Physician vs. Patient Incentives in Prescription DrugChoice

Michael J. Dickstein∗

Stanford University

This version: March 23, 2014

Abstract

In response to the rise in health spending in the US, which totaled nearly $9000 percapita in 2012, insurers and government payers use two mechanisms to direct spendingtoward the most valuable treatments. The first set, “demand-side” incentives, imposecosts on the patient to limit moral hazard. The second set, “supply-side” incentives,use the physician’s payment as a tool to minimize agency conflicts. I design a new testof the relative effectiveness of these two avenues in maximizing the value of treatmentchoices. Using variation in patients’ and physicians’ exposure to incentives, I find newevidence that physician-directed incentives may raise long-run costs. Physicians reduceoffice-based primary care in response to new payment regimes, substituting prescriptiondrugs as well as referrals for speciality care. Short-run costs fall, but patients relapse athigher rates. I discuss the likely mechanism generating this trade-off and its implicationfor disease-specific insurance design.

Keywords: physician agency, capitation, health insurance design

∗I thank David Cutler, Guido Imbens, Ariel Pakes, Dennis Yao, and seminar participants at HarvardUniversity, Yale School of Public Health, US Department of Commerce’s Bureau of Economic Analysis,and the NBER Summer Institute for helpful comments. I conducted part of this research while apostdoctoral associate at the Cowles Foundation at Yale University. Email: mjd@stanford.edu.

1 Introduction

In the past two decades, annual health spending in the United States rose from $256

billion to $2.6 trillion, accounting for nearly 18% of the Gross Domestic Product in 2011.

In response, insurers and policymakers have designed incentives to steer this spending

toward only the most cost-effective medical interventions. The incentives focus broadly

on two conflicts that arise in the market for medical care. On the demand side, many

patients purchase insurance against health risks. Facing only a small fraction of the cost

of their care, patients may demand relatively expensive treatments. On the supply side,

because patients lack the training to diagnose and treat their own condition, they must

employ physicians to oversee their care. The physician, however, has her own private

interests and may favor more intensive interventions. Both conflicts drive increases in

spending.

I compare the effects of two categories of incentives intended to address these con-

flicts: “demand-side” incentives that affect the patient’s negotiation with his physician

and “supply-side” incentives that govern the transaction between the physician and

insurer. The goal is to inform a key policy question: what set of incentives best limit

the short-run costs from demand and supply-side conflicts without harming long-run

patient health? Typical incentive schemes in this setting only correlate imperfectly with

the insurer’s long-run objective, and so they create trade-offs for designers of an incen-

tive program (Baker (1992) and Lazear (2000)). For example, placing too much power

on demand-side cost sharing can cause patients to quit treatment early, decreasing the

likelihood that they will recover and increasing the rate of relapse. With the greater de-

mand for follow-up care, long-run costs to the insurer may actually increase. Stronger

supply-side incentives, by discouraging repeated physician-patient interactions, may

lead to less tailored treatment regimens and therefore poorer health outcomes.

I measure the importance of this trade-off in setting of prescription drug choice. A

long literature in health economics, reviewed by Goldman et al. (2004), analyzes the ef-

fect of patient cost-sharing on drug choice;1 relatively little empirical work analyzes the

effect of supply-side incentives.2 I extend the supply-side literature by providing mea-

surement of a new and important elasticity: when physicians face capitated payments

for a patient’s office visits, to what degree do they substitute away from providing their

1See, for example, Aronsson et al. (2001), Foo and Cullen (2012), Gaynor et al. (2007), Hellerstein(1998), Huskamp et al. (2005), Nair et al. (2003), and Shrank et al. (2007).

2Armour et al. (2001), Bloom et al. (2002), and Ho and Pakes (2011) measure the effect of supply-side incentives on aggregate health spending and on the costs of inpatient care.

1

own office-based treatments? Under capitation schemes, physicians receive a lump sum

prospective payment to cover all outpatient services provided to an insured patient.

The physician does not bill the insurer per office visit, and so has an incentive to econ-

omize on care. I look for evidence that these physicians recommend prescription drug

treatment or medical care provided by outside specialists; in both cases, the capitated

physician does not bear the cost of the treatment.

I proceed in the paper in two steps. First, I describe a model of the patient-physician

interaction, similar to the approach of Pauly (1980) and Dranove (1988), to highlight the

possible physician responses to three classes of incentives: patient cost-sharing, man-

aged care contracts, and capitated payment schemes. Second, I measure the response of

physicians to these incentives when treating patients suffering from depression. I choose

this setting for two reasons. First, depression care is common in the United States, as

major depression affects 6.5% of adults in the US each year. In 2008, patients filled 164

million monthly prescriptions in the antidepressant class; only cholesterol treatments

and pain medicines exceeded this level of sales.3 Second, for depression diagnoses, physi-

cians choose among multiple treatment options, including office-based psychotherapy

and a large set of prescription drugs. The American Psychiatric Association’s prac-

tice guideline lists 26 unique drug compounds approved for depression treatment, with

no single drug proven superior in efficacy (Karasu et al. (2000)). Physicians thus ex-

ercise significant discretion in prescribing treatments, leaving opportunity for various

incentives to operate.

The key findings in the empirical exercises appear immediately in the raw shares.

I compare patients whose plans use different financial incentives, from the least re-

strictive Preferred Provider Organization (PPO) plans that pay physicians per visit to

Health Maintenance Organizations (HMOs) that use a variety of financial incentives

to manage the level of health spending of their enrollees. In the data, HMOs differ

importantly in whether they pay primary care physicians using capitation contracts.

I find that patients with the least restrictive PPO plans that pay physicians per visit

receive psychotherapy at a rate of 8%, with 2/3 of those patients also receiving drug

care. In HMOs plans that nonetheless pay physicians per visit, the share receiving

psychotherapy is much lower, at 2%. Strikingly, for capitated HMO patients, nearly

12% receive psychotherapy, even larger than the share in PPO plans. These differences

in share remain largely unchanged even after controlling for geographic factors, patient

3“Top Therapeutic Classes by US Sales”, IMS National Sales Perspectives. IMS Health, 2008.

2

demographics, and physician speciality in a discrete choice framework.

Across the plan types, patients also receive distinct drug treatments during visits

in which the physician prescribes medication. Patients treated under plans that pay

primary care physicians by capitation receive fluoxetine, the generic form of the popular

branded drug, Prozac, at a rate of 18%, a full 10% more often than patients with plans

that pay physicians per visit. This increase in the use of fluoxetine comes at the

expense of prescriptions for heavily-promoted patented treatments, including Lexapro

and Zoloft. Under capitation, physicians choose these branded drugs half as often as

do physicians paid by fee-for-service schemes. Logit models that condition on patient

demographics, physician specialty, region indicator variables, and other background

characteristics show similar but slightly smaller effects: capitation and patient cost-

sharing incentives are associated with an increased share for generic treatments of about

14% and 7%, respectively, holding observables constant.

Given these reduced form results, I estimate a detailed model of the physician’s

treatment recommendation to isolate the mechanism through which the incentives gen-

erate the observed cost savings and patient outcomes. Given my intention to examine

multiple incentives in a single framework, it is difficult to find the ideal research design

in which insurers randomly assign the various incentives to patients and physicians.

Without random assignment, selection of patients into insurance plans or physicians

into contracts may bias the measurement. However, several features of the data allow

me to limit concern about selection bias. On the demand-side, to deal with selection of

patients into insurance plans, I restrict my analysis to only those patients newly diag-

nosed with depression. New patients are unlikely to have chosen their original insurance

plan as a function of the specific depression treatment prescribed to them at the time

of diagnosis. In addition, I include in the model a measure of each patient’s overall

health, the Charlson Comorbidity Index, to control for the possibility that patients

with multiple illnesses select into plans with cheaper out-of-pocket costs for depression

medications. In this way, I compare the treatment choices of patients with similar ill-

ness profiles and demographics and with broadly similar insurance plans. I exploit the

“within” patient and broad plan type variation in the payment incentives physicians

face and in the specific out-of-pocket prices patients face for each drug in the choice

set. In addition, for the supply-side measurement, I introduce an auxiliary dataset that

provides evidence of physician behavior in a panel context.

I estimate this detailed choice model using a mixed logit specification that permits

unobserved heterogeneity in physician and patient preferences over price. The results

3

amplify the findings in the raw data and the simple logit analyses: promoting generic

drugs by lowering their relative copayments, holding other incentives fixed, prompts an

increase in the share of generics from 21% to 38%. The change represents an average

price elasticity of -.3, near the median of previous estimates identified in experimental

and cross-sectional studies.4 Greater copayment rates have no significant effect on the

rate of psychotherapy prescribed.

While the response of patients to copayment incentives seems uncontroversial and

follows the pattern in existing work, the response of physicians to supply-side poli-

cies requires more analysis. For drug choice, I distinguish empirically between three

theories explaining the observed correlation between capitation and generic drug use:

(1) capitated physicians choose treatments that require fewer follow-up visits, because

the physician bears the cost of repeat visits; (2) other managed care regulations cor-

related with the use of capitation contracts drive physician choices; and (3) physicians

who enter into capitation contracts have underlying preferences for cost containment.5

I test these theories using both patient-level panel data and physician-level data. I

find support for the first explanation: physicians on capitation contracts concentrate

their prescribing on precisely the set of treatments that generate less switching, holding

prices and patient demographics constant. These products are often effective generics,

including fluoxetine, which price-sensitive patients prefer. Although I control for a rich

set of patient demographics, physician prescribing may reflect either self-interest or an

attention to unobserved patient characteristics.

To test hypotheses (2) and (3), I use cross-sectional data from a survey of physi-

cians. For each physician in the sample, the data contain information on the drugs the

physician selects for all patients she treats in a given week. Physicians in this data typ-

ically treat patients enrolled in a variety of insurance plans. I compare the prescribing

habits of physicians who accept capitation for some of the patients they treat against

the habits of physicians who do not accept capitation. The drug shares are statistically

indistinguishable across these two groups. Thus, there is little evidence that physicians

who accept capitation for some but not all of their patients act in a manner wholly

distinct from those physicians who never accept capitation.

For the choice between prescription drugs, psychotherapy alone, or no formal treat-

4See Goldman et al. (2004), Huskamp et al. (2005), and Ellison et al. (1997).5In the US healthcare market, physicians rarely face the direct cost of prescription drugs. The

exception is physician-administered drugs, such as oncology treatments, which physicians may pur-chase directly. In the depression setting, changes in drug price affect the physician through insuranceincentives.

4

ment, I find that primary care physicians facing capitation elect psychotherapy treat-

ment at higher rates. Relative to non-capitated HMOs with similar plan incentives

apart from the payment mechanism, those patients in capitated plans receive formal

treatment 2.8% more often. Patients of capitated primary care physicians receive psy-

chotherapy alone at a rate of 1.1%, an increase of .86% percentage points (75%).6

Finally, I evaluate the relative effect of supply and demand-side incentives on health

in the panel data. I find stronger cost-sharing leads to a decline in patient health.

Patients more often quit treatment prematurely, and the rate of relapse in the period 3-

6 months after the initial episode increases from 5.4 to 5.8%, holding constant observed

patient and physician attributes. Supply-side incentives have a similar but stronger

effect. When physicians face capitation incentives, patients adherence to the treatment

chosen by these physicians at a rate that is 2% lower than the adherence rate for patients

treated by physicians paid per visit. The probability of relapse increases 0.6% relative

to the baseline case with no incentives. This increase is 50% greater than the change

seen from introducing cost-sharing.

I proceed in the paper as follows. In Section 2, I provide a model of the patient-

physician interaction, demonstrating the potential for conflict and the role of insurer

incentives in this relationship. I then test the model’s hypotheses in an empirical

setting. I describe the data and motivate the empirical specifications in Sections 3 and

4. I analyze the static relationship between incentives and costs in Section 5 and the

dynamic relationship between incentives and health in Section 6. Finally, in Section

7, I discuss the implications of these analyses for optimal insurance design. Section 8

concludes.

2 Model of the Patient-Physician Interaction

I model the physician’s selection of drug treatment for an ill patient to illustrate the

potential for conflict. The asymmetry in information between the trained physician and

the patient gives the physician authority to recommend a treatment that is privately

optimal but may not maximize the patient’s outcome.

I describe special case in which the patient’s health falls into one of two states:

6This share includes patients initially diagnosed with depression in a primary care office visit buttreated using psychotherapy by other providers, including psychiatrists, within 3 months of the initialdiagnoses. In the data, psychiatrists submit 75% of psychotherapy claims, general practitioners submit9%, and the balance come from other medical specialists.

5

θS or θM , where θS represents a severe illness and θM a mild illness. To treat the

illness, the physician can recommend one of two treatments. The first, dH , is a high

cost blockbuster drug, while the second, dL, is a low cost drug.7 From the patient’s

perspective, the low cost drug, dL is preferable for a mild illness; dH is preferable for a

severe illness. Formally, I assume patient‘s utility takes the following linear form:

UPatientj (θk) = Yjk − Cj (1)

Here, the patient suffers from illness k and consumes treatment j. The patient’s utility

depends on: (1) Yjk, the effectiveness of j for illness type θk, in efficacy and tolerability,

and (2) his out-of-pocket cost for the drug, Cj.

When the patient feels ill, he cannot determine his illness type to select the appro-

priate treatment. He therefore visits a trained physician. The physician does not face

the costs of the treatment but may have private benefits from a particular choice. For

example, if the physician has prescribed treatment j for many patients in the past, she’ll

save the costs of researching the proper dosing levels. Similarly, if drug j is the most

common medication prescribed, she’ll face lower malpractice risk when choosing j. For-

mally, her utility depends on her own private benefits, bkj , as well as some consideration

for the patient’s utility:

UPhysj (θk) = bkj + γ(Yjk − Cj) (2)

where bkj = bj(θk), and γ ≥ 0

Through the altruism term, γ(Yjk − Cj), the physician accounts for the patient’s

treatment outcome and drug costs, Cj, weighted by γ. The physician cares about the

patient’s utility because of a responsibility to the Hippocratic oath. As γ increases

toward infinity, the physician acts solely to maximize the patient’s outcome.

In the two drug special case above, I make four simplifying assumptions. First, CH >

CL; corresponding to the label, costs for dH exceed those for the low cost alternative, dL.

Second, I assume the private benefits to the physician of choices dH or dL do not vary

with the patient’s illness. The physician always receives higher private benefits under

the blockbuster drug, dH . Third, I let E(Y SH ) > E(Y S

L ) and E(Y SH−Y S

L ) > CH−CL > 0.

In the severe illness state where θ = θS, treatment dH is more effective on average than

dL. The blockbuster drug, dH , is also more costly, but the value of the increased

7I use drugs for clarity of exposition; one of the treatments could instead represent an office-basedtreatment.

6

effectiveness exceeds the added cost. Finally, fourth, Iet E(Y ML ) ≥ E(Y M

H ). For mild

illnesses, treatment dL is weakly more effective than dH .

