Navigating Higher Education Insurance: An Experimental Study on Demand and Adverse Selection

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Navigating Higher Education Insurance: An Experimental Study on Demand and Adverse Selection Sidhya Balakrishnan, Eric Bettinger, Michael S. Kofoed, Dubravka Ritter, Douglas A. Webber, Ege Aksu, and Jonathan S. Hartley 2024-024 Please cite this paper as: Balakrishnan,Sidhya,EricBettinger,MichaelS.Kofoed,DubravkaRitter,DouglasA.Webber, Ege Aksu, and Jonathan S. Hartley (2024). “Navigating Higher Education Insurance: An Experimental Study on Demand and Adverse Selection,” Finance and Economics Discussion Series 2024-024. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2024.024. 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.

Navigating Higher Education Insurance: An Experimental Study on Demand and Adverse Selection* SidhyaBalakrishnan†,EricBettinger‡,MichaelS.Kofoed§,DubravkaRitter¶ DouglasA.Webber|,| EgeAksu*,*andJonathanS.Hartley†† February21,2024 Abstract Weconductasurvey-basedexperimentwith2,776studentsatanon-profituniversitytoanalyzeincome insurance demand in education financing. We offered students a hypothetical choice: either a federalloanwithincome-drivenrepaymentoranincome-shareagreement(ISA),withrandomizedframingofdownsideprotections. EmphasizingincomeinsuranceincreasedISAuptakeby43%. Weobserve thatstudentsareresponsivetochangesincontracttermsandpossiblestudentloancancellation,whichis evidenceofpreferenceadjustmentoradverseselection. Ourresultsindicatethatframingspecificterms canincreasedemandforhighereducationinsurancetopotentiallyaddressriskforstudentswithvarying outcomes. *TheauthorsappreciatethehelpfulcommentsandfeedbackofNeilBhutta,JuliaCheney,BarryCynamon,AndrewHertzberg, Jerome Hodges, Caroline Hoxby, Rajeev Darolia, Robert Hunt, Jeff Larrimore, Joseph Marchand, Lois Miller, Kevin Mumford, MarshallSteinbaum,andJoshuaPrice. Wearealsogratefulforparticipantsatseminarandconferencepresentationsincludingthe AmericanEconomicAssociation,AppalachianStateUniversity,AssociationforEducationFinanceandPolicy,AssociationforPolicy AnalysisandManagement,BrighamYoungUniversity,NationalBureauofEconomicResearch(EconomicsofEducation),Providence College,SouthernEconomicAssociation,UnitedStatesAirForceAcademy,andUniversityofTennessee,Knoxville. †SidhyaBalakrishnanisthedirectorofresearchattheJainFamilyInstitute,email:sidhya.balakrishnan@jainfamilyinstitute.org ‡EricBettingeristheConleyDeAngelisFamilyProfessorofEducationatStanfordUniversityandaresearchassociateatNBER, email:ebettinger@stanford.edu. §MichaelS.KofoedisanassistantprofessorofeconomicsattheUniversityofTennessee,KnoxvilleandResearchFellowatIZA, andcorrespondingauthor:mkofoed1@utk.edu ¶DubravkaRitterisasenioradvisorandresearchfellowattheConsumerFinanceInstitute,FederalReserveBankofPhiladelphia, email:dubravka.ritter@phil.frb.org.ThisPhiladelphiaFedworkingpaperrepresentspreliminaryresearchthatisbeingcirculatedfor discussionpurposes.Theviewsexpressedinthesepapersaresolelythoseoftheauthorsanddonotnecessarilyreflecttheviewsofthe FederalReserveBankofPhiladelphiaortheFederalReserveSystem. Anyerrorsoromissionsaretheresponsibilityoftheauthors. Nostatementshereshouldbetreatedaslegaladvice. ||DouglasA.WebberisasenioreconomistattheBoardofGovernorsoftheFederalReserve,email: douglas.a.webber@frb.gov. TheanalysisandconclusionsinthispaperarethoseoftheauthorandshouldnotbeinterpretedasreflectingtheviewsoftheBoardof GovernorsortheFederalReserveSystem. **Ege Aksu is a PhD candidate at CUNY Graduate Center and fellow at the Jain Family Institute, email: ege.aksu@jainfamilyinstitute.org ††JonathanS.HartleyisaPhDcandidateatStanfordUniversity,email:hartleyj@stanford.edu

1 Introduction Insuranceproductsareimportanttoolsemployedbyindividualstohedgerisksintheirfinanciallives.Insurancemarketsallowindividualstopoolriskagainstunexpected,negativeoutcomesandarewelldevelopedin manycontexts,likehealthcareorrealestate.1 However,risk-hedgingopportunitiesarenotreadilyavailable in post-secondary education, even though college is an increasingly uncertain investment (Webber, 2022) made only once in a lifetime. While returns to college are positive on average (Lovenheim and Smith, 2022), their distribution is more nuanced (Webber, 2016, Broady and Hershbein, 2020). Financial outcomesforstudents,forexample,varyacrossinstitutiontypes(e.g.selectivevs. non-selective;four-yearvs. two-year), fields of study (e.g. education vs. engineering vs. economics), and macroeconomic conditions upongraduation(Rothstein,2023). Perhapsmoreimportantly,returnsvarywithineachofthesesegments givenunobservablestudentskill–whichispotentiallydifficultforthestudentand/ortheinstitutiontoidentify – and uncertain labor market conditions. Particularly for younger students and those entering longer degreeprograms,thereisuncertaintybothintheexpectedlevelofincomeandinitsvariability. Students,educationalinstitutions,andgovernmententitiesunderstandandbeartheseriskstodifferent degrees. For example, work by Stange (2012) shows that many students treat attending their first year of collegeakintopurchasinganoptionscontract—completinganinitialyearsoastodevelopabettersenseof theirlikelyreturns,afterwhichtheydecidewhethertoexercisetheoptionforasecondyear. Policymakers and advocates often work to transfer the riskiness of the return to taxpayers, e.g. via the free college movement,financialaidpolicy(bothgrantsandloans),ortheCovid-19studentloanrepaymentpausefor loansguaranteedbythefederalgovernment. Interestingly,individualeducationalinsurancepolicieswherestudentspayapremiumtoprotectthemselvesfromincomeriskareeithernotwelldevelopedorarenon-existent.2 Onereasonforthelowprevalenceofeducationalinsuranceinpost-secondarymarketsmaybelowdemand. Thereisevidencethatstudentscanbeover-optimisticaboutfutureearnings(e.g. Bakeretal.(2018)),failingtoadequatelyconsider insuranceriskatthetimeofenrollmentandfinancingbecauseofdifficultyinpredictingfutureincomesaccordingtomajor(Arcidiaconoetal.,2012;Bakeretal.,2018;Conlon,2021). Onthesupplyside,another explanationcouldbethepresenceofadverseselection(Einavetal.(2023))andmoralhazard(Zweifeland Manning,2000)ininsurancemarkets. 1Closesttooursetting, individualspurchaseinsurancetomitigatefinanciallosses(e.g. Arrow, 1963), bufferagainstincome shocks(e.g.ChettyandSzeidl(2007)),andforavarietyofotherreasons.GuisoandPaiella(2008)documenttheincreasingpropensity ofhouseholdstohedgeagainstlaborincomeriskinparticular,indicatingarisingawarenessofemployment/incomeuncertainties.In thecaseoffinancialmarkets,diversification,includingtheuseofderivativeinstrumentslikefuturesandoptions,remainsaprimary strategyforriskmanagementinthefaceofuncertaineconomicoutcomes(e.g.Bodie(1994);GoyalandWelch(2007)). 2Throughoutthispaper,wewillassumethattheprimaryformofinsuranceinpost-secondaryeducationisagainsttheriskoflow oruncertainincome,andwillrefertothisas"lowincomeinsurance"or"educationalinsurance." 2

Ourpapermakesauniquecontributiontoourunderstandingofstudentdemandforeducationalinsuranceandthepotentialrelevanceofadverseselectionintheviabilityoflow-incomeinsuranceineducation markets.Studentswhoareunsureabouttheirprospectsmaydemandinsurancetoprotectthemselvesagainst downsidelabormarketrisks.Adverseselectionisrelatedbutdifferent:theinsureduseinformationthatthey caneasilyconcealfromtheinsurancecompanytotakeadvantageoftheinsurance’sdownsideprotections. Ourmajorcontributionisaframingexperimentusedtoteaseoutinsurancedemand;oursurveyallowsus to see concealed information unavailable to the hypothetical insurer to test for adverse selection. Using theexperimentcoupledwiththesurvey,wecantesthowframingaffectsastudent’sdemandforeducation insuranceandseeifstudentsuseconcealedinformationtotakeuptheprotectionsatdifferentrates. We partnered with a large, non-profit university (hereafter, The University) that typically serves nontraditionallyagedstudentswhoareoftenworkingadults. Weconductedarandomizedsurveylabexperimentwith2,776studentstounderstandtheirpreferencesoverdifferenteducationalfinancingchoices. Inthesurvey,studentswereaskedtochoosebetweenahypotheticalfederalstudentloanwiththeoption ofanincome-drivenrepayment(IDR)planandahypotheticalincome-shareagreement(ISA).Bothoptions provided information on monthly loan payments that are waived for very low incomes and otherwise are capped to a fixed share of an individual’s income, with markedly different implementation and paths for satisfyingtheloanobligation. In the experiment, students were randomized into two equal groups and, similar to Abraham et al. (2020), the presented hypothetical options differed in terms of level of detail provided for each of the choices. ThefirstgroupwasshowndescriptionsofthestudentloanwithoptionofIDRandtheISAwitha risk-neutralframingthatexplainedthedifferencesinmonthlypayments,generalstructureoftheloan,the baseline payment terms, and the source of funding. The terms of the student loan with IDR and the ISA were set to be actuarially equivalent. We exposed the second group – our treatment group – to the same descriptions of the student loan with option of IDR and the ISA, but with an additional emphasis on the insurancefeatures(natureoftheincomecontingencyandmaximumrepaymentterm)ofthetwofinancing options. TherearemanydifferencesbetweenfederalstudentloanswithIDRandISAs,andmanyreasonswhy different borrowers might prefer one over the other. With federal student loans, borrowers who do well in the labor market will pay less in total by paying fixed monthly payments for the minimum number of years (120 payments, or 10 years with no gaps in payment). Since there is no prepayment penalty for federalstudentloans, theycanalsobepaidofffasterthanscheduledandmaybeparticularlyattractiveto studentswhoexpectconsistentlyhighearningsaftercollege. Toaccesstheincomecontingency,borrowers mustfollowaseriesofadministrativehurdlesinordertoqualifyforreducedmonthlypaymentscappedata 3

