Advertising and Risk Selection in Health Insurance Markets

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Advertising and Risk Selection in Health Insurance Markets Naoki Aizawa and You Suk Kim 2015-101 Please cite this paper as: Aizawa, Naoki, and You Suk Kim (2015). “Advertising and Risk Selection in Health Insurance Markets,” Finance and Economics Discussion Series 2015-101. Washington: Board of Governors of the Federal Reserve System, http://dx.doi.org/10.17016/FEDS.2015.101. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Advertising and Risk Selection in Health Insurance Markets∗ Naoki Aizawa and You Suk Kim† November 1, 2015 Abstract We study impacts of advertising as a channel of risk selection in Medicare Advantage.Weshowevidencethatbothmassanddirectmailadvertisingaretargetedtoachieve risk selection. We develop and estimate an equilibrium model of Medicare Advantage withadvertisingtounderstanditsequilibriumimpacts. Wefindthatadvertisingattracts thehealthymorethantheunhealthy. Moreover,shuttingdownadvertisingincreasespremiums by up to 40% for insurers that advertised by worsening their risk pools, which further reduces the demand of the unhealthy. We argue that risk selection may make consumersbetteroffbyimprovinginsurers’riskpools. ∗Thefirstdraftofthispaper(November2013)wascompletedwhenbothoftheauthorswereatthe UniversityofPennsylvania. WearegratefultoHanmingFang,KatjaSeim,BobTown,andKenWolpin fortheirguidanceandencouragement. WealsothankourconferencediscussantElenaKrasnokutskaya for the detailed comments and seminar participants at many places. Any remaining errors are ours. WegratefullyacknowledgefundingsupportfromtheWhartonRiskCenterRussellAckoffFellowship. KathleenNosalkindlysharedMedicareCompareDatabaseforthispaper. †NaokiAizawa: DepartmentofEconomicsattheUniversityofMinnesotaandtheFederalReserve BankofMinneapolis,aizawa@umn.edu. YouSukKim: ResearchandStatistics,theBoardofGovernorsoftheFederalReserveSystem, you.kim@frb.gov. Theviewsexpressedarethoseoftheauthors and do not necessarily reflect those of the Board of Governors, the Federal Reserve System, or the FederalReserveBankofMinneapolis.

1 Introduction ManyAmericanspurchasehealthinsuranceinprivatemarketsthatarelargelydesigned by the government. These markets, including Medicare Advantage, Medicare Part D, and health insurance marketplaces, have substantially expanded over time.1 One of theimportantgoalsforthegovernmentindesigningthesemarketsistoprovideaccess to health insurance to unhealthy individuals by mitigating insurers’ risk selection (or cream-skimming), the selective enrollment of low-cost healthy individuals. In these markets,privateinsurersareprohibitedfromdiscriminatingindividualsbasedontheir health risks in term of plan offering, premiums or plan benefits. Moreover, the government uses risk adjustment, through which insurers receives a subsidy based on an enrollee’s health risk. However, the risk adjustment is still not perfect in practice, and theincentivesforriskselectionstillremain. Previous empirical studies find the presence of risk selection (Kuziemko et al., 2013)anddiscusshowtheimperfectriskadjustmentsystemleadstoanexcessgovernment expenditure by providing excessive subsidies to insurers for enrolling low-cost healthyconsumers(Brownetal.,2014). However,inevaluatingriskselection,littleis known about its effectiveness and its effects on equilibrium outcomes. By treating the demands of individuals with different health risks differently, risk selection affects an insurer’sriskpoolandtherebyitsmarginalcost. Thus,withimperfectriskadjustment, risk selection will eventually affect an insurer’s pricing. If risk selection decreases the premium, then it may rather help unhealthy individuals purchase health insurance and improve their welfare. Of course, overall welfare impacts depend on the possibility of excessive spending on risk selection (i.e., rent-seeking) due to insurers’ competition for attracting the healthy individuals. The quantitative significance of these issues has notbeenstudiedintheexistingliterature,asthepresenceofriskselectionisexamined withoutusingmeasuresofriskselectiontools. In this paper, we empirically study advertising as a means of risk selection in the context of Medicare Advantage (MA), which offers an option for Medicare beneficiaries to choose private coverage instead of public traditional Medicare. We focus on 1In 2014, roughly 16 million elderly individuals eligible for public insurance Medicare (Medicare beneficiaries)wereinsuredbyprivateMedicareAdvantageplans. MedicarePartDprovidesprescription drug coverage to 37 million Medicare beneficiaries only through private plans. Health insurance marketplaceswereintroducedin2014duetotheAffordableCareAct. 1

advertising for several reasons. Advertising is one of the most important marketing activities of any company to target certain consumers. In particular, the significance of marketing activities and advertising by insurers in the health insurance markets has been pointed out in the literature (see, e.g., Cebul et al., 2011). Furthermore, advertising in MA is largely unregulated, unlike the design of plan benefits. Thus, advertising can be much more responsive to the risk adjustment system. In this paper, we provide thefirstempiricalanalysisofequilibriumimpactsofadvertisingonahealthinsurance market by focusing on its role as risk selection. We start our analysis by investigating whetherMAinsurerstargetadvertisingtoindividualsandregionswhereriskselection resultsingreaterprofits. ThenwestructurallyestimateamodeloftheMAmarketand quantify the effects of risk selection through advertising on the market outcomes in MAsuchasdemand,pricing,andgovernmentexpenditures. The MA market is an ideal environment in which to study risk selection for three reasons. First,anMAplanreceivesasubsidycalledacapitationpaymentfromthegovernment for an enrollee and then bears the health care costs incurred by the enrollee. Although an MA plan often charges a premium, the capitation payment accounts for most of a plan’s revenue. Moreover, the capitation payment has been known to be imperfectly adjusted to an enrollee’s health risk. Therefore, concerns for risk selection in MA have arisen.2 Second, as we describe in section 2, there is variation in the capitationpaymentacrossmarketsandoveryears,whichallowsustoidentifyhowthe incentiveofadvertisingrespondstochangesincapitationpaymentsanditsquantitative significance on market outcomes. Lastly, the large size of the MA program makes it a veryimportantmarkettostudy. We begin our empirical analysis by providing evidence that insurance companies target advertising at the market level (defined as a county-year pair) and at the individual level by exploiting unique data sets about mass advertising and direct mail advertising. We obtain the data for mass advertising by insurers from 2001 to 2005 from the AdSpender Database of Kantar Media, which includes advertising expenditures on TV, newspaper, radio, etc. The direct mail advertising data are from Mintel Comperemedia, which provide information on demographic characteristics of nationally representative households and characteristics of direct-marketing mailings they 2Notethatthegovernmenthasallowedcapitationpaymentstobecomemoreriskadjustedinthepast 10years. SeeNewhouseetal.(2012). 2

received. We first show that insurers target mass advertising at markets where risk selection (i.e. enrolling healthy individuals) in the markets is more profitable than in othermarkets. Wealsofindthat,withinamarket,directmailingsaretargetedatcertain types of individuals. Moreover, we provide evidence that the targeting of direct mailingsrespondedtotheintroductionofamorecomprehensiveriskadjustmentregimein 2004 that changed amounts of capitation payments an MA plan receive for enrolling individualswithdifferenthealthstatuses. In order to understand the impact of advertising on market outcomes, we develop and structurally estimate a consumer demand model of MA markets with advertising. Consumers make a discrete choice to enroll with one of the available MA insurers or to select traditional Medicare. The impact of advertising can differ according to theconsumer’scharacteristics,includingthepreviousinsurerchoiceandhealthstatus, whichcapturesthepossibilitythatdifferentindividualsresponddifferentlytoadvertising. Consumer preferences for an insurer depend on characteristics such as premiums and coverage benefits. We also allow that the consumer can face the switching cost associated with changing insurance choices, which is known to be important in the contextofMA(seeMiller(2014),Milleretal.(2014),andNosal(2012)).3 We estimate the demand side using data on consumer characteristics and choice fromtheMedicareCurrentBeneficiarySurveyanddataoninsurercharacteristicsfrom CMSState-County-Plan(SCP)files. Estimationisbygeneralizedmethodofmoments, in the spirit of Berry et al. (2004). We allow for insurer-year fixed effects and county fixedeffectsanduseinstrumentalvariablestoaccountfortheendogeneityofpremiums and advertising stemming from unobserved plan heterogeneity. Our estimates show thathealthierindividualsareresponsivetoadvertising,andthus,additionaladvertising attracts more healthy individuals. Moreover, sizable switching costs are associated with changing insurers. Because of the large switching costs, advertising has greater effectsonthedemandbynewMedicarebeneficiarieswhofacenoswitchingcosts. We evaluate the importance of risk selection through advertising on consumer demand, pricing, and government expenditures by conducting a counterfactual experimentthatexogenouslyshutsdowntheadvertisingactivities. Inordertoinvestigatethe supply side responses, we estimate the supply side parameters by assuming that firms 3For works on switching cost or inertia in other health insurance markets, see Handel (2013), Ho etal.(2015),andPolyakova(2014). 3

play Bertrand Nash price competition in a differentiated goods market.4 An insurer’s revenue from an enrollee equals the sum of the premium and capitation payment for the enrollee. The capitation payment is adjusted based on individual characteristics, but importantly, it is not perfectly adjusted based on individual health risks, making theinsurer’sprofitfromanenrolleevarybyindividual. Thus,theoptimalpricingtakes intoaccounttheeffectsofthesechoicesontheplan’scompositionofhealthrisks. Weinvestigatetheimpactofshuttingdownadvertisingintwocounterfactualsituations. Inone,premiumsareexogenouslyfixedattheirbaselinelevels,andintheother, insurers reoptimize their premiums in a situation without any advertising. We find that shutting down advertising lowers the overall demand for MA and that its impact is much larger for healthier consumers and new Medicare beneficiaries. Interestingly, when insurers reoptimize their premiums, the demand for MA decreases much more than when premiums are fixed exogenously. The decrease is especially pronounced, around 10%, for individuals that newly became eligible for Medicare because they do not have switching costs. The further decline is driven by a sharp increase in premiums, around 40%, among insurers that had relatively large advertising expenditures. Thekeymechanismisthatshuttingdownadvertisingmakestheinsurersunabletoengageinriskselection. Asfewerhealthierindividualswillnowobtaincoveragethrough MA,theinsurers’riskpoolswilldeteriorate. Withimperfectlyrisk-adjustedcapitation payments, the change in the risk pool will increase premiums for those insurers. At the same time, premiums decrease for other insurers that had few or zero advertising expenditures, which highlights a rent-seeking aspect of risk selection: advertising improves an insurer’s own risk pool while it negatively affects other insurers’ risk pools. Overall, shutting down advertising increases premiums on average and decreases the demand. Moreover, a wasteful advertising competition through insurers’ rent-seeking waslikelytobelimitedasmostsmallinsurersdidnotadvertise. Thus,underanimperfect risk adjustment system, risk selection through advertising may make consumers better off by lowering premiums without much inefficient spending. Although it is commonlydiscussedthatriskselectionshouldbeminimized,ourfindingsuggeststhat riskselectioncanpossiblyimprovethewelfare. 4Becauseweconductacounterfactualexperimentthatexogenouslyshutsdownadvertising,weare agnosticabouthowadvertisingisoptimallychosenintheeconomywithadvertising. 4

Related Literature This paper contributes to large literature empirically investigating selections in insurance markets. Although the majority of the literature focus on the consumer side selection, there are a few studies investigating risk selection by insurers.5 Bauhoff (2012) studies risk selection in the German health insurance market by looking at how insurers respond differently to insurance applications from regions with different profitability levels. Kuziemko et al. (2013) study risk selection among private Medicaid managed-care insurers in Texas and provide evidence that the insurers risk-select more profitable individuals. Brown et al. (2014) provide evidence that insurers engage in risk selection in MA by exploiting changes in MA risk adjustment system. Although the occurrences of risk selection are well documented in the related works, there is still little research on its channels. This paper adds to this literature by investigatingtheroleofadvertisingonriskselectionanditsequilibriumimpact.6 Thispaperisalsorelatedtonewandgrowingliteraturestudyingsupply-sidecompetition in insurance markets. Lustig (2011) studies adverse selection and imperfect competition in MA, and Starc (2014) investigates the impact of pricing regulations in Medicare supplement insurance. Recently, Cabral et al. (2014), Duggan et al. (2014), and Curto et al. (2014) study the impact of capitation payments in MA markets. Especially,Curtoetal.(2014)useMedicareadministrativerecords,whichcontainricher information about individual characteristics than we have available and which cover more recent years when capitation payment were adjusted more to variation in expected medical expenditures. They find that healthier individuals still purchase MA and it is still profitable for insurers to attract healthy individuals. However, they also arguethatinsurers’behaviorsdonotaffectitsriskpool. Theyassumethatpricingisan insurer’s only tool affecting the risk pool, and they do not find an correlation between aninsurer’spremiumanditsriskpool. Inthispaper,wefindthatpricesensitivitydoes not vary by individuals with different health status, consistent with theirs. However, wealsofindthatadvertisingisanimportantchannelofaninsurer’sriskselection,and as long as risk adjustment is not perfect, pricing decisions substantially depend on the effectiveness of risk selection through advertising. Therefore, our result suggests that 5See Chiappori and Salanie (2000), Finkelstein and McGarry (2006), and Fang et al. (2008) for consumer-sideselection. SeeVandeVenandEllis(2000)andEllisandFernandez(2013)forexcellent surveysontheconceptandissueofriskselectionandriskadjustment. 6See also Geruso and Layton (2015) who interestingly find that insurers manipulate reports to the governmentabouttherisktypesofenrolleestocapitationpayments. 5

