Who Pays For Your Rewards? Redistribution in the Credit Card Market

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Who Pays For Your Rewards? Redistribution in the Credit Card Market Sumit Agarwal, Andrea Presbitero, Andr´e F. Silva, Carlo Wix 2023-007 Please cite this paper as: Agarwal, Sumit, Andrea Presbitero, Andr´e F. Silva, and Carlo Wix (2023). “Who Pays For Your Rewards? Redistribution in the Credit Card Market,” Finance and Economics DiscussionSeries2023-007. Washington: BoardofGovernorsoftheFederalReserveSystem, https://doi.org/10.17016/FEDS.2023.007. 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.

Who Pays For Your Rewards? Redistribution in the Credit Card Market Sumit Agarwal Andrea F. Presbitero Andre´ F. Silva Carlo Wix* December 2022 Abstract We study credit card rewards as an ideal laboratory to quantify redistribution between consumers in retail financial markets. Comparing cards with and without rewards, we find that, regardless of income, sophisticated individuals profit from reward credit cards at the expense of na¨ıve consumers. To probe the underlying mechanisms,weexploitbank-initiatedaccountlimitincreasesatthecardleveland showthatrewardcardsinducemorespending,leavingna¨ıveconsumerswithhigher unpaid balances. Na¨ıve consumers also follow a sub-optimal balance-matching heuristicwhenrepayingtheircreditcards,incurringhighercosts. Banksincentivize the use of reward cards by offering lower interest rates than on comparable cards withoutrewards.Weestimateanaggregateannualredistributionof$15billionfrom less to more educated, poorer to richer, and high to low minority areas, widening existingdisparities. Keywords: householdfinance;creditcards;financialsophistication;rewards JELClassification: G21;G40;G51;G53 *Authors’ contacts: Sumit Agarwal, National University of Singapore, e-mail: ushakri@yahoo.com; Andrea F. Presbitero, International Monetary Fund and CEPR, e-mail: apresbitero@imf.org; Andre´ F. Silva, Federal Reserve Board, e-mail: andre.f.silva@frb.gov; Carlo Wix, Federal Reserve Board, e-mail: carlo.e.wix@frb.gov.WethankourdiscussantsJaredDeLisle,YaronLevi,HugoMarin,PaolinaMedina,LucianaOrozco,StephenShore,andFrancescRodriguezTousforhelpfulcommentsandsuggestions.Weare alsogratefultoAllenBerger,VitalyBord,HansDegryse,ChenLian,ThomasMosk,LakshmiNaaraayanan, Daniel Paravisini, Matthias Rodemeier, Annette Vissing-Jørgensen, Michael Weber, and participants at severalconferencesandseminarsfortheirvaluablesuggestions.Theviewsexpressedinthispapersolely reflectthoseoftheauthorsandnotnecessarilythoseoftheFederalReserveBoardortheFederalReserve System,andneithershouldbeattributedtotheInternationalMonetaryFund,itsExecutiveBoard,orits management.Firstversion:June2022. 1 Electronic copy available at: https://ssrn.com/abstract=4126641

I. Introduction Consumerslackingfinancialsophisticationoftenmakecostlymistakes(e.g.,Campbell, 2006; Gomes, Haliassos, and Ramadorai, 2021). In the consumer credit card market, such behavior can entail overindebtedness (Gross and Souleles, 2002; Heidhues and Ko˝szegi, 2010) and sub-optimal repayments (Ponce, Seira, and Zamarripa, 2017; Gathergood, Mahoney, Stewart, and Weber, 2019). Banks, in response, can design financial productstoexploitthesemistakes,combiningsalientbenefitswithshroudedpayments (DellaVigna and Malmendier, 2004; Heidhues and Ko˝szegi, 2017). Na¨ıve consumers might underestimate these payments and incur costs from usage. Sophisticated consumers, in contrast, might rake in the benefits while avoiding the payments and thus profitfromusage. Suchproductscanthereforegenerateanimplicitredistributionfrom na¨ıvetosophisticatedconsumers(GabaixandLaibson,2006)andtherebycontributeto inequality(Campbell,2016;Lusardi,Michaud,andMitchell,2017). Despite these theoretical predictions, empirically quantifying the extent of such redistribution is challenging. First, for many financial products such as mortgages, optimal behavior depends on consumers’ risk aversion, economic expectations, and other hard-to-measure variables (Campbell and Cocco, 2003; Fisher, Gavazza, Liu, Ramadorai, and Tripathy, 2021; Guiso, Pozzi, Tsoy, Gambacorta, and Mistrulli, 2021). To determinewhatconstitutesbiasedbehavioristhereforenotstraightforward. Second,linking redistributiontoindividualcharacteristicsrequiresdetailedindividual-leveldataonthe costsandbenefitsofusingafinancialproduct,whereasthelatterinparticularareoften unobservableoratleastdifficulttoquantify. Inthispaper,weusecreditcardrewardsasanideallaboratorytostudysuchredistributionbetweenconsumersinretailfinancialmarkets. Rewardcreditcards—whichoffer points, miles, or cash back to cardholders for every dollar spent—are a ubiquitous featureinAnglo-Saxonconsumercreditcardmarketsandarealsogainingmarketsharein othercountries. In2019, rewardcreditcardsaccountedfor60percentofallnewcredit 2 Electronic copy available at: https://ssrn.com/abstract=4126641

card originations in the United States (CFPB, 2019), with the largest U.S. banks paying $35billioninrewards. WeusecomprehensivecreditcarddatafromtheFederalReserve Board’sY-14Mreportswhichencompassthenear-universeofaccountsintheU.S.This datasetcontainsdetailedmonthlyaccount-levelinformationandisthereforeuniquely suited to study redistribution between different consumers. It allow us to compute a cardholder’smonthlynetreward,definedasthedollarvaluereceivedinrewardsminus interestandfeepayments,whichcapturesboththebenefitsandthecostsofcreditcard usage. We start our empirical analysis by investigating whether reward credit cards induce redistribution between consumers across the FICO score distribution. To this end, we compare the outcomes of reward cards to those of similar classic cards across cardholders in the same FICO and income percentiles, living in the same ZIP code, and who are clients at the same bank, while further controlling for an extensive set of card- and consumer-level characteristics.1 We find that for sub-prime (with a FICO score below 660)andnear-prime(660to720)cardholders,monthlynetrewardsareonaverage$5.4 and$6.8lower,respectively,onrewardcardsrelativetosimilarclassiccards. Forprime (720 to 780) and super-prime (above 780) cardholders, monthly net rewards are on average $7.3 and $16.0 higher, respectively. This result is driven by both the cost and the benefit margin of net rewards. Super-prime cardholders earn on average $9.5 in rewards and pay $7.1 less in interest on reward cards than on classic cards. In contrast, sub-prime consumers earn only $1.8 in rewards but pay $6.4 more in interest. Thus, high-FICOcardholdersonaverageearnmoneywiththeuseofrewardcardswhilelow- FICO cardholders on average lose money. In aggregate terms, we find an annualized redistributionof$15.1billioninducedbycreditcardrewards. Next,westudywhethertheredistributionacrossFICOscoresisdrivenbydifferences in cardholders’ income, suggesting a transfer from poor to rich consumers. Indeed, 1We adopt the following terminology: “Reward cards” are credit cards that earn either cash back, miles,orpoints;“classiccards”arecreditcardsthataredonotearnanyformofrewards. 3 Electronic copy available at: https://ssrn.com/abstract=4126641

credit card rewards are often framed as a “reverse Robin Hood” mechanism in which thepoorsubsidize therich.2 Ourresults, however, showthat thisexplanationisat best incomplete. Since FICO scores and income are only moderately correlated, as documentedinBeer,Ionescu,andLi(2018),wecandisentanglethesetwomargins. Wefind aredistributionfromlow-tohigh-FICOconsumersregardlessofincome. Whilesuperprime high-income consumers benefit the most from reward credit cards ($20.1 in net rewards relative to classic cards), high-income consumers with sub-prime FICO scores onaveragepaythemost(-$12.8). Meanwhile,super-primelow-incomeconsumersbenefit less ($9.7), but sub-prime low-income consumers also pay less (-$2.6). Thus, highincomeconsumerswithhighFICOscoresbenefitfromrewardcreditcardslargelyatthe expenseofhigh-incomeconsumerswithlowFICOscores. As our findings are inconsistent with the “reverse Robin Hood” hypothesis, we next investigatewhetherdifferencesincardholders’financialsophisticationcanexplainour results. Since FICO scores are based on an individual’s payment history and outstanding debt relative to available credit, they capture the same type of credit card behavior thatisassociatedwithalackoffinancialsophisticationi.e.,overindebtednessandsuboptimal repayment behavior. FICO scores might thus serve as a proxy for financial sophistication (e.g. Agarwal, Rosen, and Yao, 2016; Amromin, Huang, Sialm, and Zhong, 2018; Bhutta, Fuster, and Hizmo, 2021). Our results are consistent with this interpretation. We first provide quasi-experimental evidence that reward credit cards induce low- FICO consumers to overborrow on their credit cards. To this end, we compare the spending and borrowing responses of consumers who received a bank-initiated credit limitincreaseonrewardcardstothosewhoreceivedalimitincreaseonclassiccards. We find that the spending response is stronger for consumers with a limit increase on re- 2See,forexample,“CreditCardsTakeFromPoor,GivetotheRich”intheWallStreetJournal,andmore recently“Howcreditcardcompaniesrewardtherichandpunishtherestofus”atBrookings,and“The uglytruthbehindyourfancyrewardscreditcard”atVox. 4 Electronic copy available at: https://ssrn.com/abstract=4126641

wardcardsandthatthiseffectispresentinallFICOgroups. However,whileprimeand super-prime consumers also exhibit a proportional increase in credit card payments, this is not the case for sub-prime and near-prime consumers. As a result, following a limit increase on reward cards relative to classic cards, unpaid balances increase more forlow-FICOconsumers,whiletheyremainunchangedforhigh-FICOconsumers. This patternisconsistentwiththedocumentedtendencyofna¨ıveconsumerstooverborrow on their credit cards (Heidhues and Ko˝szegi, 2010; Lusardi and Tufano, 2015) and thus inlinewiththeinterpretationofFICOscoresasaproxyforfinancialsophistication. In a separate exercise, we also show that FICO scores are strongly correlated with mistake-based measures of financial sophistication, as suggested by Calvet, Campbell, and Sodini (2009) and Jørring (2022), and that this association is more pronounced on reward cards. Focusing on individuals with multiple credit cards at the same bank, we follow Ponce, Seira, and Zamarripa (2017) and Gathergood, Mahoney, Stewart, and Weber (2019) and calculate the share of misallocated credit card payments.3 We find thatthisshareisstronglydecreasinginFICOscoresand,forsub-primeandnear-prime cardholders, larger for reward cards than for similar classic cards. We also show that low-FICO consumers in particular tend to follow a sub-optimal (and costly) balancematching heuristic when repaying their credit cards. In line with the sub-optimal repaymentbehaviorofna¨ıveconsumers(KuchlerandPagel,2021),thesefindingsprovide further corroborative evidence that the observed redistribution across FICO scores is drivenbyfinancialsophistication. Next, we turn to the supply side and study reward credit cards from the banks’ perspective, investigating both pricing strategies and profits. Despite reward cards incurring additional expenses for banks, we find that banks offer lower annual percentage 3Giventhetotalrepaymentamount,theoptimal,interest-minimizingrepaymentbehavioristofirst maketheminimumrequiredpaymentonallcards,thenrepayasmuchaspossibleonthecardwiththe highestinterestrate,andallocatefurtherpaymentstosubsequentlycheapercards.Wecalculatetheshare ofmisallocatedpaymentsasthedifferencebetweenthisoptimalandtheactuallyobservedpaymentbehaviorasamistake-basedmeasureoffinancialsophistication. 5 Electronic copy available at: https://ssrn.com/abstract=4126641

rates(APRs)onrewardcardsthanonsimilarclassiccardsacrosstheentireFICOdistribution,suggestingthatbanksincentivizetheuseofrewardcards. Howdoesthispricing strategyaffectbanks’profitabilityofrewardandclassiccards? Wedefineabank’sprofits on a credit card as the sum of income from interest payments, fee payments, and interchangefees,minusrewardexpenses,realizedcharge-offs,andfundingcostsforrevolvingbalances. WefindthatbanksprofitfromrewardcardsacrossallFICOscores,but thatprofitsarehighestfornear-primeandprimecardholdersinthemiddleoftheFICO distribution. We further document substantial differences regarding banks’ sources of revenue between high- and low-FICO consumers. For sub-prime cardholders, more than 60 percent of banks’ revenues stem from interest income, while for super-prime cardholders,upto80percentstemfrominterchangeincome. Finally, we study the geographic distribution of net rewards across ZIP codes and investigatewhetherthelargeaggregatetransferinducedbycreditcardrewardsiscorrelatedwithsocio-demographicvariables. Wefindthataveragenetrewardsarehigherin ZIPcodeswithhighereducationlevels,withahigheraverageincome,andwithalower shareofBlackresidents. Creditcardrewardsthustransferincomefromlesstomoreeducated, from poorer to richer, and from high- to low-minority areas, thereby widening existingspatialdisparities. Our contribution to the literature is threefold. First, we empirically quantify the redistributionfromna¨ıvetosophisticatedconsumers,whichhaslargelybeenstudiedtheoretically. DellaVigna and Malmendier (2004) and Heidhues and Ko˝szegi (2010) model thecontractdesignofprofit-maximizingfirmsandshowthatfirmscanexploitthetimeinconsistent preferences of na¨ıve consumers by charging back-loaded fees. In Gabaix and Laibson (2006) and Heidhues and Ko˝szegi (2017), products with this type of pricing schemes benefit sophisticated consumers at the expense of na¨ıve consumers and the latter cross-subsidize the former. Two recent papers empirically study such redistributioninthecontextofmortgagemarkets. ForItaly,Guiso,Pozzi,Tsoy,Gambacorta, 6 Electronic copy available at: https://ssrn.com/abstract=4126641