certainpercentageoftheirincome,payingnothingiftheirincomefallsbelowasetthreshold,butpotentially extendingtheirtermupto20yearscomparedwiththestandardrepaymentplan.3 WithanISA,borrowers’ monthlypaymentsaresetasapre-agreedshareofincomebydesign,andtherepaymenttermistypically extendedtoalesserdegreethanIDRduetomonthsofnon-payment,makinganISAapotentiallyattractive proposition for borrowers with persistently low or variable earnings. On the other hand, there is no way to “refinance” out of an ISA and borrowers who end up earning high incomes will pay up to a multiple of the original loan amount, described in our experiment. Finally, borrowers may have preferences over borrowingfromthegovernmentversusaprivatelender. We find that students have a significant preference for the built-in income insurance in the ISA and thatourinsuranceframinggreatlyincreasesthedemandforthehypotheticalISA–byabout10percentage points, or 43 percent. Importantly, there is limited heterogeneity in treatment to be found along demographic, academic, or financial lines for students in our sample.4 The insurance framing has, by far, the largesteffectontake-up. Ourresultssuggestthatstudentsarenotnecessarilythinkingaboutincomerisk oraboutthepotentialbenefitsofeducationalinsurancewhentheychoosehowtofinancetheirstudies,but thateducationalandloanproviderscanhelpmakethepotentialneedforeducationalinsurancesalientfor borrowersbythoroughlyexplainingthecostsandbenefitsofsuchinsurance. Oursurveyandfollow-upquestionsalsoallowustocharacterizehowadverseselectionmayenterthe educational insurance market (Herbst and Hendren, 2021).5 The survey allows us to solicit information from the student that is unavailable to the ISA originator. Students may be confident about their income potential but can easily conceal this information from financing providers who do not have the ability to pricediscriminate(i.e. mustchargethesameinterestrateorincomesharetoall). Insuchanenvironment, studentsexpectinglowincomeswillsortintotheISAwhilestudentsexpectinghighincomeswilloptfora traditionalloan. EducationalinsurancethatlooksmorelikeanISAwillnotbeasustainablepolicychoice if students who achieve significant returns to college systematically choose student loans. Overall, there is less evidence suggestive of adverse selection across a variety of variables than we supposed ex ante. Employment uncertainty, for example, does not appear to influence take-up. However, we do find strong suggestiveevidenceofadverseselectionbasedonlikelihoodoffutureincomesbeinglow. To further test for adverse selection, we ask several follow-up questions to investigate how students 3Recent policy changes around the Saving on a Valuable Education (SAVE) IDR plan simplify some of these processes for federalstudentloans,buttheincomeprotectionisstillfarfrombuiltin.AlthoughsomecurrentlyavailableIDRplansofferedbythe DepartmentofEducationextendtherepaymenttermtoupto25years,wewantedtokeepthecomparisonsimpleforborrowersand chosethe(modal)maximumtermof20years. 4Thepreregisteredbaselinevariablesforheterogeneoustreatmenteffectsincludedrace/ethnicity(Black,Hispanic,white),gender (indicatorforfemalerespondent/recipient),householdsize,age(mediansplit),maritalstatus,riskaversion. 5Sincewecannotfollowstudentsafterthesurvey,wecannotshedlightonpotentialmoralhazardfromtheavailabilityofinsurance forstudentsinthetreatmentgroup. 4

mightchangetheiranswersiftheoffertermsweremodestlydifferent. Aftertheirinitialchoicebetweena studentloanandanISA,weofferedstudentswhooriginallyselectedastudentloanwithIDRanactuariallyequivalent alternative ISA with a lower income share and a longer term. If a student maintained their originalchoiceofstudentloaninthesecondroundaswell,wethenofferedthemanotheralternativeISAin thethirdround–thistimewithahigherincomeshareandashortertermthantheoriginalISA.Tostudents whooriginallyselectedthehypotheticalISA,weseparatelyofferedbothalternativeISAsatthesametime. We find that both original ISA choosers and original student loan choosers were equally likely to switch