evaluating the welfare impacts of risk adjustment designs requires the broader measurementofinsurers’riskselectiontools. Lastly,thispaperisalsorelatedtotheliteratureonadvertising. Manyempiricalpapersintheliteraturestudythechannelsthroughwhichadvertisinginfluencesconsumer demand,thatis,whetheradvertisinggivesinformationaboutaproductoraffectsutility fromtheproduct.7 Morerecently,researchershavestudiedtheeffectsofadvertisingin anequilibriumframeworkfordifferentcontexts: Goeree(2008)forthepersonalcomputer market; Dubois et al. (2014) for junk food markets; and Gordon and Hartmann (2013) and Moshary (2015) for the U.S. elections. A paper that is closely related to ours is Hastings et al. (2013), who also study advertising in a privatized government program (the privatized social security market in Mexico). An important difference between this paper and the related works on advertising is that advertising in health insurance markets affect the marginal cost of providing an additional insurance due to theriskselection,throughwhichpricingandconsumerwelfareareaffected. The paper is organized as follows. Section 2 describes Medicare Advantage in greater detail. Section 3 describes the data and presents results from the preliminary analysis. Section 4 outlines the model, and Section 5 discusses the estimation and identification of the model. Section 6 provides estimates of the model, and Section 7 describestheresultsfromcounterfactualanalyses. Section8concludes. 2 Background on Medicare Advantage Medicare is a federal health insurance program for the elderly (people aged 65 and older) and for younger people with disabilities in the United States. Before the introduction of Medicare Part D in 2006, which provides prescription drug coverage, Medicare had three Parts: A, B, and C. Part A is free and provides coverage for inpatient care. Part B provides insurance for outpatient care. Part C is the Medicare Advantage program, previously known as Medicare + Choice until it was renamed in 2003.8 The traditional fee-for-service Medicare comprises of Parts A and B, which reim- 7Forexamples,seeAckerberg(2001,2003);ChingandIshihara(2012);Clarketal.(2009). 8Althoughwewillfocusontheperiod2000–2003forouranalysis,wewillrefertoMedicareprivate plansasMedicareAdvantageplansinsteadofMedicare+Choiceplans. 6

burse costs of medical care utilized by a beneficiary who is covered by Parts A and B. AsanalternativetotraditionalMedicare,aMedicarebeneficiaryalsohastheoptionto receivecoverage froman MAplan runby aqualified privateinsurer. Insurerswishing toenrollMedicarebeneficiariessigncontractswiththeCentersforMedicareandMedicaid Services (CMS) describing what coverage they will provide and at what costs. The companies that participate in the MA program are usually health maintenance organizations (HMOs) or preferred provider organizations (PPOs), many of which have a large presence in individual or group health insurance markets, such as Blue Cross Blue Shield, Kaiser Permanente, United Healthcare and so on. They contract with the Center for Medicare and Medicaid Services on a county-year basis and compete for beneficiariesineachcountywheretheyoperate. The main attraction of MA plans for a consumer is that they usually offer more comprehensivecoverageandprovidebenefitsthatarenotavailableintraditionalMedicare. For example, many MA plans offer hearing, vision, and dental benefits, which are not covered by Parts A or B. Before the introduction of Part D, prescription drug coverage was available in MA plans, but not in traditional Medicare. Although a beneficiary in traditional Medicare is able to purchase Medicare supplement insurance (known as Medigap) for more comprehensive coverage than basic Medicare Parts A and B, the Medigap option is priced more expensively than a usual MA plan, many of which require no premium. Therefore, MA is a relatively cheaper option for beneficiaries who want more comprehensive coverage than traditional Medicare offers. In return for greater benefits, however, MA plans usually have restrictions on provider networks. Moreover, MA enrollees often need a referral to receive care from specialists. Incontrast,anindividualintraditionalMedicarecanseeanyproviderthataccepts Medicarepayments. PreviousworksonMAfindthathealthierindividualsaresystematicallymorelikely to enroll in an MA plan.9 Risk selection was blamed for the selection pattern. MA insurers are not allowed to charge individuals with different health statuses within a county different premiums. More importantly, capitation payments from the government do not fully account for variation in health expenditures across individuals. Until the year 2000, adjustments to capitation payments were made based only on demographic information such as an enrollee’s age, gender, welfare status, institutional 9Forexample,seeLangwellandHadley(1989);Melloetal.(2003);Batata(2004). 7

status, and location, which accounted for only about 1% of an enrollee’s expected health costs (Pope et al. 2004). During the period 2000–2003, the CMS made 10% of capitation payments depend on inpatient claims data using the PIP-DCG risk adjustmentmodel,butthefractionofvariationsinexpectedhealthcostsbythenewersystem remainedaround1.5%(Brownetal.2014). In 2004, the CMS introduced a more comprehensive risk adjustment model called the hierarchical conditional categories (HCC) model in order to reduce incentives for risk selection. The HCC model uses inpatient and outpatient claims to predict the following year’s medical expenditures. Based on this prediction, the CMS calculates an individual’s risk score with a higher score for a greater expected health expenditure. Andanindividual’sriskscore,togetherwiththecapitationbenchmarkfortheindividual’s county of residence, eventually determines the amount of capitation payment an MA insurer receives for enrolling an individual. Brown et al. (2014) find that the new HCC model reduced the returns from enrolling individuals with low risk scores. Even with the HCC model, however, enrolling a low-risk-score individual was still more profitablethan ahigh-risk-score individual. They alsofindthat MAinsurers werestill abletorisk-selectindividualswhowerehealthyindimensionsthatarenotcapturedby riskscoresintheHCCmodel. 3 Data and Preliminary Analysis 3.1 Data This paper combines data from multiple sources. We use the Medicare Current Beneficiary Survey (MCBS) for the years 2001–2005 for individual-level information on MA enrollment and demographic characteristics, including health status. Our data on massadvertising,throughmediasuchasTV,newspaperandradiobyhealthinsurersin localadvertisingmarketsfortheyears2001–2005wereretrievedfromtheAdSpender Database of Kantar Media, a leading market research firm. We obtain data for direct mail advertising for the years 2001–2005 from Mintel Comperemedia (hereafter “Mintel"). Market share data for the years 2001–2005 are taken from the CMS State- County-Plan(SCP)files,andinsurers’plancharacteristicsaretakenfromtheMedicare 8

Comparedatabasesfortheyears2001–2005.10 Although we use data sets from relatively old periods, our sample periods are highly suitable for the purposes of this paper for two reasons. First, the MCBS does not provide information on an individual’s choice of MA insurer from 2006 onward. Without this information, it would be difficult to identify how advertising affects the demand of individuals with different health types. Second, the CMS introduced more sophisticated risk adjustment of capitation payments from 2004 and on, which exogenously changed an MA insurer’s profits from enrolling individuals with different healthtypes. Thechangewaslikelytocreateincentivestotargetdifferenttypesofconsumers,whichallowsustoinvestigatehowinsurersrespondedtothispolicychangein termsofthetargetingofadvertising. 3.1.1 Individual-LevelData The MCBS is a survey of a nationally representative sample of Medicare beneficiaries, which contains information for about 15,000 Medicare beneficiaries every year. The survey is a rotating panel that tracks a Medicare beneficiary for up to four years. Thisdatasetprovidesinformationonabeneficiary’sdemographicinformationsuchas health status, age, income, education and location. An important feature of this data set is that it is linked to Medicare administrative data, which provide information on an individual’s MA insurer choice, the amount of the capitation payment paid for an MA enrollee in the sample, and the amount of Medicare claims costs for individuals intraditionalMedicare. Forouranalysis,weselectoursampleusingfourcriteria. First,weonlykeepindividuals who are eligible for Medicare solely because of their ages. Thus, we exclude individuals who are younger than 65 or who are on Medicaid.11 Second, we exclude individuals who reside in institutions such as nursing homes. We imposed the first andsecondcriteriabecausewewantedtohaverelativelyhomogeneousindividualsfor predicting an amount of capitation payment for each individual. As mentioned above, the capitation payment for an individual depends on whether he is eligible for Medicaid and whether he resides in an institution. Third, we exclude individuals whose 10WethankKathleenNosalforgenerouslysharingMedicareComparedatawithus. 11Tobeprecise,thefirstsamplecriterionalsoexcludesindividualswhoareeligibleforbothMedicare andMedicaid. 9

insurance choices last year are not available in the data.12 We have the third criterion because switching cost is found to be very important in the MA market. Although individuals who just started to receive Medicare benefits do not have a choice made last year, we still include these individuals in the final sample because we do not have any missing information about them.13 Lastly, we exclude individuals from counties wheretherewasnoavailableMAinsurers;thesearelikelytoberuralcounties. 3.1.2 AdvertisingData Mass Advertising AdSpender contains information on the annual expenditures of mass advertising by health insurers in different media such as TV, newspaper, and radio in the 100 largest local advertising markets in the United States.14 A local advertising market consists of a major city and its surrounding counties, and its size is comparable to that of a Metropolitan Statistical Area (MSA).15 AdSpender categorizes advertising across product types whenever specific product information can be detected in an advertisement, which allows us to isolate advertising expenditures for aninsurer’sMAplan. Weusethetotaladvertisingexpenditurebyaninsurerinalocal advertisingmarketasameasureoftheinsurer’sadvertisingactivityinthemarket.16 Direct Mail Advertising Mintel Comperemedia (Mintel henceforth) is a database tracking direct mail advertising in the United States. In each month, the database collects direct mailings from nationally representative households throughout the United States. These households are asked to collect and return mailings in the eight sectors monitored by Mintel, which include health insurance. The Mintel data contain infor- 12BecausetheMCBSisarotatingpaneldataset,everyindividualinthedatasetisnotsurveyedfrom thepointatwhichheorshebecomeseligibleforMedicare. Inthefirstyearanindividualissurveyed by the MCBS, we would not be able to know the individual’s choice last year, so we exclude this observationfromourfinalsample. WearestillabletoobservewhichplananindividualintheMCBS from2001hadintheyear2000becausewedohaveaccesstotheMCBSfrom2000. 13Theseindividualsaremostlikelytobe65or66yearsoldwhenfirstsurveyedbytheMCBS. 14Giventhedataperiodsofourdata,theInternetwasnotamajorchannelofadvertisingatleastfor MAinsurers. 15Intheadvertisingindustry,thislocalmarketisusuallyreferredtoasaDesignatedMediaMarket, whichisdefinedbytheNielsenCompany. 16Wedidnotuseadvertisingexpendituresindifferentmediaseparatelysinceitwouldbedifficultto estimate the effects of advertising in different media on demand separately because of a high positive correlationbetweenexpendituresindifferentmedia. 10

mation on each mailing such as the advertiser and product name, which allows us to tell whether a mailing is advertising an MA plan. Moreover, the data also provide informationofdemographiccharacteristicsoftherecipientofeachmailingsuchasages of household heads, household income, zip code, and so on. Based on the income measure provided in the Mintel data, we also created a new income variable using the five categories that were used to create a new income variable for individuals in the MCBS.Forouranalysis,weexcludedindividualsfromcountieswherethereisnoMA insurer available. Moreover, we selected households with at least one household head whoisatleast64.17 3.1.3 Firm-andMarket-LevelData The Medicare Compare Database is released each year to inform Medicare beneficiaries which private insurers are operating in their county, what plans they offer, and what benefits and costs are associated with each plan. We take a variety of plan benefit characteristics from the data, such as premiums, dental coverage, vision coverage, prescription drug coverage, and the copayments associated with primary care doctor visitsandspecialistvisits,skillednursingfacilitystays,andinpatienthospitalstays. The CMS State-County-Plan (SCP) files provide the number of Medicare beneficiaries and number of enrollees for each MA insurer. A problem with this data set is that although many insurers offer multiple plans in the same county, the aggregate enrollmentinformationisattheinsurer-county-yearlevel,notattheplan-insurer-countyyear level. We deal with this issue by taking the base plan of each MA insurer as a representativeplanbecausethebaseplanisusuallythemostpopular. Asaresult,each MAinsurerwillhaveonlyonerepresentativeplanavailableineachcountyinanalysis. In addition, we also use information on the county-year-level capitation benchmark and the county-year-level per-capita Medicare reimbursement cost for individuals in traditional Medicare, which are available from the CMS website. The capitation benchmark determines the overall amount of capitation payment for enrolling an individual in a county and in a year, and the benchmark for a county-year pair is approximately the capitation payment for an individual with the average health status 17Wechoseage64asthethresholdbecauseanindividualcanenrollinMAthreemonthsbeforethey turn65. Thus,MAinsurersarelikelytosenddirect-marketingmailto64-year-oldindividualsaswell astoolderindividuals. 11

Table1: SummaryStatisticsatCounty-YearLevel NoMassAd SmallMassAd LargeMassAd TotalAnnualMassAdvertisingExpenditure($1,000) 0 36.4 642 MATake-upRates(%) 8.84 16.7 20.3 CapitationBenchmarkperMonth($) 550 560 611 Per-CapitaperMonthMedicareReimbursementCost($) 476 474 547 NumberofMedicareBeneficiaries 41,906 51,886 115,785 AverageMonthlyPremium($) 45.7 37.7 31.9 NumberofInsurers 1.64 2.37 3.48 NumberofCounty-YearPairs 877 452 457 NumberofInsurer-County-YearCombinations 1434 1069 1589 inthecounty-year.18 3.2 Summary Statistics Table 1 presents summary statistics at the county-year level conditional on total mass advertising expenditures for each market. The first column displays summary statistics for county-year combinations without any mass advertising for MA plans. The second and third columns display summary statistics for county-year combinations with relatively small and large total mass advertising expenditures for MA plans. A county-year’s total advertising expenditure is small (large) if it is below (above) the medianoftotaladvertisingexpendituresacrosscounty-yearcombinations. We find that MA take-up rates are larger in markets with more mass advertising expenditures. The county-year combinations with large advertising tend to be larger intermsofmarketsize(i.e.,thenumberofMedicarebeneficiariesinamarket). These county-year combinations also have a higher capitation benchmark, but these markets alsotendtohavehigherhealthcarecostsmeasuredbyhigherper-capitareimbursement costs for traditional Medicare. Moreover, county-year combinations with relatively largemassadvertisingexpenditurestendtohavemoreinsurersinamarket. MAplans inthesecounty-yearcombinationstendtohavelowerpremiums. Table 2 presents summary statistics of individuals in the MCBS conditional on in- 18Theactualamountofthecapitationpaymentforanindividualistheproductoftheindividual’srisk scoreandthecapitationbenchmarkfortheindividual’scountyinayear. Becausetheaverageriskscore isnormalizedtoonefortheoverallMedicarepopulation,thecapitationbenchmarkforacounty-yearis approximatelythecapitationpaymentforanindividualwithaveragehealthstatusinthecounty-year. 12

