soc paper summary

D

G S

a

A R R A A

J I I I

K M C P I P

1

a c p c s r t a i c p

i a s a h

p c

0 h

Journal of Health Economics 32 (2013) 1345– 1355

Contents lists available at ScienceDirect

Journal of Health Economics

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e c o n b a s e

igesting the doughnut hole

eoffrey F. Joyce ∗, Julie Zissimopoulos, Dana P. Goldman chaeffer Center for Health Policy and Economics, University of Southern California, 3335 S. Figueroa, Unit A, Los Angeles, CA 90007, United States

r t i c l e i n f o

rticle history: eceived 16 May 2012 eceived in revised form 11 March 2013 ccepted 26 April 2013 vailable online 6 May 2013

EL classification: 1 13 18

a b s t r a c t

Despite its success, Medicare Part D has been widely criticized for the gap in coverage, the so-called “doughnut hole”. We compare the use of prescription drugs among beneficiaries subject to the coverage gap with usage among beneficiaries who are not exposed to it. We find that the coverage gap does, indeed, disrupt the use of prescription drugs among seniors with diabetes. But the declines in usage are modest and concentrated among higher cost, brand-name medications. Demand for high cost medications such as antipsychotics, antiasthmatics, and drugs of the central nervous system decline by 8–18% in the coverage gap, while use of lower cost medications with high generic penetration such as beta blockers, ACE inhibitors and antidepressants decline by 3–5% after reaching the gap. More importantly, lower adherence to medications is not associated with increases in medical service use.

eywords: edicare Part D

overage gap

© 2013 Elsevier B.V. All rights reserved.

g i $ u c d e e a a d t

P t g t

rescription Drugs nsurance Design, rice Elasticity

. Introduction

By most metrics, the Medicare drug benefit (Part D) has been success. Launched in January 2006, the program offers Medi- are beneficiaries the option of enrolling in a prescription drug lan administered by a private company. More than 90% of Medi- are beneficiaries have drug coverage at least as generous as the tandard Part D benefit, and about 9 out of 10 Part D enrollees eport being satisfied with their plan (Research, 2011). At the same ime, the costs of the program have been far lower than predicted – s much as 40% below what the Congressional Budget Office orig- nally projected. Lower program costs are largely attributable to ompetition between plans, high rates of generic drug use, and the reference of beneficiaries for low-premium plans1.

Yet despite its success, Medicare Part D has been widely crit- cized for the gap in coverage – the so-called “doughnut hole” – ffecting a large fraction of enrollees. Under the government’s 2012

tandard benefit design, beneficiaries not receiving subsidies face

deductible, followed by a 25% co-insurance rate. But once they ave spent a cumulative total of $2930 on prescription drugs in a

∗ Corresponding author. Tel.: +1 213 821 7958; fax: +1 213 740 3460. E-mail address: [email protected] (G.F. Joyce).

1 Lower than expected program costs may also be a attributable to overly high rojections by CBO, which had to estimate the cost of Part D in the absence of a omparable commercial plan.

t n B t o k c e s t

167-6296/$ – see front matter © 2013 Elsevier B.V. All rights reserved. ttp://dx.doi.org/10.1016/j.jhealeco.2013.04.007

iven plan year, they must start paying the full cost of their med- cations. Only after a beneficiary reaches a “catastrophic” limit of 4700 in out-of-pocket spending (or $6658 in total drug spending nder the standard benefit) does coverage resume, with minimal ost-sharing thereafter. Although most companies offering Part D rug plans modify the standard design, they usually retain the cov- rage gap. Thus, many beneficiaries with moderate to high drug xpenses, particularly the chronically ill, face breaks in coverage nd out-of-pocket costs that may alter their demand for drug ther- pies. If the gap is prompting beneficiaries to use pharmaceuticals ifferently – especially if it leads them to discontinue an effective herapy – this could have important health consequences.

While the coverage gap is being phased out under the Patient rotection and Affordable Care Act, many beneficiaries will con- inue to face a break in coverage until 2020. Further, the coverage ap provides a natural experiment to assess how seniors respond o changes in the spot price of their medications. Prior work on he demand for prescription drugs has focused on the response of on-elderly populations to increases in fixed dollar copayments. y contrast, the Part D coverage gap is temporary, changing both he spot price of a medication as well as the expected future price f a drug if the beneficiary reaches the catastrophic threshold. We now far less about the price elasticity of demand of seniors in this

ontext and how they might respond to cycling in and out of cov- rage over the course of the year. For example, how frequently do eniors switch to lower cost drugs or discontinue use of a medica- ion altogether when reaching the gap? Do gap-induced changes in

1 lth Eco

m m a t A s o

b o ( a c i r t m I e

m w h ( a d p d o ( d p g ( w e

2

2

w t o r m r p i n p

D ( ( 2 t c s c

c

2 b

c i A p T d a c w b i e m ( t a

r P g l e c F a i b b b g h o a w 2

f p e w ( i a s b b d d a

2

f t

346 G.F. Joyce et al. / Journal of Hea

edication use persist in the next year or do beneficiaries switch to ore generous plans? Further, do changes in prescription drug use

ffect the demand for medical services? Recent evidence suggests hat the introduction of Part D led to substantial reductions in Part &B spending (McWilliams et al., 2011). The reduction in medical ervice use may be even larger if beneficiaries do not cycle into and ut of drug coverage.

In this paper, we compare changes in prescription drug use efore and after reaching the coverage gap for two distinct groups f beneficiaries: (1) those eligible for the full low-income subsidy LIS), who face minimal cost-sharing throughout the year and thus re unaffected by the coverage gap; and (2) non-subsidized benefi- iaries who pay the full cost of brand name drugs or all medications n the gap. We focus on beneficiaries with diabetes given it is major isk factor for a wide range of other health conditions. More impor- antly, reductions in use of oral hypoglycemics and other diabetes

edications can lead to adverse health events in the short term. f the coverage gap has deleterious effects on health, it should be vident in this sample of beneficiaries.

We find that the coverage gap had a relatively modest effect on edication use, but the demand response was strongly correlated ith price. Use of statins ($65), anti hypertensives ($58) and oral ypoglycemics ($50/month), declined by 4–6% in the coverage gap 4–5 percentage points), while use of lower cost medications such s ACE inhibitors/ARBs ($31), beta blockers ($27) and diuretics ($8) eclined by just 2–4% in the gap (2–3 percentage points).2 This attern was consistent across other drugs classes not used to treat iabetes. For example, beneficiaries with diabetes reduced their use f high cost, brand-dominant medications such as anti-asthmatics $127), antipsychotics ($213), and other central nervous system rugs ($150) by 7–18% in the coverage gap (5–11 percentage oints). By contrast, use of less expensive medications with high eneric penetration such as antidepressants ($49) and analgesics $53) declined by 3–6% in the gap (1–4 percentage points). Further, e found no evidence that reductions in medication use in the cov-

rage gap affected medical service use that year or in the next year.

. Previous research

.1. Medicare Part D

The coverage gap changes the price of a drug in more complex ays than a typical increase in a fixed dollar copayment. During

he coverage gap, both the spot price and the expected future price f a drug are affected. Once a non-subsidized beneficiary (non-LIS) eaches the coverage gap, each prescription he fills is likely to cost ore. Yet at the same time, each fill increases the likelihood of

eaching the catastrophic threshold, which lowers the expected rice of future prescriptions that year. Moreover, any price change

n the gap is temporary since benefits reset at the beginning of the ext calendar year. How beneficiaries respond to more complex rice changes is unclear.