tothelongertermISA,with18%ofrespondentsselectingthelongertermISAovertheiroriginalchoice. Interestingly,respondentswhooriginallychosetheISAwereconsiderablymorelikely(61%)tochoosethe shortertermISAcomparedwiththerespondentswhochosethestudentloanwithIDRinboththefirstand secondrounds(17%). Wefurtherfindthattreatedstudentswhooriginallychosethestudentloanwere5.1percentagepoints (17%)morelikelytochoosetheISAoptionwithalowershareandlongerterm. Thetreatmenteffectfor those offered the alternative ISA with shorter term and higher share is not statistically significant. For students who chose the original ISA in our base experiment, treated students were 8.6 percentage points (15%) more likely to choose the alternative ISA option with a higher share and shorter term. We find no significant treatment effects on switching toward the lower share and longer term ISA for students who originally chose the ISA over the student loan with IDR. Overall, our results suggest that the insurance framinghelpedreinforcestudentpreferencesovershortermaximumrepaymentterm(12yearsforISAv. 20yearsforstudentloan),andthatstudentsselectedalternativeISAcontractsinourfollow-upquestionsin awaythatreflectedtheirstatedpreferencebetweentheinitialchoice. Separately,weaskedallstudentswhethertheywouldselectastudentloaniftherewerea20%chance that the $10,000 loan they borrowed would be forgiven. We find that students who switched from a loan to an ISA in the second round were 6.4 percentage points more likely to switch back to a loan when offeredthechanceofstudentloanforgiveness. Conversely,whenstudentswhopickedtheoriginalISAwere offeredaloanwithachanceoffuturedebtforgiveness,theprospectoffuturebalancereductiondecreased thewillingness/likelihoodofswitchingbacktothefederalstudentloan. Thestudentswhoswitchedtoan ISAinthesecondroundaremarginallyattachedtotheISAandmaybeeasilyinducedtoswitchbetween thetwofinancingchoicesgivenrelativelysmallchangestoterms,whiletheoriginalISAchoosersappear tobemoresetintheirISApreference. TakentogetherwithourresultsonstudentpreferencesoveralternativeISAvariationscomparedwithastudentloanwithIDR,ourstudycontributestotheunderstandingof optimaldesignofloanproductswithincomeinsurancefeatureswithregardtobothupsideprotections(like maximumtermortotalpaymentamount)anddownsideprotections(likeincomeshare). 5

Though focused on educational insurance markets, our paper contributes to other lines of research, includingtheliteratureonfinancialaid,educationfinance,andstudentdebt.Asthecostofhighereducation has risen and the purchasing power of public subsidies have fallen along with public financial support to universities(Webber,2017),familieshavehadtoincurdebtorforegoconsumptiontoaffordpost-secondary education. Recentresearchhasemphasizedtheburdenthatstudentloansplaceonstudents(Chakrabartiet al.,2020)-includingontheirotherconsumerspending(e.g.MezzaandSommer,2015),"lifemilestones" (Mezzaetal.,2020),andeducationaloutcomes(e.g.Blacketal.,Forthcoming,DenningandJones,2021). Because the monthly payment is proportional to income, education insurance such as IDR and ISAs can hedgeagainsttheadverseeffectsofstudentloans. Assuch, take-upofsuchproducts, particularlyamong populations where student loans have had adverse effects (e.g. students at technical or public regional collegeswithhighervarianceincollegeoutcomes),isimportant. Manyfamiliesdonotapplyforaid–meaningtheydonotcompletetherequiredfinancialaidforms– becauseofalackofinformationoruncertaintyofeligibility(e.g. Kofoed(2017),Bettingeretal.(2012)). Evenforfamiliesthatapplyforaid,theprotectionofsomeassetswithinthePelleligibilityformulaleads tolessfinancialaideligibilityforstudentsfromdisadvantagedfamilies(LevineandRitter,2023)resulting inhigherstudentloanburdensanddecreasedaccesstoselectiveinstitutions. Formsofbuilt-ineducational insurancearepotentiallymoreattractivetothesestudents,astheyautomaticallyreallocatesomeoftherisk fromthestudenttotheproviderandmayreduceuncertaintyaroundfinancialaideligibility. Additionally, our research contributes to the behavioral literature on student take-up of financial aid programs under varied framing.6 Abraham et al. (2020) and Marx and Turner (2019) demonstrate that framingmattersforgovernment-sponsoredIDRplansandtraditionalstudentloantake-up,respectively. In thisliterature,researchersmanipulatethestudents’norms,theterms,risk,andthecostscommunicatedto studentswithrespecttospecificfinancialinstruments. Ourpapercontributestothisdiscussionbyshowing thatstudentspreferincomecontingentfinancingwhenweemphasizethebuilt-ineducationalinsuranceof theISA,withtreatmenteffectsfortheinsuranceframingcomparableinmagnitudetoAbrahametal.