surancestatus. Thefirstandsecondcolumnspresentsummarystatisticsofindividuals that chose traditional Medicare and MA, respectively. We find that a majority of individualsdonotswitchbetweenthetraditionalMedicareandMA.Forthosewhochoose thetraditionalMedicare,morethan90%chosetraditionalMedicarelastyear,although only70%ofoverallMedicarebeneficiarieschosetraditionalMedicarelastyear. Likewise,about85%ofthosewhochooseMAthisyearalsochoseMAlastyear,although only 20% of the overall Medicare beneficiaries had MA last year. Also, we find that healthandincomestatusofMAenrolleesaredifferentfromthoseattraditionalMedicare. We construct a binary health status, healthy or unhealthy, based on self-reported healthstatus.19 Ourincomemeasureisconstructedasafive-levelcategoricalvariable, with five being the category for the highest income, based on the income variable in theMCBS.20 WefindthathealthyindividualsaremorelikelytochooseMA,whichis consistent with the findings of previous research on MA, as mentioned earlier. Moreover, we find that those who choose MA are more likely to have lower income and be female, although the average ages between the two groups of individuals are not very different. Table 3 presents summary statistics from Mintel. In this data set, the unit of observation is a combination of individual and month, meaning that an individual received 0.158 mailings from MA plans on average. Conditional on receiving at least one MA-related mailing, an individual received 1.24 mailings on average. We find thatthosewhoreceivedmailingstendtohavelowerhouseholdincomeandalsoreside in neighborhoods with lower average income (measured by zip-code-level).21 Those who received mailings tend to be older than those who did not. Moreover, individuals inmarketswithmoreMedicarebeneficiariesaremorelikelytoreceivemailings. 19Anindividual’shealthstatusisdefinedtobehealthyiftheself-reportedhealthstatusis“Excellent,” “Very Good,” or “Good.” An individual’s health status is defined to be unhealthy if the self-reported healthstatusis“Fair”or“Poor.” 20Although MCBS income variable has eleven categories originally, we create a new variable with five categories in order for the income measure in the MCBS to be compatible with the income measure in the Mintel data. Eventually, the new income variable we create is equal to one, two, three, four, orfiveifanindividual’sincomebelongstothefollowingfiveintervals, respectively: [0,15000), [15000,25000),[25000,35000),[35000,50000),and[50000,∞). Henceforth,whenwerefertoanindividual’sincomeintheMCBS,werefertothenewincomevariablewiththefivecategories. 21We obtain the zip-code-level mean income from the IRS,which is available at www.irs.gov/uac/SOI-Tax-Stats-Individual-Income-Tax-Statistics-Zip-Code-Data-(SOI). 13

Table2: SummaryStatisticsatIndividualLevel TraditionalMedicare(TM) MA Overall ChoseTMLastYear(%) 90.1 4.24 70.9 ChoseMALastYear(%) 1.51 85.7 20.3 NewMedicareBeneficiary(%) 7.04 5.66 6.73 Healthy(%) 83.0 84.7 83.4 Age 75.5 75.2 75.5 Income=1(%)(lowest) 2.0 1.9 2.0 Income=2(%) 30.2 37.8 31.9 Income=3(%) 32.1 35.2 32.8 Income=4(%) 18.7 15.6 18.0 Income=5(%)(highest) 17.0 9.5 15.3 Observations 16725 4986 21711 3.3 Preliminary Analysis Now, we provide evidence that advertising is related to incentives for risk selection. Asmentionedearlier,manypreviousworksfindthatanimperfectriskadjustmentprovides incentives for risk selection. Even after a more comprehensive risk adjustment regime was introduced in 2004, Brown et al. (2014) find that the new risk adjustment regime still did not account for Medicare costs for unhealthy individuals. According to their calculation, the capitation payments are estimated to be larger than their expectedMedicarecostsfor77%ofindividualsbeforeandafterthenewriskadjustment regime. In the Appendix A.1, we also show using our data that MA insurers indeed haveincentivesforriskselectionbeforeandafterthenewriskadjustmentregime. MassAdvertising Weprovideevidencethatinsurerstargettheirmassadvertisingto geographic markets with larger potential profits. The design of capitation payments duringourdataperiodprovidesatleasttworeasonsforMAinsurerstotargetadvertising at certain areas. First, there is substantial variation in the differences between the capitation payment and the expected medical cost in each market. Therefore, insurers will have more incentives to advertise in markets with larger differences, which will resultingreaterpotentialprofits. Figure1illustratestherelationshipbetweenthecapitation payment benchmark and the per-capita Medicare reimbursement cost for each 14

Table3: MintelSummaryStats Householdsw/oMAMails Householdsw/MAMails Overall NumberofMAMailings 0 1.24 0.16 Income=1(%)(lowest) 17.0 20.7 17.4 Income=2(%) 16.3 20.5 16.8 Income=3(%) 15.6 16.7 15.8 Income=4(%) 16.1 15.7 16.0 Income=5(%)(highest) 35.0 26.5 33.9 Zipcode-LevelIncome($) 48,662 47,381 48,500 AgeofFemaleHouseholdHeadifAny 67.7 71.3 68.2 AgeofMaleHouseholdHeadifAny 69.4 72.5 69.8 NumberofMedicareBeneficiaries(CountyLevel) 163,725 219,626 170,849 Observations 14,515 2,120 16,635 county-year.22 AlthoughthecapitationpaymentandtheMedicarereimbursementcost are positively correlated, there is still substantial heterogeneity among capitation paymentsconditionalonper-capitalMedicarecosts.23 Second, risk selection is potentially more profitable in some markets than others. Differencesinhealthcarecostsbetweenhealthyandunhealthyindividualswillbetypically greater in regions where health care is more expensive.24 Then an insurer will make a greater profit (loss) by enrolling healthy (unhealthy) individuals in a region withmoreexpensivehealthcare. Althoughwedonothavetheexactmeasureofhealth care prices in different regions, we have information on the per-capita Medicare reimbursement cost of each county-year, which should reflect the health care price of each county-year.25 22An important caveat is that the Medicare reimbursement cost is the health care cost only for individualswhochoosetraditionalMedicareandmayimperfectlyreflectanMAinsurer’sexpectedcost ineachcounty. However,theMedicarereimbursementwillstillprovideusefulinformationabouthow healthcarecostsvaryacrossregions. 23Onesourceofsuchvariationisbasedoncitysize: metropolitanareaswithapopulationof250,000 or more have receive an additional capitation payment that is approximately 10.5% of the premium, whichisnotavailabletoMSAsbelowthisthreshold(Dugganetal.2014). 24Anextremeexampleisahypotheticalcaseinwhichahealthyindividual’scostiszero. Inthiscase, allofthemedicalexpendituresresultfromunhealthyindividuals. Becausehealthyindividuals’health carecostisalwayszero,differencesinhealthcarecostsbetweenhealthyandunhealthyindividualswill begreaterinregionswherehealthcareismoreexpensive. 25AsreportedinTable16oftheAppendix,wefindthatanunhealthyindividualinacounty-yearwith agreaterper-capitaMedicarereimbursementcosttendstoincuragreaterMedicarereimbursementcost, comparedwithahealthyindividualinthesamecounty-year. 15

Figure 1: Cross-Market Differences between Capitation Payments and Health Care Costs )$( tnemyaP noitatipaC 0001 008 006 004 002 200 400 600 800 1000 Medicare Fee−for−Service Cost ($) Capitation Payment ($) Medicare Fee−for−Service Cost ($) Wenowtestwhetherinsurersrespondtothetwoincentivesintermsofthetargeting ofadvertisingbyestimatingthefollowingequation: ad =AvgProfit β +MC β +X β + f + f +ε , (1) jm(c)t ct 0 ct 1 jct 2 t c jct where c, j, and t represent county, insurer, and year, respectively, m(c) denotes the advertising market to which county c belongs, and ad represents each insurer’s jm(c)t advertisingexpenditure. Becauseanadvertisingmarketincludesmultiplecounties,we assume that each county within the same advertising market has the same advertising exposure. In the right-hand side of the equation, AvgProfit , MC , and X represent averct ct jct agepotentialprofit,per-capitaMedicarecostandothercovariates,respectively. AvgProfit ct is defined as the capitation benchmark minus the per-capita Medicare cost for each county-year, which captures the first incentive. Next, MC is the per-capita Medicare ct reimbursement cost for each county-year, which captures the second incentive. Because we have AvgProfit included in the regression, coefficient β will not capture ct 1 howtheoverallhealthcarecostineachcounty-yearaffectsadvertisingsimplythrough its impacts on the average potential profit in each county-year. In addition, f and f t c denote year and county fixed effects, respectively. Because we have county fixed effectsinourregression,weidentifytheeffectofthepotentialprofitonadvertisingusing within-marketvariations. 16

Table4: GeographicalTargetingofMassAdvertising DependentVariable AdvertisingExpenditure VARIABLES Coefficient Std.Error AveragePotentialProfit(AvgProfitcjt) 0.00123*** (0.000275) Per-CapitaMedicareCost(MCcjt) 0.00115*** (0.000357) Observations 4,092 Note:Inordertosavespace,wedonotreportestimateforallcoefficientshere.Table17intheAppendixprovidethecomplete results. Table 4 shows the regression result. We find that the estimates of the coefficients of potential profit and per-capita Medicare costs are both positive, which is consistent with our hypothesis that both higher average profits and higher profitability from healthyindividualsineachmarketsleadtomoreadvertising. Althoughwehaveyetto show direct evidence on how advertising can achieve risk selection, if insurers can attracthealthyindividualswithadvertising,theywillhavegreaterincentivestoadvertise more in a market where attracting a healthy individual results in greater profit. In the following sections, we will provide the evidence from our structural demand model thatadvertisingtendstoattracthealthytypesmorethanunhealthytypes. Direct Mail Although we find evidence that mass advertising is targeted based on theprofitabilityofeachcounty,insurersmayfurtherimplementsophisticatedtargeting within a county. To pursue this possibility, we investigate the second measure of advertising: directmailadvertising. Webelievethatdirectmailingsareveryusefultools from an insurer’s perspective for targeting its advertising at an individual with certain characteristics. Presumably, insurers often have access to the demographic characteristics of individuals who live at specific addresses or have access to information about the average demographic in a small geographic area such as zip code. Therefore, they may utilize sophisticated targeting to attract less costly customers. By using this data set,wecangaininsightsintowhichindividualsaremorelikelytoreceiveadvertising. We first investigate whether the targeting of direct mailings responded to the introduction of the comprehensive risk adjustment in 2004. As discussed earlier, Brown et al. (2014) find that capitation payments for individuals with lower risk scores substantially decreased after the new risk adjustment regime. Thus, although enrolling a healthy individual continues to be profitable to in the new regime, profitability from an individual with a lower risk score likely decreased compared with that from an in- 17

dividual with a higher risk score. The targeting of direct mailings was then likely to changewiththeintroductionofthenewregime. OnelimitationoftheMinteldataisthatwedonotobservehealth-relatedmeasures for individuals. Thus, we use a household’s income as a proxy for the risk scores of the household’s heads, which is motivated by the fact that an individual’s health and income are highly negatively correlated. We use two different measures for income. In the first specification, we use an individual’s income reported in the Mintel data, whichisacategoricalvariablewithfivecategoriesasmentionedbefore. Inthesecond specification,weusetheaverageincomeinanindividual’szipcode. Withthefirstspecification,werunthefollowingregression: 4 4 y =α +∑α 1[I =k]+∑α 1[t≥Oct2003]1[I =k]+X β+f +f +ε (2) it 0 1,k it 2,k it it t c(i),risk(t) it k=1 k=1 where y is the number of MA-related direct mailings that household i received in a it particular month-year t, I is a categorical variable for a household income measure, it which takes a higher value if an income is higher, and 1[I =k] is a dummy variable it thatisequaltooneifI isequaltok. Asmentionedearlier,I hasfivecategoriesfrom it it one to five, with a higher number assigned for a greater income. In (2), we normalize coefficients for the highest income to zero. That is, α = α = 0. Similarly, 1,5 2,5 1[t ≥Oct,2003]isadummyvariablethatisequaltooneforatimeinorafterOctober 2003. We chose the beginning of the fourth quarter of 2003 as the time when the new risk adjustment regime starts to affect an MA insurer’s targeting. Because its implementation was announced in March 2003, MA insurers likely adjusted their targeting even before the beginning of 2004. Moreover, X is a vector of other characteristics it of a household i, including whether there is a male or female household head, ages of male and female household heads if they exist, potential average profit as defined in equation (1) for each county-year, the number of Medicare beneficiaries in each county-year, and median household income for each county-year. Next, f represent t fixedeffectsformonth-yeart. Inaddition, f representfixedeffectsforacomc(i),risk(t) binationofhouseholdi’scountyofresidenceandriskadjustmentregime. Asdiscussed before, if t <Oct2003, then the time belongs to the old risk adjustment regime. And if t ≥Oct2003, then the time belongs to the new risk adjustment regime. Thus, each countyhastwofixedeffectsinthisregression. 18

In(2),ourmaincoefficientsofinterestareα fork=1,···,4.Thismeasureshow 2,k the change in risk adjustment in 2004 affected an insurer’s incentives to target householdswithdifferentincomes,relativetothepre-2004period. Becauseα =0bynor- 2,5 malization, coefficient α for k=1,···,4 measures how many mailings a household 2,k whoseI isequaltokreceived,comparedwithahouseholdwhoseI isequalto5(i.e., it it the highest income category group) after the new risk adjustment regime. Note that because of the fixed effects included in the regression, we are not relying on a crosscounty variation, meaning that identification of α does not come from cross-county 2k variationinpotentialprofits. Instead,theidentificationuseswithin-countyvariationin incentivestotargetdifferentindividualsbeforeandafterthepolicychange. A legitimate concern about using household income as a proxy for health risk is that income may be correlated with other unobserved heterogeneity that can have an impact on a household’s medical expenditures. This is important because an insurer’s profit will eventually depend on medical expenditures instead of health status itself. Forexample,anindividualwithahigherincomemayhaveahigherwillingnesstopay for medical care, which may result in a greater medical expenditure. Therefore, coefficientestimatesα fork=1,···,4willnotprovidegoodinformationaboutwhether 1,k MAinsurerstargethealthyindividuals. However,weareinterestedinrelativechanges in targeting induced by the policy change, which are captured by α . As long as the 2k relationshipbetweentheunobservedheterogeneityandincomedoesnotchangeatthe time when the new risk adjustment design was introduced, the concern will not apply toα . 2k Withthesecondspecification,weestimatethefollowingequation: y =α +α I +α 1[t ≥Oct,2003]I +X β+ f + f +ε (3) it 0 1,zip zip(i),t 2,zip zip(i),t it t c(i),risk(t) it where I represents the average income in the zip code of individual i’s address zip(i),t at timet. Here, the coefficient of interest is α . The concern about the unobserved 1,zip heterogeneity also applies to this specification as well and can be addressed with the sameargumentputforthinthepreviousparagraph. The results are summarized in columns (1) and (2) in Table 5, which present the results with household income and zip-code income, respectively. The results show that lower-income households are more likely to receive advertising after the new risk 19