The antidepressant class matches the assumptions of this special case. For exam-

ple, several low cost treatments outperform new patented alternatives for mildly-ill

patients. In eight studies reported by Gartlehner et al. (2007), Effexor (venlafaxine), a

popular new medication, produced a larger treatment effect for patients suffering from

major depression relative to existing products; for less severe conditions, there was no

strong evidence of improved efficacy. However, in general tolerability, Effexor was asso-

ciated with a 10% higher incidence of nausea and vomiting relative to existing generic

treatments.

With this simplified setting, I describe the treatment selections that will result in

each of the illness states.

2.0.1 One Period Model

In the severe case, both the patient and physician prefer the high cost drug, dH . It is

more effective on average, and the better expected outcome exceeds the added out-of-

pocket costs. In the mild case, there is potential for disagreement. If the patient could

diagnose his own mild condition, he would prefer the low cost drug. The physician will

recommend the low cost drug if and only if her sensitivity to patient outcomes, γ, is

sufficiently high:

E(UPhysH (θM)) < E(UPhys

L (θM))⇔ (3)

bH − bLE(Y M

L − Y MH ) + (CH − CL)

< γ ⇔ (4)

When (1) the low-cost drug is more effective, (2) the blockbuster drug is far more

expensive than the low-cost drug, and (3) (bH − bL) is small, meaning the physician

gains little in private benefits from choosing the blockbuster drug, even weakly altruistic

physicians will select the patient’s preferred choice, dL.

In this setting, there is an simple mechanism to ensure the physician chooses dL

when the patient suffers from a mild illness: pay a lump sum transfer to the physician

of (bH−bL) whenever she prescribes the low-cost drug. With this transfer, the physician

receives bH in either state. The inequality in (4) becomes γ > 0 and the patient and

physician preferences align.

This first-best scheme fails in practice because the level of the required payment

7

is often unknown, given that the private benefits to distinct alternatives can vary id-

iosyncratically by physician. Even if patients knew (bH − bL), they may still lack the

necessary funds to pay it.

With wealth constraints, the patient suffers a decline in welfare from the agency

relationship whenever γ fails to satisfy the inequality in (4). In the second-best world,

insurers can increase the likelihood that the physician chooses appropriately by increas-

ing the gap in copayments between the high and low cost drugs. They may also mimic

the first-best transfer to a degree by paying small bonuses conditional on a physician’s

prescribing behavior.

2.0.2 Two Period Model

To see role of capitation, consider a two period model. Under capitation, the physician

faces an additional cost, S, if the patient returns for a follow-up visit in the second

period. These costs place additional liability on the physician’s initial treatment choice.

In a two-period world, even less altruistic physicians select the patient’s preferred choice.

To see this, consider a slight change to the two drug setting. After realizing outcomes

under drug dL or dH in period 1, patients return early to the physician for consultation

if they experience a poor outcome.8 I denote the probability of an early reassessment

and switch as pH , when the physician chooses dH initially in the mild illness case, and

pL, when the physician (correctly) chooses the low-cost drug, dL, in the mild case.

pH >> pL, and pH , pL lie on [0, 1]. In the case of early reassessments or illness relapse,

the physician faces an additional pecuniary or non-pecuniary cost, S:

UPhysj (.) = bj + γ(Yj − Cj)− S

The physician’s utility depends on her private benefits, the patient’s utility, and S.

As in the one period example, there is a conflict when the patient suffers from a

mild illness. The expected outcome for the patient in the mild case is higher under the

low-cost treatment. The patient also pays less for the low-cost drug: CL < CH . Thus,

the low-cost drug is the efficient choice for the patient in period 1.

The forward-looking physician will prescribe the low-cost drug if the inequality

8Specifically, patients will return if the outcome causes them to update their prior beliefs on thequality of the drug sampled such that its expected utility falls below the expected utility of thealternative.

8

below holds:9

E(UPhysL + UPhys

j ) > E(UPhysH + UPhys

j )⇔ (5)

(bH − bL)− γE(UPatientL − UPatient

H ) <S ∗ (pH − pL)

2− pH − pL(6)

where pH reflects the probability of a costly reassessment when the physician initially

chooses dH and pL reflects the equivalent probability should the physician choose dL

initially. Here, UPhysj is the physician’s utility under a second period choice, j. Dividing

through by E(UPatientL − UPatient

H ) > 0:

γ >(bH − bL)

E(Y ML − Y M

H ) + (CH − CL)− S ∗ (pH − pL)

(E(Y ML − Y M

H ) + (CH − CL)) ∗ (2− pH − pL)(7)

The inequality suggests that, as in the one period case, there are ranges of γ for which

the physician will choose the blockbuster drug initially to treat mild illnesses, against

the interests of patients. The second right-hand side term in (7) is unique to the two-

period case. With this positive-valued term, the inequality is now easier to sustain for

a given γ and for non-zero S, the financial or psychic costs to the physician from a

reassessment. If capitation contracts serve to increase costs S, the inequality in (7) will

hold even for less altruistic physicians. However, capitation will have little effect if the

likelihood of a reassessment is the same regardless of the initial choice. If pH = pL, the

inequality in (7) reduces to inequality (4) from the one period case.10

I apply the insights from this simple two treatment example to the setting of antide-

pressant choice. Specifically, I test whether increasing the gap in patient copayments

between treatments leads physicians to prescribe more cost-effective drugs; I consider

the effect of changes in HMO design, which can alter the physician’s relative private

benefits from treatments; and, I examine how capitation might encourage better initial

matching to avoid treatment reassessments.

9The utility of the second period choice, labeled j, differences out of the inequality except for thecost of reassessment. The choice at period 2 would play a larger role if the horizon extended to futurechoices with a discount rate below 1.

10By focusing on the physician’s “inducement” of demand for her services given asymmetric infor-mation, the above discussion most closely matches the agency approach of Pauly (1980) and Dranove(1988), following the taxonomy described by McGuire (2000). One key distinction is that the choiceof drug treatment is observed and correlates with health status. Thus, a first-best contract, similarin spirit to health-contingent payments proposed by Arrow (1963), is possible. Such schemes breakdown, as discussed, because wealth-constrained patients cannot afford the contractual transfers orcannot determine how large this transfer must be to override the physician’s private benefits.

9

3 Data

To estimate the effects of demand and supply-side incentives, I employ health insur-

ance claims data available from Thomson Reuters’ MarketScan databases. I obtain

data for 2003-2005 from the Commercial Claims and Encounters database, which con-

tains patient level clinical utilization, expenditure, and enrollment data for inpatient,

outpatient, and prescription drug services. I link this data to patient demographics and

plan design features from the Benefit Plan Design database.11 The individuals recorded

in the data include active employees working for a group of large US firms that contract

with one of 100 participating payers; the employees’ dependents and some classes of

retirees enter the database as well.

I collect a sample of patients diagnosed in an outpatient office visit with one of five

categories of depression diagnoses: major depression; dysthymia and depression with

anxiety; prolonged depressive reaction; adjustment disorder with depressed mood; and,

depression not otherwise specified.12 Conditioning on observed diagnoses rather than

observed prescriptions has three benefits: (1) I can focus on more severe depression cat-

egories for which formal guidelines recommend medical intervention, (2) I can examine

the extensive margin choice of no treatment vs. either drug treatment or psychotherapy

and (2) I avoid episodes involving off-label use of antidepressants. However, by requir-

ing the patient receive a depression diagnosis, I may miss individuals in the broader

dataset who do not receive a formal diagnosis but may suffer from depression.13 My

analysis, therefore, pertains only to patients and physicians likely to record a depression

diagnosis during the physician office visit.

I impose the following conditions on patient backgrounds to form an appropriate

sample: patients cannot have a concurrent diagnosis of bipolar disorder or schizophrenia

or receive drugs that signal these conditions, as these illnesses require distinct treat-

ments14; the patients’ age must fall between 18 and 64, the range for which the data

are complete; patients must visit a health professional with the ability to prescribe all

11Thomson Reuters MarketScan Research Databases. Ann Arbor, MI: 2003-2005.12The depression diagnoses listed match the International Classification of Diseases (ICD-9-CM)

codes of: 296.2, 296.3, 300.4, 309.0, 309.1, and 311. Melfi et al. (1998), Pomerantz et al. (2004), andAkincigil et al. (2007) use similar diagnostic codes in selecting a sample of depression patients.

13Davidson and Meltzer-Brody (1999) discuss the widespread under-recognition and under-treatmentof depression, particularly in primary care settings.

14Patients excluded due to comorbidities have a diagnosis in one of the following classes: bipolarand manic disorders (ICD-9-CM 296.0, 296.1, 296.4-.8) and schizophrenic disorders (ICD-9-CM 295.0-295.9).

10

possible treatments; and patients must not be pregnant, a condition that raises safety

concerns for many of the treatments. I eliminate individuals dispensed medications

within the first six months of data, since the illness could be pre-existing. In such

cases, I would mistakenly interpret a second or third treatment choice as the initial

decision. For the purposes of this static choice model, I restrict the analysis to the first

prescription filled after the patient’s initial depression diagnosis. I use the patient’s

entire episode of treatment for duration model examinations. The initial filters lead

to a dataset of 98,139 unique patients observed between July 2003 through December

2005.

The data include background variables specific to the individual: (1) patient demo-

graphics, including age, gender, home county, and diagnosis; (2) the specialty of the

treating physician; and, (3) characteristics of the patient’s insurance plan, including

the required copayments. The MarketScan data also include a plan-specific “ingredi-

ent cost” for each drug, which is the insurer’s cost excluding the dispensing fee, sales

tax, and rebates from the drug manufacturer. I supplement this data on drug prices

with information on each drug’s average side effects, dosing, and efficacy from psychi-

atry textbooks, clinical practice guidelines, and a meta-analysis of published clinical

trials.15 From the MarketScan data, I also record whether a patient’s episode contains

any claims for psychotherapy treatment during the period. Finally, I collect county-level

information on reported wages, dividends, and interest income using the U.S. Internal

Revenue Service’s Statistics of Income for the 2005 tax year. I match this data to the

county reported in patient records.

Tables 1 and 2 contain summary statistics on the individual and drug-specific data.

Of the 98,139 unique enrollees, 44% visit general practitioners, 28% visit a psychiatrist,

and 28% visit other specialists, such as obstetricians. 28% suffer from major depressive

disorder, the most severe diagnosis in the depression hierarchy. Women compose 71%

of the observed sample diagnosed with depression. Of the plan types, capitated health

maintenance organizations (HMOs) cover 33% of patients in the sample; non-capitated

HMOs cover 14%; and other non-capitated plans, including preferred provider organi-

zations (PPOs), cover the remaining 53% of patients.

The plan type categorical variable plays an important role in the analysis. In panel

1 of Table 8, I list the types of incentives commonly employed by each plan category,

including whether the plan uses drug formulary tiers, utilization review, and precer-

15See Murphy et al. (2009), Karasu et al. (2000), and Gartlehner et al. (2007).

11

tification requirements for inpatient care. Capitated and non-capitated HMO plans,

as defined by the database, differ in that capitated plans pay general practitioners for

outpatient care using lump-sum payments rather than reimbursing physicians for each

visit or procedure. One can find evidence that capitated plans indeed employ such

payment schemes by looking at the types of complementary incentives included in the

alternative plan types. Crucially, the two types of HMOs both have strict drug formu-

laries and more often employ carve-out plans for mental health care relative to PPOs.

The HMO plans differ in that the capitated HMOs rarely employ utilization review

and case management–in practice, there is little need for explicit utilization review if

payment incentives cause capitated physicians to internalize the cost of their referrals.

In the sample period, physicians choose between 19 medication options. The most

common medications appear in Panel 1 of Table 1, along with their relative market

shares, the average patient copayments, and the reported insurer costs across years. The

psychotherapy share reported includes only those patients who received psychotherapy

but no drug care; if a patient received both, his treatment would contribute to the

relevant share of the drug he filled.

Using medical references, I categorize the set of treatments according to their ef-

fect on the concentration of three neurotransmitter chemicals in the brain—dopamine,

norepinephrine, and serotonin—that previous research connects to feelings of well-being

(Murphy et al. (2009)). The subclasses include: SSRIs, which selectively inhibit the re-

uptake of serotonin; NDRIs, which affect norepinephrine and dopamine; SNRIs, which

affect serotonin and norepinephrine; TCAs or tricyclic antidepressants, which unse-

lectively block the reuptake of serotonin; serotonin receptor agonists or antagonists,

labeled SM as serotonin modulators; and, NASSAs, noradrenergic and specific seroton-

ergic antidepressants, which similarly affect serotonin receptors.

The costs reported in Table 1 differ substantially across drugs. For a given drug,

there is also significant variation across insurance plans. I use the plan-specific drug

costs in estimating the choice models. In addition, two drugs entered in 2004, the

center of my sample period: citalopram, the generic version of Celexa, and a new

branded SNRI, Cymbalta (duloxetine).

12

4 Empirical Motivation

To motivate a detailed study of the physician’s choice behavior, I first present correla-

tions between insurer-designed incentives and the choice of antidepressant treatment.

In Table 2, I present the distribution of drug choices conditional on supply-side incen-

tives. Because I use insurance claims data, I only observe prescriptions actually filled.

If a patient received a prescription for an antidepressant during an office visit but never

filled it, that patient would fall into the ‘none’ treatment share.16 I separate out the

share of individuals who receive psychotherapy only, according to the insurance record.

In Table 2, individuals receiving both psychotherapy and an antidepressant fall into the

share for the relevant drug treatment.

From the observed shares, it is clear that, unconditionally, patients treated under

HMO plans that offer capitation contracts receive a far different distribution of drugs

than those patients treated under PPOs or non-capitated HMOs. Capitated physicians

use generics, before conditioning, at a rate of 36%, 13% more than physicians paid by

non-capitated HMOs and 23% more than physicians treating patients under PPO plans.

The largest difference appears in the shares of two treatments in the same ingredient

subclass: Lexapro (escitalopram), a branded treatment, and fluoxetine, the generic form

of the branded drug, Prozac. Lexapro’s share for patients treated under capitation is

8.5% relative to 15.6% for patients on non-capitated plans. The reverse is true for

fluoxetine: 17.6% of patients on capitated plans receive fluoxetine, while only 7.4% of

patients on non-capitated plans do. I separate out later whether these differences in

preferences relate to patient characteristics, supply-side incentives, or the copayments

that patients pay.

On the extensive margin, the unconditional shares in Figure 1 illustrate that supply-

side incentives affect the three broad category choices of no treatment, psychotherapy, or

prescription drug care only. Comparing across HMO plans, those that pay primary care

physicians via capitation show much greater use of psychotherapy to treat depression

patients: the rate of psychotherapy treatments is 11.5% in capitated HMOs, vs. 2.1%

for non-capitated HMOs. As a comparison, patients on the least restrictive PPO plans

in the private insurance data use psychotherapy for the depression diagnoses I study at

a rate of only 8%. The use of psychotherapy in capitated plans is largely from primary

16Similarly, if the physician offered the patient a free sample of drug A, that patient would notcontribute to the share for A unless he filled a new prescription for it at a pharmacy within 3 monthsof his initial office visit.