(2020). The lessons from our study are applicable to the design of any income-contingent education financing product and are particularly salient to ongoing policy discussions around the Department of Education’s income-drivenrepaymentplansforfederalstudentloans. Our paper is organized as follows. Section 2 reviews students’ college financing and the prospect for education insurance in financing education. Section 3 details the experimental design of our research questions.Section4laysoutourempiricalstrategyandelaboratesonthedatacollection.Section5provides 6Coxetal.(2020)examinewhystudentsdon’tchooseIDRwhentheyareworriedaboutfutureincomeexpectations. They conductalaboratoryexperimentwheretheyprovideinformationaboutIDRanddefaultstudentsintotheplan. Theyfindthatextra informationandcorrectdefaultingdoesincreaseenrollment. 6

empiricalresults. Section6offersconcludingremarksandpolicyconsiderations. 2 College Financing and Educational Insurance This section describes borrower labor market expectations and available (student loan with IDR) and relatively novel (ISAs) higher education financing options with income insurance that motivated our experiment. In Section 2.1, we describe borrowers’ income/employment trajectories and potential risks and disruptionstofutureincomeandemployment. WethenproceedtoexplainthemechanicsofanIDRoption foratraditionalstudentloaninSection2.2andthetypicalISAinSection2.3. InSection2.4,wediscuss how the features of the two financing options might influence choices between them and which types of borrowers might respond to which incentives. Finally, Section 2.5 discusses the motivation behind our experimentandthecomparisonweofferedtostudentsinourstudy. 2.1 Students’EducationalRisks,EarningsTrajectories,andRepaymentShocks Untilrecently, studentsandparentstypicallyrepaidgovernmentstudentloansinfixedmonthlypayments overagivenrepaymentperiod,akintoatraditionalmortgageloan(Karamchevaetal.,2020).Thispayment remainedconstantregardlessofage,income,employmentstatus,orfamilysituation.Mostnon-government lendershaveofferedprivateloanswithastandard,mortgage-stylepaymentschedule,thoughselectlenders arebeginningtoofferorcontemplatealternativeoptions. Some95%ofoutstandingstudentdebtisguaranteedbythefederalgovernment,sorepaymentplansdesignedandofferedbytheDepartmentofEducation dominatethesetofchoicesavailabletostudents. Itisimportanttoconsiderrisktothereturnstocollegeattendance(includingborrowingforthatattendance)whenthinkingaboutfutureincomeandemploymentprospectsforstudents(Webber,2016;BalakrishnanandCynamon,2018; HendricksandLeukhina,2018; Akers,2021). Perhapsthelargestriskfactor to the repayment of educational debt that students face involves the risk of non-completion. Historically, 6-yearcompletionrateshavehoveredaround60%, andareevenlowerfornontraditionalcollegestudents and non-selective institutions (Bowen et al., 2009). Financial circumstances, lack of academic preparedness,andahostofacademicallyorientatedchallengesmayputstudents’financialinvestmentincollegeat risk. Additionally,thereare“life”risksthatstudentsface,includingemergenciesarisingfromphysicalto emotionalhealthtofamilycircumstances. Adultlearners,inparticular,reportthatchild-careemergencies, children’shealth,andeventransportationemergenciescanderailtheireducationalcareers(Markle,2015). Even for completers, risks to income and employment are many. The average financial return for the mediangraduateofa4-yearcollegeoruniversityislarge,andhandilyoutweighstheimplicitandexplicit 7

costsofattendingcollege,whichiswhyenrollmentinapostsecondaryprogramofstudymakessensefor most students ex-ante. But ex-post, returns are heterogeneous across many dimensions including major, institution type, and institution prestige. Since student populations particularly at risk of low or negative returns to college enrollment tend to skew toward vulnerable groups, addressing the riskiness of college attendancewithproduct/programdesignandeffectivepublicpolicyisimperative. Borrowersalsofaceavarietyofincomeshocksduringrepayment,suchthatrepaymentburdenscanvary widely for individuals with variable or uncertain income and/or employment (Chapman and Lounkaew, 2015; Chapman and Dearden, 2017). Borrowers may face temporary repayment challenges (e.g. due to periods of unemployment or underemployment) or chronic repayment struggles due to low incomes (e.g. becauseofdegreenon-completion,ordegree/majorwithpoorfinancialreturnoninvestment). Fixedpaymentsovera10-yearperiodforborrowerswhorecentlycompletedordroppedoutofaprogramofstudymaynotbeoptimalgiventypicalearningstrajectories, either. Formostborrowers, student loandebtserviceratios(i.e., scheduledpaymentsasashareofaborrower’sincome)aretypicallygreater earlyintherepaymentterm,whenaborrower’sincomeislower. Thisisparticularlytrueforstudentloan borrowerswithlittleworkexperienceuponenteringrepayment,forborrowerswhotypicallybeginrepayment in lower paid early-career positions but ultimately earn substantial amounts (e.g., medical doctors), andforborrowerswithdegreesinmajorsthattraditionallyhavesteepearningstrajectories(e.g.,biology). Consideringallofthesefactors,standardrepaymentplanswithfixedscheduledmonthlypaymentsmay beburdensomeforborrowersandpoorlysuitedforasettingwithmyriadriskstoincomeandemployment. 