Table5: TargetingwithDirectMailAdvertising (1) (2) (3) (4) Dependentvariable: #ofMAmails Dependentvariable: SwitchestoMA I -0.000105 I -0.000126* zip(i),t zip(i),t 1[t≥Oct,2003]I -0.000679** 1[t≥2004]I -2.44e-05 zip(i),t zip(i),t 1[Iit=1](lowest) 0.00326 1[Iit=1](lowest) 0.00965 1[Iit=2] 0.00906 1[Iit=2] 0.0262*** 1[Iit=3] -0.00451 1[Iit=3] 0.0223*** 1[Iit=4](2ndhighest) -0.0117 1[Iit=4](2ndhighest) 0.0155** 1[t≥Oct,2003]1[Iit=1] 0.0433* 1[t≥2004]1[Iit=1] -0.0118 1[t≥Oct,2003]1[Iit=2] 0.0177 1[t≥2004]1[Iit=2] -0.00398 1[t≥Oct,2003]1[Iit=3] 0.0857*** 1[t≥2004]1[Iit=3] -0.00201 1[t≥Oct,2003]1[Iit=4] 0.0632** 1[t≥2004]1[Iit=4] -0.00315 OtherCovariates Yes Yes OtherCovariates Yes Yes County-RiskAdjustmentRegimeFE Yes Yes County-RiskAdjustmentRegimeFE Yes Yes Year-MonthFE Yes Yes Year-MonthFE Yes Yes Observations 13,430 13,317 Observations 21,836 21,448 DataSource Mintel DataSource MCBS adjustment regime in both specifications. In the first specification, we find that the number of mailings will increase the most under the new regime for households with incomesthatarenottoolowortoohigh,whichisconsistentwiththepreviousfinding that it is still unprofitable to enroll individuals with very high risk scores. When a zip-code income is used, we find that insurers tend to send more mailings to a lowerincomeneighborhoodunderthenewregime. Moreover,wedonotfindanystatistically significantpatternsintargetingbeforethenewregimeineitherspecification. Although we find that insurers target individuals with different characteristics after the new regime, it does not necessarily mean that an individual’s demand for MA responded to the different targeting. Because the Mintel data do not provide any information about an individual’s insurance choice, we cannot directly test whether the changeinthetargetingofdirectmailingsledtoaconsistentchangeindemandforMA. Instead,wetestthehypothesisindirectlyusingtheMCBS.Specifically,weinvestigate whether an individual, with characteristics targeted by MA insurers, is (i) more likely to switch to MA if the individual did not choose MA last year or (ii) more likely to switchtoadifferentMAinsurersiftheindividualchoseanMAinsurerlastyear.26 26Therefore, this approach is similar to that in Brown and Goolsbee (2002), who investigate the impactofInternetaccessonlifeinsuranceenrollment. 20

Now we define y to be a dummy variable that equals one if condition (i) or (ii) it is met. We run regressions similar to equations (2) and (3). Specifications (3) and (4) in Table 5 presents results from the two regressions. Note that none of the estimated coefficients for the interactions between incomes and the new risk adjustment regime are statistically significant. This result implies that direct mail was not very effective ininducingconsumerstoenrollinMAatleastfortheyearsconsideredinouranalysis. Becausethecostofsendingdirectmailingsisverytiny,insurerslikelyrespondedtothe change in the risk adjustment regime, expecting that direct mailings to newly targeted individuals will lead to a greater demand by them. Eventually, however, any changes indemandwerequantitativelyinsignificant. 4 Demand Model We now investigate how advertising affects consumer demand by structurally estimating a model of health insurance demand. Although we provided evidence for the targeting of both direct mail and mass advertising, we only consider the impact of mass advertising on demand. One difficulty of using the data on direct mail in the demand analysis is that it is difficult to link the direct mail data (Mintel) to the data on a consumer’s insurer choice (MCBS). The number of individuals in a county-year in the Mintel data is not large enough to construct a measure of direct mails sent to a county-year. Thus, without combined information on advertising exposure and subsequent choice, it will be difficult to estimate the effects of direct mail on demand for MA.27 Moreover, as shown in the previous section, we do not find evidence that the change in the targeting of direct mail led to a corresponding change in demand for MA. Asdiscussedinaprevioussection,MAinsurerscontractwithCMSforeachcounty (c)ineachyear(t). Asaresult,consumersindifferentcountiesanddifferentyearsface different choice sets. Thus, we will naturally define a market of MA as a combination of county-year (ct). However, each advertising decision is typically made on the basis of a local advertising market (m), which contains several counties. Thus, we assume 27OnepossibilityistoimputethenumberofMAmailingsanindividualreceivesusingcharacteristics presentinbothdatasets. Unlesswecandotheimputationprecisely,theimpactofimputedmailingson demandislikelytobeestimatedwithasubstantialbias. 21

individuals in different c but in the same m are exposed to the same advertising level by the same firm. The advertising marketm, to which countyc belongs, is denoted by m(c). Each MA market (ct) has J MA insurers available. An individual in a market ct also has the option of choosing traditional Medicare. Thus, an individual has the total J +1 options in MA market ct. An insurer j in market ct can be described by a ct combination of advertising (ad ), other observed characteristics (x ) including jm(c)t jct premiumandplancharacteristics,countyfixedeffect(µ ),aninsurer-yearfixedeffect c (ξ ), and an unobservable characteristic (∆ξ ). A consumer i can be described by a jt jct combination of health status (h), last year’s choice of insurer (d ), other observed i i,t−1 characteristics (c ), and a preference shock (ε ). We will explain each insurer’s and it ijct individual’scharacteristicsafterwedescribeanindividual’sutilityfromaninsurer. Consideranindividualilivingincountycandyeart. Consumerichoosestoenroll with one of the available J MA insurers in each c andt or in traditional Medicare. We assumethatconsumeri,livinginacountycinyeart,obtainsindirectutilityu from ijct MAinsurer j asfollows: (cid:0) (cid:1) u =ln 1+ad α +x β +φ 1[d (cid:54)= j,d ≥0]+µ +ξ +∆ξ +ε (4) ijct jm(c)t ijt jct it ict i,t−1 i,t−1 c jt jct ijct where 1 α = α +α 1[d = j]h +∑α 1[d (cid:54)= j,d ≥0]1[h =k]; ijt 0 1 i,t−1 it 2,k i,t−1 i,t−1 it k=0 β = β +β h ; it 0 1 it φ = φ +φ h +φ J +φ J2. ict 0 0 it 1 ct 0 ct A consumer’s outside option is to enroll in traditional Medicare, from which a consumerreceivesutilityofu : i0ct u =h ρ +c ρ +φ 1[d (cid:54)=0,d ≥0]+ε . (5) i0ct it 1 it 2 ict i,t−1 i,t−1 i0ct Bothanindividual’scharacteristicsandaninsurer’scharacteristicsdetermineu . ijct An individual’s characteristics included in u are individual i’s binary health status ijct h that equals to one if healthy (and zero if unhealthy), last year’s insurance choice it d , and other relevant individual characteristics c . Last year’s insurance choice i,t−1 it 22

d contains information about (i) whether individual i chose MA or traditional i,t−1 Medicare last year and (ii) which MA insurer this individual chose if MA was chosenlastyear. IncasethatindividualiisnewtoMedicare,wesetd =−1,andthus i,t−1 1[d (cid:54)= j,d ≥0]=0 for any j for new Medicare beneficiaries.28 Lastly, ε is i,t−1 i,t−1 ijct an individual i’s preference shock for insurer j, which we assume is distributed as the TypeIextremevaluedistribution. Each insurer has observable characteristics (ad and x ), county fixed effect jm(c)t jct (µ ) and an insurer-year fixed effect (ξ ), and an unobservable characteristic (∆ξ ). c jt jct First, ad denotes insurer j’s advertising expenditure in millions in advertising marjmt ket m in year t.29 30 Note that the effects of advertising diminish in its expenditure because ad enters u in logarithm.31 The effect of advertising on indirect utiljm(c)t ijct ity u is captured by α , which depends on individual i’s previous insurance status ijct ijt d and self-reported health status h . In other words, insurer j’s advertising has i,t−1 it different effects, depending on whether individuals chose the insurer last year and whether an individual is healthy. Parameter α represents the effects of advertising 0 thatareindependentofanindividual’scharacteristics. Parameterα representstheef- 1 fects of advertising for healthy consumers who chose the same insurer last year. And α and α capture the effects of advertising on unhealthy and healthy individuals 2,0 2,1 thatdidnotchooseinsurer j lastyear,respectively. We distinguish the effects of advertising on individuals who chose the advertised insurer last year and those who did not because if advertising is informative, it will be more effective for individuals who did not choose the insurer (Ackerberg, 2001). Informative advertising is likely to provide information about an insurer’s unobserved quality or simply the existence of the insurer in the market. Thus, it is plausible that 28WedefineanindividualasnewtoMedicareifheorshehasspentlessthantwoyearsonMedicare asoftheendofyeart. 29Note that advertising affects demand through the indirect utility function in our model. Alternatively,onecanmodelspecificchannelsthroughwhichadvertisingaffectsdemand: forexample,aconsumer’s awareness of a product, providing experience characteristics of product quality, or enhancing prestigeorimageofaproduct. Wedonottakethisapproach,however,becauseseparatelyidentifying differenteffectsofadvertisingischallengingwithourdata. 30This specification assumes no interaction term between advertising and price. We also estimated the version of the model allowing those interactions and also further allow interaction with them to individuallastyear’sinsurancestatus. However,noneofthemarestatisticallysignificantandtherefore wedecidedtodropforthisestimation. Estimatesforthespecificationareavailableonrequest. 31Becausead iszeroformanyinsurers,weuseln (cid:0) 1+ad (cid:1) insteadofln (cid:0) ad (cid:1) . jm(c)t jm(c)t jm(c)t 23

thistypeofadvertisingwillhavelittleeffectsonindividualswhochosetheinsurerlast year. On the other hand, if advertising has prestige or image effects, then it will likely affect both types of individuals. Moreover, advertising can be still informative for an individual who already enrolled with the advertised insurer. Unless an individual receives much medical care, the individual will not be able to know an insurer’s true unobservedquality. Advertisingcanstillprovideinformationtosuchanindividual. Moreover, we allow for the possibility that advertising has a different impact depending on h . If the impact of advertising depends on h , then advertising will evenit it tually affect an insurer’s risk pool and thereby its cost. In this case, advertising can be used for risk selection. In principle, there are two interpretations of the heterogeneous impacts of advertising depending on h : the targeting of mass advertising it at certain types of consumers and a consumer’s differential response to advertising. First, targeting refers to the possibility that an insurer targets its advertising at certain TV programs and newspapers that are more exposed to a certain health type than to another type. Note that this kind of targeting requires an insurer to employ a more sophisticated targeting strategy than targeting certain counties. Second, a consumer’s differential response to advertising refers to the possibility that a certain health type responds to advertising more than another health type. In this case, advertising can stillaffectacertaintype’sdemanddisproportionatelymoreevenwithoutsophisticated targeting. Unfortunately, we cannot clearly distinguish the two different channels becausewedonothaveinformationaboutwhichtypesofconsumerswereexposedtoan MAinsurer’smassadvertising. However, we view that the heterogeneous impacts are likely to capture the second mechanism for the following reasons.32 First, even without sophisticated targeting, health status itself can determine how much an individual is exposed to advertising. For example, mass advertising mostly appears on TV or in newspapers, and those who are able to watch TV or read newspapers are less likely to have their vision or hearingproblems. Wefindthatunhealthyindividualsaremorelikelytohavevisionor 32Incasethattheheterogeneousimpactscaptureaninsurer’stargetingtosomeextent,thenapotential problemisthatparameterα isnotpolicy-invariantforourcounterfactualanalysis. Thatisbecausean ijt insurermaytargetitsadvertisingatadifferenthealthtypewithacounterfactualchangeinitsincentives to attract different health types. In our counterfactual analysis, however, we exogenously shut down advertisinginordertoinvestigatetheimpactofadvertisingontheMAmarket. Inthiscase,parameter α will not play any role in this counterfactual analysis. Thus, results in our counterfactual analysis 1,k willnotdependonwhethertheheterogeneousimpactscapturethetargeting. 24

hearing problems in the MCBS, as shown in Table 18 in the Appendix.33 Moreover, amongthosewhohavesuchproblems,unhealthyindividualsaremorelikelytobelieve that their vision or hearing problems make it difficult for them to obtain information about Medicare, as reported in Table 18 in the Appendix.34 Thus, those who actually respond to advertising will be more likely to be healthy even without sophisticated targeting. Second, as Fang et al. (2008) argue, a health status is highly correlated withcognitionabilitiesforelderlypeople,whichmayleadtoadifferentialresponseto advertising. Third, our preliminary analysis on direct mail advertising reveals that the targetingofadvertisingatcertainindividualswithinamarketwasnotveryeffectivein attracting them to MA. The result indicates that targeting does not necessarily lead to an increase in demand by targeted consumers. Because targeting mass advertising at certainindividualsisplausiblymoredifficultthantargetingviadirectmail,webelieve that it will be difficult for insurers to risk-select through targeting mass advertising at healthytypes. The term x denotes a vector of insurer j’s observed characteristics other than jct advertising, which include the premium, copayments for a variety of medical services suchasinpatientcareandoutpatientdoctorvisits,andvariablesdescribingwhetheran insurer offers drug coverage, vision coverage, dental coverage, and so on. We define the premium to be the amount that a consumer pays in addition to the Medicare Part B premium.35 The effects of plan characteristics on utility are potentially heterogeneous depending on an individual’s health type. For example, an MA insurer offering drugcoveragemaybepreferredbyindividualswhoexpectalargeexpenditureonprescriptiondrugs,andaprivatefee-for-serviceMAinsurermaybepreferredbyacertain healthtypebecauseitsprovidernetworkisnotasrestrictiveasanHMO.Wealsoallow for the possibility that disutility from a premium depends on a healthy type because different health types may have different willingness to pay for MA. The heteroge- 33TheTable18intheAppendixpresentsresultsforregressionsofwhetheranindividualhasavision orhearingproblemonhishealthstatusandage. 34TheTable18intheAppendixalsopresentsresultsforregressionsofwhetheranindividualbelieve thathisvisionorhearingproblemsmakeitdifficulttoobtaininformationaboutMedicareonhishealth statusandage. 35WhenenrollinginanMAplan,anindividualmustpaytheMedicarePartBpremiumaswellasthe premiumchargedbytheplan. HerewedonotincludeMedicarePartBpremiumin p becausealmost jct allMedicarebeneficiaries, whoremainintraditionalMedicare, enrollinMedicarePartBandpaythe MedicarePartBpremium. 25