13

care physicians referring patients to psychiatric specialists; primary care physicians

conduct only 9% of all psychotherapy conducted in the broader dataset.

0"

10"

20"

30"

40"

50"

60"

Branded/drugs/only" Generic/drugs/only" No/treatment" Psychotherapy/(total)"

Share/(%

)"Treatment(Choice,(by(Plan(Type4

Capitated/HMO" NonGCapitated/HMO" PPO"

Figure 1: Treatment choice, by plan type

I replicate this simple comparison of treatment choices in a logit framework, to con-

trol for incentives and observable patient background characteristics that the medical

literature suggests may affect outcomes. Specifically, I condition on patient gender,

patient age summarized into five blocks, and the patient’s diagnosis. Moreover, Heller-

stein (1998), Jones et al. (2001), and Filippini et al. (2006) demonstrate that prescribing

behavior differs according to region, the physician’s specialty, and other socioeconomic

factors. I control for the specialty of the treating physician, the average income in the

patient’s home county, whether the individual lives in a metropolitan area (as defined

by the US Census Bureau), and also whether the patient lives in a state where the law

requires pharmacists to dispense generic drugs when possible.17

From the logit results shown in Panel 1 of Table 3, it is clear that holding observ-

ables constant, the main results persist: physicians with incentives under capitation

prescribe far more generics. The generic share is roughly 14% higher under capitated

HMOs relative to either PPO plans without capitation or HMO plans without capita-

17In 14 states, state law mandates that pharmacists dispense the generic form of a prescriptionwritten for an off-patent branded compound unless the physician expressly forbids substitution. In theremaining states, pharmacists exercise discretion.

14

tion. In Panel 2 of Table 3, when patients face cost-sharing, the share of physicians

recommending branded drugs falls 6.5% relative to physicians whose patient does not

face these incentives. With cost-sharing, 1.7% more patients fail to initiate treatment

all together.

For the margin of drug vs. office-based treatment, the logit results show that patients

who visit primary care physicians paid via capitation do receive more psychotherapy

than their observably similar counterparts whose plans pay physician per visit. The

rate of psychotherapy performed at some point in a patient’s episode is 0.85 percentage

points (80 percent) higher for capitated patients.18

5 Incentives and short-run costs

The summary statistics and initial discrete choice analysis suggest stronger incentives

correlate with greater use of generic drugs. I estimate the physician’s utility using

a mixed logit model to allow patients and physicians to respond idiosyncratically to

incentives.19 I then determine the likely mechanisms through which copayments and

capitation influence the physician’s treatment recommendation.

5.1 Product Choice Model

I estimate the parameters of the joint patient and physician utility function, choosing a

specification similar to the theoretical model outlined earlier. The patient cares about

an option’s effectiveness and his required copayment. In making a recommendation

to the patient, the physician accounts for the patient’s utility, to degree γ, but also

maximizes her own private benefits.

I include drug fixed effects to proxy for a medication’s expected efficacy and toler-

ability as well as features like the convenience of the required dosing.20 I measure the

18The baseline predicted rate of psychotherapy in this logit analysis is 1.06% when the primary carephysician is the first to record the patient’s depression diagnosis. If a psychiatric specialist records thefirst diagnosis, presumably after a primary care physician referred the patient for specialty care, thepredicted baseline rate of psychotherapy for the illustrative patient’s profile is higher, at nearly 15%.

19A simple conditional logit produces cross-price elasticities proportional to the prior shares of thedrugs, rather than allowing individuals in the model to substitute to a new treatment more heavily fromtreatments that are similar. This is known as the Independence of Irrelevant Alternatives property.By ‘mixing’ over the distribution of price sensitivities in a mixed logit, I break the counter-intuitiveproperty by producing correlations in the unobserved portion of utility for products with similarobserved characteristics.

20If firms choose formulary prices based on these product characteristics, including fixed effects

15

patient’s sensitivity to copayment levels by including the log of the copayment in the

specification. I allow patients to differ in their sensitivity to price depending on their

age, diagnosis, and the average income in the patient’s home county. I also introduce

random coefficients on the price variable to capture unobserved individual heterogene-

ity in the patient’s price sensitivity. For individual i on drug j in period t, the patient’s

utility equals:

UPatientijt = bi ln(copayijt) + aZPatient

ijt + εPatientijt

where unobservables that vary over time fall into a logit error term, εPatientijt . Here,

ZPatientijt contains the drug fixed effects and the interactions between the patient’s de-

mographics and the copayment level. The copayment level itself enters the model with

a random coefficient, bi.

The physician accounts for these features but may also react to the insurer’s cost,

particularly if the insurer enforces strict regulations on prescribing. I introduce an

insurer price variable with a random coefficient to account for unobserved heterogeneity

in responsiveness across the physician population. In the model, physicians may also

respond differently to insurer costs and copayments depending upon their specialty and

the characteristics of the patient’s insurance plan. These interactions enter ZPhysijt with

fixed coefficients:

UPhysijt = γUPatient

ijt + di ln(insurer costijt) + cZPhysijt + εPhys

ijt (8)

= (γbi, di)

(ln(copayijt)

ln(insurer costijt)

)+ (γa, c)Zijt + εijt (9)

= β′iXijt + α′Zijt + εijt (10)

Here, βi follows a normal distribution, with mean and variance (b,W ) to estimate. Zijt

combines ZPatientijt and ZPhys

ijt . I use a joint error term for the patient and physician, εijt,

which follows an extreme value distribution. I cannot separately identify γ from bi and

a. I include it in the above formulation to illustrate that the level of the physician’s

attention to patient interests may lessen the influence of copayments on the choice.

The strength of demand-side incentives relative to supply-side incentives is an empirical

question; the elasticities with respect to the policies depend on either the size of γ or

the size of bi and a relative to di and c.

in the specification limits the potential bias from the relationship between price and time-invariantunobservables.

16

I estimate three specifications, differing in the components of the matrix Zijt. I

proceed in estimation via a Bayesian Markov Chain Monte Carlo (MCMC) approach

in which the resulting estimates equal the average over draws from the posterior distri-

butions of the parameters of interest.21

The top panel of Table 4 contains both the fixed coefficient estimates as well as the

estimated parameters of the normally distributed random coefficients on the copayment

and the insurer cost. While I consistently estimate these population parameters, I

cannot estimate consistently any individual’s parameters, since I condition only on

the first period choice.22 As a result, in counterfactual predictions, I use Bayes’ rule

to form a conditional distribution of the individual coefficients, conditioning on the

observed first period choice. I draw from this posterior distribution using a Metropolis-

Hastings algorithm and use the draws in place of a direct estimate of the individual’s

coefficients.23

Specification (1) implies that the sensitivity to patient prices follows a normal distri-

bution with a mean equal to -.265 and a variance of .026; the insurer cost parameter has

a mean equal to -.043 and a variance of .009. In Figure 2, I illustrate the distribution

of patient and physician sensitivities to both copayments and insurer prices.24

Specification (2) includes an interaction term I create by multiplying the insurer’s

cost with a capitation indicator. The coefficient on this term suggests capitated doc-

tors care more on average about the insurer’s cost than do other physicians. Physicians

treating patients diagnosed with more severe illness categories, holding other observ-

ables constant, care far less about copayment levels. As in the theoretical model, when

the likelihood of relapse or reassessments is high regardless of the treatment selected

initially, incentives may have weaker effects on the physician’s choice. In addition, I find

21I follow an MCMC approach in estimation for two reasons. First, it reduces computation. Withinthe estimation procedure, I collect draws from the conditional distribution of the individual’s pricesensitivity parameter. I reuse these draws later to predict drug shares under counterfactual pricingpolicies. Second, in this setting, including individual-specific heterogeneity in treatment choices in theempirical model allows me to fit the data much better than using a simpler discrete choice model. Ichoose hierarchical priors to capture flexibly the relationships between heterogenous consumers. Anappendix containing convergence tests for the MCMC estimation are available from the author uponrequest.

22I would need to observe multiple choice situations for an individual, say across markets or time,to obtain a consistent estimate of the individual-level parameters (Train (2003)).

23Drawing from this distribution requires no additional computation time because I have the nec-essary draws in storage as a by-product of the MCMC procedure used to estimate the mixed logitparameters.

24The plots represent 100,000 draws from a normal distribution with the mean and variance estimatedin Specification (1).

17

−1 −0.5 0 0.50

2000

4000

6000

8000

10000

12000

14000

`i

No o

f Ind

ividu

als

Figure 2: Distribution of Mixed Logit Random CoefficientsFor Insurer Cost and Patient Cost per Day Supply

`i ins.cost`i copay

Figure 2: Distribution of Mixed Logit Random Coefficients. For insurer cost and patientcost per day supply.

that psychiatrists place significantly more emphasis on lower copayments in their choice

than do general practitioners. Physicians prescribing treatment for patients with many

comorbidities–and likely may more chronic diseases requiring drug treatment–also seem

to favor drugs with lower patient out-of-pocket costs.

To examine model fit, I compare the predicted shares under the mixed logit esti-

mates in specification (3), shown in the “base” column of Table 5, with the raw shares

reported in Table 2. The general pattern and magnitudes of the predictions match the

observed shares well. I also calculate implied price elasticities in the model, averaged

over physician and patient-specific sensitivities. The elasticities range between -.3 and

-.4, depending on the drug treatment. This is at the median of the range of elasticities

found in a meta-regression analysis by Gemmill et al. (2007).

5.2 Patient-directed incentives

I use the estimated price sensitivities to illustrate how alternative copayment policies

may affect generic drug use and overall health spending. These estimates account for

both observed and unobserved heterogeneity in price sensitivity. Both the physician and

patient’s preferences may play a role in generating the observed shares. If physicians

18

shift their prescribing toward cheaper medications under stronger incentives, this may

stem from altruism or from pressure applied by the patient. When initiation increases,

it may reflect either the physician writing more prescriptions or patients actually filling

the prescription the physician writes.

The results of three counterfactual experiments related to price incentives appear in

Table 5. In the first, I change the relative levels of the copayments such that branded

drugs cost about $50 per 30 day supply, the 95th percentile of copayments in the data.

The result is a slight shift toward generics—a 2.5% increase—with some reorientation in

shares toward drugs with lower costs and lower average side effects, such as fluoxetine.

Examination 2 shows more drastic effects. Shifting the copayment of generic drugs to

zero leads to an increase in the generic share from 20.6% to 37.6%. Part of that increase

comes from greater initiation of treatment, as the outside option loses nearly 6% from a

baseline of 24.3%. Given the same marginal cost between ‘none’ and generics, patients

and physicians shift to pharmacologic treatment, judging it to have higher expected

health returns. Examination 3 shows little effect from changing the underlying insurer

cost. Under current incentives, physicians seem attuned to patient satisfaction but

not directly to the insurer’s bottom line. These results match findings in earlier work,

including Goldman et al. (2004) and Huskamp et al. (2005).

In interpreting these results, one potential concern is that patients select into par-

ticular plans based on the plan’s copayment level. Part of the measured price response

may reflect patient characteristics that drive plan selection. I address this concern in

two ways. First, I estimate the choice model on data from private insurance claims,

restricting the analysis to patients newly diagnosed with depression. In this setting,

patients choose from a small range of plans, unlike the larger menu available to pa-

tients buying coverage on the individual market or choosing a Medicare Part D plan.

Using information on the patient’s first treatment upon diagnosis also makes it unlikely

that patients have selected into one of their employer’s insurance plans based on the

expected costs of unforeseen depression treatment. Second, I test whether the overall

health of the patient is correlated with both the level of copayments and the likelihood

of receiving particular expensive drugs. If patients select into plans based on health

status, one might expect those patients with more severe past and ongoing illnesses will

choose plans with better coverage of depression drugs. I control for a measure of the

patient’s health, the Charlson comorbidity index in specifications (2) and (3) in the top

19

panel of Table 4.25 The qualitative results change very little with these specifications.

5.3 Physician-directed incentives

The predicted shares in Examination 4 in Table 5 reveal the influence of capitation on

physician decisions. When insurers pay physicians using capitation, physicians select a

distinct distribution of treatments, much like in the raw shares in Table 2. The share of

individuals on generics increases from 20.6% to 35.5%, roughly the same increase seen

in the counterfactual world in which copayments on generic drugs equal $0. Physicians

again prefer standard generic drugs including fluoxetine, citalopram, and paroxetine.

They use fewer drugs in the NDRI and the newer SNRI sub-classes.

There are multiple channels through which capitated payments affect the physi-

cian’s drug choice. I distinguish between three major hypotheses empirically, each with

different welfare implications: (1) capitated physicians seek out drugs that cause fewer

patient follow-up visits, since physicians bear the risk of repeat consultations; (2) plans

that pay physicians using capitation impose a variety of regulations on the physician

that change her behavior apart from the reimbursement scheme; and (3) capitated plans

contract with physicians predisposed to cost containment. I look for evidence of each.

To examine the first theory, I control in the product choice model for patient diag-

nosis, age, county-level income, the specialty of the treating physician, a comorbidity

index, and an unobserved characteristic that I use to proxy for income variation in the

patient population. Controlling for these characteristics, I still find significant differ-

ences in drug shares under capitation, as reported in the final column of Table 5.

To examine whether physicians select particular treatments with a lower likelihood

of switching, I run a hazard model with a dependent variable that equals the time

to the first “switch.” A drug switch occurs when the patient switches either to a new

antidepressant or quits treatment altogether. I allow the hazard of switching to change

over time by employing a Weibull distribution in the empirical specification. I control for

drug compound fixed effects in the specification to reduce the concern that unobserved

choice characteristics introduce bias in the remaining coefficients, conditional on the

model choice. In addition to controlling for observed incentives, I also include a rich

set of patient background characteristics, including age by quantile, gender, diagnosis,

region, and county-level income, as well as the physician’s specialty. I report the model’s

25Charlson et al. (1987) define an index that predicts a patient’s ten-year mortality as a function ofprevious diagnoses, including heart disease, diabetes, HIV/AIDS, and cancer.

20

estimates in Table 6 and the predicted hazard rates by drug in Table 7.

The hazards by drug compound differ importantly and in ways partially predicted

by clinical research. Table 7 illustrates, for example, that SSRIs paroxetine (Paxil),

sertraline (Zoloft), and fluoxetine (Prozac) have some of the lower hazard rates, all

near 23-25%. Comparing the shares in Table 5 with the predicted hazards from the

Weibull model in Table 7, the drugs prescribed more often by capitated physicians are

precisely those with lower hazards of switching.

This result has two possible interpretations. First, it could reflect capitated physi-

cians choosing treatments that limit costly follow-up care out of self-interest. Second,

though I control as best as possible for patient-level attributes, the model may miss

a characteristic that explains both patient selection into capitated plans and patient

preference for drugs that cause fewer repeat visits. Physician behavior might then re-

flect a desire to match patient preferences. Thus, I judge the impact of incentives on

patient welfare not by the distribution of prescriptions alone, but also by patient health

outcomes. I show later that patients on capitated plans quit treatments early and re-

lapse at greater rates. This suggests capitation incentives may lead to choices out of

line with patient preferences.