andMistrulli(2021)reportasubsidyfromna¨ıvetosophisticatedhouseholdsof303euros per year, induced by banks steering na¨ıve households towards sub-optimal mortgages. For the United Kingdom, Fisher, Gavazza, Liu, Ramadorai, and Tripathy (2021) find that counterfactual mortgage rates without cross-subsidization would be 20 basis points higher than the teaser rates which benefit sophisticated households. Our paper,incontrast,studiesredistributioninthecreditcardmarketinducedbyrewardprograms. Ourempiricalsettingcombinedwithouruniquedataenableustoreadilyquantify the costs (interest and fee payments) and, importantly, also the benefits (rewards) of financial product usage in monetary terms, thereby allowing for a straightforward estimationoftheredistributionfromna¨ıvetosophisticatedconsumers. Second, wecontributetotheliteratureonrewardcreditcards, whichhaslargelyfocused on interchange fees as a source of funding for credit card rewards. Interchange fees get passed through to merchants, which potentially respond by increasing retail prices for all consumers. Thus, credit card rewards might to some extent be funded by cashanddebitcarduserswhopayhigherpriceswithoutreceivinganyrewardstocompensate. Hayashi (2009) provides a comprehensive overview of the market for credit card reward programs. Schuh, Shy, and Stavins (2010) study the redistribution from cashtocreditcardusersandreportanannualmonetarytransferof$149percash-using household. Felt, Hayashi, Stavins, and Welte (2020) also study the redistribution from cash to credit card users and find that they imply a transfer from low-income to highincomeconsumers. Thelegalliteraturehasalsodocumentedthisregressiveredistribution, relating it to a stronger need for consumer protection (e.g., Levitin, 2008; Sarin, 2019). In contrast, our study focuses on the redistribution within credit card users, which is, as we argue, a more important margin. We show that the relevant transfer isfromna¨ıvetosophisticatedconsumersratherthanacrossincomecohorts. Third,bydocumentingalargeredistributionthroughcreditcardsrewards,ouranalysiscontributestotheliteraturethathighlightstheroleofthefinancialsystemindriving 7 Electronic copy available at: https://ssrn.com/abstract=4126641

wealthinequality(Lusardi,Michaud,andMitchell,2017;Bach,Calvet,andSodini,2020; Campbell, Ramadorai, and Ranish, 2019). In particular, our main finding that rewards programs redistribute income from na¨ıve to sophisticated consumers is related to existing studies that link heterogeneity in asset returns with measures of financial literacy(Deuflhard,Georgarakos,andInderst,2019)andfinancialsophistication(Fagereng, Guiso,Malacrino,andPistaferri,2020). II. Credit Card Rewards Programs Creditcardrewards—intheformofcashback,miles,orpoints—areloyaltyprograms bybankswhichoffervariousbenefitstocardholdersperdollarspentonthecreditcard. Cashbackcardsrefundasmallpercentageamountofthenetpurchasevolume(usually between 0.5 and 3 percent), while miles and points cards let cardholders accrue bonus pointsthatcanberedeemedatfrequentflyerprograms(milescards)or,moregenerally, atpartneringairlines,hotels,orretailers(pointscards). Rewardcreditcardsareaubiquitous and increasingly important aspect of consumer finance, accounting for over 60 percent of all new credit card originations in the United States (CFPB, 2019). In 2019, the largest U.S. banks paid out $35 billion in rewards. For cardholders, credit card rewardsareanopportunitytoearnmoneyorperkswiththeuseoftheircreditcards. For banks, credit card rewards are an incentive scheme to induce consumers to adopt and increase the usage of the banks’ credit card products (Agarwal, Chakravorti, and Lunn, 2010;ChingandHayashi,2010). Otherthanthecardholderandthecardissuer,themarketunderlyingcreditcardpayments and rewards typically involves three parties: (i) the merchant, (ii) the merchant acquirer,and(iii)thecardnetwork.4 FollowingFelt,Hayashi,Stavins,andWelte(2020), considertheexampleofacardholdermakinga$100purchasewitharewardcreditcard. 4SeealsoHayashi(2009),ShyandWang(2011),andFelt,Hayashi,Stavins,andWelte(2020)forfurther discussionoftheunderlyingmarketstructureofcreditcardpaymentsandrewards. 8 Electronic copy available at: https://ssrn.com/abstract=4126641

This payment initially flows from the cardholder to the card-issuing bank, which in turn rewards the cardholder with, for instance, $1 in cash back, miles, or points. The card issuer then retains a $2 interchange fee and sends the remaining $98 to the merchant acquirer, which in turn pays a $0.15 network fee to the card network. The merchant acquirer subsequently sends $97.70 to the merchant, not only passing through interchangeandnetworkfees,butalsoadditionallychargingamerchantservicecharge ($0.15). Thus,merchantsonlyreceiveafractionoftheinitialpurchaseamountandcan potentiallyrespondbyincreasingretailprices, implyingthatcreditcardrewardsmight to some extent be funded by cash and debit card users who pay higher prices withoutreceivinganyrewardstocompensate(Schuh, Shy, andStavins,2010;Felt, Hayashi, Stavins,andWelte,2020). Another source of funding for credit card rewards, however, are interest payments from credit cardholders with unpaid outstanding balances as well as fees e.g., late and overlimitfees. Creditcardsasapaymentdevicehavebecomeincreasinglypopularover recent years. While in 2008 cash accounted for over 30 percent of consumer payments and credit cards for only 17 percent, in 2019 the share of credit card payments (25 percent)exceededtheshareofcashpayments(22percent)forthefirsttime(Foster,Greene, andStavins,2021). Moreover,in2019,thelargestU.S.banksreported$89.7billionininterestincomeand$9.9billioninfeeincomefromcreditcards,comparedto$41.3billion incomefrominterchangefees. Fromthebanks’perspective,interestandfeestherefore constitute asubstantially largershare ofincome thaninterchange fees. Overall, theredistributionwithincreditcardusersislikelymoreimportantthanthetransferfromcash tocardusersinrecentyears. Contrasting the $34.8 billion in rewards expenses with the combined $99.6 billion earned in interest and credit card fees suggests that credit card rewards constitute a substantial annual transfer. These aggregate numbers, however, are neither informative about the extent of the redistribution—since cardholders can simultaneously re- 9 Electronic copy available at: https://ssrn.com/abstract=4126641

ceiverewardsandpayinterestorfees—noraboutwhichtypeofconsumersbenefitand losefromusingrewardcreditcards. Inthispaper,westudythesequestionsusingcomprehensiveandgranulardataonindividualcreditcardaccounts. III. Data and Summary Statistics A. Data We obtain account-level data on consumer credit cards from the Federal Reserve Board’s FR Y-14M reports. These reports require large U.S. bank holding companies, with at least $100 billion in total assets, to report detailed information on individual credit card accounts on a monthly basis. Our data contain information on 19 banks, which cover a large portion of the market and account for 70 percent of aggregate outstandingbalancesonconsumercreditcards(CFPB,2019). Forourmainempiricalanalysis, we obtain data on cardholders’ accumulated rewards, interest and fee payments, purchasevolumes,FICOcreditscores,creditlimits,andfurthercardcharacteristics. We alsoobtaindataonthecardissuingbankaswellasthecardholders’ZIPcode. Ourmainoutcomevariableofinterestintentstocapturethebenefitsminusthecosts ofcreditcardusage. Tothisend,weconstructthevariableNetRewards whichsubtracts the amount of interest and fees paid on card i in month t from the rewards earned on thecardduringthesameperiod:5 NetRewards = Rewards −InterestPaid −TotalFees (1) i,t i,t i,t i,t 5Whileourdatasetdoesnotcontaintheamountofmonthlyrewards,weobservetheamountofaccumulatedrewardsasofthereportingmonthnetofredeemedrewards. OnlineAppendixAexplainsin detailtheestimationofmonthlyrewardsfromthevariablesinourdataset.Ourdata,byconstruction,do notcapturenon-pecuniaryrewardsassociatedwithrewardcreditcards(e.g.,accesstoairportlounges). Inthatrespect,whatwemeasureisalowerboundofcardholders’netrewards. 10 Electronic copy available at: https://ssrn.com/abstract=4126641

Cardholders with positive net rewards thus benefit from the use of credit cards, while cardholderswithnegativenetrewardspayfortheuseofcreditcards. OuranalysisfocusesonthecrosssectionofallcreditcardsinMarch2019.6 Wefocus ongeneralpurposeandprivatelabel,unsecured,consumercreditcardswitharevolving feature. We further exclude corporate credit cards and closed accounts. This sample construction procedure results in sample of about 238 million credit cards as of March 2019. B. Summary Statistics Table I presents card-level summary statistics as of March 2019 for all cards in our sample (n=237,573,278), as well as separately for reward cards (nR=119,730,353) and classic cards (nC=117,842,925). Panel A presents variables related to the calculation of net rewards. The average reward card earns $9 in monthly rewards and the average classiccard—bydefinition—zero. However,rewardcardsalsoexhibitonaveragehigher interest charges than classic cards ($18 versus $10) and higher fee payments ($3 versus $2). Thus, on aggregate, the average reward card yields a (negative) net reward of -$12—thesameastheaverageclassiccard. [TableIabouthere] Panel B presents other card-level variables. On average, reward cards have lower APRs than classic cards (18% versus 22%), yield higher bank profits per card in a given month ($23 versus $6), and have higher credit limits ($10 thousand versus $4 thousand).7 These card-level differences, however, are not necessarily due to differences between the two types of credit card products, but could conceivably be driven by differences in consumers who choose to use reward cards and classic cards, respectively. 6WefocusonMarch2019asarecentmonthbeforetheCOVID-19pandemicwhichisalsonotsubject toseasonaleffectsinconsumption(suchasDecember). 7Wedescribethecalculationofcard-levelbankprofitsindetailinSectionVII.B. 11 Electronic copy available at: https://ssrn.com/abstract=4126641

Cardholders of reward cards have, on average, higher FICO scores than cardholders of classic cards (743 versus 716) and earn a higher annual income ($98 thousand vs. $79 thousand). TheremainderofPanelBprovidesfurthersummarystatisticsforthecontrol variablesinourregressions. IV. Redistribution in the Credit Card Market A. Empirical Approach To study the extent to which credit card rewards generate a redistribution between consumers and what drives this redistribution, we compare credit card outcomes between reward cards and classic cards with similar card- and cardholder characteristics acrosstheFICOdistribution. Let Y be an outcome for credit card account i issued by bank b to individual j. Our i baselineregressionspecificationisthengivenby: (cid:88)(cid:0) (cid:1) (cid:88) (cid:88) Y = δF ×RewardCard ×DF +α + Xm + Zn +ε (2) ibj i j b,z,w,f i j ibj F m n whereRewardCard isadummyvariablewhichtakesthevalue1forrewardcardsand0 for classic cards; DF is a battery of FICO bucket dummy variables which take the value of 1 for sub-prime cardholders (with a FICO score below 660), near-prime cardholders (600-720), prime cardholders (720-780), and super-prime cardholders (above 780), respectively. To avoid endogeneity problems arising from the joint determination of net rewards and FICO scores (e.g., due to high unpaid balances), we use FICO scores as of March 2018, one year prior to our data on credit card outcomes. α are interacted b,z,w,f fixed effects at the Bank × ZIP code × Income percentile × FICO percentile level. That is, wecompare creditcardoutcomes betweenreward andclassiccards forcardholders in the same FICO percentile, the same income percentile, living in the same ZIP code, 12 Electronic copy available at: https://ssrn.com/abstract=4126641

whichareclientsatthesamebank. Wecontrolforthefollowingcard-levelcharacteristics X: the credit limit (in dollar terms), the amount past due (in dollar terms), the age ofthecard(inyears),ajointaccountindicatorwhichtakesthevalueof1iftheaccount hasmorethanoneprimaryobligor,afraudflagindicatorwhichtakesonthevalueof1if theaccountiscurrentlyfrozenduetopotentialfraud,andaworkoutprogramindicator whichtakesonthevalueof1theaccountenteredintoanytypeofworkoutprogram. We further control for cardholder-level characteristics Z: a deposit relationship indicator whichtakesonthevalueof1ifthecardholderhasadepositrelationshipwiththesame bank, a lending relationship indicator which takes on the value of 1 if the cardholder has a lending relationship with the same bank, the number of cards held by the cardholderatthesamebank,andabankruptcyindicatorwhichtakesonthevalueof1ifthe cardholderhascompletedorisinanongoingbankruptcyprocess. B. Net Rewards Figure 1 illustrates the magnitude of net rewards across the FICO distribution and point to a clear redistribution between cardholders. For both reward cards and classic cards, average net rewards are increasing in FICO scores, suggesting that low-FICO consumers pay more for credit card usage. The relative magnitudes between the two cardtypes,however,differsubstantiallyacrossFICOscores. Forcardholderswithsuperprime scores (above 780), net rewards are on average positive for reward cards and slightly negative for classic cards.8 These consumers earn money with the use of reward cards, as the monetary benefits outstrip their costs. This pattern is reversed for consumers at the lower end of the FICO distribution. For cardholders with sub-prime (below660)andnear-prime(below720)scores,netrewardsarearound-$40forreward 8Notethatthenetrewardsofclassiccardscan—bydefinition—atbestbezeroifconsumersincurno interestorfeepayments. 13 Electronic copy available at: https://ssrn.com/abstract=4126641

cards and -$25 for classic cards. On average, low-FICO cardholders lose money with rewardcards,bothinabsolutedollartermsandrelativetoclassiccards. [Figure1abouthere] Thisdescriptivepatternmightbedrivenbydifferencesbetweenindividualswithlow andhighFICOscores,regardlessofthetypeofcardtheyuse. Tocontrolforthesedifferences, Table II present the estimation of Equation (2). All specifications include cardandcardholdercontrolvariables. Tomakethecomparisonashomogeneousaspossible in terms of individual characteristics, we include, alternatively, Bank × ZIP code × Incomepercentile(column1), Bank×ZIPcode×FICOscorepercentile(column2),and Bank×ZIPcode×Incomepercentile×FICOscorepercentile(column3)fixedeffects. Allspecificationsshowthatnetrewardsaresignificantlyhigherforrewardcardsthanfor similar classic cards. The coefficient of our preferred and most stringent specification in column (3) indicates that a reward card, on average, yields a $3.5 higher net reward thanaverysimilarclassiccard. [TableIIabouthere] Thisaveragenetrewarddifferentialbetweenrewardandclassiccards,however,masks important differences between cardholders across the FICO distribution. Taking the specificationincolumn(3)asourbaseline,column(4)reportsthedifferencesinnetrewards between reward and classic cards, separately for sub-prime, near-prime, prime, andsuper-primecardholders. ConsistentwithFigure1,netrewardsforsub-primeand near-prime cardholders are between $5.4 and $6.8 lower on reward cards than on similar classic cards. On the other end of the FICO distribution, net rewards turn positive and are, on average, $7.3 and $16.0 higher for prime and super-prime cardholders, respectively. Thus, while reward cards are more beneficial than classic cards on average, 14 Electronic copy available at: https://ssrn.com/abstract=4126641