neous effects are captured by parameter β , which depends on an individual’s health it h .36 it The term φ denotes switching cost of changing insurers. Note that 1[d (cid:54)= ict i,t−1 j,d ≥ 0] is equal to one if an individual, who is not new to Medicare, chooses a i,t−1 different plan from one chosen last year. This means that new Medicare beneficiaries do not pay a switching cost for their initial choice of insurer. Notice that the switching cost makes the impact of advertising on demand depend on d . Because new i,t−1 Medicare beneficiaries do not face a switching cost, advertising will have a larger effect on them. We also allow for the possibility that φ is different, depending on h it it and J (number of available insurers in a market). We let J affect φ because the ct ct ict functional-form assumption for ε mechanically implies that an individual in a marijct ketwithmoreinsurersismorelikelytoswitchtoadifferentplanwithallothersbeing equal. The term ξ denotes insurer-year fixed effects that capture an insurer j’s brand jt effect in year t. Moreover, µ represents county fixed effects, which capture countyc specific factors that determine demand for MA in the county. An individual’s utility also depends on aspects of an insurer that are unobserved by researchers but observed by consumers and insurers. The term ∆ξ is a deviation from µ and ξ , and ∆ξ jct c jt jct captures unobserved characteristics and/or shocks to demand for this insurer. We assumethat∆ξ isknownbyconsumersandinsurerswhentheymakedecisions. jct Lastly, we discussthe specification ofutility for the outside option, which is traditionalMedicare. Notethattheconstanttermforu isnormalizedtozerobecausethe i0ct term cannot be identified in a discrete choice model. All of the terms included in u i0ct are individual characteristics such as health status, switching cost, and other characteristics denoted by c , which include age, income, and interaction between year and it previous insurance status. These individual characteristics capture different utilities from the outside option for individuals with different characteristics, relative to their utilityfromMAinsurersingeneral. 36Inordertoreducethenumberofparameterstobeestimated, wedonotinteracteveryvariablein x withhealthstatus. Weselectwhichvariablestointeractwithhealthstatusbasedontheresultsof jct thepreliminaryanalysis. AcompletelistofvariablesinteractedwithhealthstatusisreportedinTable 6. 26

5 Identification and Estimation For the discussion of identification and estimation of the model, we define θ to be a vector that contains all parameters in the model. For our discussion in this section, let θ ≡(θ ,θ ), where θ is a collection of parameters that determine the parts of utility 0 1 0 independentofindividualheterogeneityandwhereθ isacollectionofparametersthat 1 determine preference heterogeneity resulting from individual characteristics. That is, θ ≡(α ,β ),andθ containsallotherparametersinequations(4)and(5). 0 0 0 1 Mean Utility First, we discuss the identification of parameters in θ . The parts of 0 u inequation(4)thatareindependentofindividualheterogeneityareusuallycalled ijct meanutilityδ . Inotherwords, jct (cid:0) (cid:1) δ ≡ln 1+ad α +x β +ξ +µ +∆ξ . (6) jct jm(c)t 0 jct 0 jt c jct Berryetal.(1995)showthatgivenavalueforθ ,thereisauniqueδ∗ (θ )thatex- 1 jct 1 actly match predicted market shares to observed market shares. Then parameter θ is 0 estimatedusingequation(6)bytreating∆ξ asastructuralerrorterm. Awell-known jct problemregardingtheidentificationofθ isthat∆ξ ,whichmaycaptureunobserved 0 jct productcharacteristics,andendogenousplancharacteristicsincludedinthemodelare correlated. Thisproblemisatypicalendogeneityproblem,andthenasimpleordinaryleast-squaredregressionofδ∗ (θ )on(ad ,x )willresultininconsistentestimates jct 1 jmt jct of θ if (ad ,x ) contains endogenously chosen characteristics. We assume that 0 jm(c)t jct the advertising expenditure ad and the premium p , which is a part of x , are jm(c)t jct jct endogenous variables. Although almost all of the plan characteristics are potentially endogenous, we assume that these characteristics are exogenous in this estimation. A crucialreasonforthisdecisionisthatthenumberofinstrumentsrequiredforconsistent estimation should be at least as great as the number of endogenous variables included in(ad ,x ). Giventhelargenumberofplancharacteristics,itisextremelydiffijm(c)t jct culttocomeupwithinstrumentsforallofthem. Although the endogeneity problem challenges the identification, the fixed effects µ and ξ included in δ is likely to control for a significant part of the unobserved c jt jct heterogeneity of insurers. However, it is still possible that ∆ξ still contains unobjct served characteristics that are varying over insurers, counties and years. A typical 27

approachtoaccountingfortheendogeneityproblemistouseinstrumentsthatarecorrelated with the endogenous variables, but not with the unobservable. We use instruments similar to ones used by Hausman (1996) and Nevo (2001).37 In other words, we use the average advertising expenditures of the same parent companies in other advertising markets for ad and the use the average premium of the same parent jm(c)t company in other counties for p . The instruments capture the idea that an insurer’s jct marginal cost contains a component that is common to all subsidiaries of a parent company,whichisassumedtobeuncorrelatedwiththeunobservedheterogeneity. Resulting moment conditions employed in the estimation are that E[∆ξ |Γ]=0, where jct Γ is a set of instruments that includes the aforementioned two sets of instruments as wellasx . jct Preference Heterogeneity Important information for the identification of parameters for preference heterogeneity θ is an individual’s insurer choice from the MCBS 1 (theindividual-leveldata). Parameterθ willbeidentifiedbyvariationinthecharacter- 1 istics of insurers chosen by individuals having different characteristics. An important parameterinθ aretheparametersthatdeterminetheheterogeneouseffectofadvertis- 1 ing depending on an individual’s health type and last year’s choice, which are α ,α 1 2,0 andα in(4). Theseparameterswillbeidentifiedbyvariationinindividuals’switch- 2,1 ing patterns across health types, last year’s choices, and advertising expenditures by insurerstheyareswitchingto. In order to construct micro-moments for an individual’s choice and combine them with the aggregate moments, we use the score of the log-likelihood function for a choicebyanindividualobservedintheMCBS,asinImbensandLancaster(1994). The likelihood function for an individual’s choice is L=∏ i,j,c,t q jct (z i )d ijct, where q jct (z i ) is the probability that an individual with characteristics z chooses an insurer jct, and i d is an indicator variable that equals one when individual i chooses plan jct. Then ijct ourmicro-momentsare∂log(L)/∂θ =0. 1 37TownandLiu(2003)alsouseasimilarinstrumentinestimatingamodelofdemandforMAplans. 28

Table6: EstimatesforParametersofInterest Variables Estimates Std.Error Variables Estimates Std.Error (cid:0) (cid:1) log 1+adjmt ×1[di,t−1(cid:54)=j,di,t−1≥0]×1[hit=1] 1.449*** (0.612) 1[di,t−1(cid:54)=j,di,t−1≥0] -3.786*** (0.242) (cid:0) (cid:1) log 1+adjmt ×1[di,t−1=j]×hit 0.879* (0.485) 1[di,t−1(cid:54)=j,di,t−1≥0]×hit 0.016 (0.127) (cid:0) (cid:1) log 1+adjmt ×1[di,t−1(cid:54)=j,di,t−1≥0]×1[hit=0] 0.470 (0.467) 1[di,t−1(cid:54)=j,di,t−1≥0]×Jct 0.008 (0.084) log (cid:0) 1+adjmt (cid:1) -0.546** (0.268) 1[di,t−1(cid:54)=j,di,t−1≥0]×J c 2 t -0.007 (0.008) Premium -0.015** (0.006) DrugCoverage 0.358*** (0.088) Premium×hit 3.2e-4 (0.003) DrugCoverage×hit 0.0147 (0.215) 6 Demand Side Estimates Table 6 displays estimates for important parameters in the demand model. The effects of advertising on an individual’s demand is the sum of the common effects (the coef- (cid:0) (cid:1) ficient in front of log 1+ad ) and the heterogeneous effects (the coefficients for jmt interaction terms). Based on the estimates, we find that the effect of advertising on demand is much greater for healthy individuals (h =1), especially for healthy indiit viduals who are switching or new to Medicare (1[d (cid:54)= j,d ≥0]=1). In addii,t−1 i,t−1 tion,theestimatefordisutilityfromapremiumisnegativeandstatisticallysignificant, but the estimate for the interaction between a premium and the dummy variable for the healthy type is not statistically significant. This means that healthy and unhealthy consumersdonothaveverydifferentpricesensitivity. Inordertoputtheestimatesforparametersforadvertisingandpremiumsintoperspective, we calculate the semi-elasticities of demand with respect to advertising and premiums,whicharepresentedinTable7.38 Semi-elasticityofdemandwithrespectto apremiumis-0.847%,whichmeansthatadollarincreaseinaninsurer’spremiumwill leadtoadecreaseindemandfortheinsurerby0.847%. Althoughthesemi-elasticities forthetwodifferenthealthtypesareslightlydifferent,itisunlikelythatthedifference is statistically significant given the imprecise estimate for the coefficient that determines a healthy consumer’s price sensitivity relative to a unhealthy consumer. This findingisconsistentwithCurtoetal.(2014). Fortheeffectofadvertisingondemand,wecalculatethesemi-elasticityofdemand 38Semi-elasticityofdemandQwithrespecttoavariablex isdefinedas ∂Q× 1, whichmeasuresa ∂x Q percentage change in Q with a unit increase in x. We calculate semi-elasticities instead of elasticities becauseanadvertisingexpenditureandapremiumarezerofor68%and37%ofinsurers,respectively. Whenanadvertisingexpenditureiszero,elasticityofdemandwithrespecttoadvertisingbecomeszero. Thesameresultisalsotrueforelasticityofdemandwithrespecttopremiums. 29

Table7: ElasticityofDemandwithRespecttoAdvertisingandPremiums Semi-ElasticitiesofDemand Adv($2,300) Premium($1) OverallSemi-elasticity 0.066% -0.847% Semi-elasticityforhealthy 0.086% -0.851% Semi-elasticityforunhealthy -0.012% -0.943% Note:$2,300=1%ofmeanadvertisingspendingforinsurerswithpositiveamounts. withrespecttoadvertisingforanincreaseof$2,300ofadvertisingexpenditures,which is about 1% of the average advertising expenditure among insurers with positive advertising expenditures. We find that an additional $2,300 in an insurer’s advertising expenditure increases demand for the insurer by 0.066% on average. Semi-elasticities fordifferenthealthtypesshowthattheeffectsofadvertisingaresubstantiallydifferent across different health types. A healthy consumer’s average semi-elasticity is 0.086% whereasanunhealthyconsumer’ssemi-elasticityisclosetozero. Variables other than advertising and premiums are also important in determining demandforanMAinsurer. Wefindthattheswitchingcostisveryimportantinexplaining an individual’s demand, although the cost is not very different across individuals with different health types and in different markets. The important role of the switching cost in our results is consistent with findings by other papers on health insurance markets.39 In addition, the provision of drug coverage has a positive and significant effect on demand, which reflects the fact that during our data period (2001–2005), the drug coverage would not be available if an individual chose traditional Medicare. However, the interaction of drug coverage and the healthy dummy is not significant. Lastly, estimates for all other parameters are reported in Table 19 and 20 in the Appendix. 7 Counterfactual Experiments With the estimated model, we conduct counterfactual analyses to understand the impact of risk selection through advertising on the MA market. In order to quantify the impact of advertising on the MA market, we simulate the model by exogenously setting every insurer’s advertising expenditure to zero. We simulate the model under two 39Forexample,seeHandel(2013),Hoetal.(2015),Miller(2014),Milleretal.(2014),Nosal(2012), andPolyakova(2014). 30

different counterfactual scenarios. In the first scenario, we assume that a premium is fixed exogenously at its level in the baseline. In the second scenario, we assume that insurers can reoptimize their premiums in the counterfactual environment. In this case, we investigate the equilibrium impact of shutting down advertising on market outcomes. Webelievethatmodelingequilibriumpriceresponsesisimportanttobetter understanding advertising as playing a role in risk selection and its implications for consumerdemandandultimatelywelfare. Inordertoobtainitsimpactonequilibrium,wefirstneedtospecifyamodelofhow MA insurers choose their premiums. Therefore, we discuss the model of the supply sidebeforesimulatingourmodel. 7.1 Model of the Supply Side Weassumethatinsurersplayasimultaneousgameinchoosingoptimalpricingineach market (county-year).40 When insuring an individual i with characteristic z, insurer i jct expectstoincuramarginalcostc (z)asfollows: jct i c (z)=E [m(z,x ,ω ;λ)]+η , (7) jct i ω i jct ijct jct where m(z,x ,ω ;λ) is a realized reimbursement cost for an individual with chari jct ijct acteristic z who choose plan jct. The term ω represents a random chock to the i ijct reimbursement cost, and λ represents parameters to be estimated. Next, η is a jct insurer-county-year-specific shock to marginal cost that is constant across individuals having different z. We assume that η is observed by all insurers when making i jct a pricing decision. Note that the expected marginal cost c (z) depends on the conjct i sumer’s characteristic z, which includes health status. Therefore, an insurer’s costs i willeventuallydependonwhatkindsofindividualsareenrolledwiththeinsurer. Weestimatethemarginalcostparametersλ usingtheindividual-levelinformation from the MCBS on how much an individual’s MA insurer paid for the individual’s medical care in a year. Details on the exact functional form of m(z,x ,ω ;λ) and i jct ijct estimationofλ arereportedinAppendixA.2. 40BecausewedonotconsidercounterfactualsituationswhereMAinsurersre-optimizetheiradvertisingexpenditures,wedonotconsidertheoptimaladvertisingdecisionhere. 31