As a second rationalization for the choices of capitated physicians, I test whether

coincident incentives employed by capitated HMO plans enter the physician’s decision

process. In addition to prospective payment, insurers may issue warnings to physicians

when they prescribe high volumes of expensive drugs or monitor their actions via elec-

tronic medical records. To examine this hypothesis, I collect information at the plan

level on the types of incentives employed across capitated HMO plans, non-capitated

HMO plans, and PPO plans. In Table 8, I list the frequency of each tool by plan type.

For example, the two categories of HMO plans both use formulary incentives more

often than PPOs and charge higher prices for “non-preferred” drugs. The capitation

effect described earlier appears when comparing capitated HMOs with non-capitated

HMOs. This within-HMO comparison controls for the effects of correlated incentives,

suggesting the main findings relate importantly to the reimbursement scheme itself.

Finally, I consider whether the selection of providers into the plan network explains

the capitation results. Physicians electing to accept capitated payments for medical

services may be those with better knowledge of available drug alternatives or may have

a higher sensitivity to patient costs and outcomes. To provide some evidence on this

hypothesis, I employ a physician-level dataset, the National Ambulatory Medical Care

21

Survey (NAMCS), which samples patient visits to office-based physicians.26 The sur-

vey collects information on the patient’s background and the physician’s characteristics,

for all patients who visit the surveyed physician in a specific recording interval. The

sample provides detailed information on the physician’s managed care contracts, the

characteristics of her practice and patient population, and on the physician’s prescrip-

tion choices across patients. However, unlike the main dataset, it lacks detail on the

payment scheme for any particular patient. That is, when observing the choice for a

patient, I cannot identify whether that patient’s plan pays the physician using capita-

tion. I therefore analyze variation in the behavior of a cross-section of physicians in the

NAMCS data who differ in their capitation share. I collect the NAMCS data for the

same sample period and for the same diagnosis classes chosen in the main analysis.

Panel 1 of Table 9 shows the correlations in the data across the physician practice

characteristics on which I condition in the choice model. Panel 2 shows results from a

logit model that varies the characteristic, “physician accepts new patients on capitated

plans,” while holding constant other physician and patient background characteristics.

If underlying, stable preferences govern a physician’s prescription choices independent of

financial incentives, I would expect physicians who accept capitation for some patients

to write fewer branded prescriptions and more generic prescriptions for all of the patients

they treat. In addition, I estimate a choice model that controls for capitation and

other physician attributes, such as whether the insurer directly employs the physician.

If adding these controls diminishes the effect from capitation, then it may be these

attributes—and not the reimbursement scheme alone—that explains observed behavior.

I present the predictions from this logit exercise in Panel 2 of Table 9. Physicians

who accept capitation for some patients prescribe substantially the same treatments

across their entire patient panel as do physicians paid only via fee-for-service schemes.

Physicians accepting some capitated patients write slightly fewer branded prescriptions

for Lexapro, but the difference in share is far smaller than the differences in prescribing

observed in the main analysis. Overall, the physician-level data provides little evi-

dence that physicians who accept capitated plans are predisposed to cost containment,

independent of the financial incentives.

26National Ambulatory Medical Care Survey, National Center for Health Statistics, Hyattsville, MD:2003-2005.

22

6 Incentives and health

Improving the quality of depression care requires addressing both the cost and health

dimension of outcomes. I measure the effect on health using a hazard model of exit

from drug care and a conditional logit predicting rates of relapse in the sample.27

6.1 Adherence and relapse

I estimate a hazard model to predict the time to exit as a function of incentives. I use

a Weibull distribution in the empirical specification to allow the hazard to vary within

a patient episode. The estimated scale parameter in Table 10 implies an increasing

hazard—the pace of exit from care accelerates the further one is into an episode. I

control for a large set of patient demographics as well as the physician’s specialty.

To capture differing hazards by drug, I include drug indicator variables as well as

an indicator for whether a drug requires two or more daily doses. Finally, I control

for managed care incentives, for the level of the patient’s drug copay for his initial

treatment, and for the patient’s plan type—capitated HMO, non-capitated HMO, or

PPO. This analysis includes the prescription paths for the same 98,139 unique patients

appearing in the product choice model.

The estimates from the hazard model appear in Table 10. The dependent variable

reflects the number of monthly decision points that elapse before the observed exit

from treatment.28 Fixing other patient background characteristics, more severely-ill

patients and those in higher income counties remain on treatment longer. Relative to

general practitioners, patients treated by psychiatrists remain on treatment longer, as

do patients treated by other specialists. The characteristics of the initial drug chosen

also matter: patients starting on drugs that require 1-2 doses per day as opposed to

2-3 show greater adherence to treatment.

Copayments and capitation both importantly increase the hazard of exit conditional

on drug fixed effects and background variables. I illustrate this pattern for capitation in

Figure 3. The estimated hazard rates begin around 20% and increase to 30-35% three

27The clinical literature distinguishes between relapse, the return of symptoms in an ongoing episode,and recurrence, a new episode following a recovery period of several months. Lacking detailed medicalrecords, I cannot separately apply these designations. I use “relapse” to describe both cases in mysample.

28I use the coarser monthly measure, rather than the observed number of days before a reevaluationor exit, since the daily variation may simply reflect the difficult of scheduling a consultation with aphysician.

23

-2 -1.5 -1 -0.5 0 0.5 1 1.5 20

0.5

1

1.5

2

2.5x 10

4

βi

No

of In

divi

dual

s

Figure 2: Distribution of Mixed Logit Random CoefficientsFor Insurer Cost and Patient Cost per Day Supply

βi ins.cost

βi copay

1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 60.2

0.25

0.3

0.35

0.4

0.45

Figure 3: Predicted Exit Hazards from Weibull Duration ModelConditional on Insurance Plan Type (95% CI Shown)

For Individuals Beginning on Fluoxetine

Number of 30-day Prescriptions Filled

Haz

ard

Cap.HMOPPO

Figure 3: Predicted exit hazards of exit from Weibull duration model. Conditionalon insurance plan type (95% confidence interval shown), for individuals beginning onFluoxetine

months into treatment. The American Psychiatric Association’s guideline for depression

treatment recommends patients diagnosed with depression follow a treatment course

for at least 6 months (Karasu et al. (2000)). The data make clear that the existing

incentive structure fails to encourage the recommended treatment regimen, particularly

for individuals diagnosed with major depression.

The rates of adherence found match results in Pomerantz et al. (2004), Melfi et al.

(1998) and Akincigil et al. (2007). Poor adherence is a problem for health if patients

continue to suffer from depression symptoms or relapse. I test for the effect of incentives

on relapse using a logit model. I report the results in Table 11.

I define a relapse as a new office visit or prescription filled after a gap of 3 months

from the last observed office visit or from the exhaustion of the patient’s final pre-

scription, whichever is later. For example, if a patient exhausted the supply of his last

prescription on January 1, 2004, I examine the next six months of data. If the patient

did not receive care for depression in January, February, and March, but did appear for

depression care sometime after April 1, I consider the office visit a relapse. To simplify

the problem of censoring, I look only for new depression care occurring between 3 and

6 months after the initial active treatment period ended. Overall, 33,657 unique indi-

viduals had episodes with a 3 month gap after the last treatment, followed by at least 3

24

more months of data to allow me to identify whether the patient relapsed during that

period. In this subsample, 5.4% of patients suffered a relapse.

Karasu et al. (2000) find that 50% of patients relapse over two years. This is

equivalent to a constant rate of relapse of 2.9% per month over 24 months. Over three

months, the literature would predict that 8.3% of patients relapse. The rate in my

sample is slightly lower because I do not measure relapses that occur when a patient

seeks inpatient care.

6.2 Patient vs. physician-directed incentives

I use the logit estimates to predict changes in relapse as a function of incentives. I

report these examinations in panel 2 of Table 11. I hold constant the length of the

initial treatment episode—which can proxy for illness severity and/or the quality of the

initial depression care—as well as a rich set of patient background characteristics. Both

higher copayments and the introduction of capitation lead to statistically significant

increases in the rate of relapse in outpatient and prescription care. Capitation has

an effect 50% stronger than demand-side incentives. The relapse rate increases from

to 5.4% to 6.0% with all patients on capitated plans; doubling copayments leads to a

0.4% increase in relapse. Over two years, assuming a constant rate of relapse, 35.9% of

patients will relapse in the baseline case without incentives. Given the logit estimates,

38.0% of patients facing increased copayments and over 39% facing capitation incentives

will likely suffer a relapse.

7 Optimal insurance design

The results from both the static and dynamic analyses provide insight into the policy-

maker’s incentive design problem. In the product choice model, I predict how alterna-

tive pricing policies would affect the initial choice of the physician. I calculate now, in

dollars, how these alternative policies impact the insurer’s profit. Looking only at pre-

scription drug costs, I sum the individual-specific insurer costs by patient, subtracting

off the copayments these patients pay for their treatment.29 With existing incentives,

insurers’ costs for the first 30 days of drug treatment for the roughly 100,000 individ-

uals in the sample observed equals $3.0 million. Increasing the branded copayment up

29The negotiated payment between the insurer and pharmacy typically accounts for the copaymentrevenue collected by the pharmacy. See Levy (1999).

25

to the 95th percentile reduces the insurers’ costs 24% to $2.3 million, as insurers take

in more revenue from copayments and physicians prescribe cheaper generics.30 More

modest increases in branded copayments lead to smaller cost savings for the insurer. A

policy that decreases generic copayments to $0 actually raises costs by about $79,000

or 3%. The composition of use shifts toward generics, but 5.9% more patients initiate

treatment. Finally, paying all physicians by capitation decreases costs by 19.6%.

Capitation appears to have a stronger effect on short-run drug costs than changes

to the relative copayments. However, these savings neglect the costs from relapse and

from the small increase in referrals to specialists’ care for psychotherapy. The dynamic

analysis suggests that stronger incentives lead to greater need for follow-up care. With

lower initiation of drug treatment or a change in the quality of the patient-drug match,

patients relapse more often. The increase in relapse is 50% greater under supply-side

incentives. This finding suggests the optimal insurance design may vary by disease. For

conditions like depression that may become chronic, the insurer can minimize long-run

costs by employing weaker incentives and focusing on demand-side tools rather than

capitation. For acute conditions with little risk of relapse, using capitation, possibly in

combination with demand-side incentives, may minimize costs.

8 Conclusion

The interrelationships between patients, physicians, and insurers grew out of two char-

acteristics of the market for prescription drugs: patients lack the knowledge to select

their own treatment and also face uncertainty in their future health status. The theoret-

ical model predicts that when physicians possess more information than their patients,

the physician’s treatment selection may diverge from the patient’s preference. Ab-

sent first best contracts, insurers operating in this market can encourage physicians to

prescribe cost-effective drugs by employing supply and demand-side incentives.

In the data, the application of these incentives involves trade-offs. Conditional on

recommending drug treatment, both capitation and copayment policies encourage the

selection of cheaper alternatives within the choice set, shifting prescribing from popular

branded drugs to a common, effective generic. Higher copayments on the demand-

side, however, can cause price-sensitive patients to quit the recommended treatment

30Patients may also leave an insurer’s plan in response to high cost-sharing. The feedback of in-centives on the insurer’s overall enrollment can also hurt profitability. I lack data on patient planenrollment changes, and so neglect this dimension of the insurer’s maximization in my analysis.

26

course prematurely. Capitation, by placing the risk of follow-ups on the physician,

encourages physicians to concentrate prescribing on drugs with simpler dosing and

moderate effectiveness for a broad population. Under less tailored care, relapse rates

increase.

The relative benefits of using demand and supply side incentives depends on disease

characteristics, including the likelihood of costly relapse. If follow-up costs are low,

strong supply-side policies provide a sharp reduction in short-run costs. If the illness

may become chronic, minimizing long-run costs requires weaker supply-side incentives.

27

References

Akincigil, A., J. R. Bowblis, C. Levin, J. T. Walkup, S. Jan, and S. Crys-

tal (2007): “Adherence to antidepressant treatment among privately insured pa-

tients diagnosed with depression,” Medical Care, 45, 363. [10, 24]

Armour, B. S., M. M. Pitts, R. Maclean, C. Cangialose, M. Kishel,

H. Imai, and J. Etchason (2001): “The effect of explicit financial incentives

on physician behavior,” Archives of Internal Medicine, 161, 1261. [1]

Aronsson, T., M. A. Bergman, and N. Rudholm (2001): “The impact of generic

drug competition on brand name market sharesEvidence from micro data,” Review

of Industrial Organization, 19, 423–433. [1]

Arrow, K. J. (1963): “Uncertainty and the welfare economics of medical care,” The

American Economic Review, 53, 941–973. [9]

Baker, G. P. (1992): “Incentive Contracts and Performance Measurement,” The

Journal of Political Economy, 100, 598–614. [1]

Bloom, J. R., T. Hu, N. Wallace, B. Cuffel, J. W. Hausman, M. L. Sheu,

and R. Scheffler (2002): “Mental health costs and access under alternative cap-

itation systems in Colorado,” Health services research, 37, 315–340. [1]

Charlson, M. E., P. Pompei, K. L. Ales, and C. R. Mackenzie (1987): “A new

method of classifying prognostic comorbidity in longitudinal studies: Development

and validation* 1,” Journal of Chronic Diseases, 40, 373–383. [20]

Davidson, J. R. T. and S. E. Meltzer-Brody (1999): “The underrecognition

and undertreatment of depression: What is the breadth and depth of the problem?”

Journal of Clinical Psychiatry, 60, 4–9. [10]

Dranove, D. (1988): “Demand inducement and the physician/patient relationship,”

Economic Inquiry, 26, 281–288. [2, 9]

Ellison, S. F., I. Cockburn, Z. Griliches, and J. Hausman (1997): “Char-

acteristics of demand for pharmaceutical products: An examination of four

cephalosporins,” The Rand journal of economics, 28, 426. [4]

28

Filippini, M., G. Masiero, and K. Moschetti (2006): “Socioeconomic determi-

nants of regional differences in outpatient antibiotic consumption: Evidence from

Switzerland,” Health Policy, 78, 77–92. [14]

Foo, P. K. and M. R. Cullen (2012): “The Effect of Prescription Drug Cost-

Sharing on Generic Substitution,” Stanford University Working Paper. [1]

Gartlehner, G., R. A. Hansen, P. Thieda, A. M. DeVeaugh-Geiss, B. N.

Gaynes, E. E. Krebs, L. J. Lux, L. C. Morgan, J. A. Shumate, and L. G.