onlyhigh-FICOconsumersgainfromthem,whilelow-FICOconsumerswouldbebetter offchoosingclassiccards,otherthingsequal.9 Robustness. While our baseline results compare very similar cardholders by using a granularsetoffixedeffects,ourresultscouldstillbedrivenbyremainingheterogeneity acrosscardsandcardholders. AsshowninTableI,rewardcardstendtohavelowerAPRs and higher credit limits than classic cards. Individuals might therefore chose to hold reward cards to access more credit at a cheaper price and our results might be driven by such differences in consumer preferences. To alleviate these concerns, columns (1) and (2) of Table III augment our baseline specification with credit limit percentile and APR percentile fixed effects. While the sample size is now substantially smaller, due tothe increasednumber offixedeffects, we obtainsignificantand qualitativelysimilar results, albeit smaller in magnitude. In columns (3) and (4), we replicate our baseline specification on the sample used in columns (1) and (2) and find that the change in magnitudesislargelydrivenbysampleselectioneffects. [TableIIIabouthere] Our dataset further contains a unique individual identifier within banks which allows us to compare credit card outcomes between reward and classic cards within the samecardholderj. Restrictingoursampletothesetofindividualswhoownatleastone rewardcardandoneclassiccardatthesamebank,wecanestimateourbaselinespecificationwithcardholderfixedeffects,thuscomparingtheoutcomesofrewardandclassic cardswithinthesameindividual. Asshownincolumns(5)and(6)ofTableIII,weobtain quantitatively similar results as in our baseline specification in Table II. One limitation ofourdatasetistheimpossibilitytotrackindividualsacrossbanks. Thus,theinterpretation of these results is subject to the caveat that individuals might hold additional, 9ToshowthatourregressionresultsarenotdrivenbyourthresholdvaluesforthedifferentFICObuckets,FigureA1inOnlineAppendixCprovidesacoefficientplotwhichplotsthecoefficientsδF alongside the95%confidenceintervalswhenestimatingEquation(2)with50insteadof4differentFICObuckets. 15 Electronic copy available at: https://ssrn.com/abstract=4126641

unobservedcreditcardsatotherbanks. Furthermore,whilethewithin-individualcomparison has the advantage of controlling for all unobservable individual heterogeneity (likedifferencesintastesandpreferences),itignoresthepotentialspillovereffectsthat other(rewardorclassic)creditcardscouldhaveontheoutcomesoftheobservedcards. Aggregateredistribution. Ourresultsshowthatcreditcardrewardsinducearedistribution from low- to high-FICO consumers. To illustrate the aggregate size, we sum up the net rewards of reward cards with positive and of reward cards with negative net rewards,bothacrossallcardholdersandwithineachFICObucket. Theeconomicmagnitudeissubstantial. Cardholderswithnegativenetrewardsinaggregatepay$4.1billion fortheuseofrewardcardsandcardholderswithpositivenetrewardsearn$1.3billion.10 Themonthly$1.3billionpositivenetrewardstranslateintoanannualizedredistribution of$15.1billioninducedbyrewardcreditcards. Ofthe$4.1billionthatarepaidbycardholders with negative net rewards, $1.0 billion come from sub-prime, $1.6 billion from near-prime, $1.1 billion from prime, and only $0.4 billion from super-prime cardholders. Ofthe$1.3billionearnedbycardholderswithpositivenetrewards,only$35million gotosub-prime,$134milliontonear-prime,$407milliontoprime,and$680millionto super-prime cardholders. Thus, while sub-prime and near-prime cardholders are the largest source of funding for credit card rewards, prime and super-prime cardholders are the biggest beneficiaries. Reward credit cards therefore constitute a substantial aggregatetransferfromlow-tohigh-FICOscoreconsumers. C. Net Rewards Components Wenextexaminethethreeindividualcomponentsofnetrewards—rewards,interest charges, and total fee charges. The differences in net rewards along the FICO distribution suggests that these costs and benefits also vary across FICO scores. Figure 2 illus- 10Table A1 in Online Appendix D summarizes our aggregate findings. The difference of $2.9 billion constitutesbankincome.Westudythebanks’perspectiveonrewardcreditcardinSectionVII. 16 Electronic copy available at: https://ssrn.com/abstract=4126641

tratesthatthisisthecase. RewardsareincreasinginFICOscores(PanelA)andhighest forsuper-primecardholders,whereasinterestchargesarehump-shapedinFICOscores (PanelB)andlowestforsuper-primecardholders.11 Whileinterestchargesaregenerally higher for reward cards than for classic cards, this difference is largest for near-prime cardholdersintheleftpartofthedistribution. [Figure2abouthere] We substantiate this descriptive evidence by estimating Equation (2) with rewards, interest charges, and total fee charges as outcome variables. Results are shown in Table IV. Rewards are on average $6.4 higher on reward cards than on classic cards (column 1) but this difference increases along the FICO distribution, ranging from $1.8 for sub-primecardholdersto$9.5forsuper-primecardholders(column2). High-FICOconsumersdonotonlyearnmoremoneyinrewards,theyalsoincurlowerinterestcharges. For sub-prime and near-prime cardholders, interest charges are on average $6.4 and $10.9higheronrewardcardsthanonsimilarclassiccards, whileforsuper-primecardholdersinterestchargesare$7.1lower(column4). Finally,feechargesareeconomically less relevant: the difference between reward and classic card is less than a US dollar and is quite similar along the FICO distribution (columns 5 and 6). These results show howhigh-FICOconsumersrakeinthebenefitswhileavoidingthecostsofrewardcredit cards and therefore profit from usage, while low-FICO consumers incur high costs due tohighinterestcharges. [TableIVabouthere] 11Figure A3 in Online Appendix C additional illustrates total fee charges, which are substantially smallerinmagnituderelativetointerestcharges. 17 Electronic copy available at: https://ssrn.com/abstract=4126641

V. The Reverse Robin Hood Hypothesis WenextinvestigatewhetherdifferencesinnetrewardsacrossFICOscoresaredriven byunderlyingdifferencesincardholders’income,whichwouldsuggestaredistribution from poor to rich consumers. If FICO scores are positively correlated with income and high-income consumers spend more money, then they will earn higher rewards. Indeed, in the financial press, credit card rewards are often framed as a “reverse Robin Hood”mechanisminwhichthe“poorfootmuchofthebillforcreditcardpoints,miles, andcashback”(Stewart,2021).12 Our results, however, show that this explanation is at best incomplete. First, FICO scores and income are only moderately correlated, as documented in Beer, Ionescu, and Li (2018). This allows us to study net rewards across the FICO distribution within differentincomegroups.13 Wesplitcardholdersintotercilesoflow-incomecardholders with an annual income below $44 thousand, middle-income cardholders with an annual income between $44 thousand and $79 thousand, and high-income cardholders withanannualincomeabove$79thousand. Figure 3 illustrates the magnitude of net rewards for reward cards across the FICO distribution for the three income groups.14 All income groups exhibit a pattern similar to what is observed in the whole sample, suggesting that FICO scores still play a key role in shaping the distribution of net rewards, regardless of income. For super-prime individuals, the distribution of average net rewards across income groups is consistent with a “reverse Robin Hood” hypothesis. High-income consumers with high FICO 12Seealso“CreditCardsTakeFromPoor,GivetotheRich”(Derby,2010)intheWallStreetJournal. 13Figure A4 shows that, while the distributions of FICO scores shifts to the right when moving from low-tohigh-incomecardholders,theystronglyoverlap,suggestingthatwithingivenFICObucketsthere areindividualswithverydifferentincomelevels. 14Foreaseofexposition,Figure3onlyplotsnetrewardsforrewardcards.PanelAofFigureA5inOnline AppendixCadditionallyplotsnetrewardsforclassiccards. Additionally,PanelBofFigureA5showsthe coefficientplotwhichtracesthecoefficientsδF alongsidethe95%confidenceintervalswhenestimating Equation(2)with50insteadof4differentFICObucketsforthethreedifferentincomebuckets,respectively.FigureA6furtherillustratesthemagnitudeofnetrewardsacrossincomepercentiles,showingthat thereisnoclearpatterninnetrewardsacrosstheincomedistribution. 18 Electronic copy available at: https://ssrn.com/abstract=4126641

scores benefit the most from reward credit cards compared to mid- and low-income consumers with high FICO scores. At the lower end of the FICO distribution, however, thispatternisreversed. Onaverage,netrewardsarefarmorenegativeforhigh-income consumerswithlowFICOscoresthanformiddle-andlow-incomeconsumerswithlow FICOscores. [Figure3abouthere] Table V shows that these patterns hold when including the granular set of fixed effects used in the baseline analysis and controlling for card- and cardholder-specific characteristics. Columns (1), (3), and (5) show that net rewards are higher for reward cardsthanforclassiccardsinallincomegroups. Whileaveragenetrewardsareincreasingwithincome,theyremainpositivealsointhebottomtercileoftheincomedistribution($1.9),inconsistentwiththenarrativethatthepoorpayforthepositivenetrewards oftherich. Instead,columns(2),(4),and(6)showthattherelevantredistributionoccurs from low- to high-FICO cardholders, regardless of the income level. In fact, sub-prime cardholdersinthehighestincometercilehavemorenegativenetrewards(-$12.7)than sub-primecardholdersinthemiddle-income(-$4.9)andlow-incometercile(-$2.6),respectively. Bycontrast,primeandsuper-primecardholdersexhibitpositivenetrewards acrossallincomegroups. High-incomesuper-primecardholdersearnonaverage$20.1 innetrewards,whilemiddle-andlow-incomesuper-primecardholdersearnonaverage $13.6and$9.7,respectively. [TableVabouthere] The combined results in Figure 3 and Table V show that, on average, high-income consumerswithhighFICOscoresbenefitfromrewardcreditcardslargelyattheexpense ofhigh-incomeconsumerswithlowFICOscores.15 Hence,ourfindingsarenotprimar- 15TableA4inOnlineAppendixDfurthersubstantiatesthisfindingbyshowingverysimilarresultsfor thetop10%and5%oftheincomedistribution,respectively. 19 Electronic copy available at: https://ssrn.com/abstract=4126641

ily driven by income and therefore inconsistent with a “reverse Robin Hood” mechanism. VI. Credit Card Rewards and Financial Sophistication We next investigate whether our results can be explained by underlying differences in financial sophistication. Financial sophistication refers to the ability of consumers tomakeinformeddecisionsandavoidmistakesintheuseoffinancialproducts(Calvet, Campbell, and Sodini, 2009; Lusardi and Mitchell, 2014). Conversely, low financial sophistication is often linked to behavioral biases, such as over-indebtedness (Meier and Sprenger, 2010; Gathergood, 2012) and sub-optimal repayments (Kuchler and Pagel, 2021). The financial behavior of consumers is reflected in their FICO scores, which are largely based on an individual’s payment history and outstanding debt relative to availablecredit.16 Consequently,individualswithhigher(lower)FICOscoreshavebeen found to incur lower (higher) interest payments, fee payments, and charge-offs (Agarwal, Chomsisengphet, Mahoney, and Stroebel, 2015). FICO scores thus capture the same type of credit card behavior that is associated with a lack of financial sophistication,namelyoverindebtednessandsub-optimalrepaymentbehavior. Therefore,alarge streamoftheexistingliteratureusesFICOscoresasameasureforfinancialsophistication(Agarwal,Rosen,andYao,2016;Amromin,Huang,Sialm,andZhong,2018;Bhutta, Fuster,andHizmo,2021). A. Overindebtedness We first study whether reward cards induce consumers to incur higher levels of unpaid balances relative to classic cards and whether, consistent with the interpretation of FICO scores as a proxy measure for financial sophistication, this effect is stronger 16https://www.myfico.com/credit-education/whats-in-your-credit-score 20 Electronic copy available at: https://ssrn.com/abstract=4126641

for low-FICO cardholders. While there is anecdotal evidence that reward cards induce higher spending and borrowing, causal identification of such an effect is empirically challenging.17 The ideal experiment would randomly assign a reward feature to a classiccardandthentrackchangesincreditcardoutcomesovertime. Weapproximatethis experiment by studying the differential spending and borrowing responses of reward and classic cards to increases in credit card limits and therefore an increase in credit supply(GrossandSouleles,2002;Aydin,2022). We collect all credit cards which received a bank-initiated credit limit increase in March2019,themonthofourcross-sectionalanalysis.18 Wethenobtaindataonspending, repayments, and unpaid balances for these cards in a 1-year time window around thecreditlimitincreaseandcomparetheoutcomechangesofrewardcardstotheoutcomechangesofclassiccardsinastandarddifference-in-differencessetting: (cid:88)(cid:0) (cid:1) (cid:88) (cid:88) ∆Y = δF ×RewardCard ×DF +α + Xm + Xn +ε (3) i(±6m) i z,b i j i F m n Thedependentvariableisthechangeinaveragespending,repayments,orunpaidbalancesbetweenthe6-monthperiodbeforeandthe6-monthperiodafterthecreditlimit increase. We calculate credit card outcomes by aggregating over all cards owned by theindividualwhichreceivedacreditlimitincrease.19 AsinEquation(2), RewardCard takesthevalue1forrewardcardsand0forclassiccards,andDF isasetofFICObucket dummyvariablesforsub-primecardholders(withaFICOscorebelow660),near-prime cardholders(600-720),primecardholders(720-780),andsuper-primecardholders(above 780). WeincludeBank×ZIPcodefixedeffects,thestandardsetofcard-andcardholder- 17Forexample,thepopularcomparisonwebsiteFinderwarnsthat“thepotentialfortravelperks,cash back and bonus points could cause you to spend more than normal, potentially resulting in high fees andinterestonthosepurchases”.Similarly,arecentarticleonnasdaq.comcautionsagainst“consistently overspendinginthehopesofgettingrewards”. 18Ourdatasetallowsustodistinguishbetweencreditlimitincreasesinitiatedbythebankandthose requested by the cardholder. We focus on the former to rule out anticipated changes in spending and borrowing. 19TableA6inOnlineAppendixDprovidesarobustnesscheckwhichonlyconsidersthecardswitha creditlimitincrease,findingqualitativelysimilarresults. 21 Electronic copy available at: https://ssrn.com/abstract=4126641

level control variables, and further income, FICO scores, spending, and payments, all measuredbytheirpre-treatmentaverages.20 TableVIpresentstheestimationresultsofEquation(3)withspending, repayments, andunpaidbalancesasoutcomevariables. Acrossallcardholdersinoursample,wefind that the spending response to a credit limit increase is higher on reward than on similar classic cards (column 1). The difference is economically meaningful and amounts to $76, which corresponds to about 9% of average monthly spending. We also find a differential increase in repayments, albeit smaller in magnitude ($32, column 3). As a result, unpaid balances on reward cards increase compared to similar classic cards ($19),suggestingthatanincreaseincreditlimitsonrewardcardsinducesconsumersto overborrowrelativetoclassiccards. [TableVIabouthere] As before, these average results mask important differences across the FICO distribution. While credit limit increases on reward cards induce all cardholders to spend more,withtheeffectbeinglargerforhigh-FICOconsumers(column2),onlyprimeand super-prime cardholders also increase their repayments (column 4). In contrast, for low-FICOconsumerstheincreaseinpaymentsisstatisticallyinsignificantandcloseto zero in magnitudes. As a result, credit limit increases on reward cards yield a significant increase in unpaid balances for sub-prime ($33.8) and near-prime ($25.3) consumers, while unpaid balances do not change significantly for high-FICO consumers (column6). Theseresultssuggestthatcreditcardrewardsinducesub-andnear-prime consumerstooverspendandsubsequentlyoverborrowontheircreditcards,consistent withtheinterpretationofFICOscoresasameasureforfinancialsophistication(Grubb, 2015;LusardiandTufano,2015). 20AsoursampleisnowlimitedtocardswithacreditlimitincreaseinMarch2019,wecannotestimate themodelwiththesamesetofgranularfixedeffectsusedinthebaselineanalysis,assuchaspecification wouldyieldaverysmallandnon-representativesample. 22 Electronic copy available at: https://ssrn.com/abstract=4126641