Insurer j’sprofitfromacountycinyeart isgivenby ˆ π =M (p +cp(z)−c (z))q (z)dF (z)−C(ad ), (8) jct ct jct i jct i jct i ct i jm(c)t zi whereM isthepopulationofthosewhoareatleast65yearsoldincountycinyeart, ct and p isthepremiumchargedbyinsurer j incountycinyeart.41 Next,cp(z)isthe jct i expected capitation payment for an individual having characteristics z. We calculate i cp(z) based on the relationship between the observed amount of capitation payment i and z in the MCBS. The details about the calculation can be found in Appendix A.1. i Lastly, q (z) is the probability of choosing insurer j by an individual having charjct i acteristics z. Lastly, C(ad ) denotes the advertising cost for each firm, which i jm(c)t capturesboththevariableandfixedcostsassociatedwithad . jct Withtheprofitequation,itisclearhowriskadjustmentandadvertisingaffectprofits. With a perfect risk adjustment of capitation payment, cp(z)−c (z) is constant i jct i across z. In this case, a healthy individual will not cost less than an unhealthy indii vidual, and advertising will affect an insurer’s profit just by increasing the overall de- ´ mand for the insurer q (z)dF (z). With an imperfect risk adjustment, in contrast, z jct i ct cp(z)−c (z) will depend on z and will be typically larger for healthy individuals, i jct i i whichisthecasefortheMAmarket. Inthiscase,advertisingaffectsaninsurer’sprofit ´ throughaninsurer’scost c (z)q (z)aswellasthroughtheoveralldemand. z jct i jct i The Nash equilibrium condition for the optimal pricing game for insurers is that insurers’ choices maximize their profits given choices made by other insurers. Thus, we have the following condition for each p jct such that ∂π jct /∂p jct =0. We can solve for η in a way that is similar to Berry et al. (1995). Appendix A.2 provides details jct onhowwesolveforη . jct 41Oneimportantassumptionwemadeisthatfirmsaremyopic. Withanindividual’sswitchingcost, anMAinsurerpotentiallyhasadynamicpricingincentive. Miller(2014)isthefirstattempttoestimate a model with forward-looking insurers in the MA market. One alternative is to follow the approach by Decarolis et al. (2015), who also estimate an equilibrium insurance market model with switching costs and myopic firms. In order to correct for possible bias resulting from ignoring dynamic pricing incentives,theydoarobustnesscheckonwhethertheestimatesofmarginalcostarebiasedatacertain level. Fully characterizing the dynamic pricing decision is a very challenging task and left to future work. 32

Table8: Counterfactual: IndividualDemand MarketswithSmallAdv MarketswithLargeAdv MarketswithAnyAd HealthType Baseline PartialEq FullEq Baseline PartialEq FullEq Baseline PartialEq FullEq Panel1:ConsumersThatAreNewtoMedicare:Pr(SwitchingtoMA) Healthy 0.177 0.175 0.171 0.228 0.205 0.199 0.208 0.193 0.188 Unhealthy 0.196 0.196 0.190 0.212 0.212 0.206 0.207 0.207 0.201 Panel2:ConsumersThatChoseTraditionalMedicareLastYear:Pr(SwitchingtoMA) Healthy 0.0170 0.0167 0.0162 0.0174 0.0151 0.0141 0.0172 0.0158 0.0150 Unhealthy 0.0148 0.0148 0.0144 0.0133 0.0133 0.0125 0.0139 0.0139 0.0133 Panel3:ConsumersThatChoseaMAPlanLastYear:Pr(SwitchingtodifferentMA) Healthy 0.0368 0.0365 0.0365 0.0716 0.0679 0.0706 0.0605 0.0578 0.0597 Unhealthy 0.0351 0.0349 0.0348 0.0709 0.0667 0.0702 0.0596 0.0567 0.0591 Note:MarketswithSmallAdvreferstoasetofmarketswheremarket-leveltotaladvertisingexpendituresarebelowthemedianof market-leveltotaladvertisingexpenditures;MarketswithLargeAdvreferstoasetofmarketswheremarket-leveltotaladvertising expendituresareabovethemedianofmarket-leveltotaladvertisingexpenditures. 7.2 Simulation Results We now evaluate the effects of shutting down advertising on the MA market under two different scenarios. First, we assume that a premium is fixed at its baseline level exogenously. Second,weassumethatinsurerscanreoptimizetheirpremiums. Henceforth, we will refer to the first and second counterfactual scenarios as the “Partial Eq” andthe“FullEq”counterfactual. Table8summarizestheeffectsofshuttingdownadvertisingonaconsumer’sswitching patterns, depending on a consumer’s insurance choice last year. For each group of consumers, we calculate the effects on demand separately for markets with different levels of advertising expenditures. First, we compare results in the baseline and those inthecounterfactualwherepremiumsarefixed,whicharepresentedunderthecolumns labeled“PartialEq.” Asexpected,theprobabilityofswitchingtoMAwillbelowerin the counterfactual situation than in the baseline, regardless of a consumer’s insurance status last year. Moreover, the probability of switching to MA is much greater for a new Medicare beneficiary because a new Medicare beneficiary does not face switching costs. Therefore, the effect of advertising on demand is much greater for a new Medicare beneficiary in terms of an absolute change in probabilities of switching to MA. This indicates that it is important to look at flows instead of stocks in order to understand the effect of advertising on demand. We also find that the decrease in the 33

probabilities will be greater in markets with larger advertising expenditures. This result shows that an insurer’s geographical targeting of advertising plays an important roleinexplainingcross-marketdifferencesindemandforMA. Next,weinvestigatetheimpactofadvertisingondemandinthe“FullEq”counterfactual. Compared with the results in the other counterfactual situation where premiumsarefixed,wefindmoresubstantialdeclinesinoverallprobabilitiesofswitchingto MAforindividualsthatnewtoMedicareandthosewhochosetraditionalMAlastyear. Incontrast,theprobabilityofswitchingforthosewhochoseaMAinsurerlastyearin the“FullEq”counterfactualisgreater,comparedwiththe“PartialEq”counterfactual. In order to understand the difference between the results in the two counterfactual situations,weanalyzehownewequilibriumpremiumsinthe“FullEq”counterfactual aredifferentfromthepremiumsinthebaseline. Table9reportsequilibriumpremiums and market shares in different counterfactual situations. We report the results for two groupsofinsurers,dependingonwhethertheyadvertiseatallinthebaselineeconomy. First, insurers with positive baseline advertising expenditures will increase premium in the “Full Eq” counterfactual. We find that the increase in the average premium will be much larger in the markets with relatively large baseline advertising expenditures than for the markets with relatively small baseline expenditures. In the former groupofmarkets,insurerswithpositivebaselineadvertisingexpenditurewillincrease monthly premiums by about 40% from $20.6 to $28.8 (or from $247.2 to $345.6 annually). Such a large increase in premiums will keep individuals who did not choose MA last year from switching to MA in the “Full Eq” counterfactual. Second, insurers with zero baseline advertising expenditures will decease their monthly premiums by about 19% from $18.6 to $15.0 (or from $223.2 to $180.0 annually). Because of the premium decrease, individuals who chose MA last year will be more likely to switch to a different MA insurer in the “Full Eq” counterfactual, compared with the “Partial Eq” counterfactual. Overall, the average premium across all insurers increase in both groupofmarkets. Inmarketswithlargeradvertisingexpenditures,theaveragemonthly premiumincreasesbyabout11%from$19.5to$21.6(orfrom$234.0to$259.2annually). Inmarketswithsmalleradvertisingexpenditures,theaveragemonthlypremium increasesbyabout2%from$41.5to$42.5(orfrom$498to$510annually). The changes in market shares reported in Table 9 are consistent with the changes in premiums. Compared with the predicted changes in premiums, however, market 34

Table9: Counterfactual: Market-LevelOutcomes MarketswithSmallAd MarketswithLargeAd Baseline PartialEq FullEq Baseline PartialEq FullEq Insurerswith AverageMonthlyPremium($) 37.1 37.1 39.3 20.6 20.6 28.8 PositiveAd AverageMarketShare(%) 0.099 0.098 0.097 0.078 0.076 0.074 Insurerswith AverageMonthlyPremium($) 45.1 45.1 45.1 18.6 18.6 15.0 ZeroAd AverageMarketShare(%) 0.047 0.047 0.047 0.027 0.027 0.028 AverageMonthlyPremium($) 41.5 41.5 42.5 19.5 19.5 21.6 AllInsurers AverageMarketShare(%) 0.070 0.070 0.069 0.052 0.051 0.050 Note:MarketswithSmallAdvreferstoasetofmarketswheremarket-leveltotaladvertisingexpendituresarebelowthemedianof market-leveltotaladvertisingexpenditures;MarketswithLargeAdvreferstoasetofmarketswheremarket-leveltotaladvertising expendituresareabovethemedianofmarket-leveltotaladvertisingexpenditures. shares will not decrease as much. The main reasons for this result for market shares are because a market share is a stock, as opposed to a flow, and because we calculate changes in market shares in a single year. Although advertising has a large impact on a new Medicare beneficiary’s transition to MA as reported in Table 8, new Medicare beneficiaries are only a small fraction of the entire Medicare beneficiaries, most of whomfacelargeswitchingcost. Asaresult,theeffectofadvertisingonmarketshares inasingleyearwillbelimitedinbothcounterfactualsituations. Shutting down advertising has qualitatively different effects on premiums dependingonwhetheraninsurerhadapositivebaselineadvertisingexpenditure. First,insurerswithpositivebaselineadvertisingexpenditureswillincreasepremiumsbecausethe fractionofunhealthyenrolleeswithsuchinsurerswillincreasewithoutadvertising. In contrast, the other type of insurers will decrease premium because more healthy individuals enroll with them. In the baseline, insurers that advertised took away healthy consumers at the expense of insurers with advertising. Therefore, shutting down the advertising leads to transfer of risk pools across insurers, which highlights a rentseekingaspectofriskselection. Theresultimpliesthatinsurerscanpotentiallyengage in a wasteful advertising competition in order to take away healthy consumers. However, about 68% of insurers, who are mostly small in terms of market shares, did not haveadvertisingexpendituresinthedata. Itindicatesthattheseinsurersfacehighfixed costsofadvertising. Thus,theextentofthewastefulcompetitionwaslikelytolimited. It is important to recognize that changes in risk pools of insurers can impact premiums only when risk adjustment is imperfect. From (8), we can see that if risk ad- 35

Table10: Counterfactual: OverpaymentforMAEnrollees MarketswithSmallAd MarketswithLargeAd MarketswithAnyAd Baseline PartialEq FullEq Baseline PartialEq FullEq Baseline PartialEq FullEq Panel1:OverpaymentforConsumersThatAreNewtoMedicare($,perMonth) OverpaymentperMAenrollee($) 113.8 112.9 113.0 106.6 98.1 96.7 109.1 103.4 102.6 NetOverpayment($) -15.3 -16.3 -16.1 10.3 1.8 0.4 -0.1 -5.8 -6.6 Panel2:OverpaymentforallMAEnrollees($,perMonth) OverpaymentperMAenrollee($) 151.7 151.5 151.5 155.9 154.0 154.0 154.5 153.2 153.2 NetOverpayment($) -2.3 -2.5 -2.5 9.9 8.0 8.1 5.3 4.0 4.0 justment were perfect, then c (z)−cap (z) will be constant across z. On the other jct ct hand, if risk adjustment is imperfect, then an insurer’s risk selection through advertising can result in a lower premium because an insurer will be able to construct a better risk pool by attracting healthier customers. Therefore, shutting down advertising may substantiallyincreasepremiums,whichwillpotentiallylowertheconsumer’swelfare. Next,weinvestigatetheeffectsofadvertisingonoverpaymentforanMAenrollee bythegovernment. Wedefineoverpaymentasthepredictedcapitationpaymentminus predictedMedicarereimbursementcostfortheindividual. Table10presentspredicted overpaymentforindividualswhowouldchooseMAinthebaselineandcounterfactual situations. Because a MA insurers will receive overpayment even for enrolling an individualwiththeaveragehealthasillustratedinFigure1,wealsoreporttheamount of overpayment net of the amount of overpayment for enrolling an individual with the average health, which we call “net overpayment.” We find that if advertising is shut down, then overpayment for overall MA enrollees will decrease because healthy consumers will be less likely to switch to MA without advertising. The change in the amount of overpayment is not very large and is smaller than the difference in the average premiums between the baseline and the “Full Eq” counterfactual. The result implies that the decrease in premiums resulting from advertising does not lead to too much government expenditures. Moreover, the effects of advertising on overpayment aremuchgreaterforindividualswhoarenewtoMedicarebecauseofthelargereffects of advertising on them. In fact, net overpayment for the new Medicare beneficiaries willbealmosteliminatedinthemarketswithrelativelylargeadvertisingexpenditures. Lastly, we investigate the effects of advertising on the per-capita government ex- 36