2.2 Income-DrivenRepaymentPlansforFederalStudentLoans In response to these earnings patterns, repayment shocks, and increasing debt service ratios, the Department of Education expanded income-contingent repayment programs (sometimes referred to as incomedriven repayment, or IDR) for federal student loans. IDR plans reduce student loan payments to 5-15% of discretionary income, defined as the amount of adjusted gross income (AGI) above a multiple of the FederalPovertyLevel(FPL).Asofthetimeofourexperiment,thedominantIDRplan(andoneonwhich our student loan with IDR option was modeled) was Revised Pay As You Earn (REPAYE) plan. Under REPAYE, borrowers owed 10% of income over 150% of the FPL for up to 20 (undergraduate loans) or 25(graduateloans)years. ThenewestIDRplanintroducedin2023, theSavingforaValuableEducation (SAVE)IDRplan,setsscheduledpaymentsto5-10%ofincomeover225%oftheFPL.Asof2019,about 35% of borrowers with federal Direct student loans enrolled in one of the several IDR plans offered by the Department of Education (Trends in Student Aid, 2019), with the rest ineligible for IDR (e.g., parent 8

borrowers)orpreferringtoremainwiththestandardplan. Aborrower’spaymentinanIDRplanisdefinedbelow:  0 ifDI it−1 ≤0 P = (1) it rDI it−1 ifDI it−1 >0 where last year’s discretionary income is DI = AGI −mFPL (n), n is the applicant’s family it−1 it−1 t size,m∈{150%,225%},andr ∈{0.05,0.10,0.15}. TheIDRobligationissatisfiedwhena)theoriginal balanceandaccumulatedinterestarerepaid,orb)theborrowerhasreachedthemaximumrepaymentperiod of20or25years(dependingontheparticularIDRplan),whichevercomesfirst. Bydesign,anIDRplan retainsabalance-trackingfeature,suchthatborrowerswhoseIDRpaymentislessthanthemonthlyinterest experienceinterestcapitalizationandcanseetheirbalancesgrow(evenballoon),asfrequentlyreportedin thepopularpress. Again, the hypothetical student loan with IDR in our experiment is modeled after the REPAYE plan, suchthatm=150%∗FPL(n),r =0.10,andthemaximumrepaymentperiodis20years.Wenotethatthe equationabovedeterminesthemaximumIDRpayment,andthataborrower’spaymentinaparticularmonth may amount to less than this maximum payment, such that the IDR option acts as insurance against low income at the potential cost of a higher aggregate payment amount and/or an extended repayment period relativetothe10yearsinthestandardplan. Yet, enrollment rates in these repayment plans remain low for myriad reasons such as administrative hurdles, poor design, and servicer misconduct (e.g. Mueller and Yannelis, 2022). Herbst, 2023 uses randomized variation in loan servicer outreach to find that enrollment in an IDR plan reduces student loan delinquenciesby22percentagepointsanddecreasesoutstandingbalanceswithinayearoftake-up. Many students donot qualifyfor federal studentloans —perhaps because theyhave exhaustedthe federallifetime undergraduate borrowing limit (approximately $30,000), or because they are ineligible for federal loans(e.g.duetoparentalincome/assets,duetoattendinganon-eligibleinstitution,orduetonotbeingUS citizens). Otherstudents(andtheirparents)takeoutloansthatareineligibleforIDRaltogether. 2.3 AlternativeIncome-ContingentFinancing: IncomeShareAgreements One alternative to student loans are ISAs, or income share agreements. Both states and institutions have experimented with ISAs in the past decade. ISA providers claim to fill funding gaps faced by students andtopushprogramstoalignincentiveswithstudentsbecauseinstitutionsreceiverepaymentasgraduates succeedinthelabormarketbecauserepaymentratesaredependentonstudents’post-collegeemployment 9