B. Sub-Optimal Repayment Behavior Arecentstreamofliteraturefurtherattemptstoquantifythefinancialsophistication ofhouseholdsbymeasuringtheextenttowhichtheymakewell-definedmistakesinthe use of financial products (Calvet, Campbell, and Sodini, 2009; Jørring, 2022). Specifically,wefollowPonce,Seira,andZamarripa(2017)andGathergood,Mahoney,Stewart, and Weber (2019) and calculate the share of misallocated repayments for consumers with multiple credit cards at the same bank.21 This measure can be interpreted as the shareofpaymentsthatwereincorrectlymadeonacheapercardthatshouldhavebeen madeonmoreexpensivecards. We first plot the share of misallocated payments at the borrower level across the FICO distribution, aggregatedover both rewardcards and classiccards. Panel A ofFigure4showsthatmisallocatedpaymentsaredecreasinginFICOscores,consistentwith high-FICOconsumersbeingmorefinanciallysophisticated. PanelBofFigure4further shows that misallocated payments are higher on reward cards, especially for low-FICO consumers. For super-prime cardholders, the misallocated payment share is as low as 6 percent on both reward cards and classic cards. Sub-prime cardholders, in contrast, misallocate up to 14 percent of all credit card repayments on reward cards and around 8percentonclassiccards. [Figure4abouthere] We next estimate Equation (2) with the share of misallocated payments as the outcome variable. Table VII presents the results for this analysis when imposing increasingly stricter sample restriction criteria. In the most restrictive sample in columns (5) and (6), we consider cards with different APRs owned by individuals with at least two cardswithunpaidbalances,whomademinimumpaymentsonallcards,andmorethan 21Theoptimalrepaymentruleistofirstmaketheminimumpaymentdueonallcards,thenpayoffin full the card with the highest APR, and subsequently pay off cheaper cards in order of their APRs. The misallocated payment share is the difference between optimal and actual payments as a share of total payments.WedescribethecalculationofthemisallocatedpaymentshareindetailinOnlineAppendixB. 23 Electronic copy available at: https://ssrn.com/abstract=4126641

the minimum on at least one card. In this sample, we find that the share of misallocated payments is almost 2 percentage points higher on reward than on classic cards (column 5). This result is exclusively driven by low-FICO cardholders. While we find a 4.2 percentage point higher share of misallocated payments on reward cards for subprime cardholders, there is no significant difference between reward and classic cards for prime- and super-prime cardholders. Thus, reward cards do not only induce low- FICOconsumerstooverborrow,butalsotoengageinsub-optimalrepaymentbehavior. These results also hold true when relaxing some of the sample restrictions (columns 1-4).22 [TableVIIabouthere] Finally, we follow Gathergood, Mahoney, Stewart, and Weber (2019) and show that cardholdersfollowasub-optimalbalance-matchingheuristicwhenrepayingtheircredit cards. Rather than optimally allocating repayments across cards based on their APRs, individuals tend to repay their cards proportional to outstanding balances. We calculate the theoretical repayment amount based on three different rules: (i) the optimal repaymentrule,(ii)thebalance-matchingheuristic,and(iii)anequalallocationacross all cards (the 1/N heuristic). As shown in Panel A of Table VIII, actual payments are moststronglycorrelatedwiththebalance-matchingheuristic,inlinewithGathergood, Mahoney, Stewart, and Weber (2019). Again, there is substantial heterogeneity across FICO scores. We find that the correlation between actual payments and the balancematching heuristic is stronger for sub-prime (Panel B) and near-prime (Panel C) cardholders, while prime (Panel D) and super-prime (Panel E) cardholders exhibit repayment behavior most strongly correlated with the optimal allocation rule. Thus, suboptimalrepaymentbehaviortendstobemoresevereforlow-FICOconsumers. [TableVIIIabouthere] 22Resultsarealsorobusttorestrictingthesampletoindividualswithonlytwocards—seeTableA7in OnlineAppendixD. 24 Electronic copy available at: https://ssrn.com/abstract=4126641

Overall, our findings in Section VI are consistent with the hypothesis that reward cardsexploittheover-borrowingandsub-optimalrepaymentbehavioroflow-FICOconsumers and that FICO scores are a reasonable proxy measure for financial sophistication. Ourresultsthereforesuggestthatcreditcardrewardprogramsinducearedistributionfromna¨ıvetosophisticatedconsumers. Thisinterpretationofourresultswarrants some discussion. While we define financial sophistication as the ability of consumers toavoidmistakesintheuseoffinancialproducts(Calvet,Campbell,andSodini,2009), weremainagnosticregardingthesourceofthisability. Alackoffinancialsophistication might therefore reflect individuals’ unawareness about their time-inconsistent preferences(DellaVignaandMalmendier,2004),lowlevelsoffinancialliteracyduetoloweducational attainment (Lusardi and Mitchell, 2014), attentional neglect due to resource scarcity(Shah,Mullainathan,andShafir,2021),oracombinationthereof. Thesefactors all yield a higher propensity for individuals to make financial mistakes, but disentanglingthesefactorsisbeyondthescopeofthispaper. VII. The Banks’ Perspective: Pricing and Profits Our analysis so far focuses on the perspective of cardholders. In this section, we investigate the perspective of banks and study both their pricing strategies and profits inthecreditcardmarket,bothacrosscardtypesandacrosstheFICOdistribution. A. Pricing Wefirststudytheinterestratesofferedbybanksonrewardcardsrelativetocomparable classic cards. Panel A of Figure 5 shows that the average annual percentage rate 25 Electronic copy available at: https://ssrn.com/abstract=4126641

(APR) of interest on reward cards is systematically lower than interest rates on classic cardsacrosstheentireFICOdistribution.23 [Figure5abouthere] Thispatternisconfirmedinourstandardregressionsetting,estimatingEquation(2) withAPRsastheoutcomevariable. Columns(1)and(2)ofTableIXpresenttheresults. Acrossallcardholders,APRsonrewardcardsareonaverage1.0percentagepointslower than on comparable classic cards. This interest rate differential between reward and classic cards is larger for high- than for low-FICO cardholders. For sub-prime cardholders, banks on average offer 0.2 percentage points lower interest rates on reward cards, while for super-prime cardholders the difference is 1.7 percentage points. This evidenceindicatesthatbanksincentivizeconsumerstoadoptrewardcardsbyoffering betterpricingterm. [TableIXabouthere] B. Bank Profits Atprimafacie,offeringlowerinterestratesonrewardcardsthanoncomparableclassiccardstoincreasethenumberofrewardcardsmaynotappearasaprofit-maximizing strategy. However,theevidenceonhigherinterestandfeechargesforrewardcards(Figure2)suggeststhat,evenifwithlowerprices,theseproductscouldgeneratemoreprofitsforbanks. Toinvestigatemoreformallyhowthispricingstrategytranslatesintoprofitability,wedefineabank’sprofitoncreditcardias: Profit = InterestPaid +TotalFees +InterchangeIncome (4) i i i i −Rewards −RealizedCharge-Offs −WACC×UnpaidBalances (5) i i i 23Giventhatallcreditcardaccountsinthesampleareinitiatedatleast12monthspriortoMarch2019, thelowerAPRonrewardcardsrelativetoclassiccardsdoesnotreflectzeroorlowAPRsduringpotential promotionalperiods. 26 Electronic copy available at: https://ssrn.com/abstract=4126641

The variables Interest Paid, Total Fees, and Rewards are defined as in Section III. Whereas interest and fees represent payments from the cardholder’s perspective, they represent income from the bank’s perspective. Conversely, whereas rewards represent income from the cardholder’s perspective, they represent costs from the bank’s perspective. Our analysis of bank profitability also introduces three new terms which are not included in the previous analysis: Interchange Income, Realized Charge-Offs, and WACC×Unpaid Balances. As discussed in Section II, when consumers pay with their creditcard,bankschargeaninterchangefeefromthemerchantacquirer,whichgenerallyrangesfrom1to3percentofthepurchaseprice(GAO,2009). Weassessinterchange income at the card level to be 1.5 percent of the purchase volume for classic cards and 2.5 percent for reward cards. Realized charge-offs are an expense incurred by the bank onaccountsthatremaindelinquentfor180daysandforwhichtheoutstandingbalance cannolongerbeconsideredanassetonthebalancesheet(CFPB,2019). Fromthecardholder’s perspective, charge-offs do not matter for the net cash flow on a credit card. Fromabank’sperspective,however,realizedcharge-offsareanimportantdeterminant of the ex-post profitability of an account and we therefore include them in the definition of banks’ profits. The third term captures banks’ cost of financing revolving credit card balances. We assess these costs at a conservative 5 percent weighted average cost ofcapital(WACC). Panel B of Figure 5 shows that bank profits are hump-shaped in FICO scores and substantially higher on reward than on classic cards across the entire FICO distribution. Columns(3)and(4)ofTableIXpresenttheestimationresultsofEquation(2)with bankprofitsastheoutcomevariable. Acrossallcardholders,bankprofitsareabout$7.4 higher on reward cards than on comparable classic cards. While banks profit from reward cards across the entire FICO distribution, profits are not uniformly distributed, as shown in column (4). We find that bank profits per card are highest for near-prime ($15.3) and prime ($9.0) cardholders in the middle of the FICO distribution. For sub- 27 Electronic copy available at: https://ssrn.com/abstract=4126641

prime cardholders, which tend to incur the highest charge-offs, profits are also higher onrewardscards,butwiththedifferentialbeingsmallerinmagnitude($4.1). Forsuperprime cardholders, which tend to earn a lot of rewards and also incur low interest payments,bankprofitsareonly$1.3higheronrewardthanonclassiccards. Thus,fromthe banks’ perspective, near-prime and prime cardholders are the largest source of profits inthemarketforrewardcreditcards. There are also substantial differences in banks’ sources of revenue across the FICO distribution. Figure 6 illustrates the average revenue share of interest income, fee income, and interchange income as a percentage of total card revenue across the FICO distribution. For low-FICO cardholders, banks’ revenues largely stem from interest income. Forhigh-FICOcardholders,ontheotherhand,banks’revenueslargelystemfrom interchangeincome. Feeincomerepresentsthesmallestrevenuesourceofbanksacross theFICOdistribution. [Figure6abouthere] VIII. The Geography of Net Rewards Our analysis so far focuses on the redistribution from na¨ıve to sophisticated consumers at the individual level. In this section, we focus on the aggregate implications andanalyzethereward-inducedredistributionacrossregionsintheUnitedStates. Figure7plotstheaveragenetreward(PanelA)andtheaverageFICOscore(PanelB) across counties. The figure illustrates the high level of spatial correlation between the two variables and confirm, at the aggregate level, the redistribution from na¨ıve to sophisticatedconsumersinthecreditcardmarket. Regionswithhighaveragenetrewards (thenortheast,thenorth,andthewestcoast)tendtoberegionswithhighaverageFICO scores. Conversely,regionswithlowaveragenegativenetrewards(thesouth)tendtobe regionswithlowaverageFICOscores. 28 Electronic copy available at: https://ssrn.com/abstract=4126641

[Figure7abouthere] A relevant concern is whether this redistribution is penalizing areas with specific socio-demographiccharacteristics, potentiallywideningexistingspatialdisparities. To answer this question we regress card-level net rewards on various ZIP code-level characteristicsandestimatethefollowingregressionspecification: (cid:88) NetReward = βkXk +γ ×CreditScore +ε (6) i,z z z i,z k wheretheoutcomevariableisthenetrewardofcardiinZIPcodezandwhereXkarethe z followingZIPcode-levelcharacteristics: i)thepercentageofresidentswithahighschool diploma (but no more), as a measure for low educational attainment; ii) the median individualincome;andiii)thepercentageofresidentswhoreporttheirraceasBlackor African American. Since these socio-demographic characteristics are likely correlated withaverageFICOscore, wereportallcoefficientswithandwithoutcontrollingforthe averageFICOscoreinZIPcodez. As shown in columns 1,3, and 5 of Table X, higher net rewards are associated with a higherlevelofeducationalattainment, withahighermedianincome, andwithalower share of Black residents. These results suggest that credit card rewards are a potential channel that can exacerbate existing socio-economic disparities across regions in the UnitedStates,astheyimplyatransferfromlesstomoreeducated,frompoorertoricher, and from high- to low-minority areas, thereby widening existing spatial disparities.24 Columns 2,4, and 6 illustrate that all coefficients become statistically insignificant and close to zero in magnitude when controlling for a ZIP code’s average FICO score, indicatingthatdifferencesinfinancialsophisticationaretheunderlyingmechanismdriving ourgeographicalresults. 24Although FICO scores and income are only moderately correlated, as discussed in Section V, high FICOscoresarestillmoreprevalentamonghigh-incomecardholders,asshowninFigureA4.Thus,while ourcard-levelresultsarenotdrivenbydifferencesinincome,westillfindapositivecorrelationbetween netrewardsandincomeinouraggregateZIPcode-levelanalysis. 29 Electronic copy available at: https://ssrn.com/abstract=4126641