Most research to date examines how the introduction of Part affected prescription drug use and beneficiaries out-of-pocket

OOP) expenditures (Lichtenberg and Sun, 2007); (Yin et al., 2008); Ketcham and Simon, 2008); (Levy and Weir, 2008); (Joyce et al., 009); (Kaestner and Khan, 2012). Most of these studies find that he availability of government subsidized drug coverage is asso-

iated with higher pharmaceutical use and lower out-of-pocket pending, with more pronounced effects for low-income benefi- iaries. The exception is work by Levy and Weir (Levy and Weir,

2 Prices reflect the average price paid in the sample for a 30-day supply of medi- ation.

c p e a ( e p

nomics 32 (2013) 1345– 1355

008), who find that Part D has only a small effect on medication use ased on patient self-reports in the Health and Retirement Study.

Only a few studies have directly examined the impact of the overage gap. Zhang et al. (2009a, b) compare Medicare beneficiar- es in employer-sponsored drug plans with enrollees in a Medicare dvantage prescription drug (MA-PD) plan, looking at changes in rescription drug use before and after entering the coverage gap. he majority of enrollees in the MA-PD plan have some generic rug coverage while in the gap. They find that in comparison to

group with employer-provided coverage, seniors without gap overage reduce their monthly use of prescription drugs by 14%, hereas those with generic coverage in the gap reduce their use

y just 3%. The primary limitation of this study is that the sample s restricted to those that remain in the coverage gap through the nd of the year. Thus, the highest users of prescription drugs and ost price inelastic beneficiaries are excluded from the analysis

i.e. those most likely to reach catastrophic threshold). As a result, heir estimates likely overstate the impact of the coverage gap on ll beneficiaries.

Polinski et al. (2011) use data from CVS Caremark to assess ates of medication switching and discontinuation. They compare art D beneficiaries “exposed” and “unexposed” to the coverage ap, where the latter are defined as those receiving some form of ow-income subsidy (partial or full). They find that beneficiaries xposed to the coverage gap have twice the hazard rate of dis- ontinuing a drug, but do not present baseline rates of cessation. urther, they find that beneficiaries exposed to the coverage gap re much less likely to switch medications. This finding is counter- ntuitive given the potentially large increases in spot prices facing eneficiaries exposed to the gap. The authors suggest that exposed eneficiaries may switch to lower-cost brand or generic versions efore they reach the threshold to prevent or delay entering the ap. However, they do not present any evidence to support this ypothesis. Nor is it consistent with beneficiary surveys that show nly 40% of beneficiaries were aware of a coverage gap in 2006, nd those that were, had little understanding of how it worked or hether they were personally at risk of entering the gap (Hsu et al.,

008). Hoadley et al. (2011) use data from IMS Health to measure the

raction of Part D enrollees who reach the coverage gap and how rescription drug use changes during the gap. They compare ben- ficiaries who do not receive the low-income subsidy (non-LIS) ith two control groups; beneficiaries who receive the subsidy

LIS) and commercially insured seniors. They find that nearly one n five non-LIS enrollees (19%) reached the coverage gap in 2009, nd one in six of those (3%) reached the catastrophic threshold. Pre- cription drug use, as measured by the number of scripts, declined y 7–8% in the coverage gap, while total drug spending declined y 13–16%. The primary limitation of this analysis is that the IMS ata do not capture the universe of Part D claims, although their ifference-in-differences approach should mitigate the extent of ny measurement error.

.2. Cost offsets

If the gap is prompting beneficiaries to use pharmaceuticals dif- erently – especially if it leads them to discontinue an effective herapy – this could have important health consequences. In fact, ycling into and out of coverage may be more disruptive to care lans than a stable benefit with higher coinsurance. There is limited vidence on the link between cost-sharing for prescription drugs

nd health. While initial studies find mixed evidence on this issue Johnson et al., 1997); (Motheral and Fairman, 2001); (Fairman t al., 2003), several recent studies find that that increasing co- ayments for drugs increases the use of other medical services.

lth Eco

G t c t i c t r a 2 f d v t F s a w s i

3

3

u ( t D B a m D e t

t d c m s 9 i b f a v t s r

l m i y i i d D e o M i

3

s i p d i s s b e g s g t o r t b

3

d i s g o h n f s d d s a a

f F p T t ( u c m fi t b

e a o p o L t T a

G.F. Joyce et al. / Journal of Hea

aynor et al. (2007) examine the effects of changes in pharmaceu- ical co-payments by private employers. They find that increasing o-payments leads to a decrease in drug spending, but about one- hird (35%) of the savings in drug expenses are offset by increases n medical spending. Moreover, the demand response to higher opayments was stronger in the next year. Chandra et al. (2007) ake a similar approach in examining the price responsiveness of etired public employees in California. They find that moving from

$0 to a $10 co-payment for prescription drugs is associated with a 0% reduction in physician visits. Further, increasing co-payments or physician visits (by an average of $6) reduces use of prescription rugs by 20%. They also find that higher co-payments for outpatient isits and prescription drugs are associated with increases in hospi- alization rates, with the largest effects among the sickest patients. inally, Zhang et al. (2009a, b) examine changes in spending on pre- cription drugs and other medical services in the two years before nd after Part D. They find that enrollment in Part D is associated ith increases in prescription drug use and reductions in medical

pending for those with no or minimal drug coverage before the mplementation of Part D.

. Data and methods

.1. Data and study sample

We use a 20% sample of Medicare beneficiaries through a re- se agreement with the National Bureau of Economic Research NBER). This dataset links enrollment and Part A and B claims for raditional fee-for-service Medicare enrollees (1992–2008) to Part

claims from 2006 to 2008. The additional years of Part A and data improve our measurement of disease incidence/prevalence nd risk adjustment. The pharmacy data include all the key ele- ents related to prescription drug events (e.g., drug name; National rug Code (NDC); dosage; supply; date of service; cost of ingredi- nts; dispensing fees; excluded expenses; and payments made by he beneficiary, plan, and other third-party coverage).

The Part A data include information about inpatient hospi- al stays, including length of stay, diagnosis-related group (DRG), epartment-specific charges, and up to ten individual procedure odes and diagnostic codes. Part B information includes claims sub- itted by physicians and other health providers and facilities for

ervices reimbursed by Part B. Each claim contains diagnostic (ICD- -CM) and procedure (CPT-4) codes, dates of service, demographic

nformation on beneficiaries, and a physician identification num- er. Data from outpatient hospital stays, stays at skilled nursing acilities, home hospice care, and durable medical equipment are lso included. All of these claims are linked with the beneficiary’s ital status. The denominator file contains demographic informa- ion about every beneficiary ever entitled to Medicare, including tate and county codes, zip code, date of birth, date of death, sex, ace, and age.

We identify seniors (age 65+) with diabetes on the basis of at east one inpatient or skilled nursing facility diagnosis, or two or

ore outpatient diagnoses of diabetes at least five days apart. Once dentified, individuals are assumed to have diabetes in subsequent ears. We also assume that any enrollee with a Part D claim for nsulin has diabetes. We restrict our analysis to individuals enrolled n traditional fee-for-service Medicare and a Part D prescription rug plan (PDP). Individuals are required to have the same Part

contract/plan for the entire year. In 2008, 29.8% of beneficiaries

nrolled in the same PDP plan for the whole year had diabetes based n these definitions. The final study sample consists of 609,723 edicare beneficiaries with diabetes in 2006, 673,646 beneficiaries

n 2007, and 714,403 beneficiaries in 2008.

b A p p

nomics 32 (2013) 1345– 1355 1347

.2. Phases of coverage

Each pharmacy claim includes the amount of the low-income ubsidy; the true out-of-pocket (TrOOP) amount; and a field that ndicates in which benefit phase a claim was made: deductible, re-coverage gap, coverage gap, or catastrophic phase (or strad- les two phases). For each individual, the particular coverage phase

n their part D plan is clearly defined in the data on the basis of pending levels: Entry into the gap is based on total annual drug pending, whereas exit (and entry into the catastrophic phase) is ased on the beneficiary’s TrOOP amount. Thus we can identify the xact date that non-LIS beneficiaries enter and exit the coverage ap, and the dates LIS beneficiaries not subject to the gap reach the ame levels of spending associated with entrance and exit from the ap. Most analyses include all beneficiaries, regardless of whether heir spending is high enough to reach the catastrophic thresh- ld. While beneficiaries that reach the catastrophic threshold may espond differently to the coverage gap, excluding them is likely o overestimate the true impact of the coverage gap on Medicare eneficiaries.