Table11: Per-CapitaMonthlyGovernmentExpenditures($) Baseline PartialEq FullEq ConsumersThatAreNewtoMedicare 307.4 304.9 304.2 AllMAEnrollees 452.6 451.8 451.5 penditure. The total government expenditure is defined as the sum of capitation payments for MA enrollees and Medicare reimbursement costs for traditional Medicare enrollees. Table11showsthatalthoughadvertisingdoesnothavealargeeffectforthe overall population because of switching cost, shutting down advertising will decrease theexpenditurebyabout1%to$304.2perindividualpermonthamongnewMedicare beneficiaries. Wealsofindthatthedifferenceingovernmentexpendituresbetweenthe baseline and the “Full Eq” counterfactual is smaller than the difference in the average premiums between the two situations, again implying that the decrease in premiums resultingfromadvertisingdoesnotleadtotoomuchgovernmentexpenditures. In summary, we find that advertising mainly affects the demand for consumers who become newly eligible in Medicare, and at the same time, we find a substantial increase in premiums. These equilibrium impacts are often ignored when researchers are interested in measuring the welfare impact of risk selection under various risk adjustmentsystems,whichsofaremphasizesexcessgovernmentexpenditureduetorisk selection (see Brown et al. 2014). Although a more complete welfare analysis is left tothefuturework,ourresultshighlightthatitisimportanttoendogenizeandquantify theriskselectiontoolsofinsurersinordertounderstandriskadjustmentdesigns.42 8 Conclusion This paper quantifies the impacts of advertising as risk selection on equilibrium marketoutcomesinMA.Wefirstdocumentevidencethatbothmassadvertisinganddirect mail advertising are targeted in order to risk-select, attracting healthier individuals. In the main analysis, we develop and estimate an equilibrium model of the MA market withadvertisinginordertounderstandtheimpactofadvertisingonconsumerdemand. 42Animportantreasonwedidnotattempttoconductacompletewelfareanalysisisthatsuchananalysiswithadvertisingheavilydependsonhowwespecifythewayadvertisingaffectsdemand.Forexample,informativeadvertisingcanbewelfare-improving,whereaspersuasiveadvertisingcanbewasteful intermsofsocialwelfare. 37

Ourestimatesdemonstratethatadvertisinghaspositiveeffectsonoveralldemand,but amuchlargereffectonthedemandsofthehealthy. Then,weconductacounterfactual experiment that shuts down advertising to quantitatively evaluate the importance of risk selection through advertising on market outcomes. We find that the equilibrium premium increases on average up to 40% for insurers that had relatively large advertisingexpenditures,astheirriskpoolsdeteriorate. Althoughwefindthatriskselection throughadvertisinghasarent-seekingaspect,itdidnotlikelyleadtoawastefuladvertisingcompetition. Therefore,riskselectionthroughadvertisingmaymakeconsumers betteroffbyloweringpremiumswithoutmuchinefficientspending. An important future work is to quantify the welfare impact of risk selection and investigate the optimal design of risk adjustment. The main challenge in our context is to develop a coherent framework in which to measure the impact of advertising on consumer welfare. This requires an explicit modeling and identification of various mechanisms of the impact of advertising, for example informative, persuasive, and signaling roles, which are known to be challenging. Another important avenue is to consider other instruments for conducting risk selection. These extensions will allow ustoconductamorecompletewelfareassessmentofriskselection. References Ackerberg, D. A., 2001. Empirically distinguishing informative and prestige effects of advertising. RANDJournalofEconomics,316–333. Ackerberg, D. A., 2003. Advertising, learning, and consumer choice in experience good markets: an empiricalexamination.InternationalEconomicReview44(3),1007–1040. Batata,A.,2004.Theeffectofhmosonfee-for-servicehealthcareexpenditures: Evidencefrommedicarerevisited.Journalofhealtheconomics23(5),951–963. Bauhoff, S., 2012. Do health plans risk-select? an audit study on germany’s social health insurance. JournalofPublicEconomics96(9),750–759. Berry, S., Levinsohn, J., Pakes, A., 1995. Automobile prices in market equilibrium. Econometrica, 841–890. Berry,S.,Levinsohn,J.,Pakes,A.,2004.Differentiatedproductsdemandsystemsfromacombination ofmicroandmacrodata: Thenewcarmarket.JournalofPoliticalEconomy112(1),pp.68–105. Brown, J., Duggan, M., Kuziemko, I., Woolston, W., 2014. How does risk selection respond to risk adjustment? new evidence from the medicare advantage program. American Economic Review 104(10),3335–64. 38

Brown, J. R., Goolsbee, A., 2002. Does the internet make markets more competitive? evidence from thelifeinsuranceindustry.JournalofPoliticalEconomy110(3),pp.481–507. Cabral, M., Geruso, M., Mahoney, N., 2014. Does privatized health insurance benefit patients or producers? evidence from medicare advantage. Working Paper 20470, National Bureau of Economic Research. Cebul,R.D.,Rebitzer,J.B.,Taylor,L.J.,Votruba,M.E.,2011.Unhealthyinsurancemarkets: Search frictions and the cost and quality of health insurance. The American Economic Review 101 (5), 1842–71. Chiappori, P.-A., Salanie, B., 2000.Testingforasymmetricinformationininsurancemarkets.Journal ofpoliticalEconomy108(1),56–78. Ching, A. T., Ishihara, M., 2012. Measuring the informative and persuasive roles of detailing on prescribingdecisions.ManagementScience58(7),1374–1387. Clark,C.R.,Doraszelski,U.,Draganska,M.,2009.Theeffectofadvertisingonbrandawarenessand perceived quality: An empirical investigation using panel data. Quantitative Marketing and Economics7(2),207–236. Curto,V.,Einav,L.,Levin,J.,Bhattacharya,J.,2014.Canhealthinsurancecompetitionwork?evidence frommedicareadvantage.WorkingPaper20818,NationalBureauofEconomicResearch. Decarolis,F.,Polyakova,M.,Ryan,S.P.,2015.Thewelfareeffectsofsupply-sideregulationsinmedicarepartd.WorkingPaper21298,NationalBureauofEconomicResearch. Dubois,P.,Griffith,R.,O’Connell,M.,2014.Theeffectsofbanningadvertisinginjunkfoodmarkets. Unpublishedmanuscript,ToulouseSchoolofEconomics. Duggan,M.,Starc,A.,Vabson,B.,2014.Whobenefitswhenthegovernmentpaysmore? pass-through inthemedicareadvantageprogram.WorkingPaper19989,NationalBureauofEconomicResearch. Ellis, R. P., Fernandez, J. G., 2013. Risk selection, risk adjustment and choice: Concepts and lessons fromtheamericas.InternationalJournalofEnvironmentalResearchandPublicHealth10(11),5299– 5332. Fang, H., Keane, M. P., Silverman, D., 2008. Sources of advantageous selection: Evidence from the medigapinsurancemarket.JournalofPoliticalEconomy116(2),pp.303–350. Finkelstein, A., McGarry, K., 2006. Multiple dimensions of private information: evidence from the long-termcareinsurancemarket.AmericanEconomicReview96(4),938–958. Geruso,M.,Layton,T.,2015.Upcoding:Evidencefrommedicareonsquishyriskadjustment.Working Paper21222,NationalBureauofEconomicResearch. Goeree,M.S.,2008.Limitedinformationandadvertisingintheuspersonalcomputerindustry.Econometrica76(5),1017–1074. Gordon,B.R.,Hartmann,W.R.,2013.Advertisingeffectsinpresidentialelections.MarketingScience 32(1),19–35. Handel, B. R., 2013. Adverse selection and inertia in health insurance markets: When nudging hurts. AmericanEconomicReview103(7),2643–82. 39

Hastings, J. S., Hortaçsu, A., Syverson, C., 2013. Advertising and competition in privatized social security: thecaseofmexico. Hausman, J.A., 1996.Valuationofnewgoodsunderperfectandimperfectcompetition.In: Theeconomicsofnewgoods.UniversityofChicagoPress,pp.207–248. Ho, K., Hogan, J., Morton, F. S., 2015. The impact of consumer inattention on insurer pricing in the medicarepartdprogram.WorkingPaper21028,NationalBureauofEconomicResearch. Imbens,G.W.,Lancaster,T.,1994.Combiningmicroandmacrodatainmicroeconometricmodels.The ReviewofEconomicStudies61(4),655–680. Kuziemko,I.,Meckel,K.,Rossin-Slater,M.,2013.Doinsurersrisk-selectagainsteachother? evidence frommedicaidandimplicationsforhealthreform.NBERWorkingPaper(w19198). Langwell, K. M., Hadley, J. P., 1989. Evaluation of the medicare competition demonstrations. Health CareFinancingReview11(2),65. Lustig, J., 2011.Measuringwelfarelossesfromadverseselectionandimperfectcompetitioninprivatizedmedicare.Manuscript. Mello, M. M., Stearns, S. C., Norton, E. C., Ricketts, T. C., 2003. Understanding biased selection in medicarehmos.Healthservicesresearch38(3),961–992. Miller, K., 2014. Do private medicare firms face lower costs? Unpublished manuscript, University of Minnesota. Miller, K., Petrin, A., Town, R., Chernew, M., 2014. Health plan changes in response to medicare advantagepaymentrates.Unpublishedmanuscript,UniversityofMinnesota. Moshary,S.,2015.Pricediscriminationacrosspacsandtheconsequencesofpoliticaladvertisingregulation.Unpublishedmanuscript,MIT. Nevo,A.,2001.Measuringmarketpowerintheready-to-eatcerealindustry.Econometrica69(2),307– 342. Newhouse, J. P., Price, M., Huang, J., McWilliams, J. M., Hsu, J., 2012. Steps to reduce favorable riskselectioninmedicareadvantagelargelysucceeded,bodingwellforhealthinsuranceexchanges. HealthAffairs31(12),2618–2628. Nosal, K., 2012. Estimating switching costs for medicare advantage plans. Unpublished manuscript, UniversityofMannheim. Polyakova, M., 2014. Regulation of insurance with adverse selection and switching costs: Evidence frommedicarepartd.Unpublishedmanuscript,StanfordUniversity. Pope, G. C., Kautter, J., Ellis, R. P., Ash, A. S., Ayanian, J. Z., Ingber, M. J., Levy, J. M., Robst, J., etal.,2004.Riskadjustmentofmedicarecapitationpaymentsusingthecms-hccmodel. Starc,A.,2014.Insurerpricingandconsumerwelfare: evidencefrommedigap.TheRANDJournalof Economics45(1),198–220. Town,R.,Liu,S.,2003.Thewelfareimpactofmedicarehmos.RANDJournalofEconomics,719–736. VandeVen,W.P.,Ellis,R.P.,2000.Riskadjustmentincompetitivehealthplanmarkets.Vol.1,PartA ofHandbookofHealthEconomics.Elsevier,pp.755–845. 40

A Appendix A.1 Incentives for Risk Selection Usingthedataavailableinthispaper,weinvestigatewhetherMAinsurershaveincentives to risk-select by calculating potential expected profits from enrolling a healthy and an unhealthy individual. Although recent papers make use of an individual’s risk score (e.g., Brown et al. 2014 and Curto et al. (2014)), we do not have access to such information with our data. However, the MCBS still provides useful information that can shed light on potential profits for insurers from enrolling individuals of different health types. We make use of the two variables in the MCBS in order to calculate the potential profits from a healthy and an unhealthy individual. First, the MCBS contains information on how much an MA insurer received for enrolling an individual includedinthedata. Second,weusetheamountofMedicarereimbursementcostsfor individualsenrolledintraditionalMedicare. A possible measure of potential profit for an individual is the difference between theexpectedcapitationpaymentiftheindividualenrollsinMAandtheexpectedMedicare reimbursementcost ifthe individual enrollsin traditional Medicare. However, an importantlimitationofthetwovariablesisthattheyarenon-missingonlyforindividualsdependingontheirinsurancechoice. Therefore,weimputetheexpectedcapitation payment and the expected Medicare reimbursement cost using their relationship with anindividual’sobservedcharacteristics,whichisestimatedwithindividualswhohave non-missingvaluesforthetwovariables. Forthecapitationpayment,werunthefollowingregressionusinginformationfrom individualswhoenrolledinMA: cp = f(Age ,Female ,Health ,Benchmark )β +ε , (9) it it it it county(i),t it where cp denotes the amount of the monthly capitation payment for individual i and it year t, and f(Age ,Female ,Health ,Benchmark ) is a function that generit it it county(i),t ates interactions between an individual’s age, gender, health status, and the capita- 41

tion benchmark of the individual’s county in year t (Benchmark ). An indicounty(i),t vidual health status Health is a binary variable that is equal to one if individual i is it healthy as defined when we described the MCBS in Section 2. Because of the introduction of the new risk adjustment regime in 2004, the relationship between cp and it f(Age ,Female ,Health ,Benchmark ) may have changed during the year. it it it county(i),t Thus, we run separate regressions for the years before 2004 and after 2003. The regression results are reported in Table 14 in the Appendix. Using the estimates, we simulatetheexpectedcapitationpaymentforeachindividualincludedintheMCBS. For the expected Medicare reimbursement cost, we run a similar regression using informationfromindividualswhoenrolledintraditionalMedicare: mr = f(Age ,Female ,Health ,Cost )β +ε , (10) it it it it county(i),t it wheremr denotestheMedicarereimbursementcostforindividualiinyeart averaged it over twelve months, and f(Age ,Female ,Health ,Cost ) is a function that it it it county(i),t generatesinteractionsbetweenanindividual’sage,gender,healthstatus,andper-capita Medicarereimbursementcostintheindividual’scountyinyeart (Cost ). Note county(i),t thatinformationforCost doesnotcomefromtheMCBSbutdirectlyfromthe county(i),t CMS.Thus,Cost istheexactper-capitaMedicarecostforthecountyinyeart. county(i),t The regression results are reported in Table 15 in the Appendix. Using the estimates, we simulate the expected Medicare reimbursement cost for each individual included intheMCBS. Once we calculate the expected capitation payment and Medicare cost for each individual in the MCBS, we calculate the potential profit for an MA insurer from enrollingeachindividual. Thepotentialprofitπ isdefinedas it π =E[cp ]−E[mr ]. it it it Table 12 presents the average monthly potential profits depending on an individual’s health status before and after the introduction of the more comprehensive risk adjustmentregimeafter2003. Notethatthepotentialprofitfromahealthyindividualissubstantially larger than that from an unhealthy individual, regardless of risk adjustment regimes. Thedifferencesbetweenthepotentialprofitsfromahealthyandanunhealthy 42