[TableXabouthere] IX. Conclusion Creditcardrewardprograms provideanideallaboratorytostudythe redistribution across consumers in retail financial markets. Using comprehensive and granular data from the Federal Reserve’s Y-14M reports, we find that high-FICO consumers benefit from reward programs at the expense of low-FICO consumers and estimate an annual redistribution of of $15.1 billion. This redistribution is driven by both the cost and the benefit margin of reward credit cards. Super-prime and prime consumers spend more moneyandthusearnhigherrewards,buttheyalsopaybacktheirbalancesintimeand thusincurlowerinterestpayments. Conversely,sub-primeandnear-primeconsumers earn lower rewards and incur higher interest payments due to higher outstanding balancesonrewardcards. Notably, our results are not driven by income, as they hold within the sub-samples of low-, middle- and high-income individuals. In particular, high-FICO high-income consumers benefit the most from reward credit cards, but they do so at the expense of low-FICO high-income consumers. While credit card rewards are often framed as a “reverseRobinHood”mechanisminwhichthepoorsubsidizetherich,ourresultsshow thatthisexplanationisatbestincomplete. Werationalizeourfindingsintermsoffinancialsophistication,meaningthatreward cardsconstitutearedistributionfromna¨ıvetosophisticatedconsumers. Wearguethat FICOscorescanbeinterpretedasameasureoffinancialsophisticationand,consistent withthat, weshowthatFICOscoresarecorrelatedwithconsumers’financialmistakes. First,weprovidequasi-experimentalevidencethatrewardcreditcardsinducelow-FICO consumers to overborrow on their credit cards. Second, we show that FICO scores are 30 Electronic copy available at: https://ssrn.com/abstract=4126641

strongly correlated with the share of misallocated credit card payments, especially for sub-primeandnear-primecardholders. We further show that banks incentivize consumers to use reward cards by offering lower interest rates than on comparable classic cards. Banks profits from reward cards arehighestfornear-primeandprimeconsumersinthemiddleoftheFICOdistribution. We conclude by documenting that the costs and benefits of credit card rewards are unequallydistributedacrossgeographiesintheUnitedStates. Creditcardrewardstransfer income from less to more educated, from poorer to richer, and from high- to lowminorityareas,therebywideningexistingspatialdisparities. 31 Electronic copy available at: https://ssrn.com/abstract=4126641

REFERENCES Agarwal, Sumit, Sujit Chakravorti, and Anna Lunn, 2010, Why do banks reward their customerstousetheircreditcards?, FederalReserveBankofChicagoWorkingPaper 2010-19. Agarwal, Sumit, Souphala Chomsisengphet, Neale Mahoney, and Johannes Stroebel, 2015,Regulatingconsumerfinancialproducts: Evidencefromcreditcards,Quarterly JournalofEconomics130,111–164. Agarwal,Sumit,RichardJ.Rosen,andVincentYao,2016,Whydoborrowersmakemortgagerefinancingmistakes?,ManagementScience62,3494–3509. Amromin, Gene, Jennifer Huang, Clemens Sialm, and Edward Zhong, 2018, Complex mortgages,ReviewofFinance22,1975–2007. Aydin, Deniz, 2022, Consumption Response to Credit Expansions: Evidence from ExperimentalAssignmentof45,307CreditLines,AmericanEconomicReview112,1–40. Bach, Laurent, Laurent E. Calvet, and Paolo Sodini, 2020, Rich Pickings? Risk, Return, andSkillinHouseholdWealth,AmericanEconomicReview110,2703–2747. Beer, Rachael, Felicia Ionescu, and Geng Li, 2018, Are income and credit scores highly correlated?,FEDSNotesAugust13,2018. Bhutta, Neil, Andreas Fuster, and Aurel Hizmo, 2021, Paying too much? Borrower sophistication and overpayment in the us mortgage market, CEPR Discussion Paper DP14924. Calvet, Laurent E., John Y. Campbell, and Paolo Sodini, 2009, Measuring the financial sophisticationofhouseholds,AmericanEconomicReview99,393–398. Campbell,JohnY.,2006,Householdfinance,JournalofFinance61,1553–1604. 32 Electronic copy available at: https://ssrn.com/abstract=4126641

Campbell,JohnY.,2016,Restoringrationalchoice: Thechallengeofconsumerfinancial regulation,AmericanEconomicReview: Papers&Proceedings106,1–30. Campbell, John Y., and Joao F. Cocco, 2003, Household risk management and optimal mortgagechoice,QuarterlyJournalofEconomics118,1449–1494. Campbell, John Y., Tarun Ramadorai, and Benjamin Ranish, 2019, Do the Rich Get Richer in the Stock Market? Evidence from India, American Economic Review: Insights1,225–240. CFPB,2019,Theconsumercreditcardmarket,Report. Ching,AndrewT.,andFumikoHayashi,2010,Paymentcardrewardsprogramsandconsumerpaymentchoice,JournalofBanking&Finance34,1773–1787. DellaVigna, Stefano, and Ulrike Malmendier, 2004, Contract design and self-control: Theoryandevidence,QuarterlyJournalofEconomics119,353–402. Derby,MichaelS.,2010,Creditcardstakefrompoor,givetotherich,WallStreetJournal July27,2010,https://www.wsj.com/articles/BL-REB-11033. Deuflhard, Florian, Dimitris Georgarakos, and Roman Inderst, 2019, Financial Literacy andSavingsAccountReturns,JournaloftheEuropeanEconomicAssociation17,131– 164. Fagereng, Andreas, Luigi Guiso, Davide Malacrino, and Luigi Pistaferri, 2020, HeterogeneityandPersistenceinReturnstoWealth,Econometrica88,115–170. Felt, Marie-He´le`ne, Fumiko Hayashi, Joanna Stavins, and Angelika Welte, 2020, Distributional Effects of Payment Card Pricing and Merchant Cost Pass-through in the United States and Canada, Federal Reserve Bank of Kansas City, Research Working Paper20-18. 33 Electronic copy available at: https://ssrn.com/abstract=4126641

Fisher, Jack, Alessandro Gavazza, Lu Liu, Tarun Ramadorai, and Jagdish Tripathy, 2021, Refinancing cross-subsidies in the mortgage market, Bank of England Staff Working Paper948. Foster, Kevin, Claire Greene, and Joanna Stavins, 2021, The 2020 survey of consumer payment choice: Summary results, Federal Reserve Bank of Atlanta Research Data Report. Gabaix, Xavier, and David Laibson, 2006, Shrouded attributes, consumer myopia, and informationsuppressionincompetitivemarkets,QuarterlyJournalofEconomics121, 505–540. GAO, 2009, Government accountability office: Rising interchange fees have increased costs for merchants, but options for reducing fees pose challenges, Report to CongressionalAddressees. Gathergood, John, 2012, Self-control, financial literacy and consumer overindebtedness,JournalofEconomicPsychology33,590–602. Gathergood,John,NealeMahoney,NeilStewart,andJo¨rgWeber,2019,Howdoindividuals repay their debt? The balance-matching heuristic, American Economic Review 109,844–875. Gomes, Francisco, Michael Haliassos, and Tarun Ramadorai, 2021, Household finance, JournalofEconomicLiterature59,919–1000. Gross, David B., and Nicholas S. Souleles, 2002, Do liquidity constraints and interest rates matter for consumer behavior? Evidence from credit card data, Quarterly JournalofEconomicsVolume117,149–185. Grubb, Michael D., 2015, Overconfident Consumers in the Marketplace, Journal of EconomicPerspectives29,9–36. 34 Electronic copy available at: https://ssrn.com/abstract=4126641

Guiso, Luigi, Andrea Pozzi, Anton Tsoy, Leonardo Gambacorta, and Paolo Emilio Mistrulli, 2021, The cost of steering in financial markets: Evidence from the mortgage market,JournalofFinancialEconomics143,1209–1226. Hayashi, Fumiko, 2009, Do U.S. consumers really benefit from payment card rewards?, EconomicReview94,37–63. Heidhues, Paul, and Botond Ko˝szegi, 2010, Exploiting na¨ıvete about self-control in the creditmarket,AmericanEconomicReview100,2279–2303. Heidhues, Paul, and Botond Ko˝szegi, 2017, Na¨ıvete´-based discrimination, Quarterly JournalofEconomics132,1019–1054. Jørring, Adam, 2022, Financial sophistication and consumer spending, Journal of Financeforthcoming. Kuchler, Theresa, and Michaela Pagel, 2021, Sticking to your plan: The role of present biasforcreditcardpaydown,JournalofFinancialEconomics139,359–388. Levitin, Adam J., 2008, The Social Costs of Credit Card Merchant Restraints, Harvard JournalonLegislation45,1–58. Lusardi, Annamaria, Pierre-Carl Michaud, and Olivia S. Mitchell, 2017, Optimal financialknowledgeandwealthinequality,JournalofPoliticalEconomy125,431–477. Lusardi,Annamaria,andOliviaS.Mitchell,2014,Theeconomicimportanceoffinancial literacy: Theoryandevidence,JournalofEconomicLiterature52,5–44. Lusardi, Annamaria, and Peter Tufano, 2015, Debt literacy, financial experiences, and overindebtedness,JournalofPensionEconomicsandFinance14,332–368. Meier,Stephan,andCharlesSprenger,2010,Present-biasedpreferencesandcreditcard borrowing,AmericanEconomicJournal: AppliedEconomics2,193–210. 35 Electronic copy available at: https://ssrn.com/abstract=4126641

Ponce, Alejandro, Enrique Seira, and Guillermo Zamarripa, 2017, Borrowing on the wrong credit card? Evidence from Mexico, American Economic Review 107, 1335– 1361. Sarin,Natasha,2019,Makingconsumerfinancework,ColumbiaLawReview119,1519– 1596. Schuh, Scott, Oz Shy, and Joanna Stavins, 2010, Who gains and who loses from credit card payments? Theory and calibrations, Federal Reserve Bank of Boston Public PolicyDiscussionPapers10-03. Shah, Anuj K., Sendhil Mullainathan, and Eldar Shafir, 2021, Some Consequences of HavingTooLittle,Science338,682–685. Shy,Oz,andZhuWang,2011,Whydopaymentcardnetworkschargeproportionalfees?, AmericanEconomicReview101,1575–1590. Stewart, Emily, 2021, The ugly truth behind your fancy rewards credit card, vox.com June 3, 2021, https://www.vox.com/the-goods/22454885/who-pays-for-credit-cardrewards. 36 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure1. NetRewardsAcrossFICOScorePercentiles. This figure illustrates the dollar magnitude of average net rewards across the FICO distribution, separately for reward cards(solidredline)andclassiccards(dashedblueline). Foreachcardtype,weplotthe averagenetrewardfor100equal-sizedFICObucketsbetween480and830. Thedashed vertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime, prime, and super-prime cardholders, respectively. The graph is based on our baseline sampleof238millioncreditcardsinMarch2019. sdraweR teN ylhtnoM egarevA 01 0 01- 02- 03- 04- 480 660 720 780 830 Credit Score Reward Cards Classic Cards 37 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure 2. Net Reward Components Across FICO Score Percentiles. This figure illustrates the dollar magnitude of average rewards (Panel A) and interest charges (Panel B) across the FICO distribution, separately for reward cards (solid red line) and classic cards (dashed blue line). For each card type, we plot the average reward and interest charges for 100 equal-sized FICO buckets between 480 and 830. The dashed vertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime, prime, and super-prime cardholders, respectively. The graph is based on our baseline sample of238millioncreditcardsinMarch2019. (A)Rewards sdraweR ylhtnoM egarevA 51 01 5 0 480 660 720 780 830 Credit Score Reward Cards Classic Cards (B)InterestCharges segrahC tseretnI ylhtnoM egarevA 04 03 02 01 0 480 660 720 780 830 Credit Score Reward Cards Classic Cards 38 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure 3. Net Rewards Across FICO Score Percentiles by Income Groups. This figure illustratesthedollarmagnitudeofaveragenetrewardsonrewardcardsacrosstheFICO distributionbyincomegroups. Theredlineplotstheaveragenetrewardforborrowers withanannualincomebelow44thousand,theyellowlineforborrowerswithanannual income between 44 thousand and 79 thousand, and the green line for borrowers with anannualincomeabove79 thousand. Foreachincomegroup, weplottheaveragenet reward (in dollar) for 100 equal-sized FICO buckets between 480 and 830. The dashed vertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime, prime, and super-prime cardholders, respectively. The graph is based on our baseline sampleof238millioncreditcardsinMarch2019. sdraweR teN ylhtnoM egarevA 02 0 02- 04- 06- 480 660 720 780 830 Credit Score Low Income Middle Income High Income 39 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure 4. Share of Misallocated Payments Across FICO Score Percentiles. This figure illustrates the average percentage share of misallocated payments across the FICO distribution at the borrower level (Panel A) and separately for reward cards (solid red line) and classic cards (dashed blue line) (Panel B). In each panel, we plot the average shareofmisallocatedpaymentsfor100equal-sizedFICObucketsbetween480and830. The dashed vertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime, prime, and super-prime cardholders, respectively. The graph is based on oursampleof34millioncreditcardsofborrowerswhoholdmultiplecreditcardsatthe samebankinMarch2019. (A)BorrowerLevel )% ni( stnemyaP detacollasiM fo erahS 21 01 8 6 4 480 660 720 780 830 Credit Score (B)RewardCardsversusClassicCards )%( stnemyaP detacollasiM fo erahS egarevA 41 21 01 8 6 480 660 720 780 830 Credit Score Reward Cards Classic Cards 40 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure5. APRsandBankProfitsAcrossFICOScorePercentiles. This figure illustrates the average annual percentage rate (APRs) (Panel A) and the average dollar magnitude of bank profits per card (Panel B) across the FICO distribution, separately for reward cards (solid red line) and classic cards (dashed blue line). For each card type, we plot theaverageAPRandbankprofitfor100equal-sizedFICObucketsbetween480and830. The dashed vertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime, prime, and super-prime cardholders, respectively. The graph is based on ourbaselinesampleof238millioncreditcardsinMarch2019. (A)APRs RPA dethghieW egarevA 42 22 02 81 61 480 660 720 780 830 Credit Score Reward Cards Classic Cards (B)BankProfits tiforP knaB egarevA 03 02 01 0 01- 480 660 720 780 830 Credit Score Reward Cards Classic Cards 41 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure 6. Bank Revenue Shares Across FICO Score Percentiles. This figure illustrates the average bank revenue share across the FICO distribution for 100 equal-sized FICO buckets between 300 and 850, separately for reward cards (Panel A) and classic cards (Panel B). For each card type, we plot the share of interchange income (black), fee income(darkgray),andinterestincome(lightgray)asapercentageoftotalcardrevenue. The dashed vertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime,prime,andsuper-primecardholders,respectively. Thegraphsarebasedon ourbaselinesampleof238millioncreditcardsinMarch2019. (A)Rewardcards 100 80 60 40 20 0 )% ni( erahS euneveR knaB 480 660 720 780 830 Credit Score Intercharge income Fee income Interest income (B)Classiccards 100 80 60 40 20 0 )% ni( erahS euneveR knaB 480 660 720 780 830 Credit Score Intercharge income Fee income Interest income 42 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure 7. The Geography of Net Rewards and FICO Scores. This figure illustrates the average dollar amount of net rewards (Panel A) and the average FICO score (Panel B) across counties in the United States. The graph is based on our baseline sample of 238 millioncreditcardsinMarch2019. (A)AverageNetRewardsAcrossCounties -14.74 - 0.76 -17.05 - -14.74 -18.82 - -17.05 -20.69 - -18.82 -23.01 - -20.69 -46.83 - -23.01 No data (B)AverageFICOScoresAcrossCounties 750.9 - 774.3 745.8 - 750.9 740.3 - 745.8 733.7 - 740.3 726.0 - 733.7 687.2 - 726.0 No data 43 Electronic copy available at: https://ssrn.com/abstract=4126641

