5 percent sample of Medicare claims. The first step is to link each drug in the formulary files to the mostly likely medical diagnosis associated with that drug. This model is identical to our Equation 2, but Carey uses the largest estimated probit coefficient from the model to indicate the most likely diagnosis, whereas we use the entire matrix of coefficients that describe the full joint distribution. The second step is to aggregate treatment costs by patient in the Part D claims data, and regress treatment costs on the set of indicators of predicted diagnoses associated with the drugs taken by each person. This step is analogous to our Equation 1, except that we also control for county fixed effects and demographic characteristics of beneficiaries. Carey (2017) shows that Part D plans respond to these Part D selection incentives in the design of their formularies, consistent with the general type of strategic behavior that we investigate in this paper, but she does not test for differential incentives between integrated and stand-alone drug plans.

The predictions from this model yield expected treatment costs by diagnosis, which can then be compared to the formulaic risk-adjusted payments by diagnosis that are set by Medicare, and the difference between the two is the expected profit or loss associated with each medical condition. Figure 3 shows the distribution of risk-adjusted Part D profits by drug NDC, for all drugs that appear on Part D formularies. The distribution has a mean value of −$68, but has substantial mass away from zero, suggesting that risk-adjustment has not fully eliminated selection incentives.

Distribution of Part D Risk-Adjusted Profit by Drug