Monroe (2007): Comparative effectiveness of second-generation antidepressants in

the pharmacologic treatment of adult depression, Agency for Healthcare Research and

Quality (US). [7, 11]

Gaynor, M., J. Li, and W. B. Vogt (2007): “Substitution, spending offsets, and

prescription drug benefit design,” in Forum for Health Economics & Policy, vol. 10,

4. [1]

Gemmill, M. C., J. Costa-Font, and A. McGuire (2007): “In search of a cor-

rected prescription drug elasticity estimate: a meta-regression approach,” Health

Economics, 16, 627–643. [18]

Goldman, D. P., G. F. Joyce, J. J. Escarce, J. E. Pace, M. D. Solomon,

M. Laouri, P. B. Landsman, and S. M. Teutsch (2004): “Pharmacy Benefits

and the Use of Drugs by the Chronically Ill,” JAMA: The Journal of the American

Medical Association, 291, 2344–2350. [1, 4, 19]

Hellerstein, J. K. (1998): “The Importance of the Physician in the Generic versus

Trade-Name Prescription Decision,” The RAND Journal of Economics, 29, 108–136.

[1, 14]

Ho, K. and A. Pakes (2011): “Do physician incentives affect hospital choice? A

progress report,” International Journal of Industrial Organization, 29, 317–322. [1]

Huskamp, H. A., R. G. Frank, K. A. McGuigan, and Y. Zhang (2005): “The

impact of a three-tier formulary on demand response for prescription drugs,” Journal

of Economics and Management Strategy, 14, 729. [1, 4, 19]

Jones, M. I., S. M. Greenfield, and C. P. Bradley (2001): “Prescribing

new drugs: qualitative study of influences on consultants and general practitioners,”

British Medical Journal, 323, 378. [14]

29

Karasu, T. B., A. Gelenberg, A. Merriam, and P. Wang (2000): Practice

Guideline for the Treatment of Patients With Major Depressive Disorder: Second

Edition, Arlington, VA: American Psychiatric Association. [2, 11, 24, 25]

Lazear, E. P. (2000): “Performance Pay and Productivity,” The American Economic

Review, 90, 1346–1361. [1]

Levy, R. (1999): The pharmaceutical industry: a discussion of competitive and

antitrust issues in an environment of change, Washington, DC: Bureau of Eco-

nomics/Federal Trade Commission staff report. [25]

McGuire, T. G. (2000): “Physician agency,” Handbook of Health Economics, 1, 461–

536. [9]

Melfi, C. A., A. J. Chawla, T. W. Croghan, M. P. Hanna, S. Kennedy, and

K. Sredl (1998): “The effects of adherence to antidepressant treatment guidelines

on relapse and recurrence of depression,” Archives of General Psychiatry, 55, 1128.

[10, 24]

Murphy, M. J., R. L. Cowan, and L. I. Sederer (2009): Blueprints: Psychiatry,

Philadelphia, PA: Lippincott Williams Wilkins, 5 ed. [11, 12]

Nair, K. V., P. Wolfe, R. J. Valuck, M. M. McCollum, J. M. Ganther, and

S. J. Lewis (2003): “Effects of a 3-tier pharmacy benefit design on the prescription

purchasing behavior of individuals with chronic disease,” Journal of Managed Care

Pharmacy, 9, 123–133. [1]

Pauly, M. V. (1980): “Doctors and their Workshops: Economic Models of Physician

Behavior,” National Bureau of Economic Research Monograph. [2, 9]

Pomerantz, J. M., S. N. Finkelstein, E. R. Berndt, A. W. Poret, L. E.

Walker, R. C. Alber, V. Kadiyam, M. Das, D. T. Boss, and T. H. Ebert

(2004): “Prescriber intent, off-label usage, and early discontinuation of antidepres-

sants: a retrospective physician survey and data analysis,” Journal of Clinical Psy-

chiatry, 65, 395–404. [10, 24]

Shrank, W. H., M. Stedman, S. L. Ettner, D. DeLapp, J. Dirstine, M. A.

Brookhart, M. A. Fischer, J. Avorn, and S. M. Asch (2007): “Patient,

30

physician, pharmacy, and pharmacy benefit design factors related to generic medica-

tion use,” Journal of general internal medicine, 22, 1298–1304. [1]

Train, K. (2003): Discrete Choice Methods with Simulation, Cambridge Univ Press.

[17]

31

Tab

le1:

Sum

mar

ySta

tist

ics

onP

roduct

and

Pat

ient

Char

acte

rist

ics

!""#

!""$

!""%

!""#

!""$

!""%

!""#

!""$

!""%

&'(

)&'(

)&'(

)*

*++++++++

*++++++++

*++++++++

*++++++++

*++++++++

*++++++++

!,-.

+++++++

!/-,

+++++++

!.-"

+++++++

01234'

54)678

201234'

54)678

2&'(

)*

*++++++++

*++++++++

*++++++++

*++++++++

*++++++++

*++++++++

"-"

+++++++++

"-!

+++++++++

#-9

+++++++++

:;57

<'867=

+>?6

:)<)@7

AABC

D)1

9!-/

+++++++

9!-$

+++++++

./-%

+++++++

!#-E

+++++++

#!-/

+++++++

#%-!

+++++++

$-.

+++++++++

#-"

+++++++++

"-E

+++++++++

:;57

<'867=

+>?6

:;57

<'867=

+>?6

AABC

&'

*++++++++

#/-#

+++++++

!!-,

+++++++

*++++++++

/-%

+++++++++

/-/

+++++++++

*++++++++

"-!

+++++++++

$-$

+++++++++

F13;57<'867=

+G@7<75)

H)@786'

AABC

D)1

.#-#

+++++++

.!-"

+++++++

."-E

+++++++

!#-#

+++++++

!,-.

+++++++

!/-"

+++++++

E%-E

+++++++

E!-/

+++++++

E!-/

+++++++

I<J'

@)5;(

)+>:<

I<J'

@)5;(

)+>:<

AABC

&'

#"-.

+++++++

!E-.

+++++++

!"-/

+++++++

E"-#

+++++++

9-9

+++++++++

/-!

+++++++++

9-9

+++++++++

E!-"

+++++++

EE-"

+++++++

I<J'

@)5;(

)+>:<

06'K73

AABC

D)1

E#%-%

+++++

E!$-,

+++++

E.$-#

+++++

#%-#

+++++++

#"-#

+++++++

#.-/

+++++++

"-$

+++++++++

"-9

+++++++++

"-!

+++++++++

076'@)5;(

)+>:<

076'@)5;(

)+>:<

AABC

&'

.,-E

+++++++

%E-,

+++++++

#9-9

+++++++

EE-,

+++++++

,-"

+++++++++

,-!

+++++++++

E-/

+++++++++

$-%

+++++++++

$-/

+++++++++

076'@)5;(

)+>:<

07@;<+:

BAABC

D)1

/E-.

+++++++

/$-E

+++++++

9.-"

+++++++

!!-"

+++++++

!"-#

+++++++

!$-E

+++++++

$-,

+++++++++

#-$

+++++++++

E-%

+++++++++

A)6567<;(

)+>:<

L'<'M5

AABC

D)1

/"-"

+++++++

9,-!

+++++++

9.-/

+++++++

!E-$

+++++++

!/-.

+++++++

!9-,

+++++++

E#-E

+++++++

E"-$

+++++++

E"-!

+++++++

NJ<'@)5;()+>:<

:2=

O7<57

A&BC

D)1

*++++++++

E""-%

+++++

,9-,

+++++++

*++++++++

##-/

+++++++

##-"

+++++++

*++++++++

"-%

+++++++++

!-.

+++++++++

P)(<7M7@;()+>:<

FMM)@'6

A&BC

D)1

9$-9

+++++++

9E-E

+++++++

9#-E

+++++++

!E-"

+++++++

!%-#

+++++++

!.-!

+++++++

"-%

+++++++++

"-.

+++++++++

"-%

+++++++++

P)(<7M7@;()+>:<

FMM)@'6*QB

A&BC

D)1

E",-#

+++++

EE$-!

+++++

E".-%

+++++

!E-,

+++++++

!E-"

+++++++

!%-"

+++++++

,-E

+++++++++

/-"

+++++++++

9-$

+++++++++

?J86'8

;'(+>:<

?J86'8

;'(+>:<

&NBC

&'

#9-9

+++++++

%%-$

+++++++

%,-!

+++++++

E"-9

+++++++

,-#

+++++++++

E"-#

+++++++

"-#

+++++++++

#-.

+++++++++

$-9

+++++++++

?J86'8

;'(+>:<

R)<<OJ56;(+QH

&NBC

D)1

E"$-%

+++++

E"#-%

+++++

,,-E

+++++++

!!-9

+++++++

!E-#

+++++++

!.-!

+++++++

E-.

+++++++++

$-.

+++++++++

%-9

+++++++++

S=;56;852<;(

)+>:<

S=;56;852<;(

)+>:<

T:S

&'

%-.

+++++++++

$-!

+++++++++

$-"

+++++++++

9-"

+++++++++

%-%

+++++++++

%-%

+++++++++

"-,

+++++++++

"-,

+++++++++

"-/

+++++++++

U;657K78;()

U;657K78;()

&7AAS

&'

.E-E

+++++++

$E-/

+++++++

#!-"

+++++++

EE-.

+++++++

/-!

+++++++++

/-.

+++++++++

"-9

+++++++++

"-/

+++++++++

"-9

+++++++++

T67K'V

'()+>:<

T67K'V

'()+>:<

AU&'

9-,

+++++++++

.-%

+++++++++

.-.

+++++++++

9-9

+++++++++

.-"

+++++++++

.-#

+++++++++

E-/

+++++++++

!-#

+++++++++

!-"

+++++++++

E!W#.,

+++$$WE$E

+++$EW."!

+++

07()<+!X+T7O<)+'M+;(V

;Y;VJ7

<*<)Y)<+3'Y

76;75)1+;(+86'VJ

35+34'

;3)+='V

)<

%Z!%Z

%"Z

9%Z

,%Z

E"-#

EE-#

"-"

!%!-$

"-"

"-"

E"-"

E%-"

#"-"

$!-E

E!-$

E/-"

.$-"

!"-"

##-"

$#-"

%!-"

.E-"

$!-.

EE-"

E/-9

/.-$

!/-%

#%-E

$E-!

$.-.

.#-,

!9-%Z

9E-EZ

$$-"Z

!/-EZ

,/WEE!

++++++

T'57<+'

O1)6Y75;'

(1+O2+2)76X

P76;7O<)

GMM;3)+P;1;5+:'8

7207

5;)(5+S[)

SY[

-+:'J

(52+C(3'=)+\"""1]

E^N;7[(

'1;1W+U

7_'6+N)8

6)11;'(`

T'57<+a+'M+G

O1)6Y75;'

(1X

U7@+

07()<+EX+0

6;3)1+M'6+'

()+=

'(54b1+1J88

<2+7(V

+147

6)1W+O2+86'V

J35

E^I)=7<)`

E^c)()67<+06735;5;'(

)6+`

E^01234;756;15`

C(1J6)6+:

'15+\de='(

54+

1J88

<2]

:'8

72=)(5+\de='(

54+

1J88

<2]

U76f)5+A

476)

gJ7

(5;<)1

U)7(

A5V+N)Y

U;(

C([6)V

;)(5+&

7=)

06'V

J35+&

7=)

AJO*:<711

?67(

Vh

&'5)1Xi

E-+A'J

63)X+T4'

=1'(+B)

J5)61b+U76f)5A37(+GJ587

5;)(5+7(V

+N6J[+:<7;=

1+N757O71)1W+2)761+!""#*!""%-++A7=8<)+1JO1)5+5'

+;(V;Y;VJ

7<1+V;7[('

1)V+j;54

+'()+

'M+M'

J6+3<711)1+'M+V

)86)11;'(X+=

7_'6+V)8

6)11;'(W+V21542=

;7W+V

)86)11;Y)+6)735;'

(1W+7(V

+V)8

6)11;'(+('

5+'54)6j;1)+18)3;k)V

-++T4

)+17=8<)+)@3<JV

)1+

86)[(7

(5+j

'=)(W+;(V

;Y;VJ7

<1+j

;54+3'(

3J66)(5+V

;7[(

'1)1+'M+O

;8'<76+V;1'6V)

6+'6+134;K'8

46)(;7-i

!-+U

)7(+3'87

2=)(51+6)

l)35+54

)+87

5;)(5b1+'J

5*'M*8

'3f)5+3'15+M'6+7+#"*V7

2+1J88

<2+'M+=

)V;375;'(-++T

4)+=

)7(+;(1J6)6+3

'15+;1+54)+6)8'

65)V

+3'15+

6)18'(

1;O;<;52+'M+54

)+87

5;)(5b1+;(1J6)6+M'6+7+#"*V7

2+1J88

<2-+T

4;1+86;3)+)@3<JV

)1+6)

O75)1+3'=='(

<2+'m)

6)V+5'+;(

1J6)61+O2+V6J[

+=7(

JM735J6)61-+i

#-+T4)+V6J[+1JO3<711)1+6)

8'65)V

+V;m)6+733'6V;([

+5'+54

)+;([6)V

;)(5b1+)m

)35+'

(+54)+3'(3)(5675;'(+'M+()J

6'567(

1=;n)6+34)=;37<1+;(+54

)+O67;(-++AAB

C1+

1)<)35;Y)<2+;(4;O;5+54)+6)

J857f)+'M+1)6'5'(

;(o+A&BC1+;(4;O;5+54)+6)

J857f)+'M+1)6'5'(

;(+7(V

+('6)8

;()8

46;()o+&

NBC1+7m

)35+(

'6)8

;()8

46;()+7(

V+V'

87=;()o+

T:S1W+56;323<;3+7(5;V)8

6)117(

51W+J

(1)<)35;Y

)<2+O<'3f+54)+6)J8

57f)+'M+1)6'5'(

;(o+AU1+='V

J<75)+1)6'5'(;(o+7(V

+&7AAS

1+735+O

2+7(

57['

(;K;([

+Y76;'J1+

7V6)()6[;3+7(V

+1)6'5'(

;(+6)

3)85'61-i

32

Tab

le2:

Dis

trib

uti

onof

Tre

atm

ents

Pre

scri

bed

,C

ondit

ional

onC

ovar

iate

s

!"#$

%&'

(")*

$+,)-

./0%

1&2)33

45.")22

6)7

8')'.$+

9:4

,#*

;6)7

8')'.$+

9:4

!)'8.

*'+

&#5.".$+1<

+!!

4

=".)'.$+1<

+>.*.")2+

!")&'8'8#*."