TableI.SummaryStatistics This table presents card-level summary statistics as of March 2019, for all call cards in our sample (Columns 1 to 3), and separately for reward and classic cards (Columns 4 and 5). Panel A presents variables related to the calculation of net rewards(asdescribedinSectionA).PanelBpresentsothercard-leveloutcomeandcontrolvariablesusedinouranalysis. AllCards RewardCards ClassicCards (1) (2) (3) (4) (5) Mean Median SD Mean Mean PanelA.NetRewardVariables Rewards(in$) 4.69 0.00 20.42 9.30 0.00 InterestCharges(in$) 14.38 0.00 37.91 18.34 10.36 FeeCharges(in$) 2.64 0.00 11.01 3.33 1.93 NetRewards(in$) -12.33 0.00 44.41 -12.37 -12.29 PanelB.OtherVariables APR(in%) 20.63 21.49 7.15 18.64 22.64 BankProfits(in$) 14.53 1.11 232.94 22.54 6.39 FICOScore 729.60 742.00 75.65 743.22 715.77 BorrowerIncome(in$k) 88.44 60.00 1863.36 98.02 78.71 CreditLimit(in$k) 7.37 5.00 7.90 10.42 4.28 AmountPastDue(in$) 10.26 0.00 172.45 8.19 12.37 AgeofCard(inyears) 7.43 4.83 7.36 7.61 7.24 JointAccount(0/1) 0.02 0.00 0.15 0.03 0.02 FraudFlag(0/1) 0.00 0.00 0.06 0.00 0.00 DepositRelationshipWithSameBank(0/1) 0.19 0.00 0.39 0.28 0.10 LendingRelationshipWithSameBank(0/1) 0.08 0.00 0.27 0.11 0.05 No. CardsWithSameBank(0/1) 2.11 2.00 1.25 1.89 2.34 WorkoutProgram(0/1) 0.01 0.00 0.07 0.00 0.01 BankruptcyFlag(0/1) 0.00 0.00 0.05 0.00 0.00 Observations 237,573,278 119,730,353 117,842,925 Electronic copy available at: https://ssrn.com/abstract=4126641 44

TableII.NetRewards: BaselineResults Thistablepresentstheestimationresultsfordifferencesinnetrewardsbetweenreward cards and classic cards from Equation (2) in Section IV.A, where the outcome variable isthenetrewardofcardiasdefinedinEquation(1)inSectionIII.ThevariableReward Card takes on the value of 1 if card i is a reward card, and 0 otherwise. Cards are clusteredinthefollowingFICOscoregroups: sub-prime(below660),near-prime(660-720), prime (720-780), and super-prime (above 780). Card controls include the credit limit, theamountpastdue,thecardage,ajointaccountindicator,afraudflagindicator,and a workout program indicator. Cardholder controls a deposit relationship indicator, a lending relationship indicator, thenumber ofcards held bythe cardholderat thesame bank, and a bankruptcy indicator. Borrower income and FICO scores are defined as of March2018i.e.,oneyearpriortotheoutcomevariable. Standarderrorsareclusteredat thebank-statelevel. *,**,and***indicatestatisticalsignificanceatthe10%,5%,and1% levels,respectively. NetRewards (1) (2) (3) (4) RewardCard 4.66*** 3.88*** 3.48*** (0.30) (0.37) (0.38) RewardCard×Sub-Prime -5.37*** (0.67) RewardCard×Near-Prime -6.80*** (0.69) RewardCard×Prime 7.28*** (0.44) RewardCard×Super-Prime 16.05*** (0.93) CardControls Y Y Y Y CardholderControls Y Y Y Y FE:Bank×Zip×Income Y N - - FE:Bank×Zip×FICO N Y - - FE:Bank×Zip×Income×FICO N N Y Y Observations 237,573,278 45 Electronic copy available at: https://ssrn.com/abstract=4126641

TableIII.NetRewards: RobustnessTests Thistablepresentsrobustnesschecksfortheestimationresultsfordifferencesinnetrewardsbetweenrewardcardsand classic cards. The outcome variable is the net reward of card i as defined in Equation (1) in Section III. The variable RewardCard takesonthevalueof1ifcardiisarewardcard, and0otherwise. CardsareclusteredinthefollowingFICO scoregroups: sub-prime(below660),near-prime(660-720),prime(720-780),andsuper-prime(above780). Cardcontrols includethecreditlimit,theamountpastdue,thecardage,ajointaccountindicator,afraudflagindicator,andaworkout program indicator. Cardholder controls a deposit relationship indicator, a lending relationship indicator, the number of cardsheldbythecardholderatthesamebank,andabankruptcyindicator. BorrowerincomeandFICOscoresaredefined asofMarch2018i.e.,oneyearpriortotheoutcomevariable. Columns1and2additionallyincludecreditlimitpercentile andAPRpercentilefixedeffects. Columns3and4estimateourbaselinespecificationfromEquation(2)onthesampleof columns1and2. Columns3and4includecardholderfixedeffects. Standarderrorsareclusteredatthebank-statelevel. *,**,and***indicatestatisticalsignificanceatthe10%,5%,and1%levels,respectively. NetRewards (1) (2) (3) (4) (5) (6) RewardCard 0.62*** 1.94*** 1.77*** (0.15) (0.51) (0.37) RewardCard×Sub-Prime -0.49*** -1.02*** -5.53*** (0.09) (0.16) (1.07) RewardCard×Near-Prime -0.95*** -1.79*** -8.53*** (0.35) (0.53) (0.96) RewardCard×Prime 1.20*** 2.89*** 4.08*** (0.30) (0.44) (0.47) RewardCard×Super-Prime 2.62*** 6.50*** 14.09*** (0.34) (1.20) (1.03) CardControls Y Y Y Y Y Y CardholderControls Y Y Y Y - - FE:Bank×Cardholder - - - - Y Y FE:Bank×Zip×Income×FICO - - Y Y - - FE:Bank×Zip×Income×FICO×Limit×APR Y Y - - - - Observations 12,381,801 65,513,743 Electronic copy available at: https://ssrn.com/abstract=4126641 46

TableIV.NetRewardComponents This table presents the estimation results for differences in net reward components between reward cards and classic cards from Equation (2) in Section IV.A. The outcome variables are the dollar amount of rewards (columns 1 an 2), the dollar amount of interest charges (column 3 and 4), and the dollar amount of total fee charges (column 5 and 6). The variableRewardCard takesonthevalueof1ifcardiisarewardcard,and0otherwise. Cardsareclusteredinthefollowing FICO score groups: sub-prime (below 660), near-prime (660-720), prime (720-780), and super-prime (above 780). Card controls include the credit limit, the amount past due, the card age, a joint account indicator, a fraud flag indicator, and a workout program indicator. Cardholder controls a deposit relationship indicator, a lending relationship indicator, the numberofcardsheldbythecardholderatthesamebank,andabankruptcyindicator. BorrowerincomeandFICOscores are defined as of March 2018 i.e., one year prior to the outcome variable. Standard errors are clustered at the bank-state level. *,**,and***indicatestatisticalsignificanceatthe10%,5%,and1%levels,respectively. Rewards InterestCharges TotalFeeCharges (1) (2) (3) (4) (5) (6) RewardCard 6.38*** 2.20*** 0.70*** (0.35) (0.18) (0.08) RewardCard×Sub-Prime 1.79*** 6.38*** 0.78*** (0.14) (0.69) (0.10) RewardCard×Near-Prime 4.83*** 10.86*** 0.78*** (0.27) (0.75) (0.12) RewardCard×Prime 8.39*** 0.34 0.77*** (0.31) (0.24) (0.08) RewardCard×Super-Prime 9.45*** -7.09*** 0.50*** (0.38) (0.64) (0.06) CardControls Y Y Y Y Y Y CardholderControls Y Y Y Y Y Y FE:Bank×Zip×Income×FICO Y Y Y Y Y Y Observations 237,573,278 Electronic copy available at: https://ssrn.com/abstract=4126641 47

TableV.NetRewardsbyIncomeGroups This table presents the estimation results for differences in net rewards between reward cards and classic cards from Equation (2) in Section IV.A, estimated separately for three different income groups: low-income cardholders with an annualincomebelow$44thousand;middle-incomecardholderswithanannualincomebetween$44-79thousand;and high-income cardholders with an annual income above $79 thousand. The outcome variable is the net reward of card i as defined in Equation (1) in Section III. The variable Reward Card takes on the value of 1 if card i is a reward card, and 0 otherwise. Cards are clustered in the following FICO score groups: sub-prime (below 660), near-prime (660-720), prime (720-780), and super-prime (above 780). Card controls include the credit limit, the amount past due, the card age, a joint account indicator, a fraud flag indicator, and a workout program indicator. Cardholder controls a deposit relationshipindicator,alendingrelationshipindicator,thenumberofcardsheldbythecardholderatthesamebank,and abankruptcyindicator. BorrowerincomeandFICOscoresaredefinedasofMarch2018i.e.,oneyearpriortotheoutcome variable. Standarderrorsareclusteredatthebank-statelevel. *,**,and***indicatestatisticalsignificanceatthe10%,5%, and1%levels,respectively. LowerTercileof MiddleTercileof UpperTercileof IncomeDistribution IncomeDistribution IncomeDistribution (1) (2) (3) (4) (5) (6) RewardCard 1.86*** 2.73*** 5.36*** (0.20) (0.28) (0.61) RewardCard×Sub-Prime -2.56*** -4.88*** -12.75*** (0.34) (0.59) (1.18) RewardCard×Near-Prime -2.36*** -5.80*** -13.15*** (0.45) (0.58) (0.77) RewardCard×Prime 5.93*** 6.29*** 8.70*** (0.33) (0.37) (0.58) RewardCard×Super-Prime 9.71*** 13.60*** 20.10*** (0.60) (0.71) (1.03) CardControls Y Y Y Y Y Y CardholderControls Y Y Y Y Y Y FE:Bank×Zip×Income×FICO Y Y Y Y Y Y Observations 75,159,536 79,540,729 82,873,013 Electronic copy available at: https://ssrn.com/abstract=4126641 48

TableVI.Overindebtedness: Difference-in-DifferencesAnalysis Thistablepresentstheestimationresultsforthedifference-in-differencesregressioninEquation(3)inSectionVI.A.Wecomparechangesincredit cardoutcomesofconsumerswhoreceivedabank-initiatedcreditlimitincreaseonrewardcardstothosewhoreceivedalimitincreaseonclassic cardsinatimewindow6monthsbeforeandafterthecreditlimitincrease. Theoutcomevariablesarechangesinspendingvolumes(columns 1and2), creditcardpayments(columns3and4), andunpaidbalances(columns5and6). Theanalysisconsidersallcardsofconsumerswho receivedabank-initiatedcreditlineincreasehas. ThevariableRewardCard takesonthevalueof1ifcardiisarewardcard, and0otherwise. CardsareclusteredinthefollowingFICOscoregroupsD:sub-prime(below660),near-prime(660-720),prime(720-780),andsuper-prime(above 780).CardcontrolsincludetheFICOscore,thecreditlimit,theamountpastdue,thecardage,ajointaccountindicator,afraudflagindicator,and aworkoutprogramindicator.Cardholdercontrolsincome,adepositrelationshipindicator,alendingrelationshipindicator,thenumberofcards held by the cardholder at the same bank, a bankruptcy indicator, and average spending and payments in the pre-treatment period. Borrower incomeandFICOaredefinedasofMarch2018i.e., oneyearprior. Standarderrorsareclusteredatthebank-statelevel. *, **, and***indicate statisticalsignificanceatthe10%,5%,and1%levels,respectively. ∆Spending ∆Payments ∆UnpaidBalances (1) (2) (3) (4) (5) (6) RewardCard 75.77*** 31.96*** 19.17** (6.83) (3.72) (8.79) RewardCard×Sub-Prime 59.75*** 5.06 33.82*** (6.43) (3.12) (11.24) RewardCard×Near-Prime 62.88*** 4.53 25.25* (7.18) (4.29) (13.53) RewardCard×Prime 89.03*** 73.19*** 4.83 (7.98) (6.17) (12.16) RewardCard×Super-Prime 164.85*** 153.22*** -28.20 (14.14) (13.22) (25.26) CardControls(Pre-Period) Y Y Y Y Y Y CardholderControls(Pre-Period) Y Y Y Y Y Y IncomeandFICO(Pre-Period) Y Y Y Y Y Y SpendingandPayments(Pre-Period) Y Y Y Y Y Y FE:Bank×Zip Y Y Y Y Y Y Observations 1,236,604 Electronic copy available at: https://ssrn.com/abstract=4126641 49