.3. Beneficiary groups

The Part D data categorize beneficiaries into four groups: (1) ual-eligibles (individuals who qualify for both Medicare and Med-

caid), (2) full low-income subsidy (LIS), (3) partial low-income ubsidy (Partial LIS), and (4) non-low-income subsidy (Non-LIS). In eneral, the dual-eligibles are a particularly vulnerable subgroup f Medicare beneficiaries. They tend to be poor, younger, in worse ealth, and much more likely than other beneficiaries to have cog- itive or mental impairments. For these reasons, we exclude them

rom our analysis. We also exclude beneficiaries receiving Part D ubsidies on a sliding scale (the partial LIS group). Partial subsi- ies cover 25%, 50%, or 75% of the regional benchmark threshold, epending on the beneficiary’s individual income level, or if he or he is part of a couple, the couple’s combined income. Although the verage coinsurance rate for Part D drugs is about 15% for this group s a whole, there is considerable heterogeneity across beneficiaries.

In this paper, we focus on group (2), beneficiaries eligible for the ull low-income subsidy (LIS), and group (4), non-LIS beneficiaries. or beneficiaries receiving the full low-income subsidy (group 2), atient cost-sharing is minimal and constant throughout the year. hese beneficiaries are not subject to the coverage gap even when heir level of drug spending reaches the coverage gap threshold e.g. $2400 in 2007). As a result, we do not expect their medication se to change before and after reaching the various (hypothetical) overage thresholds. We use the LIS as controls and compare their edication use before and after reaching the gap to non-LIS bene-

ciaries who can face vastly different spot prices over the course of he year and spending distribution. We briefly discuss each group elow.

LIS beneficiaries. To protect low-income seniors from high drug xpenses, the federal government pays their monthly premiums nd deductibles, as well as almost all of their drug costs through- ut the year. In 2010, individuals with incomes less than 150% of the overty level and modest assets (less than $8100 for an individual r $12,910 for a couple) qualified for the low-income subsidy. These IS beneficiaries may choose to enroll in any Part D plan. However, heir premiums are covered only up to a “benchmark” amount. his amount is calculated separately for each of 34 PDP regions cross the country and is based on the average premium bid for the

asic benefit by stand-alone prescription drug plans and Medicare dvantage prescription drug plans. Stand-alone PDPs with monthly remiums below the benchmark amount are called “benchmark lans.” These plans qualify to automatically enroll LIS beneficiaries

1 lth Eco

a e t i i t d i m s

p c D a e S B w h a g 2 g o

3

g d l g t d d h d s a s c t s n i h t e c p

3

e t

c t n

i P

t b e c g a t

Y

w c fi o m i p t c k a i d

i a p o a D d s a e i p e g

d a d m o e i a o r t i a

348 G.F. Joyce et al. / Journal of Hea

nd receive the full subsidy for their premiums. If LIS beneficiaries nroll in a non-benchmark prescription drug plan, they must pay he amount of the premium above the benchmark. LIS beneficiar- es pay modest copayments for each medication included on their nsurance plan’s formulary and the full cost of any drugs not on he formulary. In 2010, their copayments were $2.50 for generic rugs and $6.30 for brand-name drugs, up to $6440 in total spend-

ng. After that point, their co-payments dropped to $0. As of 2011, ore than 10 million beneficiaries were receiving the low-income

ubsidy (Kaiser Family Foundation, 2011). Non-LIS beneficiaries. Although non-LIS beneficiaries enroll in

lans with a variety of benefit designs, almost all of them face the overage gap in some form. Under the government’s standard Part

benefit, non-LIS beneficiaries are responsible for 100% coinsur- nce payments while in the coverage gap. In 2006, and to a lesser xtent 2007, several prominent insurers, including Humana and ierra Rx, offered gap coverage of brand-name drugs in some plans. ut due to adverse selection, these plans suffered large losses and ere discontinued after a year. Since then, any coverage in the gap as generally applied to generic drugs only.3 About 26% of stand- lone prescription drug plans currently offer some coverage for the ap – primarily limited to generic drugs (Kaiser Family Foundation, 011). Given the heterogeneity in gap coverage within the non-LIS roup, we examined beneficiaries separately, according to whether r not they had some coverage in the gap.

.4. Prescription drug use

We compare medication use before and after reaching the ap for LIS and non-LIS groups. We focus on beneficiaries with iabetes because nearly one out of every three Medicare dol-

ars is spent treating the disease or its sequelae. If the coverage ap has deleterious effects on health, it should be evident in his sample of beneficiaries. For each subgroup, we estimate the emand response for diabetes-related and nondiabetes-related rug classes. The set of nine diabetes-related classes include: oral ypoglycemic agents, ACE inhibitors, calcium channel blockers, iuretics, beta blockers, angiotensin II receptor blockers (ARBs), tatins, loop diuretics, digitalis glycosides, and a combination of ntihypertensives.4 We combine ACE inhibitors and ARBs into a ingle class because they are commonly considered therapeuti- ally interchangeable. The set of other drugs we analyze consist of he nine most prevalent nondiabetes-related classes used by this et of beneficiaries: antidepressants, antipsychotics, other central ervous system medications, antiasthmatics, platelet aggregation

nhibitors, antiulcerants, anticonvulsants, opioid analgesics, and ormones/synthetics/modifers. We estimate changes in prescrip- ion drug use before and after reaching the coverage gap, for ach therapeutic class. We restrict the sample to brand-dominant lasses when examining switching behavior from brand to generic roducts.

.5. Empirical approach

We use a before and after design with a comparison group to xamine the effect of the coverage gap on medication use. The con- rol group consists of full LIS beneficiaries who are unaffected by

3 A few plans offered brand coverage in 2008, but for all intents and purposes, overage of generics only was the norm in 2007–2008. As such, we heretofore refer o the two treatment groups as the: (1) non-LIS with generic gap coverage; and (2) on-LIS without gap coverage. 4 Insulin is used to identify beneficiaries with diabetes. However, we exclude

nsulin from demand models because injectable drugs are generally covered under art B and not Part D.

i r t b r f

c o

nomics 32 (2013) 1345– 1355

he coverage gap (it is simply a hypothetical threshold for these eneficiaries). There are two treatment groups: (1) non-LIS ben- ficiaries with generic-only coverage in the gap (heretofore gap overage”); and (2) non-LIS beneficiaries with no coverage in the ap (heretofore “no gap coverage”). A difference-in-differences pproach can be incorporated into a regression framework to con- rol for observed differences across groups.

ijt = ˇ0 + ˇ1(coverage gap)it + ˇ2(gap coverage)it + ˇ3(no coverage)it + ˇ4(CG × gap coverage)it + ˇ5(CG × no coverage)it + ˇ6Xit + ˇ7Zt + �ijt

here Yit reflects medication use by person i, in class j, at time (t); overage gap is a binary indicator that equals 1 for prescriptions lled after reaching the coverage gap and 0 before; Xit is a vector f patient characteristics and health conditions; and Z is a vector of onth dummies. The two treatment groups are captured by binary

ndicators for “gap coverage” and “no coverage” in the gap. The key arameters of interest, ˇ4 and ˇ5, capture differences in medica- ion use before and after reaching the coverage gap for the gap overage and no coverage groups, relative to LIS beneficiaries. The ey outcome measures are (1) monthly medication use, by class nd drug type (brand/generic); and (2) Part A and B utilization, ncluding inpatient admissions, outpatient visits, and emergency epartments visits.