=".)'.$+1<

+!3<&?8)'"83'

@8)A*

#383+#B+

:)C#"+

@.7

".338#*

,#*

.;

,#*

.DEFEG

++++++++++++

DEFGE

++++++++++++

HIFDJ

++++++++++++

DEFHK

++++++++++++

DLFML

++++++++++++

DEFIN

++++++++++++

DGFMM

++++++++++++

!3<&?#

'?.")7

<+#*

2<;

,#*

.NFLE

++++++++++++++

DFHJ

++++++++++++++

IFHH

++++++++++++++

NFMH

++++++++++++++

IFNJ

++++++++++++++

MFIE

++++++++++++++

DFEJ

++++++++++++++

O.P)7"#

Q.3

00RS

NHFNN

++++++++++++

KFGE

++++++++++++++

NMFEH

++++++++++++

NMFLH

++++++++++++

NLFDL

++++++++++++

JFDM

++++++++++++++

NDFND

++++++++++++

T2%#

P.'8*

.+96O

,#

00RS

NNFIM

++++++++++++

NEFMM

++++++++++++

EFHM

++++++++++++++

EFGE

++++++++++++++

KFED

++++++++++++++

NDFML

++++++++++++

NNFGH

++++++++++++

U#2#B'

Q.3

00RS

NIFEI

++++++++++++

KFNM

++++++++++++++

NDFIM

++++++++++++

NNFJL

++++++++++++

NDFMI

++++++++++++

KFLE

++++++++++++++

NIFII

++++++++++++

VBB.P#";WR

Q.3

0,RS

EFKK

++++++++++++++

LFNJ

++++++++++++++

KFML

++++++++++++++

KFJN

++++++++++++++

JFND

++++++++++++++

LFKM

++++++++++++++

KFHG

++++++++++++++

X.221%'"8*+WO

Q.3

,@RS

GFLJ

++++++++++++++

HFMN

++++++++++++++

HFDE

++++++++++++++

LFIK

++++++++++++++

MFNG

++++++++++++++

GFLN

++++++++++++++

MFLN

++++++++++++++

!)"#P.'8*

.+96O

,#

00RS

GFHI

++++++++++++++

MFKH

++++++++++++++

HFNL

++++++++++++++

HFLI

++++++++++++++

GFNL

++++++++++++++

GFNE

++++++++++++++

GFII

++++++++++++++

(%7"#7

8#*+96O

,#

,@RS

HFLK

++++++++++++++

MFMK

++++++++++++++

HFHI

++++++++++++++

DFML

++++++++++++++

HFIH

++++++++++++++

GFJI

++++++++++++++

GFIE

++++++++++++++

68')

2#7")-

+9("

,#

00RS

NFJK

++++++++++++++

DFEG

++++++++++++++

NFDD

++++++++++++++

NFEJ

++++++++++++++

NFJH

++++++++++++++

DFDN

++++++++++++++

DFIE

++++++++++++++

6.2.P)

Q.3

00RS

NFJL

++++++++++++++

NFKH

++++++++++++++

DFEJ

++++++++++++++

NFLN

++++++++++++++

NFLJ

++++++++++++++

DFHL

++++++++++++++

DFNL

++++++++++++++

JKYNHJ

++++++++++

HGYELI

++++++++++

NHYDDE

++++++++++

GDYLHE

++++++++++

GHYNKL

++++++++++

DEYMLJ

++++++++++

DLYJGE

++++++++++

=#')2+Z+#B+S*$

858$%)

23+8*

+0%1

7#7%

2)'8#

*+,#'.3[\

NF+0#%

"&.[+=?#

-3#*+R.

%'."3]+:)"^.'0&)*+4%'7)

'8.*'+)*$

+@"%A+62)8-

3+@)')1)3.3Y+<.)"3+DIIH;DIIMF+0.2.&'.$+7"#$

%&'3+3?#

_*F\

DF+!)'8.*'3+8*+.8'?

."+&)'.A#"<+#B+9

.)2'?

+:)8*'.*)*

&.+4"A)*

8`)'8#*+a9

:4b+-

%3'+&?#

#3.+B"#-

+)+7)"'8&%2)"+283'+#B+7

"#58$.

"3+B#

"+)22+*#

*;.-

."A.*&<+&)".F+V

)&?+7)

'8.*'+&?#

#3.3+)+7"8-)"<+&)".+7?<

38&8)*

+'#+-

)*)A.+?83+&)".Y+_

8'?+37.

&8)283'+&)".+1<

+".B."")2+#*2<F++6

)78')

'.$+

72)*

3Y+%*28^.+'?.+*#

*;&)78')'.$+5."38#*3Y+7

)<+B#

"+A.*.")2+7")&'8'8#*

."+3."58&.3+7"#37.

&'85.2<+1<

+.*"#22..F++!

".B."".$

+!"#58$.

"+4"A)*

8`)'8#*3+a!

!43b+$#+*#

'+".c%8".+)+7"8-)"<+&)".+7?<

38&8)*

+#"+".B."")23Y+)*$

+".8-

1%"3.+7?

<38&8)*3+".

'"#37.

&'85.2<F\

HF+:

)C#"+$.7

".338#*+aS6

@;J+&#$

.3+DJLFDY+DJLFHb+83+#*

.+#B+B#

%"+$.7

".338#*+&)'.A#

"8.3+&#-

7#38*A

+'?.+)*

)2<383+$)')3.'F\

GF+=?.+#7'8#*+],#*

.]+".

7#"'3+'?.+3?)".+#B+8*

$858$%

)23+$8)A*#

3.$+_8'?

+$.7

".338#*+8*+)*+#%

'7)'8.*'+#d&.+5838'+_?#

+*.5."+e22+)+

7".3&"87'8#

*+B#"+)

*+)*

'8$.7

".33)*

'F+++=

?83+&)*+#&&%"+_

?.*+'?.+7?

<38&8)*+$#

.3+*#'+7".3&"81.+$"%A+'".)'-

.*'+#

"+_?.*+'?.+7)

'8.*'+B)

823+'#

+e22+)+7".3&"87'8#*F++SB+'?.+7)'8.*'+?)$

+*#+#1

3."5.$

+$"%A+&2)8-3+1%

'+$8$+".

&.85.+73<&?#

'?.")7

<+'".)'-

.*'Y+?83+.P7.

"8.*&.+_

#%2$+

&#*'"81%

'.+'#

+'?.+#7

'8#*+2)1.2.$+f!3<&?#'?.")7<

+#*2<fF++S*$858$%

)23+".&.858*A+1#

'?+73<&?#'?.")7<

+)*$

+)+$"%A+'".)'-

.*'+_

#%2$+)77

.)"+

8*+'?

.+3?)".+#B+'?

.+37.&8e&+$"%A

+'".)'-

.*'+$

837.

*3.$

F\

33

Tab

le3:

Mult

inom

ial

Log

itE

xam

inat

ion

ofD

rug

Cat

egor

yC

hoi

ce

Pred

icted(Prob

abilitie

s((in

(%)

Bran

ded(drug

(on

lyGeneric(drug(

only

Psycho

therap

yNo(recorded

(treatm

ent

Prob

.(HMO,(cap

itated

37.67

((((((((((((((((

43.05

((((((((((((((((

1.06

((((((((((((((((((

18.22

((((((((((((((((

Prob

.(HMO,(n

onIcap

itated((POS)

49.28

((((((((((((((((

29.48

((((((((((((((((

0.22

((((((((((((((((((

21.02

((((((((((((((((

Prob

.(PPO

/Com

prehensive

50.57

((((((((((((((((

29.27

((((((((((((((((

1.41

((((((((((((((((((

18.75

((((((((((((((((

HMO,(cap

.(I((H

MO,(n

onIcap

.(11.62)

(((((((((((((((

13.58

((((((((((((((((

0.85

((((((((((((((((((

(2.80)

(((((((((((((((((

Std.(error(of(cha

nge

0.71

((((((((((((((((((

0.72

((((((((((((((((((

0.06

((((((((((((((((((

0.51

((((((((((((((((((

HMO,(cap

.(I(PPO

/Com

p.(12.91)

(((((((((((((((

13.79

((((((((((((((((

(0.35)

(((((((((((((((((

(0.53)

(((((((((((((((((

Std.(error(of(cha

nge

0.56

((((((((((((((((((

0.58

((((((((((((((((((

0.05

((((((((((((((((((

0.38

((((((((((((((((((

Prob

.(w/Cop

ay(In

cent,(N

o(Coin(Incent

50.57

((((((((((((((((

29.27

((((((((((((((((

1.41

((((((((((((((((((

18.75

((((((((((((((((

Prob

.(w/N

o(Cop

ay(In

cent,(N

o(Coin(Incent

57.10

((((((((((((((((

25.24

((((((((((((((((

0.60

((((((((((((((((((

17.06

((((((((((((((((

Cop

ay(In

centI(N

o(Cop

ay(In

cent,(N

o(Coin

(6.52)

(((((((((((((((((

4.03

((((((((((((((((((

0.81

((((((((((((((((((

1.69

((((((((((((((((((

Std.(error(of(cha

nge

0.75

((((((((((((((((((

0.72

((((((((((((((((((

0.07

((((((((((((((((((

0.49

((((((((((((((((((

Treatm

ent(T

ype

Panel&1:&Plan&Type/Physician&Payment&Regime

Panel&2:&Cost&Sharing&Regimes

Notes:T

1.(Sou

rce:(Tho

mson(Re

utersV(MarketScan(Outpa

tient(and

(Drug(Claim

s(Database,(years(2003I2005.(T

2.(Exercise(sets(re

maining

(backg

roun

d(variables(to(th

e(follo

wing(values:(Spe

cialty=G

eneral(Practition

er;(

Diagn

osis=m

ajor(dep

ression;(Gende

r=Female;(Region=

West;(Age(Quintile=40thI60th;(In(MSA

?=Ye

s;(

Restrictive(Ann

ual(C

ap(on(Psych.(Visits=N

o;(Psychiatric(CarveIO

ut=N

o;(M

anda

tory(Gen.(Sub

s(State?=N

o;(

Office(Visit(Cop

ayment(=

($10;(C

ounty(Average(In

come(=(Med

ian.T

34

Table 4: Mixed Logit Estimates

Variable Estimate Std.Error Estimate Std.Error Estimate Std.Errorlog(insurer.cost) 60.0431 0.0024 0.0562 0.0104 60.0919 0.0034log(patient.drug.copay) 60.2646 0.0053 60.4375 0.0197 60.2111 0.0103variance,.log(insurer.cost) 0.0087 0.0006 0.0907 0.0047 0.0206 0.0014variance,.log(patient.drug.copay) 0.0257 0.0030 0.1512 0.0100 0.2308 0.0128Age*log(copay) 0.0021 0.0006Office.visit.copay*log(copay) 0.0036 0.00081{MDD}.*log(copay) 0.2196 0.01741{Psychiatrist}.*log(copay) 60.2098 0.01291{Cap..HMO}*log(insurer.cost) 60.0630 0.00671{Non6Cap..HMO}*log(insurer.cost) 60.0161 0.00631{high.charlson.index}*ln(copay) 60.0798 0.0137 60.0759 0.01411{high.charlson.index}*ln(insurer.cost) 0.0042 0.0074 0.0205 0.0055Drug.effects.included?Interactions.of.drug.effects.with.capitation?

Random.Coefficient Mean Std.Dev.log(insurer.cost) 60.0432 0.0932log(patient.drug.copay) 60.2649 0.1594

(3)

No No Yes

(1) (2)

Yes Yes Yes

67.8%

%.of.draws.with.coefficient.below.zero

95.1%

Distribution.of.coefficients.in.the.population:Calculated.using.100,000.draws.from.a.normal.distribution.with.the.estimated.parameters.from.specification.(1)

Notes:\1..Source:.Thomson.Reuters^.MarketScan.Outpatient.and.Drug.Claims.Databases,.years.200362005.\2..Mixed.logit.estimated.via.Markov.Chain.Monte.Carlo.methods...The.estimates.come.from.a.sequence.of.20,000.draws.collected.after.a.burn6in.sequence.of.75,000.draws.to.ensure.convergence.to.the.relevant.posterior.distributions.\3..Log.price.variables.have.random.coefficients...I.report.above.both.the.mean.estimates.and.the.variance.estimates...The.variances.come.from.the.main.diagonal.of.the.covariance.matrix.calculated.using.draws.from.the.posterior.distribution.\

35

Tab

le5:

Mix

edL

ogit

-P

redic

ted

Dru

gC

hoi

ceP

robab

ilit

ies

for

Sel

ecte

dA

nti

dep

ress

ants

Estim

ate

Std.+Error

Estim

ate

Std.+Error

Estim

ate

Std.+Error

Estim

ate

Std.+Error

Estim

ate

Std.+Error

No+drug

+treatm

entNon

eN

24.33

++++++++++

0.16

++++++++++++

26.61

++++++++++

0.19

++++++++++++

18.47

++++++++++

0.20

++++++++++++

24.53

++++++++++

0.16

++++++++++++

28.11

++++++++++

0.21

++++++++++++

Celexa

SSRI

Y2.13

++++++++++++

0.06

++++++++++++

2.01

++++++++++++

0.07

++++++++++++

1.71

++++++++++++

0.13

++++++++++++

2.12

++++++++++++

0.06

++++++++++++

1.65

++++++++++++

0.10

++++++++++++

Cita

lopram

+HBr

SSRI

N1.93

++++++++++++

0.05

++++++++++++

2.15

++++++++++++

0.04

++++++++++++

3.56

++++++++++++

0.04

++++++++++++

1.95

++++++++++++

0.05

++++++++++++

2.67

++++++++++++

0.05

++++++++++++

Lexapro

SSRI

Y16.36

++++++++++

0.06

++++++++++++

15.16

++++++++++

0.06

++++++++++++

13.08

++++++++++

0.05

++++++++++++

16.24

++++++++++

0.06

++++++++++++

9.65

++++++++++++

0.04

++++++++++++

Fluo

xetin

e+HCL

SSRI

N8.53

++++++++++++

0.14

++++++++++++

9.56

++++++++++++

0.15

++++++++++++

15.52

++++++++++

0.17

++++++++++++

8.60

++++++++++++

0.14

++++++++++++

18.29

++++++++++

0.18

++++++++++++

Paroxetin

e+HCL

SSRI

N4.02

++++++++++++

0.02

++++++++++++

4.52

++++++++++++

0.02

++++++++++++

7.52

++++++++++++

0.04

++++++++++++

4.05

++++++++++++

0.02

++++++++++++

5.89

++++++++++++

0.03

++++++++++++

Paxil+C

RSSRI

Y3.63

++++++++++++

0.02

++++++++++++

3.22

++++++++++++

0.01

++++++++++++

2.89

++++++++++++

0.01

++++++++++++

3.61

++++++++++++

0.02

++++++++++++

2.22

++++++++++++

0.03

++++++++++++

Zoloft

SSRI

Y12.97

++++++++++

0.08

++++++++++++

11.97

++++++++++

0.07

++++++++++++

10.35

++++++++++

0.06

++++++++++++

12.88

++++++++++

0.08

++++++++++++

9.29

++++++++++++

0.07

++++++++++++

Cym

balta

SNRI

Y1.85

++++++++++++

0.05

++++++++++++

1.77

++++++++++++

0.05

++++++++++++

1.47

++++++++++++

0.04

++++++++++++

1.84

++++++++++++

0.05

++++++++++++

0.76

++++++++++++

0.07

++++++++++++

EffexorLXR

SNRI

Y9.71

++++++++++++

0.04

++++++++++++

8.61

++++++++++++

0.04

++++++++++++

7.70

++++++++++++

0.03

++++++++++++

9.64

++++++++++++

0.04

++++++++++++

6.61

++++++++++++

0.04

++++++++++++

Buprop

ion+HCL

NDRI

N2.85

++++++++++++

0.03

++++++++++++

3.21

++++++++++++

0.03

++++++++++++

5.38

++++++++++++

0.04

++++++++++++

2.87

++++++++++++

0.03

++++++++++++

5.12

++++++++++++

0.05

++++++++++++

Wellbutrin+XL

NDRI

Y5.34

++++++++++++

0.06

++++++++++++

4.76

++++++++++++

0.07

++++++++++++

4.23

++++++++++++

0.15

++++++++++++

5.30

++++++++++++

0.07

++++++++++++

3.61

++++++++++++

0.11

++++++++++++

Amitriptylin

e+HCL

TCA

N0.83

++++++++++++

0.16

++++++++++++

0.93

++++++++++++

0.19

++++++++++++

1.38

++++++++++++

0.20

++++++++++++

0.84

++++++++++++

0.16

++++++++++++

0.63

++++++++++++

0.21

++++++++++++

Trazod

one+HCL

SMN

1.33

++++++++++++

0.13

++++++++++++

1.49

++++++++++++

0.13

++++++++++++

2.28

++++++++++++

0.14

++++++++++++

1.34

++++++++++++

0.13

++++++++++++

1.87

++++++++++++

0.18

++++++++++++

20.59

++++++++++

23.09

++++++++++

37.63

++++++++++

20.76

++++++++++

35.47

++++++++++

55.08

++++++++++

50.31

++++++++++

43.90

++++++++++

54.71

++++++++++

36.42

++++++++++

Overall+Generic+Sha

reOverall+Bran

ded+Sh

are

(4)

Origina

l+sam

ple

Cop

ay+of+b

rand

ed+

drug

s+=+95th+pctile+($

50+

for+3

0+da

ys)

Cop

ay+of+g

eneric+drugs+

set+to+$0

Insurer+c

osts+fo

r+bran

ded+drug

s+increase+

by+20%

All+pa

tients+enrolled+in+

capitated+HMO+plans

Prod

uct+N

ame

Class

Bran

d?