Table12: IncentivestoTargetHealthyConsumers E[π |Health =1]($) E[π |Health =0]($) Difference($) it i it i Before2004 214.4 -303.8 518.2 After2003 252.4 -214.2 466.6 individual are $518.2 and $466.4 before and after the new risk adjustment regime, respectively. Although E[π |Health = 1] increased after 2003, E[π |Health = 0] init i it i creased even more, and the difference decreased after 2003. Therefore, we find that enrolling healthy individuals is much more profitable for MA insurers before and after the new risk adjustment regime, although relative potential profits from healthy individualsslightlydecreasedafter2003. The fact that we find that enrolling healthy individuals continues to be profitable even after 2003 may seem inconsistent with the finding that the new risk adjustment regimesubstantiallyreducesthecapitationpaymenttoindividualswithlowriskscores, whoareconsideredhealthieraccordingtotheriskscoresystem(seeTable3inBrown et al. (2014)). However, we argue that our finding is not necessarily contradictory to the finding by Brown et al. (2014) for two reasons. First, they also find that the new risk adjustment regime still does not account for Medicare costs for unhealthy individuals. In other words, the capitation payment for an individual with a lower risk score is still greater than the individual’s expected Medicare cost. In fact, Brown et al. (2014) find that for 77% of individuals, the capitation payments are estimated to be larger than their expected Medicare costs before and after the new risk adjustment regime. Because Health is equal to one for about 83% of individuals as shown in it Table2,itislikelythatoverall,healthyindividualsoverallcontinuetoresultingreater profits for MA insurers. Second, the capitation benchmark increased when the new risk adjustment regime was introduced after 2003. As a result, the capitation payment foreveryindividualincreased,althoughtherelativecapitationpaymentchanged. 43

A.2 Details on the Supply Side A.2.1 EstimationoftheExpectedHealthCosts WeassumethatanMAenrollee’srealizedhealthreimbursementcostforinsurer jct is givenby (cid:0) (cid:1) ln 1+m(z,x ,ω ;λ) =zλ +x λ +λ ω i jct ijct i z jct x w ijct whereω isassumedtobeastandardnormalrandomvariable. ijct The realized reimbursement cost for an MA’s enrollee in a given year is available from the MCBS Cost and Use module. Because we observed an individual’s characteristics z and those of the insurer the individual chose x , estimating parameter i jct λ is straightforward and can be done independently of the demand model. Table 13 presentsestimatesforλ. A.2.2 Solvingforη jct The profit function (8) will lead to the first order condition for the optimal pricing as follows: ´ Q + (cid:0) p +cp(z)−E [m(z,x ,ω ;λ)] (cid:1)∂q jct (z) dF (z) η = jct z i jct i ω i jct ijct ∂p jct ct i . (11) jct ∂Q jct ∂p jct Because parameter λ can be estimated outside the demand model and because both ∂q jct (z) and ∂Q jct canbecalculatedbasedonparameterestimatesforthedemandmodel, ∂p jct ∂p jct η canbecalculatedusingequation(11)byassumingobservedpremiumsinthedata jct areatequilibrium. 44

A.3 Tables Table13: EstimatesforHealthReimbursementCosts VARIABLES Coeff Std.Error hit -1.048*** (0.0674) Age 0.420*** (0.0812) Age2 -0.00262*** (0.000516) Female 0.234*** (0.0524) Per-CapitaMedicareReimbursementCostsinCounty-Year 0.00121*** (0.000240) Copayfor10InpatientDays -1.36e-05 (5.26e-05) Copayfor20DaysatSkilledNursingFacility -0.000161*** (6.16e-05) Coinsurancefor20DaysatSkilledNursingFacility 0.210*** (0.0276) CopayforSpecialistVisit -0.00420 (0.00273) CopayforPrimaryCarePhysicianVisit -0.0244*** (0.00600) CoinsuranceforSpecialistVisit -0.0888* (0.0486) CoinsuranceforPrimaryCarePhysicianVisit -0.177 (0.185) Dummy:DentalCoverage -0.249*** (0.0848) Dummy:HearingExam 0.561** (0.251) Dummy:HearingAid 0.0453 (0.0633) Dummy:RoutineEyeExam 0.300*** (0.0914) Dummy:DrugCoverage 0.154** (0.0645) Dummy:HMO 0.0586 (0.556) Dummy:PPO -1.388** (0.566) Dummy:PrivateFeeforService -1.016 (0.709) Observations 4,890 R-squared 0.097 Note1: Thevariable“Copayfor10InpatientDays”referstotheamountofcopaymentswhenapatientstays10daysatan inpatientfacility.Othervariableswithsimilarformatscanbeinterpretedinasimilarway. Note2: Inadditiontothevariablesincludedinthetable,wealsoincludedvariabledummyvariablesforinsurerswithmissing informationineachbenefit. Forexample,someinsurershaveacoinsuranceforaspecialistvisitinsteadofacopayment. Inthis case,weincludedadummyvariablethatequalstooneifinformationaboutcopaymentdoesnotexist. 45

Table14: CapitationPaymentsforMAEnrollees Before2004 After2004 VARIABLES Coeff StdErr Coeff StdErr 1[Health =1] -3,040 (3,539) -2,318 (4,549) it 1[Health =1]×Age -11.74 (15.88) -51.74 (32.55) it 1[Health =0]×Age -86.87 (87.70) -119.5 (115.5) it 1[Health =1]×Age2 0.0856 (0.102) 0.350* (0.210) it 1[Health =0]×Age2 0.550 (0.550) 0.832 (0.758) it 1[Health =1]×Benchmark -5.516*** (1.082) -7.350*** (2.011) it 1[Health =0]×Benchmark -9.376* (5.521) -9.524 (7.292) it 1[Health =1]×Age×Benchmark 0.148*** (0.0279) 0.197*** (0.0520) it 1[Health =0]×Age×Benchmark -0.000809*** (0.000179) -0.00114*** (0.000333) it 1[Health =1]×Age2×Benchmark 0.244* (0.140) 0.267 (0.191) it 1[Health =0]×Age2×Benchmark -0.00141 (0.000879) -0.00167 (0.00124) it Female 16.45 (12.01) 5.777 (21.12) Female×Benchmark -0.169*** (0.0206) -0.156*** (0.0349) Observations 6,258 2,592 ***p<0.01,**p<0.05,*p<0.1 Note:ThesampleforthisanalysisconsistsofindividualsintheMCBSwhochoseMA. Table15: ReimbursementCostsforTraditionalMedicareEnrollees Before2004 After2004 VARIABLES Coeff StdErr Coeff StdErr 1[Health =1] 1,666 (2,327) -5,824* (3,027) it 1[Health =1]×Age 71.88*** (12.97) 44.36** (21.52) it 1[Health =0]×Age 106.2* (57.72) -102.0 (74.30) it 1[Health =1]×Age2 -0.407*** (0.0829) -0.212 (0.137) it 1[Health =0]×Age2 -0.621* (0.367) 0.708 (0.472) it 1[Health =1]×Cost 0.547*** (0.105) 0.698*** (0.145) it 1[Health =0]×Cost 2.180*** (0.278) 1.470*** (0.370) it Female 34.00 (65.91) 174.9* (95.80) Female×Cost -0.194 (0.147) -0.419** (0.192) Observations 23,890 12,058 ***p<0.01,**p<0.05,*p<0.1 Note:ThesampleforthisanalysisconsistsofindividualsintheMCBSwhostayedwithtraditionalMedicare. 46

Table16: CorrelationbetweenMeanandVarianceofHealthExpenditures DependentVariable MedicareReimbursementCost (1) (2) Coefficient Std. Error Coefficient Std. Error 1[Health =0] -61.85 (156.7) -35.28 (158.3) it Age -8.837 (5.714) -10.43* (5.804) Dummy: Female? 106.4 (81.03) 135.5* (82.26) Per-CapitaMedicareCost -2.229** (0.872) -2.753*** (1.063) 1[Health =0]×Per-CapitaMedicareCost 1.140*** (0.311) 1.094*** (0.313) it Age×Per-CapitaMedicareCost 0.0373*** (0.0115) 0.0406*** (0.0117) Female×Per-CapitaMedicareCost -0.307* (0.162) -0.363** (0.165) YearFE? Yes Yes CountyFE? No Yes Observations 16,525 16,525 R-squared 0.070 0.095 ***p<0.01,**p<0.05,*p<0.1 Note:ThesampleforthisanalysisconsistsofindividualsintheMCBSwhostayedwithtraditionalMedicare. Table17: GeographicalTargetingofMassAdvertising DependentVariable AdvertisingExpenditure VARIABLES Coefficient Std. Error AveragePotentialProfit(AvgProfit ) 0.00123*** (0.000275) cjt Per-CapitaMedicareCost(MC ) 0.00115*** (0.000357) cjt PopulationofMedicareBeneficiaries(apartofX ) 2.64e-06** (1.13e-06) cjt NumberofMAInsurers(apartofX ) -0.00777 (0.00745) cjt Dummy: InsurerthatEnteredaCountyThisYear? (apartofX ) -0.0200 (0.0133) cjt Dummy: InsurerthatEnteredaCountyLastYear? (apartofX ) -0.0199 (0.0166) cjt Dummy: InsurerthatExitedaCountyintheEndofThisYear? (apartofX ) -0.0167 (0.0175) cjt YearFE?(f ) Yes t CountyFE?(f ) Yes c Observations 4,092 ***p<0.01,**p<0.05,*p<0.1 47

gniraeHdnanoisiVhtiwsmelborPdnahtlaeH :81elbaT 2sutatSgniraeH 1sutatSgniraeH 2sutatSnoisiV 1sutatSnoisiV raVtnednepeD rrEdtS ffeoC rrEdtS ffeoC rrEdtS ffeoC rrEdtS ffeoC )3920.0( ***7690.0 )9210.0( ***031.0 )0330.0( ***641.0 )6410.0( ***302.0 ]0= htlaeH[1 ti )8220.0( ***3670.0- )5210.0( ***1160.0- )2620.0( ***741.0- )5410.0( ***4950.0egA )741000.0( ***425000.0 )50-e20.8( ***784000.0 )761000.0( ***40100.0 )50-e43.9( ***944000.0 derauqSegA 923,6 118,12 750,5 437,12 snoitavresbO 1.0<p*,50.0<p**,10.0<p*** .noisivs’enohtiwsmelborpsuoireseromgnitneserperrebmunrehgihahtiwnoisivfosutatss’laudividninasebircsedtahtelbairavlacirogetaca=1sutatSnoisiV:1etoN rofylnodetalupopsielbairavsihT .noisivhtiwmelborpaoteuderacideMtuobatuodnfiotlaudividninatlucfifidwohsebircsedtahtelbairavlacirogetaca=2sutatSnoisiV :2etoN .2002morfelbaliavaemacebelbairavsihtdnA.1sutatSnoisiVrofnoisivhtiwsmelborpdahyehttahtdetacidnitahtslaudividni .2002morfelbaliavaemacebosla2sutatSgniraeHdnA.yawralimisanidenfiedera2dna1sutatSgniraeH:3etoN 48

Table19: EstimatesforParametersinMeanUtility(δ ) jmt VARIABLES Coefficient StdError Premium -0.0146** (0.00606) (cid:0) (cid:1) ln 1+ad -0.546** (0.268) jm(c)t Dummy:DrugCoverage 0.358*** (0.0878) CopayperInpatientStay 0.000488 (0.000383) Copayfor10InpatientDays 0.000112 (0.000160) Copayfor40InpatientDays -0.000218 (0.000137) Copayfor90InpatientDays 9.10e-05* (5.00e-05) CopayperStayatSkilledNursingFacility 0.000285 (0.000193) Copayfor5DaysatSkilledNursingFacility 0.000373 (0.000622) Copayfor20DaysatSkilledNursingFacility 0.000165 (0.000227) Copayfor50DaysatSkilledNursingFacility -9.36e-05* (5.47e-05) CopayforSpecialistVisit 2.57e-05 (0.00494) CopayforPrimaryCarePhysicianVisit -0.0431*** (0.0107) CoinsuranceperInpatientStay -0.0350 (0.0621) CoinsuranceperStayatSkilledNursingFacility -0.786 (0.612) Coinsurancefor20DaysatSkilledNursingFacility -0.766 (0.713) Coinsurancefor100DaysatSkilledNursingFacility 0.236 (0.210) CoinsuranceforSpecialistVisit -0.0350 (0.0621) CoinsuranceforPrimaryCarePhysicianVisit -0.786 (0.612) Dummy:DentalCoverage 0.233 (0.158) Dummy:HearingExam -0.386 (0.299) Dummy:HearingAid 0.345** (0.146) Dummy:RoutineEyeExam -0.0525 (0.134) Insurer-YearFE? Yes MarketFE? Yes Observations 3,955 ***p<0.01,**p<0.05,*p<0.1 Note1: Thevariable“Copayfor10InpatientDays”referstotheamountofcopaymentswhenapatientstays10daysatan inpatientfacility.Othervariableswithsimilarformatscanbeinterpretedinasimilarway. Note2: Inadditiontothevariablesincludedinthetable,wealsoincludedvariabledummyvariablesforinsurerswithmissing informationineachbenefit. Forexample,someinsurershaveacoinsuranceforaspecialistvisitinsteadofacopayment. Inthis case,weincludedadummyvariablethatequalstooneifinformationaboutcopaymentdoesnotexist. 49

Table20: EstimatesforParametersofPreferenceHeterogeneity Variables Estimates Std.Error (cid:0) (cid:1) log 1+adjmt ×1[di,t−1=j]×hit 0.879* (0.485) (cid:0) (cid:1) log 1+adjmt ×1[di,t−1(cid:54)=j,di,t−1≥0]×1[hit=0] 1.449*** (0.467) (cid:0) (cid:1) log 1+adjmt ×1[di,t−1(cid:54)=j,di,t−1≥0]×1[hit=1] 0.470 (0.612) Premium×hit 3.2e-4 (0.003) 1[di,t−1(cid:54)=j,di,t−1≥0] -3.786*** (0.242) 1[di,t−1(cid:54)=j,di,t−1≥0]×hit 0.016 (0.127) 1[di,t−1(cid:54)=j,di,t−1≥0]×NumberofFirmsinMarket 0.008 (0.084) 1[di,t−1(cid:54)=j,di,t−1≥0]×NumberofFirmsinMarketSquared -0.007 (0.008) hit×MA 0.181 (0.215) Income×MA 0.631*** (0.234) Income2×MA -0.131*** (0.035) Age×MA -17.76** (8.271) 65 (cid:16) (cid:17)2 Age ×MA 6.698* (3.467) 65 DrugCoverage×hit 0.0147 (0.215) PrivateFee-for-ServicePlan×hit -0.832 (0.540) TraditionalMedicareLastYear×MA -0.338 (0.195) MALastYear×MA -0.463*** (0.180) NewtoMedicare×MA -1.709*** (0.196) 50