We measure changes in medication adherence using the Med- cation Possession Ratio (MPR), which is the fraction of days that

patient “possess” or has access to medication, as measured by rescription fills. For example, a patient who fills a 30-day script n April 1st and refills the prescription on May 10th would have n MPR of 75%, since they possessed 30 pills over a 40 day span. ue to the nature of claims data, this measure does have some rawbacks; for instance, patients are not observed using the pre- cription, only filling it. However, using a difference-in-difference pproach should mitigate any potential bias due to measurement rror. For each drug class, we compute the total days supply of med- cations before and after reaching the coverage gap to compute the ercentage of compliant days for each individual in the sample. We stimate changes in overall medication use (MPR), as well as the eneric dispensing rate (GDR), overall and by therapeutic class.

Poor adherence to medications can come about through three ifferent behavioral pathways: reduced initiation of drug ther- pies, worse adherence among existing users, or more frequent iscontinuation of therapy. In this paper, we focus on changes in edication adherence, conditional on use (i.e. we ignore the impact

f the coverage gap on the decision to initiate a new therapy). We xamine the fraction of beneficiaries that reduce medication use or ncrease use of generic drugs after reaching the coverage gap. We lso measure the fraction of beneficiaries who stop using a class f medication after reaching the gap and the extent to which they estart use in the first 90 days of the next year. We measure discon- inuation by comparing medication use within a therapeutic class n the 90 days prior to a beneficiary’s index date (gap date) and fter reaching the gap. For example, a beneficiary observed tak- ng an oral hypoglycemic, an antihypertensive, and a statin before eaching the gap, but only an oral hypoglycemic and an antihyper- ensive after entering the gap (for the remainder of the year) would

e categorized as having discontinued one medication within the elevant classes.5 Conditional on stopping, we then measure the raction of beneficiaries that resume use in the first 90 days after

5 We restrict our analyses of stopping and resuming to beneficiaries reaching the overage gap before November 15th to allow sufficient time to measure cessation f a therapy.

G.F. Joyce et al. / Journal of Health Economics 32 (2013) 1345– 1355 1349

as a no

c t a

b a 2 i a u m w r p

c i i t m c p p b

t m c b i d

3

s f u o p L c f d

c t h b o p n t

4

P n i $ p c s b a i i l w l T b h f

b d fi o p r a w i

Fig. 1. Standard part D benefit h

overage resumes the following year. We also examine the extent o which patients switch plans and move to more generous plans fter reaching the gap in the prior year.

Given that 2006 was the initial year of the program and that eneficiaries could enroll up to May 15th (partial year of cover- ge), we restrict most of our analyses of gap behavior to 2007 and 008. Nonetheless, we use the 2006 data to categorize beneficiar-

es and for risk adjustment, as well as in assessing plan switching nd enrollment decisions. We examine changes in medical service se by type of service (inpatient, outpatient and emergency depart- ent) using the primary ICD-9 diagnosis code. In some analyses, e examine diabetes-related utilization only (e.g., neuropathies,

etinopathy, amputation and cardiovascular outcomes) to more recisely target the effects of the coverage gap.

We include binary indicators for the most common comorbid onditions on the basis of the presence of ICD-9 diagnostic codes n the medical claims. These include twenty conditions defined n the Chronic Conditions Warehouse (CCW), as well as hyper- ension, hyperlipidemia, asthma, gastro-intestinal disorders. The

odels also include a set of monthly time dummies, a binary indi- ator for beneficiary type (generic-only and no gap coverage), and lan fixed-effects. We then used the results from these models to redict the monthly use of prescription drugs by LIS and non-LIS eneficiaries overall, and for each of nine therapeutic classes.

Many states have pharmaceutical assistance programs (SPAPs) o help their residents pay for prescription drugs. Some states ter-

inated their SPAPs after the introduction of Part D, while others ontinued to offer prescription drug subsidies as a supplemental enefit or as “wrap around” coverage to Part D. We include binary

ndicators for the (14) states that offer some form of supplemental rug coverage to Part D.

.6. Alternative control group

Patients fail to adhere to their medications for numerous rea- ons. The simplest and most common explanation is that they orget. Other important factors for non-adherence include sched- le disruptions, side-effects, and out-of-pocket costs. The validity f the difference-in-difference approach assumes that variation in rescription drug use between beneficiary groups (LIS and non-

IS) before and after reaching the coverage gap only reflects price hanges. However, it is possible that there are unobserved dif- erences in medication adherence across groups that is changing ifferentially over time. To test the appropriateness of the LIS as a

e w d a

n-linear price schedule (2007).

ontrol group, we re-estimate the models using alternative con- rols: seniors with employer-provided insurance. These seniors ave drug coverage at least as generous as the standard Part D enefit, but are not subject to the coverage thresholds or any ther aspect of Part D. As such, they face the same out-of-pocket rice over the course of the year and their medication use should ot change when drug spending reaches the Part D coverage gap hresholds.

. Results

Fig. 1 illustrates the non-linear price schedule of the standard art D benefit. Under the 2007 standard design, beneficiaries ot receiving subsidies face a deductible, followed by a 25% co-

nsurance rate. But once they have spent a cumulative total of 2400 on prescription drugs in a given year, they must start aying the full cost of their medications. Only after a benefi- iary reaches a “catastrophic” limit of $3850 in out-of-pocket pending (or $5451 in total drug spending under the standard enefit) does coverage resume, with minimal cost-sharing there- fter. A completely myopic beneficiary would view the spot price n each phase as the true price. By contrast, a future look- ng individual with moderate to high drug expenses that are ikely to exceed the coverage gap (or catastrophic threshold)

ould be less sensitive to fluctuations in the spot price and ess likely to change his use of prescription drugs in response. hus, prescription drug use and spending of the 41% of non-LIS eneficiaries who reach the coverage gap in 2007 depends on ow they take into account the non-linear price schedule they

ace. Fig. 2 shows the average coinsurance rates faced by different

eneficiaries in 2007 across the four phases of the Part D benefit esign: deductible, pre-gap, coverage gap, and catastrophic. The gure illustrates the variation we utilize to understand the effects f discrete changes in out-of-pocket costs on beneficiaries use of rescription drugs and medical services. The average coinsurance ate of LIS beneficiaries is at or below 5% across all phases; in the cat- strophic phase, the rate is zero. In contrast, non-LIS beneficiaries ith generic gap coverage have an average coinsurance rate of 59%

n the deductible phase, 36% in the pre-gap phase, 86% in the cov-

rage gap, and 7% in the catastrophic phase. Non-LIS beneficiaries ithout gap coverage face the highest coinsurance rates: 90% in the eductible phase, 32% in the pre-gap phase, 99% in the coverage gap nd 7% in the catastrophic phase. The reason the coinsurance rate

1350 G.F. Joyce et al. / Journal of Health Economics 32 (2013) 1345– 1355

Table 1 Beneficiary characteristics, by coverage group.

LIS Non-LIS, gap coverage Non-LIS, no gap coverage Employer-provided

Demographics (%) Age (years) 74.8 74.6 75.6 75.4 Male 29.9 43.4 42.4 52.1 White 46.5 91.4 87.8 N/A Black 19.1 4.0 5.8 N/A Hispanic 21.6 2.6 4.2 N/A

By region (%)a

Northeast 20.2 14.2 21.8 23.2 Midwest 15.3 33.8 25.7 40.5 South 39.7 37.6 37.7 30.8 West 24.6 14.2 13.6 5.0

Comorbidities No. of conditions (mean) 6.6 6.3 5.9 N/A

Number of observations 264,716 74,503 334,427 72,456

Utilization in 2005b (mean) No. of office visits 8.6 9.5 8.5 8.2 No. of ED visits 0.7 0.5 0.4 0.6 No. of inpatient stays 0.4 0.4 0.3 0.3 No. of Inpatient days 2.4 2.2 1.7 2.1

Total spending in 2005b (mean $) Total 11,006 10,588 8500 15,867 Inpatient 4000 4080 3127 4982 Outpatient 4949 5064 4255 8204 Other 2057 1444 1117 2681

Number of observations 224,832 67,895 311,949 58,677

N bid co ort re on-LI

i g c p i p g t g p

b w b t s e r t t

4 b d a r g a

o i g g b 2 c w

ote: Sample is individuals with diabetes and ages 65 and older. Year: 2007; comor a Percentages may not add up to 100 because a small number of cases did not rep b If covered by FFS Medicare Part A&B in 2005 for LIS, Non-LIS, Gap coverage, and N

s less than 100% during the deductible phase and in the coverage ap for non-LIS beneficiaries is due to the prerogative of insurance ompanies to vary the plan design from the government’s standard lan. For example, some Part D plans that provide generic coverage

n the gap also cover the cost of generic drugs during the deductible hase. In addition, beneficiaries who are ineligible for the federal overnment’s low-income subsidy may qualify for state programs hat provide coverage for a subset of medications. Such state pro- rams lower coinsurance rates, on average, to 85% in the deductible hase and 89% in the coverage gap.