Pred

icted+Sh

ares

Base

(1)

(2)

(3)

Notes:`

1.+Sou

rce:+Tho

mson+Re

utersa+MarketScan+Outpa

tient+and

+Drug+Claim

s+Databases,+years+2003L2005.`

2.+Predicted

+sha

res+rely+on+coeffi

cients+from

+spe

cification+(3)+rep

orted+in+Table+4.`

3.+Stand

ard+errors+com

puted+using+draw

s+from

+the+po

sterior+d

istribution,+con

ditio

ning

+on+the+ob

served

+first+p

eriod+choice.++Th

e+compu

tatio

n+invo

lves+fu

rther+

simulation+to+accou

nt+fo

r+the+sam

pling+distribu

tion+of+th

e+estim

ated

+coefficients+th

emselves.`

36

Table 6: Estimates from a Weibull Duration Model on Drug Switches

!"#$"%&' ()'** +,-./##)# 012,",

3'$%4&&.25"6'.6"#"7','#.8"&65"9 :;<<: :;::= >?;@

A)B2,"B, 1C;CDE :;:=: E>;@

!"#$%&'()*+,-)'&'+."&*/.*+/CFG"&'HCFIJ'.K4"B,$&'.L.M:;:N:;ED9 :;CDO :;:C? >;@

CFG"&'HCFIJ'.K4"B,$&'.L.M:;EDN:;D:9 :;:E@ :;:C> C;O

CFG"&'HCFIJ'.K4"B,$&'.L.M:;DN:;?D9 1:;:EC :;:C? C;E

CFG"&'HCFIJ'.K4"B,$&'.L.M:;?DNC;:9 :;:OE :;:C? <;O

CFP'7"&'HCFIJ'.K4"B,$&'.L.M:;:N:;ED9 :;CDC :;:C< CC;>

CFP'7"&'HCFIJ'.K4"B,$&'.L.M:;EDN:;D:9 :;:E> :;:C< E;E

CFP'7"&'HCFIJ'.K4"B,$&'.L.M:;?DNC;:9 :;:OC :;:C< =;O

CMQ'J$)B.L.3'2,R :;C<: :;:C: C<;:

A)4B,S.$BA)7'N.$B.C:::2 1:;::< :;::: >;:

CFGTT.-$"JB)2$2R 1:;:=> :;:CC =;=

CFU2SA5$",#$2,R 1:;:<: :;:CE E;D

CFV,5'#.+6'A$"&$2,R 1:;:C: :;:C: C;:

01+"1.*2"/CF("6$,",'-.WGV.6&"BR :;C:= :;:C: @;@

CFX)B1("6$,",'-.WGV.6&"BR :;:DO :;:CE =;O

CFA"#Y'.)4,R :;:=< :;::> D;C

A)6"S.6'#.-"S.2466&S :;:O: :;:C: O;<

!&3%,-)'&'+."&*/.*+/CFT)2$BJ.E1<Z.6'#.-"SR :;<C@ :;:E: CD;O

CF[46#)6$)B.83'&&%4,#$B9R 1:;<?: :;:<@ @;O

CF($,"&)6#"7.8('&'Z"9R 1:;<D< :;:<@ >;@

CFT4&)Z',$B'.8(S7%"&,"9R 1:;<=C :;:=O ?;D

CF/2A$,"&)6#"7.8\'Z"6#)9R 1:;=OD :;:<? CE;=

CFP&4)Z',$B'.8U#)]"A9R 1:;<=< :;:<? @;<

CFG$#,"]"6$B'.8Q'7'#)B9R 1:;CD: :;:D< E;>

CFX)#,#$6,S&$B'.8U"7'&)#9R 1:;E:> :;:O< <;<

CFU"#)Z',$B'.8U"Z$&9R 1:;<?E :;:<> @;@

CF+'#,#"&$B'.8^)&)*,9R 1:;=OC :;:<> CE;<

CF0#"])-)B'.8T'2S#'&9R 1:;:E? :;:=> :;O

CF!'B&"*"Z$B'.8/**'Z)#9R 1:;OC> :;:<> CO;<

X),'2_`C;.+)4#A'_.05)72)B.Q'4,'#2a.G"#b',+A"B.T","%"2'N.S'"#2.E::<1E::D;`E;.T",".42'-.*)#.'2,$7",$)B.$BA&4-'2.)B&S.$B-$Y$-4"&2.6#'2A#$%'-.".-#4J.,#'",7'B,.$B.,5'.c#2,.6'#$)-;..05'.-4#",$)B.)*.,5'.26'&&.'B-2.d5'B.,5'.6",$'B,.,'#7$B",'2.,#'",7'B,.%S.2d$,A5$BJ.,).".B'd.-#4J.,#'",7'B,.)#.,).,5'.)4,2$-'.J))-;`<;./ZA&4-'-.-#4J.A)76)4B-.$B-$A",)#_.CFI7$,#$6,S&$B'.8/&"Y$&9R`

37

Table 7: Predicted hazard rates of switching away from initial treatment choice. Esti-mates from Weibull duration model.

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

M%1N)&-.&7$'"% FT!&'$0";%$,)U!(0(+$VW ?4@B ?4??B ?4C@ ?4??D ?4CD ?4??GM%1N)&-.&7$'"% FT*37&'$0";%$,)UX(+$;%"VW ?4@C ?4??E ?4@A ?4??L ?4CE ?4??LM%1N)&-.&7$'"% FTR01"+('&-()U8%"Y$7VW ?4@L ?4??E ?4C@ ?4??B ?4CB ?4??BM%1N)&-.&7$'"% FT8$%"+('&-()U8$+&0VW ?4@L ?4??L ?4CF ?4??B ?4CB ?4??DM%1N)&-.&7$'"% FT5(%'%$0&-()UZ"0"2'VW ?4@C ?4??E ?4@A ?4??L ?4CE ?4??BM%1N)&-.&7$'"% FTM10"+('&-()U!:,[$0'$VW ?4@D ?4??G ?4CE ?4?F? ?4CG ?4?FFM%1N)&-.&7$'"% FT/(-0$2$+&-()U*22(+"%VW ?4@? ?4??E ?4@B ?4??L ?4@A ?4??LM%1N)&-.&7$'"% FT\1;%";&"-)U](00[1'%&-VW ?4@A ?4??L ?4CB ?4??B ?4EF ?4??DM%1N)&-.&7$'"% FTQ,&'%&;':0&-()U*0$#&0VW ?4CB ?4?FE ?4EL ?4?FD ?4LF ?4?@?M%1N)&-.&7$'"% FTJ"%'%&;':0&-()U8$,(0"%VW ?4@A ?4?FB ?4CB ?4?@? ?4E@ ?4?@@M%1N)&-.&7$'"% FT=&%'$Y$;&-()U^(,(%"-VW ?4CF ?4?FC ?4CA ?4?FB ?4EE ?4?FAM%1N)&-.&7$'"% FT9%$Y"."-()UM(3:%(0VW ?4EG ?4?FE ?4B? ?4?FD ?4BA ?4?@?

M%1N)!",;"1-.3

="-'I)F ="-'I)@ ="-'I)C

J"'(3_`F4)5"1%7(_)9I",3"-)^(1'(%3a)=$%b('57$-)M$'$[$3(c):($%3)@??CK@??L`@4)\$3(0&-()7"#$%&$'(3)3(')'"_)$N()S1$-'&0(de4L?c4DLVc)N(-.(%d2(,$0(c)%(N&"-)d"1'3&.()](3'c);0$-)':;(d88>c)3;(7&$0':dN(-(%$0);%$7'&'&"-(%c).&$N-"3&3)d),$P"%).(;%(33&"-c)2%(S1(-7:df@+H.$:c).%1N)7",;"1-.)d)R01"+('&-(c).%1N)7";$:H7"1-':)&-7",()3(')'"),($-)0(#(04`C4)M$'$)13(.)2"%)(3'&,$'&"-)&-701.(3)"-0:)&-.&#&.1$03);%(37%&[(.)$).%1N)'%($',(-')&-)'I()g%3');(%&".4))9I().1%$'&"-)"2)'I()3;(00)(-.3)hI(-)'I();$'&(-')'(%,&-$'(3)'%($',(-')[:)3h&'7I&-N)'")$)-(h).%1N)'%($',(-')"%)'")'I()"1'3&.()N"".4`E4)M(7&3&"-)-".(3)UiV)%(j(7')&-.&#&.1$0K3;(7&g7)%(g00);"&-'3)"%)"k7()#&3&'34))9I(),".(0)."(3)-"')13().1%$'&"-)&-).$:3)$3)'I().(;(-.(-')#$%&$[0(c)3&-7()'I&3),&NI')7"-213()'I()'&,()"-)'%($',(-')h&'I)'I().&k710':)"2)37I(.10&-N)$)2"00"hK1;)#&3&'4))J"'()'I$')2%(()3$,;0(3).")-"')$;;($%)&-)'I().$'$4))62)'I();$'&(-')g003)$);%(37%&;'&"-)2"%)'I()3$,;0(.).%1N)h&'I&-)A?).$:3)"2)I&3)&-&'&$0)"k7()#&3&'c)'I()&-.&#&.1$0)h&00)$;;($%)&-)'I()3I$%()2"%)'I$').%1N4`!!

38

Tab

le8:

Obse

rved

supply

and

dem

and-s

ide

ince

nti

ves

acro

ssb

enefi

tpla

ns,

by

bro

adpla

nty

pe

!"#

$%&

'()#*+&

!"#

$%&

'()#*+&

,$-&$./)&0123

41%&$&4/-154+%016*37

&41-3/$0+4#$

-&8-39

#:;&$.0<

=>?@A

BC?>A

BC?DA

C?EA

F?>F

C?GA

F?FB

H0&1321234;+*#4:1./&

40123

4154+%1-39#

:;&$.0

ED?@A

C>?IA

EE?IA

(B?GA

F?>G

(I?EA

F?GD

'0:-"/#.4/-10&4)/-&01-#4)&513+

.GI?EA

GI?GA

>F?>A

F?IA

F?EF

>I?DA

F?FF

'#./&

$.1-30.1234190:-"/#.4/-13+.9#

./&$.10&4)/-&01/$-4&#0&017/."

1J1321)

/0/.0

I>?>A

>G?IA

I>?=A

(II?=A

F?IF

(F?@A

F?GE

'4&-&4./2/-#./3$14&K+

/4&;

&$.0123

41/$9

#./&$.1-#4&

GD?=A

B>?BA

CI?CA

(DB?=A

F?FF

(I@?@A

F?F>

H./*/L#./3$14&)/&7

12341/$9

#./&$.1-#4&

I@?IA

>@?CA

DG?CA

(IF?EA

F?IE

(I@?EA

F?FF

!#0&1;#$

#%&;

&$.12341/$

9#./&

$.1-#4&

I=?EA

DI?FA

EE?EA

(II?EA

F?IG

(DC?FA

F?FF

'4&0-4/9./3

$1-39#

:M1%&$&4/-16N

8;3$

."10+9

9*:<

B?G@

111111111111111111

B?IC

111111111111111111

@?IB

111111111111111111

F?DI

F?DC

I?>=

F?FD

'4&0-4/9./3

$1-39#

:M1O4#$5

1$#;

&16N8;

3$."10+9

9*:<

I@?CB

1111111111111111

IB?@@

1111111111111111

ID?=E

1111111111111111

(I?IF

F?>G

D?@D

F?F>

'4&0-4/9./3

$1-39#

:M1$3$

(234;

+*#4:1O4#$

516N8;

3$."10+9

9*:<

DI?C@

1111111111111111

DC?FC

1111111111111111

>I?GE

1111111111111111

(G?GF

F?I>

IF?>I

F?FF

P$$

+#*1;

#Q/;

+;123

413+.9#

./&$.190:-"/#.4/-1)/0/.016J1321)

/0/.0<

DB?GG

1111111111111111

D@?DI

1111111111111111

D=?B@

1111111111111111

I?IG

F?GF

(I?GD

F?DI

R+;

O&413

219*#$013O0&4)&51321&#-"1.:9&

SGI

>=>CD

R3$

&T!

PRUV,

WWV,

WRV,

RPWWP

XYU

'43O

17/."

1-#4)&13+.M1'

0:-"/#.4/0.

>E?B=A

>?DGA

IG?DGA

GG?FGA

@?CIA

F?BGA

G?=GA

'43O

17/."

3+.1-#4)&13+.M11'0:-"/#.4/0.

>E?GEA

>?DDA

IG?DCA

GG?GDA

@?=GA

F?@DA

G?@CA

Z/."

1-#4)&(3+.1(17/."

3+.1-#4)&(3+.M1'

0:-"/#.4/0.

F?GGA

F?FIA

(F?FIA

(F?GFA

(F?DDA

F?IIA

F?IBA

W.51&443413

21-"#

$%&

F?DIA

F?IDA

F?DFA

F?D@A

F?IBA

F?F@A

F?>>A

'#$&*1IS1Y

O0&4)&510+99

*:1#$5

15&;

#$5(0/5&

1/$-&$./)&01#-43001O&$&2/.19*#$0M1O:19*#$

1.:9&

'*#$

1,$-&$./)&0

'*#$

1T:9

&U/22&4&$-&01#-43001T:9

&0!#9

/.#.&51

[\Y

R3$

(!#9

/.#.&51

[\Y

''Y8!3;

94&"

&$0/)&

!#9

?1(1R

3$(-#9

?1[\Y

!#9

/.#.&51[\Y1('

'Y

'#$&*1>S1\

+*./$

3;/#*1*3%

/.194&5

/-./3

$1&Q&4-/0&M1-#4)&13+.1/$

-&$./)&?