and income (Ritter and Webber, 2019). Over the past decade, ISAs have covered the spectrum of postsecondaryeducation,fromlargepublicsystemsandprivatenon-profitcollegestocertificateprogramsand small,for-profitcodingbootcamps(RitterandWebber,2019,Zaberetal.,2023). ISAsareneitheratraditionalloanproductnoratraditionalfinancialaidinstrument. Theyareintentionallydesignedtosharesomeoftherisksofpursuingpost-secondaryeducationbetweenthestudentandthe ISAprovider,includingriskstograduationoradverselifeevents.UnderanISAagreement,studentspledge aproportionoftheirfutureearningsinlieuofpayingtuitioninthepresent. ISAstypicallyinclude,asbuiltin features of the financial product, downside protections (a minimum income threshold below which no paymentsaredue,amaximumrepaymentwindow,and/orandmaximumnumberofpayments)andupside protections(maximumcapfortheamountthatcanberepaidbeforetheobligationissatisfied). Aborrower’spaymentinanISAisdefinedasbelow:  0 ifI it ≤I min P = (2) it a∗s∗I it ifI it >I min where P is the scheduled payment under the ISA plan for borrower i in time t, I is the minimum it min incomethreshold,aisthetotalamountborrowed,andsistheincomeshareperdollarborrowed.Differently fromthemaximumIDRpaymentspecifiedinEquation1,whichisdeterminedirrespectiveoftheamount borrowed,theISApaymentscaleswiththeamountborrowed. In terms of the exit criteria from the obligation, ISAs do not track balances, by design. Instead, the providertestswhethertheborrower’scumulativeamountofpaymentsisaboveamaximumcapormultiplier oftheborrowedamount(e.g. 2Xoriginalborrowedamount)ineachperiod, orwhethertheborrowerhas reachedthemaximumnumberofpaymentsorthemaximumtermlength. Inotherwords,thetotalamount ofpaymentsissubjecttothefollowingrestriction: t (cid:88) CumulativePayments = P ≤c∗a (3) t t 1 s.t. NumPayments<MaxNumPaymentsandt<TimeCap. Asbefore,aisthetotalamountborrowed,whilecisthemultiplierormaximumcap. Inotherwords, an ISA obligation is satisfied if a) the borrower reaches the maximum cumulative payment amount, b) theborrowermakestherequirednumberofnon-zeropayments, orc)theborrowerreachesthemaximum repayment window. In our experiment, the hypothetical ISA presented to the students closely resembles theoneconsideredbyTheUniversity-anincomeshareof2.3%ofmonthlyincome,maximumnumberof nonzeropaymentsof120(10years),maximumtermof13years,andamaxcapof2Xtheoriginalborrowed 10

amount. Fromaconsumerprotectionperspective, ISAsarecurrentlynotastightlyregulatedastraditionalstudent loans, with a hodgepodge of state-level approaches to enforcing consumer protection and prudential regulations and an emerging federal regulatory framework for ISAs. The Consumer Financial Protection Bureau considers ISAs private student loans for regulatory purposes, but has provided limited guidance regarding how existing regulations apply to ISAs. For example, the Truth in Lending Act governs the disclosuresofAPR,amongthings,whichisadistinctlyloan-specifictermanddoesnotstraightforwardly applytoanISA-stylecontract.However,althoughtheseimportantconsumerprotectionsconcernsexist,we havechosentoabstractfromtheminourexperimentasthetypicalborrowerwouldbehighlyunlikelytobe awareofanyregulatoryorconsumerprotectiondistinctionsbetweenthetwofinancingoptions. 2.4 BorrowerChoiceBetweenFinancingOptions WhilemanycomparableinsuranceprotectionsarepresentlyavailableunderIDRplansforfederalstudent loans,theyarenotthedefaultoptionintheUnitedStates. Borrowershavetobearthetimeandeffortcost ofenrollingandoftenfulfillmanyrequirementsinordertoremainenrolled,includingburdensomeannual recertification processes for the most generous IDR plans. Additionally, the menu of IDR programs has changedmultipletimessincetheintroductionofthefirstrepaymentprogramofthisstyle–asrecentlyas summerof2023–meaningthattheavailabilityofaparticularrepaymentplanforfederalloansisuncertain, atbest. Thereforesomestudentsmayfindattractivetheex-anteagreementtorepaymenttermsthatwillbe in effect for the duration of the contract inherent with an ISA. Monthly payments vary with income by design, and there is no interest rate, per se, which could be welcome (as payments will never be overly burdensome)orunwelcome(becausepaymentsmaybeseenasunpredictable)forborrowers. Turning to the monthly payment incentives to choose a student loan with IDR versus an ISA, there are several factors which may lead to differing payment levels considering the parameters introduced in Equations1and2.Thisistruebothatapointintimeandintermsofthecumulativepaymentamountdefined above,whichcouldleadborrowerstopreferoneproductovertheother.Forone,thetotaltimeobligationof eachpotentialloanproductispotentiallydifferent. UnderanISA,thelengthisdeterminedbasedon(1)the numberofnon-zeromonthlypayments,and(2)thetotalnumberofmonthselapsedincludingmonthswhere no payment is required. For simplicity, in the hypothetical ISA we offered to students in our experiment these terms were both set at 10 years.7 Therefore borrowers who have a strong preference against long maximumtermsmightprefertheISA. 7Inpractice,whilethemaximumtermlengthandmaximumnumberofpaymentscouldalign,mostISAcontractsdifferentiate betweentotalpaymentlengthandthenumberofnon-zeropayments. 11