Table 1 summarizes the characteristics of the LIS and non-LIS eneficiaries, as well as our alternative control group of seniors ith employer-provided insurance. On average, LIS and non-LIS

eneficiaries are of similar age: 75 years old on average. However, he income and asset requirements to qualify for the full Part D ubsidy suggest that these groups are different in other ways. For

xample, LIS beneficiaries are more likely to be female, non-white, egionally located in the West, and less healthy (as measured by he number of comorbid conditions) than non-LIS beneficiaries and he privately insured. Only 30% of LIS beneficiaries are male and

i y a y

Fig. 2. Percent of drug costs paid by b

nditions are any diagnosis years 2002 – 2007. gion. S, No gap coverage groups. Employer-provided group covered by private insurance.

7% are white compared to 42% and 88%, respectively for non-LIS eneficiaries without gap coverage. Low-income beneficiaries with iabetes average 6.6 comorbidities on average, compared to 6.3 nd 5.9 for non-LIS beneficiaries with gap and no gap coverage, espectively. Utilization of medical services and total spending is enerally higher for LIS beneficiaries compared to non-LIS benefici- ries although spending is the highest among the privately insured.

Table 2 shows the percent of beneficiaries that reach the thresh- ld level of spending for the coverage gap and catastrophic phase n 2007 and 2008. In 2007, LIS (58%) and non-LIS beneficiaries with eneric coverage (56%) are much more likely to reach the coverage ap threshold, and reach the gap earlier in the year than non-LIS eneficiaries without coverage (40%). This pattern is repeated in 008. In both years, LIS beneficiaries are more likely to reach the atastrophic level of spending (21–22%) compared to the non-LIS ith generic coverage and no coverage. Complicating these results

s the fact that beneficiaries can change categories from year to ear. For example, non-LIS beneficiaries that reach the gap in 2007 re more likely to switch to a plan with gap coverage in the next ear. As a result, some of the cross-year differences are likely to

eneficiaries, by coverage phase.

G.F. Joyce et al. / Journal of Health Eco

Table 2 Percent reaching coverage gap (CG) and catastrophic phase by group and year.

LIS Non-LIS, gap coverage

Non-LIS, no gap coverage

2007 (n = 264,716) (n = 74,503) (n = 334,427) Reached coverage gap 57.6% 55.8% 39.7% Median month (entry

CG) 7.4 8.0 8.7

Reached catastrophic phase

21.4% 8.6% 5.6%

2008 (n = 270,579) (n = 69,650) (n = 374,174) Reached coverage gap 56.9% 58.9% 32.9% Median month (entry

CG) 7.3 8.0 8.8

Reached catastrophic phase

21.9% 8.6% 5.4%

2007 (conditional on same group)

(n = 208,138) (n = 53,298) (n = 284,364)

Reached coverage gap 57.5% 58.2% 38.8% Median month (entry

CG) 7.4 8.0 8.8

Reached catastrophic phase

20.9% 9.0% 5.2%

2008 (conditional on same group)

(n = 208,138) (n = 53,298) (n = 284,364)

Reached coverage gap 58.6% 59.2% 33.7% Median month (entry

CG) 7.3 8.0 8.8

Reached catastrophic phase

22.6% 8.9% 5.6%

Note: Sample is individuals with diabetes and ages 65 and older. Conditional on s g

r f t i b n t s w

4

g m i a t d 2 t j T o t a g l r ( e B e

a a e m d h d g g g i g

t d o t a c p c F ( l i t t u t a w b

l r h ( t o a s a u r f m e t C d c

s p b t u F t i i

ame group sample consists of individuals staying in the same treatment or control roup in both 2007 and 2008. Years: 2007 and 2008.

eflect compositional changes in the beneficiary groups. To control or this, the last two panels of Table 2 condition on remaining in he same beneficiary group over the two year period. These find- ngs suggest that non-LIS beneficiaries without gap coverage (in oth years) are more likely to reduce their drug spending in the ext year after reaching the coverage gap. For example, 38.8% of he no coverage group reaches the gap in 2007, while just 33.7% do o in 2008. In contrast, a slightly higher fraction of LIS and non-LIS ith gap coverage reach the gap in 2008 compared with 2007.

.1. Use of prescription drugs

To shed light on specific behavioral responses to the coverage ap, we estimate the fraction of beneficiaries that decrease their edication use or increase their rate of generic use after reach-

ng the coverage gap. We define a decrease in medication use as t least a 2.5 percentage point decrease in MPR after reaching he coverage gap (relative to pre-gap). Similarly, a beneficiary is efined as increasing their generic rate if their GDR increases by .5 percentage points or more after reaching the gap (relative to heir GDR before the gap). The results shown in Fig. 3 reflect unad- usted difference-in-difference estimates at the drug class-level. he bars on the left-side of Fig. 3 reflect the additional fraction f non-LIS beneficiaries reducing MPR after reaching the gap (rela- ive to the LIS), and the bars on the right-side of Fig. 3 measure the dditional fraction increasing their generic rate after reaching the ap (relative to the LIS). Drug classes are ordered from highest to owest average price to highlight the correlation between demand esponse and the average out-of-pocket cost in the coverage gap

see also Table A1). Panel A (Fig. 3) display results for non-LIS ben- ficiaries with gap coverage relative to LIS beneficiaries and Panel

(Fig. 3) show results for non LIS beneficiaries without gap cov- rage compared to LIS beneficiaries. We find, for example, that an

n a a n

nomics 32 (2013) 1345– 1355 1351

dditional 9% of non-LIS beneficiaries reduce their use of statins fter reaching the coverage gap (relative to LIS statin-users). How- ver, the demand response is more muted (1–3%) for lower cost edications such as beta blockers ($27/month), diuretics ($8) and

igitalis glycosides ($7). Both non-LIS groups reduce their use of igher cost medications after reaching the coverage gap, but the emand response is larger for those without any coverage in the ap. Depending on the class of medication and the availability of enerics, a higher fraction of non-LIS beneficiaries increase their eneric rate compared to the LIS (Fig. 3). Further, and not surpris- ngly, the response is larger for those with generic coverage in the ap.