U4+%1!*#00

R3.&0S]

I?1W3+

4-&S1T"3

;03$1V&

+.&40^1\#4_&.W-#$1`&

$&a.1'*#$1U&0/%$1U#.#O#0&M1:&#401>FFD(>FFE?]

>?1'#./&$.01/$1&/."

&41-#.&%34:1321[

&#*."

1\#/$.&$#$

-&1Y4%#$

/L#./3$16[

\Y<1;

+0.1-"3

30&1243;

1#19#4./-+*#41*/0.13219

43)/5&

40123

41#**1$3

$(&;

&4%&$-:1-#4&?1b

#-"1

9#./&

$.1-"3

30&01#1'4/;#4:1!#4&1'"

:0/-/#$1.31;

#$#%&1"/01-#4&M17

/."109&

-/#*/0.1-#4&1O:

14&2&44#*13$*:?11!

#9/.#

.&519*#$

0M1+$*/_&1."&1$3

$(-#9/.#.&51)&40/3$0M19

#:123

41#**19

":0/-/#$

10&4)/-&0194309&

-./)&*:1O:

1&$43**&&?]

D?1V&;

#/$/$%

1)#4/#O*&01/$

1;+*./$

3;/#*1*3%

/.10&.1.3S1P%&S1GF."(CF."c1d

&$5&

4S1e&;#*&c1'*#$1.:9&

S1!#9

?1[\Yc1!

39#:;&$.1,$-&$./)

&0S1f

&0c1!

3/$0+4#$

-&1,$

-&$./)&0S1

R3c1,$

1\WP

ghf&

0c1V&%/3$h

Z&0.c1Yi-&1)/0/.1-39#

:;&$.hNIF?]

39

Table 9: Examination of Physician Characteristics on Prescribing Behavior

!"#$"%&'

()*+$,$"-.+/("0$'-0/1'23/4)"#'/5#67/8"-"9':/

;"#'/<=>?@()*+$,$"-/"-/

A7B&6*''

()*+$,$"-.+/("0$'-0/1'23/4)"#'/5#67/(#$2"0'/C-+3/

<=>?@

()*+$,$"-/4B',$"&0*/=/

D'-'#"&/(#",0$0$6-'#

()*+$,$"-/4B',$"&0*/=/(+*,)$"0#$+0

()*+$,$"-/E,,'B0+/F'G/("0$'-0+/6-/;"B$0"0':/

(&"-+H3??? I?3H?J ?3JKJ I?3?HL I?3?JL ?3?HJ

<3???H <3???H ?3MHH ?3HMK ?3>NLOH3??? ?3?K> ?3HH> I?3H>O ?3HLP

?3??H <3???H <3???H <3???HH3??? I?3?MK ?3?LK ?3?KK

?3??P ?3LPM ?3??HH3??? I?3NJK ?3HJ?

<3???H <3???HH3??? I?3HP?

<3???HH3???

8'"- M?3>K@ LP3M>@ PK3H>@ PO3LP@ PH3OL@ MJ3JM@FQ7%'#/65/6%+'#2"0$6-+ HRN?O////////////// HRN?O////////////// HRN?O////////////// HRN?O////////////// HRN?O////////////// HRN?O//////////////

()*+$,$"-/E,,'B0+/;"B3/

(&"-+/S@T

()*+$,$"-/U6'+/F60/

E,,'B0/;"B3/(&"-+/S@T

E,,'B0+/;"B3!/U6'+-.0/

E,,'B0/S@T

40"-:"#:/A##6#/65/

U$55'#'-,'

E,,'B0+/;"B3!/U6'+-.0/

E,,'B0/S@T

40"-:"#:/A##6#/65/

U$55'#'-,'F6-' HP3H//////////////// HJ3M//////////////// ?3O////////////////// H3>////////////////// H3J////////////////// H3M//////////////////;$0"&6B#"7/VW#/S;'&'X"T P3O////////////////// M3J////////////////// SH3>T///////////////// H3H////////////////// S?3NT///////////////// ?3M//////////////////A+,$0"&6B#"7/YX"&"0'/SZ'X"B#6T HP3N//////////////// H>3N//////////////// SL3OT///////////////// L3J////////////////// SJ3MT///////////////// L3K//////////////////[&Q6X'0$-'/V;Z/S(#6\",T O3O////////////////// H?3K//////////////// SH3PT///////////////// H3L////////////////// S?3OT///////////////// ?3O//////////////////("#6X'0$-'/V;Z/S("X$&T K3L////////////////// O3?////////////////// ?3H////////////////// L3H////////////////// ?3>////////////////// L3L//////////////////4'#0#"&$-'/V;Z/S]6&650T HJ3J//////////////// HJ3M//////////////// ?3L////////////////// H3K////////////////// ?3J////////////////// H3M//////////////////!'-&"5"X$-'/V;Z/SA55'X6#R/^1T HJ3M//////////////// H?3>//////////////// L3P////////////////// L3H////////////////// H3J////////////////// L3L//////////////////WQB#6B$6-/V;Z/S_'&&%Q0#$-R/^Z`41T H>3P//////////////// HP3P//////////////// H3N////////////////// H3L////////////////// H3K////////////////// H3J//////////////////a#"\6:6-'/V;Z/S4'#\6-'T M3K////////////////// M3H////////////////// ?3P////////////////// ?3P////////////////// ?3J////////////////// ?3P//////////////////

("-'&/Hb/;6##'&"0$6-+/$-/()*+$,$"-/(#",0$,'/;)"#",0'#$+0$,+R/#'B6#0':/G$0)/BI2"&Q'+

()*+$,$"-.+/("0$'-0/1'23/4)"#'/5#67/8"-"9':/;"#'/<=>?@()*+$,$"-/"-/A7B&6*''

()*+$,$"-.+/("0$'-0/1'23/4)"#'/5#67/(#$2"0'/C-+3/<=>?@()*+$,$"-/4B',$"&0*/=/D'-'#"&/(#",0$0$6-'#()*+$,$"-/4B',$"&0*/=/(+*,)$"0#$+0

()*+$,$"-/E,,'B0+/F'G/("0$'-0+/6-/;"B$0"0':/(&"-+

U#Q9/;67B6Q-:

_$0)/W",c9#6Q-:/"-:/C-,'-0$2'/;6-0#6&+

("-'&/Lb/(#':$,0$6-/AX'#,$+'+R/8Q&0$-67$"&/Z69$0/6-/U#Q9/;67B6Q-:/;)6$,'

Y%+'#2':/U"0" _$0)/W",c9#6Q-:/;6-0#6&+

F60'+bdH3/46Q#,'b/F"0$6-"&/E7%Q&"06#*/8':$,"&/;"#'/4Q#2'*R/*'"#+/L??JIL??>3dL3/('"#+6-/,6##'&"0$6-/,6'e,$'-0+/#'B6#0':/$-/("-'&/H3//(I2"&Q'/#'B6#0':/56#/-Q&&/)*B60)'+$+/0)"0/,6'e,$'-0/=/?3dJ3/W",c9#6Q-:/,6-0#6&+/$-/7Q&0$-67$"&/&69$0b/D'-:'#/=/['7"&'f/1",'/=/_)$0'f/E9'/gQ"-0$&'/=/h3PR3Mif/("0$'-0/B"*/0*B'/=/B#$2"0'/$-+Q#"-,'f/4B',$"&0*/=/D'-'#"&/(#",0$0$6-'#3dP3/C-,'-0$2'/,6-0#6&+/$-/7Q&0$-67$"&/&69$0b/()*+$,$"-/#'2'-Q'/5#67/7"-"9':/,"#'/<>?@j=k'+f/()*+$,$"-/"-/'7B&6*''j=F6f/()*+$,$"-/#'2'-Q'/5#67/B#$2"0'/$-+Q#"-,'<>?@=k'+3d!

40

Table 10: Estimates from a Weibull Duration Model on Exit

!"#$"%&' ()'**+ ,-.+/0##)# 123-"-4'$%5&&/36"7'/7"#"8'-'#/9"&76": ;+<<= ;+;;> ?@+AB)C3-"C- 2A+@<; ;+;>; ><+A!"#$%&'()*+,-)'&'+."&*/.*+/ADE"&'FADGH'/I5"C-$&'/J/K;+;L;+M=: ;+MA? ;+;A@ AM+=ADE"&'FADGH'/I5"C-$&'/J/K;+M=L;+=;: ;+;@M ;+;A? >+;ADE"&'FADGH'/I5"C-$&'/J/K;+=L;+@=: ;+;A< ;+;A@ ;+@ADE"&'FADGH'/I5"C-$&'/J/K;+@=LA+;: ;+;@> ;+;A@ >+<ADN'8"&'FADGH'/I5"C-$&'/J/K;+;L;+M=: ;+A?O ;+;A< A>+@ADN'8"&'FADGH'/I5"C-$&'/J/K;+M=L;+=;: ;+;>@ ;+;A< <+PADN'8"&'FADGH'/I5"C-$&'/J/K;+@=LA+;: ;+;<O ;+;A< <+;AKQ'H$)C/J/4'3-R ;+AAP ;+;A; AA+PB)5C-S/$CB)8'L/$C/TA;;;3 2;+;;< ;+;;; ?+;ADEUU/.$"HC)3$3R 2;+;@O ;+;AA @+<ADV3SB6$"-#$3-R 2;+;P= ;+;AM =+<ADW-6'#/,7'B$"&$3-R 2;+;A? ;+;A; A+?01+"1.*2"/AD("7$-"-'./XEW/7&"CR ;+;O< ;+;A; ?+OADY)C2("7$-"-'./XEW/7&"CR ;+;<; ;+;AM M+=ADB"#Z'/)5-R ;+;@; ;+;;? ?+>B)7"S/7'#/."S/3577&S ;+;>P ;+;;O >+O!&3%,-)'&'+."&*/.*+/ADU)3$CH/M2<[/7'#/."SR ;+M<= ;+;M; AA+?AD($-"&)7#"8/9('&'[":R 2;+;<= ;+;<O ;+OAD03B$-"&)7#"8/9\'["7#):R 2;+;>> ;+;<@ A+MADN&5)['-$C'/9V#)]"B:R ;+;== ;+;<@ A+=ADV"#)['-$C'/9V"[$&:R 2;+A;A ;+;<? M+@AD,'#-#"&$C'/9^)&)*-:R 2;+;AP ;+;<? ;+>ADU5&)['-$C'/9(S8%"&-":R 2;+;AA ;+;>P ;+<AD!'C&"*"[$C'/90**'[)#:R 2;+MMM ;+;<? =+OAD_57#)7$)C/94'&&%5-#$C:R 2;+AA; ;+;<? M+OADY)#-#$7-S&$C'/9V"8'&)#:R 2;+;>@ ;+;P< ;+@ADE$#-"]"7$C'/9Q'8'#)C:R ;+;>O ;+;=< ;+OAD1#"]).)C'/9U'3S#'&:R 2;+A@A ;+;>? <+PY)-'3`aA+/,)5#B'`/16)83)C/Q'5-'#3b/E"#c'-,B"C/U"-"%"3'L/S'"#3/M;;<2M;;=+aM+/U"-"/53'./*)#/'3-$8"-$)C/$CB&5.'3/)C&S/$C.$Z$.5"&3/7#'3B#$%'./"/.#5H/-#'"-8'C-/$C/-6'/d#3-/7'#$).+//16'/.5#"-$)C/)*/-6'/37'&&/'C.3/e6'C/-6'/7"-$'C-/-'#8$C"-'3/-#'"-8'C-/%S/3e$-B6$CH/-)/-6'/)5-3$.'/H)).+a<+/0[B&5.'./.#5H/B)87)5C./$C.$B"-)#`/ADG8$-#$7-S&$C'/90&"Z$&:Ra

41

Table 11: Rates of Illness Relapse Conditional on Incentives and Demographics

!"#$%&'(&)*+,-&./-,0"-$/

1"2,"3%$ ./-,0"-$ 4-56&.22*2 78/-"-9:$2"+$&

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

HH\IPN&&&&&&&

!"#$%&E(&!2$5,=-,*#&.]$2=,/$/

@"/$

!2*36&*U&

2$%"D/$E@"/$&8&

<"/$%,#$4-56&.22*2&*U&@A"#+$

O3/$2:$5&=*:"2,"-$/ P6JE^ 8&&&&&&&&&&&&& 8&&&&&&&&&&&&&9%%&D"-,$#-/&-2$"-$5&?#5$2&="D,-"-$5&[TO&D%"#/ P6FS^ G6PI^ G6EJ^9%%&D"-,$#-/&3$+,#&52?+&-2$"-0$#-&*#&"#&44RC P6'F^ 8G6EH^ G6'P^;2?+&=*D"L0$#-/&,#=2$"/$&'GG^&U2*0&3"/$%,#$ P6SE^ G6JG^ G6EG^9%%&D%"#/&,#:*%:$&0$#-"%&A$"%-A&="2:$8*?-/ J6'S^ 8'6EJ^ G6'I^

BC#=%?5$/&5$0*+2"DA,=&=*#-2*%/(&"+$&W?"#-,%$/&,#-$2"=-$5&Z,-A&+$#5$2\&'KD"-,$#-&%,:$/&,#&T49M\&D"-,$#-&A*0$&2$+,*#&_Q*2-A&."/-\&Q*2-A&@$#-2"%\&4*?-A\&`$/-a\&D"-,$#-&5,"+#*/,/&_3L&2$D*2-$5&C@;8F&=*5$a\&":$2"+$&,#=*0$&,#&D"-,$#-b/&A*0$&=*?#-L6

Q?03$2&*U&*3/$2:"-,*#/&,#&$/-,0"-,*#(

Q*-$/(c'6&4*?2=$(&7A*0/*#&R$?-$2/b&T"2>$-4="#&O?-D"-,$#-&"#5&;2?+&@%",0/&;"-"3"/$/\&L$"2/&EGGH8EGGP6cE6&R$%"D/$&d&'&,U&D"-,$#-&2$-?2#$5&-*&"&DAL/,=,"#&U*2&5$D2$//,*#&="2$&3$-Z$$#&-A2$$&"#5&/,]&0*#-A/&"U-$2&"#&,#,-,"%&-2$"-0$#-&$D,/*5$6cH6&O0,e$5&@"-$+*2,$/&,#&$/-,0"-,*#(&;,"+#*/,/&,/&f5$D2$//,*#&#*-&*-A$2Z,/$&/D$=,g$5fh&4D$=,"%-L&,/&i$#$2"%&!2"=-,-,*#$2h&!%"#&,/&!!Oh&C#,-,"%&52?+&,#&44RC&=%"//h&i$#5$2Y"+$&W?,#-,%$&,/&'KV$0"%$MB'K9+$&W?,#-,%$&d&jG6J\G6IaMh&R$+,*#&,/&`$/-6c

42