TableVII.ShareofMisallocatedPayments Thistablepresentstheestimationresultsfordifferencesintheshareofmisallocatedpayments(asdefinedinEquationA5 inSection B)betweenrewardcardsandclassiccardsfromEquation(2)inSectionIV.A.ThevariableRewardCardtakeson thevalueof1ifcardiisarewardcard,and0otherwise. CardsareclusteredinthefollowingFICOscoregroups: sub-prime (below660),near-prime(660-720),prime(720-780),andsuper-prime(above780). Cardcontrolsincludethecreditlimit, the amount past due, the card age, a joint account indicator, a fraud flag indicator, and a workout program indicator. Cardholder controls a deposit relationship indicator, a lending relationship indicator, the number of cards held by the cardholderatthesamebank,andabankruptcyindicator. BorrowerincomeandFICOscoresaredefinedasofMarch2018 i.e., one year prior to the outcome variable. Standard errors are clustered at the bank-state level. *, **, and *** indicate statisticalsignificanceatthe10%,5%,and1%levels,respectively. ShareofMisallocatedPayments (1) (2) (3) (4) (5) (6) RewardCard 1.24*** 1.71*** 1.74*** (0.28) (0.33) (0.37) RewardCard×Sub-Prime 2.65*** 3.74*** 4.18*** (0.20) (0.25) (0.30) RewardCard×Near-Prime 0.76*** 1.15*** 1.08*** (0.28) (0.34) (0.35) RewardCard×Prime 0.14 0.35 0.13 (0.37) (0.41) (0.42) RewardCard×Super-Prime 0.07 0.30 0.12 (0.41) (0.44) (0.47) Restrictions: Atleasttwocardswithrevolvingdebtatthesamebank Y Y Y Y Y Y Notfullypaidbalanceonallcardswithrevolvingdebt Y Y Y Y Y Y Minimumpaymentonallcardswithrevolvingdebtandmorethantheminimumonatleastone N N Y Y Y Y DifferentAPRsonallcardswithrevolvingdebt N N N N Y Y CardControls Y Y Y Y Y Y FE:Cardholder×Bank Y Y Y Y Y Y Observations 21,288,917 16,136,165 12,858,916 Electronic copy available at: https://ssrn.com/abstract=4126641 50

TableVIII.MisallocatedPaymentsandHeuristics Thistablecomparestheactualpaymentamountstothetheoreticalpaymentamountsbasedonthreedifferentheuristics as discussed in Section VI.B: (i) the optimal repayment rule, (ii) the balance-matching heuristic, and (iii) an equal allocation across all cards (the 1/N heuristic). The table presents the mean shares and correlation coefficients between the differentpaymentamounts,separatelyforrewardcards(columns1and2)andforclassiccards(1and2). PaymentonRewardCard(s) PaymentonClassicCard(s) Mean ρ Mean ρ (1) (2) (3) (4) PanelA:AllCardholders(n=21,288,917) ActualShareofPayment 48.7% 35.9% OptimalShareofPayment 47.0% 0.50 37.5% 0.49 BalanceMatchingHeuristicShareofPayment 47.5% 0.52 37.0% 0.54 1/NHeuristicShareofPayment 42.8% 0.38 41.4% 0.35 PanelB:Sub-primeCardholders(n=7,469,187) ActualShareofPayment 47.0% 38.8% OptimalShareofPayment 43.9% 0.39 41.6% 0.43 BalanceMatchingHeuristicShareofPayment 47.3% 0.47 38.6% 0.49 1/NHeuristicShareofPayment 43.6% 0.36 41.9% 0.42 PanelC:Near-primeCardholders(n=7,482,795) ActualShareofPayment 47.8% 34.6% OptimalShareofPayment 46.8% 0.51 35.6% 0.49 BalanceMatchingHeuristicShareofPayment 47.9% 0.55 34.6% 0.54 1/NHeuristicShareofPayment 41.8% 0.41 40.0% 0.40 PanelD:PrimeCardholders(n=4,412,700) ActualShareofPayment 50.8% 34.3% OptimalShareofPayment 49.9% 0.55 35.3% 0.51 BalanceMatchingHeuristicShareofPayment 47.7% 0.53 37.3% 0.51 1/NHeuristicShareofPayment 42.8% 0.39 42.0% 0.32 PanelE:Super-primeCardholders(n=1,924,235) ActualShareofPayment 53.8% 32.9% OptimalShareofPayment 52.1% 0.63 34.5% 0.58 BalanceMatchingHeuristicShareofPayment 46.8% 0.56 39.9% 0.53 1/NHeuristicShareofPayment 43.4% 0.37 43.2% 0.26 Electronic copy available at: https://ssrn.com/abstract=4126641 51

TableIX.AnnualPercentageRates(APR)ofInterestandBankProfits This table presents the estimation results for differences in net reward components between reward cards and classic cardsfromEquation(2)inSectionIV.A.Theoutcomevariablesaretheannualpercentagerateofinterest(APR)(columns 1 an 2) and the dollar amount of bank profits per card as defined in Equation 5 in Section VII.B (column 3 and 4). The variableRewardCard takesonthevalueof1ifcardiisarewardcard,and0otherwise. Cardsareclusteredinthefollowing FICO score groups: sub-prime (below 660), near-prime (660-720), prime (720-780), and super-prime (above 780). Card controls include the credit limit, the amount past due, the card age, a joint account indicator, a fraud flag indicator, and a workout program indicator. Cardholder controls a deposit relationship indicator, a lending relationship indicator, the numberofcardsheldbythecardholderatthesamebank,andabankruptcyindicator. BorrowerincomeandFICOscores are defined as of March 2018 i.e., one year prior to the outcome variable. Standard errors are clustered at the bank-state level. *,**,and***indicatestatisticalsignificanceatthe10%,5%,and1%levels,respectively. APR Profit (1) (2) (3) (4) RewardCard -0.96*** 7.48*** (0.19) (0.71) RewardCard×Sub-Prime -0.20** 2.66* (0.09) (1.41) RewardCard×Near-Prime -0.47*** 13.10*** (0.16) (1.06) RewardCard×Prime -1.34*** 9.80*** (0.26) (0.49) RewardCard×Super-Prime -1.65*** 3.98*** (0.27) (0.43) CardControls Y Y Y Y CardholderControls Y Y Y Y FE:Bank×Zip×Income×FICO Y Y Y Y Observations 237,573,278 Electronic copy available at: https://ssrn.com/abstract=4126641 52

TableX.TheGeographyofNetRewards This table presents the estimation results for net rewards at the ZIP code-level from Equation (6) in Section VIII. The outcome variable is the net reward of card i in ZIP codez andwhereX arethefollowingZIPcode-levelcharacteristics: thepercentageof k residentswithabachelor’sdegreeasaproxyforeducation,themedianincomeofindividualsintheZIPcode,andthepercentageofresidentswhoreporttheirraceasBlackor AfricanAmerican. Standarderrorsareclusteredatthestatelevel. *,**,and***indicate statisticalsignificanceatthe10%,5%,and1%levels,respectively. NetRewards (1) (2) (3) (4) (5) (6) Education 0.29*** -0.01 (0.02) (0.02) Income 0.21*** 0.00 (0.02) (0.02) BlackPopulationShare -0.14*** 0.00 (0.01) (0.01) CreditScore 0.19*** 0.18*** 0.19*** (0.01) (0.00) (0.00) Observations 237,573,278 53 Electronic copy available at: https://ssrn.com/abstract=4126641

Online Appendix A. Estimating Monthly Net Rewards Whilerewardcreditcardsallowconsumerstoearnmoneythroughtheuseofcredit cards, cardholders may also incur costs in the form of interest payments and fees. To measure the monthly net cash flow on a credit card, we construct the variable Net Rewards which subtracts the amount of interest and fees paid on card i in month t from therewardsearnedonthecardduringthesameperiod: NetRewards = Rewards −InterestPaid −TotalFees (A1) i,t i,t i,t i,t In our dataset, we directly observe the dollar amounts of Interest Paid and Total Fees. However,wedonotobservetheamountofmonthlyrewards,butonlytheaccumulated rewardsasofthereportingmonth,netofredeemedrewards,thatis: CumulativeRewards = CumulativeRewards +Rewards −Redemptions (A2) i,t i,t−1 i,t i,t We have data on the stocks Cumulative Rewards, but not on the flows Rewards and on Redemptions. To calculate the monthly net rewards in Equation (1), we estimate the monthlyvariableRewards. First,weestimatetheeffectiverewardrateofcardibydividingthemonth-to-monthchangeincumulativerewardsbythepurchasevolumeofcard iduringthegivenmonth: ∆CumulativeRewards i,t Card-SpecificRewardRate = (A3) i,t PurchaseVolume i,t This estimated reward rate is correct if redeemed rewards in month t are zero. For example,ifcumulativerewardsoncardiincreaseby12dollarsinmonthtandifthecard 54 Electronic copy available at: https://ssrn.com/abstract=4126641

exhibitsapurchasevolumeof$1000duringthesamemonth,thentheestimatedeffectiverewardrateequals1.2percent. If,however,thecardholderredeemsrewardsduring themonth,thenthiswillunderestimatethecard-specificrewardrate. Inthecasewhen allrewardsare(automatically)redeemedinmontht,wewouldestimateacard-specific rewardrateofzero. To filter out these card-specific idiosyncrasies in redemption behavior, we estimate rewardratesattheindividualcreditcardproduct-level. Tothisend,weclusterallcards inoursampleintogroupsbasedonthefollowingvariables: bank,creditcardtype,product type, card network, reward type, fee type, and fee level.25 Within each cluster, we calculatethemedianrewardrateusingonlycardswithapositivechangeincumulative rewards,thatiscardsforwhich∆CumulativeRewards > 0. Wethenusetheestimated i,t rewardratetocalculatethemonthlyrewardsofcardiinmonthtas: Rewards = EstimatedRewardRate ×PurchaseVolume (A4) i,t i,t i,t Intherawsample,thismethodologyyieldsanaveragemonthlyrewardof$13.34perrewardcard,whichimpliesanextrapolatedaverageannualrewardof$160.08. Thisfigure isveryclosetothe$167inannualrewardsperaccountreportedinCFPB(2019),thereby confirmingthevalidityofourapproach. Furthermore,wecalculatethevariableTotalFees asthesumoflate,overlimit,nonsufficientfunds(NSF),cashadvance,debtsuspension,balancetransfer,other,andmonthly fees. Combining the data on total fees and interest paid with the estimated amount of monthly rewards from Equation (A4) allows us to calculate the monthly net rewards of cardiinmonthtasdefinedinEquation(1). 25Thisprocedureyields380individualcreditcardproductclusters.TableA2intheappendixdescribes allthevariablesusedinthecalculationofthevariableNetRewards. 55 Electronic copy available at: https://ssrn.com/abstract=4126641

B. Share of Misallocated Payments This appendix describes the calculation of the share of misallocated payments, followingPonce,Seira,andZamarripa(2017)andGathergood,Mahoney,Stewart,andWeber (2019). Given the amount of total funds used to pay off credit cards, the optimal, interest-cost-minimizing repayment rule is as follows. First, make the minimum payments due on all cards. Second, pay off in full the card with the highest interest rate. Third, subsequently allocate further repayments to cheaper cards ranked in order of their interest rates. Based on this rule, we calculate the misallocated payment (MP) share for borrower b on card i as the minimum between zero (if the actual payment is equal or lower than the optimal one) and the difference between the optimal paymentamount(OPA)andtheactualpaymentamount(APA)scaledbythetotalpayment amount:     ActualPayment T A o m ta o l u P n a t y i m ,b − en O t p A t m im o a u l n P t aymentAmount i,b if APA i,b > OPA i,b MPShare = i,b (A5)   0 if APA i,b ≤ OPA i,b Thismeasurecanbeinterpretedastheshareofpaymentsthatwereincorrectlymade onacheapercardthatshouldhavebeenmadeonmoreexpensivecards. Figure4illustrates the share of misallocated payments across the FICO distribution. The misallocatedpaymentshareisstronglydecreasinginFICOscores. Whilelow-FICOconsumers misallocate more than 6 percent of all credit card repayments, the misallocated paymentshareislessthan2percentforhigh-FICOconsumers. 56 Electronic copy available at: https://ssrn.com/abstract=4126641

C. Additional Figures FigureA1. CoefficientPlot: NetRewardsAcrosstheFICODistribution. This figure illustratesthedifferentialdollarmagnitudeofaveragenetrewardsbetweenrewardcards andclassiccardsacrosstheFICOdistribution. ThefigureplotsthecoefficientsδF alongside the 95% confidence intervals when estimating Equation (2) with 50 instead of 4 different FICO buckets. The dashed vertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime, prime, and super-prime cardholders, respectively. The graph is based on our baseline sample of 238 million credit cards in March 2019. 57 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure A2. Net Rewards Across FICO Score Percentiles by Reward Type. This figure illustratesthedollarmagnitudeofaveragenetrewardsonrewardcardsacrosstheFICO distribution by reward type. The red line plots the average net reward for borrowers withanannualincomebelow44thousand,theyellowlineforborrowerswithanannual income between 44 thousand and 79 thousand, and the green line for borrowers with anannualincomeabove79 thousand. Foreachincomegroup, weplottheaveragenet reward (in dollar) for 100 equal-sized FICO buckets between 480 and 830. The dashed vertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime, prime, and super-prime cardholders, respectively. The graph is based on our baseline sampleof238millioncreditcardsinMarch2019. sdraweR teN ylhtnoM egarevA 02 0 02- 04- 06- 480 660 720 780 830 Credit Score Cash Back Miles Points 58 Electronic copy available at: https://ssrn.com/abstract=4126641

FigureA3. FeeChargesAcrossFICOScorePercentiles. Thisfigureillustratesthedollar magnitude of average fee charges across the FICO distribution, separately for reward cards(solidredline)andclassiccards(dashedblueline). Foreachcardtype,weplotthe averagefeechargefor100equal-sizedFICObucketsbetween480and830. Thedashed vertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime, prime, and super-prime cardholders, respectively. The graph is based on our baseline sampleof238millioncreditcardsinMarch2019. segrahC eeF ylhtnoM egarevA 6 4 2 0 480 660 720 780 830 Credit Score Reward Cards Classic Cards 59 Electronic copy available at: https://ssrn.com/abstract=4126641