Table 3 shows the results of difference-in-difference regressions hat model changes in medication adherence (MPR) and generic ispensing (GDR) after reaching the coverage gap, controlling for bserved differences in patient demographics, comorbid condi- ions and Part D plans. Over the 9 diabetes-related drug classes, verage medication use declines in the coverage gap by 2.9 per- entage points for non-LIS with gap coverage and by 3.3 percentage oints for non-LIS without coverage, relative to the LIS. Moreover, lass-level changes are highly correlated with the price of the drug. or example, the use of higher cost medications such as statins $65/month) falls by 5.0 percentage points in the gap, compared to ess than 2 percentage points for beta blockers ($27) and diuret- cs ($8). In practical terms, these changes translate into statin users aking their medication as prescribed 78% of the time after reaching he gap compared to 83% before the gap. While overall medication se declines, the fraction of drugs dispensed as generic increases in he coverage gap. For example, use of generic statins is 5.4 percent- ge points higher after reaching the gap for non-LIS beneficiaries ith gap coverage and 3.7 percentage points higher for non-LIS

eneficiaries without gap coverage, relative to the LIS group. Table 4 presents results from similar models, but for the 9

argest classes of nondiabetes-related medications. Similar to the esults in Table 3, the demand response to the coverage gap is ighly correlated with the price of the drug. Use of antipsychotics $213/month) fall by nearly 10 percentage points after reaching he coverage gap (relative to the LIS), with similar declines in ther high cost classes such as antiasthmatics ($127) and anti- ggregants ($123). In contrast, the demand response is markedly maller for drugs costing less than $100 per month, including ntidepressants, analgesics, and anticonvulsants. Rates of generic se increase only slightly across all 9 classes, partly due to baseline ates of generic penetration in these classes. There are relatively ew generic products available in some classes (antipsychotics, CNS

edications, antiasthmatics) and high generic penetration in oth- rs (opioid analgesics, anticonvulsants). The coverage gap is likely o have a smaller impact in classes with high generic penetration. onversely, in brand dominant classes (e.g. antipsychotics), the emand response is larger for those with (generic) gap coverage ompared to the no coverage group.

Lower medication adherence may reflect behaviors such as tretching a prescription over more days (e.g. pill-splitting) or stop- ing a medication altogether. We estimate the fraction of non-LIS eneficiaries that discontinue use of a medication after reaching he coverage gap, as well as the fraction of stoppers who restart se after coverage resumes in the next year, all relative to the LIS. ig. 4 shows the differential rates of starting and stopping for the hree most expensive therapeutic classes, where the coverage gap s likely to have its biggest impact (results for all 9 classes are shown n Table A1). Consistent with prior results, a higher percentage of

on-LIS beneficiaries discontinue use of statins, antihypertensives, nd oral hypoglycemics after reaching the coverage gap. However,

greater fraction resume use within the first three months of the ext year. In the case of statins, about 4 percent more non-LIS than

1352 G.F. Joyce et al. / Journal of Health Economics 32 (2013) 1345– 1355

Table 3 Adjusted difference-in-difference in MPR and GDR for diabetes-related drug classes.

Number of observationsa

Average price ($)

Average MPR

Average GDR

MPR GDR

Non-LIS, gap coverage vs. LIS

Non-LIS, no gap coverage vs. LIS

Non-LIS, gap coverage vs. LIS

Non-LIS, no gap coverage vs. LIS

All Classes −0.029*** −0.033*** 0.031*** 0.020*** Statins 469,450 65.47 0.826 0.319 −0.050*** −0.050*** 0.054*** 0.037*** Anti-hypertensives, combo 143,972 58.33 0.828 0.239 −0.035*** −0.037*** 0.043*** 0.011 Oral hypoglycemics 378,372 49.73 0.866 0.652 −0.035*** −0.038*** 0.040*** 0.022*** Calcium channel blockers 227,924 45.58 0.854 0.483 −0.012** −0.024*** −0.020** −0.020*** Anti-hypertensives, other 46,784 37.37 0.759 0.869 −0.019 −0.026*** N/A N/A ACE/ARB 360,822 30.55 0.838 0.536 −0.025*** −0.027*** 0.020*** 0.012** Beta blockers 285,208 27.31 0.835 0.551 −0.015*** −0.022*** 0.050*** 0.051*** Diuretics 201,384 7.93 0.740 0.990 −0.019*** −0.030*** N/A N/A Digitalis glycosides 56,014 7.14 0.851 0.299 −0.011 −0.013* −0.004 −0.004

Note: Sample is individuals with diabetes, age 65 and older that reach coverage gap in 2007. Differences are in percentage points. Results are from regression models. Model includes age, sex, race, and indicators for ending in catastrophic phase, age-squared, and co-morbid conditions.

a Number of observations for MPR calculations. Significance levels are indicated as the following:

.

L a o

4

p f a f i a

b f A a o a t t s

F i

* p < 0.050. ** p < 0.010.

*** p < 0.001. Average cost is empirically derived, for 30-day equivalent. Year: 2007

IS beneficiaries stop taking their cholesterol-lowering medication fter reaching the coverage gap, but a greater fraction resume use nce coverage commences in the next year.

.2. Plan switching

Another potential response to the coverage gap is to switch lans the next year, particularly to a more generous plan with some orm of gap coverage. Some degree of plan switching is inevitable,

s the set of Part D plans change over time. This is particularly true or the LIS group, who receive full premium support if they enroll n a “benchmark” plan, where the benchmark is determined by the verage premium bid for the basic benefit in each region. If LIS

c c 2 r

ig. 3. (a) Behavioral change to coverage gas: Non-LIS, gap coverage vs. LIS (% of individu ndividuals).

eneficiaries enroll in a non-benchmark plan, they are responsible or paying the premium amount above the benchmark threshold. s a result, switching rates are quite high (31%) among the LIS s the set of benchmark plans changes from year to year based n plan bids. Among the non-LIS, 17% of those with gap coverage nd 11% without coverage changed plans in 2008. However, among hose reaching the coverage gap in 2007, 14.5 and 13.5%, respec- ively, switched plans the next year (Table 5). Moreover, by doing o, they lowered their out-of-pocket expenses (both premiums and

ost-sharing) by an average of $458 (gap coverage) and $153 (no overage group). Among switchers with generic gap coverage in 007, the reduction in out-of-pocket expenses was primarily a esult of switching to lower premiums plans: 89% of switchers with

als). (b) Behavioral change to coverage gas: non-LIS, no gap coverage vs. LIS (% of

G.F. Joyce et al. / Journal of Health Economics 32 (2013) 1345– 1355 1353

Table 4 Adjusted difference-in-difference in MPR and GDR for nondiabetes-related drug classes.

Number of observationsa

Average price ($)

Average MPR

Average GDR

MPR GDR

Non-LIS, gap coverage vs. LIS

Non-LIS, no gap coverage vs. LIS

Non-LIS, gap coverage vs. LIS

Non-LIS, no gap coverage vs. LIS

All classes −0.058*** −0.057*** 0.016*** 0.005* Antipsychotics, other 31,974 212.56 0.790 0.054 −0.096*** −0.076*** 0.005 0.003 CNS medications, other 73,346 149.54 0.762 0.019 −0.062*** −0.051*** 0.000 0.000 Antiasthma, other 64,588 126.87 0.617 0.002 −0.111*** −0.079*** 0.004*** 0.003*** Platelet aggregation

inhibitors 134,966 123.08 0.828 0.164 −0.086*** −0.070*** −0.017*** −0.026***

Gastric medications, other 232,756 107.92 0.747 0.241 −0.109*** −0.102*** 0.071*** 0.042*** Hormones/synthetics/modifiers,

other 104,964 90.04 0.757 0.104 −0.072*** −0.061*** 0.006 0.004

Anticonvulsants 79,738 61.19 0.725 0.804 0.004 −0.030*** 0.009 0.002 Opioid analgesics 201,672 53.37 0.340 0.964 −0.010* −0.020*** 0.005 0.004 Antidepressants, other 165,736 48.54 0.783 0.619 −0.028*** −0.038*** 0.024** 0.009

Note: Sample includes beneficiaries age 65 and older with diabetes that reach coverage gap in 2007. Differences are in percentage points. Drug classes sorted by average price for a 30-day script. Results are from regression models. Model includes age, sex, race, and indicators for ending in catastrophic phase, age-squared and co-morbid conditions.

a Number of observations for MPR calculations. Significance levels are indicated as the following: * p < 0.050.

** p < 0.010. *** p < 0.001. Average price is empirically derived, for a 30-day equivalent script. Year: 20

Table 5 Switching plans in response to reaching the coverage gap.

Non-LIS, gap coverage

Non-LIS, no gap coverage

Switch plans (%) 14.5 13.5 Change in OOP ($) −457.95 −152.86 Change in OOP (%) −16.8 −7.3

Do not swith plans (%) 85.5 86.5 Change in OOP ($) −14.09 −54.67 Change in OOP (%) −1.0 −2.0

Plan change type (%) Add gap coverage N/A 16.3 Lowered premium 89.0 24.0 Lower deductible 0.0 25.9 Other 11.0 33.8

Note: Sample includes beneficiaries age 65 and older with diabetes that reach coverage gap in 2007. Changes are from 2007 to 2008. Changes in out-of-pocket e

g a p p

4

m t r b s ( u

xpenditures (OOP) include premiums. i r F d

Fig. 4. Effect of coverage gap on stoppin

07.

ap coverage moved to a lower premium plan in 2008. However, mong switchers without gap coverage in 2007, 16% moved to a lan with gap coverage in 2008, 24% moved to a lower premium lan, and 26% enrolled in a lower deductible plan.

.3. Use of medical services

The primary concern of the coverage gap is that reductions in edication use may adversely affect health, which may increase

he use of non-pharmacy services. The panels in Table 6 show the esults from regression analyses of changes in medical service use efore and after the coverage gap, relative to the LIS group. Irre- pective of service type (inpatient, outpatient, ED), or condition diabetes-related or all), we find no difference in medical service se before and after reaching the coverage gap. Nor do we observe

ncreases in Part A&B use in the next year (results available upon equest). There are several possible explanations for these results. irst, the demand response to the coverage gap is modest for most iabetes-related medications. While a higher fraction of non-LIS

g therapy and resuming therapy.

1354 G.F. Joyce et al. / Journal of Health Eco

Table 6

A: Adjusted medicare Part A and B utilization, diabetes-related conditions

Non-LIS, gap coverage vs. LIS

Non-LIS, no gap coverage vs. LIS

Emergency department visits per month 0.0003 −0.0001 Outpatient visits per month 0.0005 −0.0025 Inpatient stays per month −0.0005 −0.0017*** Inpatient days per month −0.0008 −0.0076*

B: Adjusted medicare Part A and B utilization, all conditions

Non-LIS, gap coverage vs. LIS

Non-LIS, no gap coverage vs. LIS

Emergency department visits per month 0.0013 −0.0003 Outpatient visits per month 0.0016 −0.0056 Inpatient stays per month −0.0014 −0.0027** Inpatient days per month −0.0164 −0.0206**

Note: Sample (n = 292,200) includes beneficiaries age 65 and older with diabetes that reach coverage gap after January and before December in 2007. Individuals had to be enrolled and alive throughout 2007. Significance levels are indicated as the following:

b g c e e c p a s e t i m m i r e

m n

t p d o w T s s t i e r

5

i t t e a t i d h i c u b a i l s u y

w o b

T A

N M

* p < 0.050. ** p < 0.010.

*** p < 0.001. Year: 2007.

eneficiaries stop taking a statin, antihypertensive or oral hypo- lycemic after reaching the gap (Fig. 4), many reinitiate use once overage resumes in the next year, mitigating potential adverse ffects. Second, the marginal beneficiary is only subject to the cov- rage gap for a limited period of time. Beneficiaries that enter the overage gap early in the year are likely to reach the catastrophic hase, where insurance coverage is nearly complete (5% coinsur- nce). Thus, unless they are perfectly myopic, the demand response hould be moderated. On the other hand, fewer than 2% of those ntering the gap in the last half of the year reach the catastrophic hreshold. While they are more likely to reduce their use of med- cations, the median beneficiary is only subject to the gap for 3

onths. Third, the adverse effects of poor medication adherence ay take more time to manifest. We find no differences in med-

cal service use in the current year or the next year among those eaching the gap, but we do not capture longer term effects from

pisodic coverage.

As shown in Table 1, the LIS are different from the non-LIS in any ways: they are obviously poorer, more likely to be female,

on-white, and sicker on average. Thus, there may be some concern

fi o m c

able 7 djusted difference-in-difference in MPR and GDR by drug for employer-provided sampl

Number of observationsa MPR

Non-LIS, gap coverage vs. employer-provided

All Classes −0.031*** Statins 314,212 −0.048*** Anti-hypertensives, combo 88,904 −0.039*** Oral hypoglycemics 254,462 −0.033 Calcium channel blockers 140,588 −0.003 Anti-hypertensives, other 26,514 −0.006 ACE/ARB 237,134 −0.057*** Beta blockers 188,080 −0.013** Diuretics 127,760 0.005 Digitalis glycosides 39,466 −0.010 ote: Sample is individuals with diabetes and ages 65 and older that reach coverage gap odel includes age, sex, race, and indicators for ending in catastrophic phase, age-square a Number of observations for MPR calculations. Significance levels are indicated as the * p < 0.050.

** p < 0.010. *** p < 0.001. Average cost is empirically derived, for 30-day equivalent. Year: 2007.

nomics 32 (2013) 1345– 1355

hat they differ in unobserved ways that make them an inappro- riate control group although their constant level of prescription rug use before and after the coverage gap threshold suggests therwise. To test this, we re-estimate the models using seniors ith employer-provided insurance as an alternative control group.

hese seniors have drug coverage at least as generous as the tandard Part D benefit, and like the LIS, their demand for pre- cription drugs should be not be affected by the Part D coverage hresholds. The results presented in Table 7 confirm this. Averag- ng over the 9 diabetes-related classes, the difference-in-difference stimates for changes in MPR and GDR are very similar to those eported in Table 3, using the LIS group as a control group.

. Conclusion

Despite its success, Medicare Part D continues to be widely crit- cized for the gap in coverage, supported by several recent studies hat find the gap is associated with reduced adherence to drug herapies (Zhang et al., 2009a, b); (Hoadley et al., 2011), (Polinski t al., 2011). While reductions in prescription drug use are widely ssumed to adversely affect health, there is little reliable evidence o support this contention. We find that the coverage gap does, ndeed, disrupt the use of prescription drugs among seniors with iabetes. But the declines in usage are modest and concentrated in igher cost, brand-dominant classes. Demand for high cost med-

cations such as antipsychotics, antiasthmatics, and drugs of the entral nervous system decline by 7–18% in the coverage gap, while se of lower cost medications with high generic penetration such as eta blockers, ACE inhibitors and antidepressants decline by 2–5% fter reaching the gap. More importantly, lower adherence to med- cations is not associated with increases in medical service use, at east in the short run. Part D beneficiaries with diabetes display the ame patterns of inpatient, outpatient and emergency department se before and after reaching the coverage gap, as well as the next ear, both in absolute levels and relative to the LIS.

These findings do not imply that offering an insurance benefit ith a gap in coverage reflects sound public policy. Cycling in and

ut of coverage may be more disruptive to care plans than a stable enefit with higher coinsurance and closing the gap will reduce the

nancial risk associated with pharmaceutical spending. However, ur results suggest that behavioral responses to the coverage gap ay mitigate potential short-term health effects. After reaching the

overage gap, non-LIS beneficiaries are more likely to use lower

e.

GDR

Non-LIS, no gap coverage vs. employer-provided

Non-LIS, gap coverage vs. employer-provided

Non-LIS, no gap coverage vs. employer-provided

−0.035*** 0.038*** 0.028*** −0.048*** 0.035*** 0.018*** −0.040*** 0.043*** 0.01 −0.035 0.045*** 0.027*** −0.015*** −0.019* −0.019** −0.013 N/A N/A −0.059*** 0.012 0.003 −0.019*** 0.095*** 0.095*** −0.006 N/A N/A −0.012 −0.04** −0.04**

in 2007. Differences are in percentage points. Results are from regression models. d, and co-morbid conditions.

following:

G.F. Joyce et al. / Journal of Health Economics 32 (2013) 1345– 1355 1355

Table A1 Adjusted difference-in-difference in rates of stopping and resuming therapy in response to coverage gap.