FigureA4. FICOScoreDistributionsbyIncomeGroups. Thisfigureillustratesthedistribution of FICO scores across the full sample (solid red line) and three different income groups: low-income cardholders with an annual income below $44 thousand; middle-income cardholders with an annual income between $44-79 thousand; and high-incomecardholderswithanannualincomeabove$79thousand. Thedashedvertical lines mark FICO scores of 660, 720, and 780, our cut-off scores for near-prime, prime, and super-prime cardholders, respectively. The graph is based on our baseline sampleof238millioncreditcardsinMarch2019. 60 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure A5. Net Rewards Across the FICO Distribution by Income. Panel A plots the dollarmagnitudeofaveragenetrewardsacrosstheFICOdistribution,separatelyforreward cards (solid lines) and classic cards (dashed lines), and for threedifferent income groups (below 44 thousand, 44 thousand and 79 thousand, and above 79 thousand). Panel B plots the coefficients δF alongside the 95% confidence intervals when estimatingEquation(2)with50insteadof4differentFICObucketsseparetelyforthesamethree differentincomebuckets. Inbothpanels,thedashedverticallinesmarkFICOscoresof 660,720,and780,ourcut-offscoresfornear-prime,prime,andsuper-primecardholders, respectively. The graph is based on our baseline sample of 238 million credit cards inMarch2019. (A)RewardCardsversusClassicCards sdraweR teN ylhtnoM egarevA 02 0 02- 04- 06- 480 660 720 780 830 Credit Score Reward: Low In. Reward: Middle Income Reward: High Income Classic: Low Income Classic: Middle Income Classic: High Income (B)CoefficientPlot 61 Electronic copy available at: https://ssrn.com/abstract=4126641

Figure A6. Net Rewards Across Income Percentiles. This figure illustrates the dollar magnitudeofaveragenetrewardsacrosstheincomedistribution,separatelyforreward cards(solidredline)andclassiccards(dashedblueline). Foreachcardtype,weplotthe average net reward for 100 equal-sized income buckets between $3,000 and $400,000. The dashed vertical lines mark income levels of $44,000 and $79,000, denoting the tercile values in our dataset. The graph is based on our baseline sample of 238 million creditcardsinMarch2019. sdraweR teN ylhtnoM egarevA 0 5- 01- 51- 02- 52- $44,000 $79,000 Income Reward Cards Classic Cards 62 Electronic copy available at: https://ssrn.com/abstract=4126641

D. Additional Tables TableA1. AggregateNetRewards This table presents the aggregate sum of net rewards (in USD million) for reward cards withnegative(column1)andpositive(column2)netrewards,bothfortheentiresample (first row) and across different FICO buckets (second to last row). In the second to last row, cards are clustered in the following FICO score groups: sub-prime (below 660), near-prime(660-720),prime(720-780),andsuper-prime(above780). Thetableisbased onoursampleof91millionrewardcardsinMarch2019. NegativeRewards PositiveRewards ∆ (1) (2) (3) AllRewardCards -4140 1260 -2880 Sub-Prime -1030 35 -996 Near-Prime -1630 134 -1496 Prime -1130 407 -723 Super-Prime -361 680 319 63 Electronic copy available at: https://ssrn.com/abstract=4126641

TableA2. CreditCardCategories Thistablereportsthedetailedcategoriesusedforcreditcardclusteringattheindividual productlevelinthecalculationofnetrewardsinSectionIII.A.Ourprocedureyields380 individualcreditcardproductcluster. Variable Categories Bank 19banks CreditCardType GeneralPurpose PrivateLabel ProductType Co-brand OilandGasCo-Brand Affinity Student Other NetworkType Visa MasterCard AmericanExpress Discover Other RewardType Cash Miles Other None FeeType Nofee Annualfee Monthlyfee AnnualizedFeeAmount 0dollar 0-60dollar 60-120dollar 120+dollar 64 Electronic copy available at: https://ssrn.com/abstract=4126641

TableA3. FeeComponents This table presents the estimation results for differences in annual fee, late payment fee, and other fee charges between rewardcardsandclassiccardsfromEquation(2)inSectionIV.A: (cid:88)(cid:0) (cid:1) (cid:88) (cid:88) Y = δF ×RewardCard ×DF +α + Xm + Xn +ε i i b,z,w,f i j i F m n The variable Reward Card takes on the value of 1 if card i is a reward card, and 0 otherwise. Cards are clustered in the following FICO score groups D: sub-prime (below 660), near-prime (660-720), prime (720-780), and super-prime (above 780). Card characteristics include the credit limit, amount past due, card age, a joint account indicator, and a fraud dummy. Borrower characteristics including a deposit relationship indicator, a lending relationship dummy, the total number of cards the consumer has with the bank, a workout program dummy, and a bankruptcy indicator. Borrower incomeandFICOaredefinedasofMarch2018i.e.,oneyearprior. Standarderrorsareclusteredatthebank-statelevel. *, **,and***indicatestatisticalsignificanceatthe10%,5%,and1%levels,respectively. AnnualFeeCharges LatePaymentFeeCharges OtherFeeCharges (1) (2) (3) (4) (5) (6) RewardCard 0.51*** 0.14*** 0.06*** (0.05) (0.03) (0.02) RewardCard×Sub-Prime 0.56*** 0.14* 0.08*** (0.04) (0.08) (0.02) RewardCard×Near-Prime 0.35*** 0.19*** 0.24*** (0.07) (0.04) (0.03) RewardCard×Prime 0.57*** 0.15*** 0.05* (0.06) (0.02) (0.03) RewardCard×Super-Prime 0.54*** 0.08*** -0.12*** (0.05) (0.01) (0.02) CardControls Y Y Y Y Y Y CardholderControls Y Y Y Y Y Y FE:Bank×Zip×Income×FICO Y Y Y Y Y Y Observations 237,573,278 237,573,278 237,573,278 237,573,278 237,573,278 237,573,278 Electronic copy available at: https://ssrn.com/abstract=4126641 65

TableA4. NetRewardsbyIncomeGroups—TopIncomeDistribution This table presents the estimation results for differences in net rewards between reward cards and classic cards from Equation(2)inSectionIV.A: (cid:88)(cid:0) (cid:1) (cid:88) (cid:88) Y = δF ×RewardCard ×DF +α + Xm + Xn +ε i i b,z,w,f i j i F m n We reports results separately for three different annual income groups. The variable Reward Card takes on the value of 1 if card i is a reward card, and 0 otherwise. Cards are clustered in the following FICO score groups D: sub-prime (below660),near-prime(660-720),prime(720-780),andsuper-prime(above780). Cardcharacteristicsincludethecredit limit, amount past due, card age, a joint account indicator, and a fraud dummy. Borrower characteristics including a depositrelationshipindicator,alendingrelationshipdummy,thetotalnumberofcardstheconsumerhaswiththebank, a workout program dummy, and a bankruptcy indicator. Borrower income and FICO are defined as of March 2018 i.e., one year prior. Standard errors are clustered at the bank-state level. *, **, and *** indicate statistical significance at the 10%,5%,and1%levels,respectively. Top10%of Top5%of IncomeDistribution IncomeDistribution (1) (2) (3) (4) RewardCard 6.96*** 7.70*** (0.86) (0.96) RewardCard×Sub-Prime -21.97*** -25.61*** (1.50) (1.72) RewardCard×Near-Prime -18.35*** -19.43*** (1.00) (1.15) RewardCard×Prime 10.65*** 11.77*** (0.76) (0.86) RewardCard×Super-Prime 22.33*** 22.24*** (1.14) (1.16) CardControls Y Y Y Y CardholderControls Y Y Y Y FE:Bank×Zip×Income×FICO Y Y Y Y Observations 26,600,689 26,600,689 14,754,880 14,754,880 Electronic copy available at: https://ssrn.com/abstract=4126641 66

TableA5. NetRewardsbyTypeofRewardCard This table presents the estimation results for differences in net rewards between reward cards and classic cards from Equation(2)inSectionIV.A: (cid:88)(cid:0) (cid:1) (cid:88) (cid:88) Y = δF ×RewardCard ×DF +α + Xm + Xn +ε i i b,z,w,f i j i F m n We reports results separately for the three types of reward cards i.e., miles, cash back, and points. The variable Reward Card takes on the value of 1 if card i is a reward card of a given type, and 0 if it is a classic card. Cards are clustered in the following FICO score groups D: sub-prime (below 660), near-prime (660-720), prime (720-780), and super-prime (above 780). Card characteristics include the credit limit, amount past due, card age, a joint account indicator, and a fraud dummy. Borrower characteristics including a deposit relationship indicator, a lending relationship dummy, the totalnumberofcardstheconsumerhaswiththebank,aworkoutprogramdummy,andabankruptcyindicator. Borrower incomeandFICOaredefinedasofMarch2018i.e.,oneyearprior. Standarderrorsareclusteredatthebank-statelevel. *, **,and***indicatestatisticalsignificanceatthe10%,5%,and1%levels,respectively. MilesCards CashBackCards PointsCards (1) (2) (3) (4) (5) (6) RewardCard -4.52*** 7.25*** 1.57*** (1.30) (0.73) (0.270) RewardCard×Sub-Prime -26.84*** -2.57*** -6.42*** (2.30) (0.47) (0.46) RewardCard×Near-Prime -23.63*** -2.07*** -8.03*** (2.85) (0.49) (0.64) RewardCard×Prime 0.47 12.41*** 4.10*** (1.41) (0.70) (0.31) RewardCard×Super-Prime 12.62*** 22.48*** 10.04*** (1.09) (1.30) (0.47) CardControls Y Y Y Y Y Y CardholderControls Y Y Y Y Y Y FE:Bank×Zip×Income×FICO Y Y Y Y Y Y Observations 113,283,147 113,283,147 153,206,808 153,206,808 158,481,157 158,481,157 Electronic copy available at: https://ssrn.com/abstract=4126641 67

TableA6.Overindebtedness:Difference-in-differencesAnalysis,OnlyCardswithaBank-initiatedCreditLineIncrease Thistablepresentstheestimationresultsforthedifference-in-differencesregressioninEquation(3)inSectionVI.A.Wecomparechangesincredit cardoutcomesofconsumerswhoreceivedabank-initiatedcreditlimitincreaseonrewardcardstothosewhoreceivedalimitincreaseonclassic cardsinatimewindow6monthsbeforeandafterthecreditlimitincrease. Theoutcomevariablesarechangesinspendingvolumes(columns1 and2),creditcardpayments(columns3and4),andunpaidbalances(columns5and6). Theanalysisconsidersonlycardswithabank-initiated creditlineincrease.ThevariableRewardCardtakesonthevalueof1ifcardiisarewardcard,and0otherwise.Cardsareclusteredinthefollowing FICO score groups D: sub-prime (below 660), near-prime (660-720), prime (720-780), and super-prime (above 780). Card controls include the FICOscore,thecreditlimit,theamountpastdue,thecardage,ajointaccountindicator,afraudflagindicator,andaworkoutprogramindicator. Cardholdercontrolsincome,adepositrelationshipindicator,alendingrelationshipindicator,thenumberofcardsheldbythecardholderatthe samebank,abankruptcyindicator,andaveragespendingandpaymentsinthepre-treatmentperiod. BorrowerincomeandFICOaredefinedas ofMarch2018i.e.,oneyearprior.Standarderrorsareclusteredatthebank-statelevel.*,**,and***indicatestatisticalsignificanceatthe10%,5%, and1%levels,respectively. ∆Spending ∆Payments ∆UnpaidBalances (1) (2) (3) (4) (5) (6) RewardCard 75.21*** 43.76*** 48.28*** (6.70) (3.61) (11.70) RewardCard×Sub-Prime 57.21*** 18.83*** 46.95*** (6.27) (2.82) (12.13) RewardCard×Near-Prime 62.65*** 21.26*** 68.61*** (6.86) (3.47) (16.35) RewardCard×Prime 89.06*** 77.15*** 37.70*** (8.10) (5.73) (13.88) RewardCard×Super-Prime 169.17*** 156.26*** -12.77 (13.02) (11.72) (26.40) MeanY 860.315 851.559 1922.45 CardControls(Pre-Period) Y Y Y Y Y Y CardholderControls(Pre-Period) Y Y Y Y Y Y IncomeandFICO(Pre-Period) Y Y Y Y Y Y SpendingandPayments(Pre-Period) Y Y Y Y Y Y FE:Bank×Zip Y Y Y Y Y Y Observations 1,236,604 1,236,604 1,236,604 1,236,604 1,236,604 1,236,604 Electronic copy available at: https://ssrn.com/abstract=4126641 68

TableA7. ShareofMisallocatedPayments—Two-cardSample Thistablepresentstheestimationresultsfordifferencesintheshareofmisallocatedpayments(asdefinedinEquationA5 in Section B) between reward cards and classic cards from Equation (2) in Section IV.A. The analysis considers only individuals with two credit cards. The variable Reward Card takes on the value of 1 if card i is a reward card, and 0 otherwise. Cards are clustered in the following FICO score groups: sub-prime (below 660), near-prime (660-720), prime (720-780),andsuper-prime(above780). Cardcontrolsincludethecreditlimit,theamountpastdue,thecardage,ajoint account indicator, a fraud flag indicator, and a workout program indicator. Cardholder controls a deposit relationship indicator,alendingrelationshipindicator,thenumberofcardsheldbythecardholderatthesamebank,andabankruptcy indicator. Borrower income and FICO scores are defined as of March 2018 i.e., one year prior to the outcome variable. Standarderrorsareclusteredatthebank-statelevel. *,**,and***indicatestatisticalsignificanceatthe10%,5%,and1% levels,respectively. ShareofMisallocatedPayments (1) (2) (3) (4) (5) (6) RewardCard 1.15*** 1.64*** 1.74*** (0.34) (0.41) (0.46) RewardCard×Sub-Prime 2.96*** 4.11*** 4.60*** (0.28) (0.31) (0.38) RewardCard×Near-Prime 0.40 0.79** 0.83** (0.29) (0.34) (0.39) RewardCard×Prime -0.34 -0.12 -0.22 (0.38) (0.43) (0.47) RewardCard×Super-Prime -0.18 0.10 0.00 (0.50) (0.55) (0.59) Restrictions: Atleasttwocardswithrevolvingdebtatthesamebank Y Y Y Y Y Y Notfullypaidbalanceonallcardswithrevolvingdebt Y Y Y Y Y Y Minimumpaymentonallcardswithrevolvingdebtandmorethantheminimumonatleastone N N Y Y Y Y DifferentAPRsonallcardswithrevolvingdebt N N N N Y Y CardControls Y Y Y Y Y Y FE:Cardholders×Bank Y Y Y Y Y Y Observations 13,080,528 13,080,528 9,909,754 9,909,754 8,862,432 8,862,432 Electronic copy available at: https://ssrn.com/abstract=4126641 69