Statins Anti- hypertensive, combo

Oral hypoglycemic

Calcium channel blockers

Anti- hypertensives, other

ACE/ARB Beta blockers

Diuretics Digitalis glycosides

Stop Non-LIS, gap coverage vs. LIS 3.8% 2.0% 1.6% −1.5% −2.3% −0.9% −0.9% −1.0% −0.5% Non-LIS, No gap coverage vs. LIS 4.5% 1.7% 2.7% −0.1% −0.9% −0.2% 0.4% 0.5% 0.1%

Resume Non-LIS, gap coverage vs. LIS 8.0% 4.6% 5.4% −0.4% −3.5% 5.6% 4.4% 1.9% 1.0% Non-LIS, no gap coverage vs. LIS 8.8% 6.9% 7.9% 3.7% −1.5% 5.8% 4.9% 0.7% 3.7%

Number of observations 245,482 76,586 195,949 122,675 27,835 207,418 153,582 117,598 29,731

N older u ar. Re u

c t a o

s p r p d t c b i o e p t

c h p r a w fi g g m t

A

t h a r a t h C a r

A

R

C

F

G

H

H

H

J

J

K

K

K

L

L

M

M

P

R

T

Y

ote: All results are relative to the LIS group. Sample includes beneficiaries age 65 and se of any drug in the class for at least 45 days without resuming in the calendar ye se in the first quarter of 2008. Years: 2007, 2008.

ost (generic) medications and to switch to more generous plans he next year. And while the non-LIS are more likely to stop taking

medication in the gap, they are more likely to resume therapy nce coverage resets in January.

Our results may overstate the impact of the coverage gap on pre- cription drug use if beneficiaries obtain free samples from their roviders or pay cash for medications at discount outlets after eaching the gap (Tseng et al., 2004). An increasing number of retail harmacies (e.g. Wal-Mart, Target) sell a broad range of generic rugs at $4 per prescription. While there is little empirical data on he extent of this behavior, one study finds that 6% of enrollees pur- hase prescriptions outside of their plan after reaching the annual enefit limit (Hsu et al., 2006). We observe a substantial and rapidly

ncreasing number of $4 claims in the Part D data, thus the extent f bias from uncaptured claims is likely to be small. Further, since ntry into the catastrophic phase is based on accumulating out-of- ocket expenses, beneficiaries have an incentive to purchase all of heir medications – even $4 scripts – through the Part D program.

Our results capture the effects of the coverage gap for a spe- ific group of beneficiaries and drug classes. Moreover, some of the igh cost classes in 2006-2008 are far less expensive today due to atent expiration, e.g. statins. Future work should examine demand esponses across different beneficiary groups, including minorities nd low-income seniors who do not qualify for subsidies. More ork is also needed in understanding the extent to which bene- ciaries change their prescription drug use in anticipation of the ap and whether these behaviors change over time as beneficiaries ain experience navigating the program. The Part D coverage gap ay have long-term consequences on beneficiaries’ health, but at

his point they are not evident.

cknowledgements

Support was provided by the National Institute on Aging, hrough grants 7R01 AG029514 and P01 AG033559. The sponsor ad no role in the design and conduct of the study; collection, man- gement, analysis, and interpretation of the data; and preparation, eview, or approval of the manuscript. All authors had full access to ll of the data in the study and take responsibility for the integrity of he data and the accuracy of the data analysis. None of the authors ave any relevant conflicts of interests. We are grateful to Patty St lair, Bryan Tysinger, Laura Gascue, Alejandro Bugacov, Raj Mehta nd Deborah Testa for expert programming, and to the editor and eviewers for helpful comments.

ppendix A.

See Table A1.

Z

Z

with diabetes that reach coverage gap in 2007. Stopping is defined as discontinuing sumption is defined as stopping a drug in the coverage gap in 2007 and resuming

eferences

handra, A., Gruber, J. et al., 2007. Patient cost-sharing, hospitalization offsets, and the design of optimal health insurance for the elderly. National Bureau of Eco- nomic Research, Inc, NBER Working Papers 12972.

airman, K.A., Motheral, B.R., et al., 2003. Retrospective, long-term follow-up study of the effect of a three-tier prescription drug copayment system on pharma- ceutical and other medical utilization and costs. Clinical Therapeutics 25 (12), 3147–3161, discussion 3144–3146.

aynor, M., Li, J., et al., 2007. Substitution, spending offsets, and prescription drug benefit design. Forum for Health Economics and Policy 10 (2).

oadley, J., Summer, L., et al., 2011. Understanding the effects of the Medicare Part D coverage gap in 2008 and 2009: Costs and consequences prior to improvements in coverage established by the 2010 Health Reform Law. Program on Medicare Policy. The Kaiser Family Foundation, pp. 1–46.

su, J., Fung, V., et al., 2008. Medicare beneficiaries’ knowledge of Part D prescription drug program benefits and responses to drug costs. Journal of the American Medical Association 299 (16), 1929–1936.

su, J., Price, M., et al., 2006. Unintended consequences of caps on Medicare drug benefits. New England Journal of Medicine 354 (22), 2349–2359.

ohnson, R.E., Goodman, M.J., et al., 1997. The effect of increased prescription drug cost-sharing on medical care utilization and expenses of elderly health mainte- nance organization members. Medical Care 35 (11), 1119–1131.

oyce, G.F., Goldman, D.P., et al., 2009. Medicare Part D after 2 years. American Journal of Managed Care 15 (8), 536–544.

aestner, R., Khan, N., 2012. Medicare Part D and its effect on the use of prescrip- tion drugs and use of other health care services of the elderly. Journal of Policy Analysis and Management 31 (2), 253–279.

aiser Family Foundation, 2011. The Medicare prescription drug benefit. Fact Sheet. K. F. Foundation. Kaiser Family Foundation, Washington, DC.

etcham, J.D., Simon, K.I., 2008. Medicare Part D’s effects on elderly patients’ drug costs and utilization. American Journal of Managed Care 14 (11 Suppl.), SP14–SP22.

evy, H., Weir, D., 2008. The impact of Medicare Part D on drug utilization and out-of-pocket spending: Evidence from the health and retirement study.

ichtenberg, F.R., Sun, S.X., 2007. The impact of Medicare Part D on prescription drug use by the elderly. Health Affairs 26 (6), 1735–1744.

cWilliams, J.M., Zaslavsky, A.M., et al., 2011. Implementation of Medicare Part D and nondrug medical spending for elderly adults with limited prior drug coverage. Journal of the American Medical Association 306 (4), 402–409.

otheral, B., Fairman, K.A., 2001. Effect of a three-tier prescription copay on pharmaceutical and other medical utilization. Medical Care 39 (12), 1293–1304.

olinski, J.M., Shrank, W.H., et al., 2011. Changes in drug utilization during a gap in insurance coverage: an examination of the medicare Part D coverage gap. PLoS Medicine 8 (8), e1001075.

esearch, K., 2011. Seniors’ Opinions About Medicare Rx: Sixth Year Update. K. Research. KRC Research, Washington, DC.

seng, C.W., Brook, R.H., et al., 2004. Cost-lowering strategies used by medi- care beneficiaries who exceed drug benefit caps and have a gap in drug coverage. Journal of the American Medical Association 292 (8), 952–960.

in, W., Basu, A., et al., 2008. The effect of the Medicare Part D prescription ben- efit on drug utilization and expenditures. Annals of Internal Medicine 148 (3), 169–177.

hang, Y., Donohue, J.M., et al., 2009a. The effect of Medicare Part D on drug and medical spending. New England Journal of Medicine 361 (1), 52–61.

hang, Y., Donohue, J.M., et al., 2009b. The effects of the coverage gap on drug spending: a closer look at Medicare Part D. Health Affairs 28 (2), w317–w325.