From an income trajectory perspective, the shorter obligation period of ISA would be preferable to borrowerswhobelievetheywillfallbelowtheincomethresholdearlyon(forinstanceiftheyhaveprivate knowledgetheywilleitherbeoutofthelabormarket),butmaybesubstantiallyabovethethresholdinlater years. Apart from the length of obligation, the hard discontinuity in ISA payment calculations close to thethresholdcanalsocreateastrongincentivetopreferonecontractovertheother. IDRpaymentshares only apply to discretionary income, e.g. the marginal income above the 1.5*FPL threshold (for a single borrower,intheneighborhoodof$14,000in2022). AndalthoughtheISAminimumincomethresholdof $36,000peryearborrowerishigherthantheIDRminimumincomethreshold,aborrowermakingasingle dollarabovetheISAminimumincomethresholdwillmakeapaymentbasedontheirentireincome,rather thanthemarginalamountabovethethreshold. Thus,borrowersexpectingtoearnincomesbetweenthetwo thresholdsmightbeparticularlymotivatedtochoosethestudentloanwithIDR. Finally, we consider borrowers who may anticipate doing particularly well in the labor market and benefiting from the upside protections in either financing option. Borrowers who might like to have the optionofpayingtheirloansbackmorequicklythantheinitialrepaymentschedulemightpreferthestudent loan because the repayment term in the standard plan is less than 10 years for borrowers making excess paymentsabovescheduledamount. Similarly, borrowersexpectingtopotentiallyearnhighamountsmay be attracted by the lower cumulative amount paid on the standard plan ($12,720 for $10,000 borrowed given typical interest rates) compared with the relatively higher maximum cap on the ISA ($20,000 per $10,000borrowedinourhypotheticalISA).Ofcourse,thispotentiallylowercumulativeamountpaidgiven ahighincomeoutcomeistradedoffagainstthepossibilityofalowincomeoutcomewheretheborrower experiencesincomecapitalizationandmayberequiredtomakepaymentsforupto20years. Ultimately, students’ decisions between financing options are ultimately about earnings expectations (bothlevelandtimepath),preferencesovermonthlypaymentsandupside/downsideprotections,risktolerance,andriskaversion,asdescribedabove. Theyareverysimilartothedecisionsmadeincontemplating othertypesofinsurance. 2.5 MotivationfortheExperiment Student choices between available financial options could provide information useful not only to educationalprogramsandstudents,butalsotoregulatorsastheydecidehowtoapplyoraugmentregulationsin the coming years. Given the similarities between ISAs and IDR plans for federal student loans, insights into student preferences over different features of income-contingent financing can be highly relevant to plans for the expansion or redesign of IDR. As mentioned previously, the Department of Education has 12

recently undertaken the design of a new version of an IDR plan for Direct student loans, the SAVE IDR planlaunchedinAugust2023,soinsightsfromourpaperareparticularlytimely. 8 Despite the recent proliferation of these financing products, there is a surprising dearth of evidence regardingstudentpreferencesoverfeaturesofincome-contingentfinancinggenerallyorISAsinparticular. Currently, regarding the United States, there are no studies exploring experimentally why students may choose an ISA or a product like it over a more traditional student loan. Even globally, there are limited studies(i.e. Herbstetal.,2022)discussingwhichtypesofstudentsmaybenefitfromanISA,oraddressing theeffectsanISAmighthaveonstudents’educationalsuccessandfinancialwell-being. Understandingthe preferences and characteristics of students likely to take up an ISA sheds light on which of them can be helpedorharmedbyeducationalinsuranceandwhy. Finally, ISAs offer a useful laboratory for applying lessons on income-contingent financing to public policyonstudentloans. OurresultsshedlightonIDRplansofferedbytheDepartmentofEducationand otherentities–especiallygiventhatfederalloansgenerallyandIDRplansinparticulararenotavailableto manystudentsandprograms. Thus,theexperimentalevaluationweconductinthepresentpaperhashigh practicalimportance. 3 Research Questions and Experimental Setting 3.1 ResearchQuestions Westudythefollowingresearchquestions. First,howcanframingoremphasizingafinancingproductas educationalinsuranceaffectstudenttake-up? Second,whatfactorsinfluenceastudent’sdecisiontotakeup educationalinsurance? Third,whatisthenatureofadverseselectionineducationalinsurancemarkets? Thefirstquestioninvolvesbehavioraltestingofwhetherexplicitlyemphasizingthelowincomeinsurance features of the ISA affects take-up relative to a federal student loan with the option of IDR. Many studentsareunawareofthedifferencesinupsideanddownsideriskprotectionsthatIDRsandISAsoffer. Intheexperimentalportionofourpaper,werandomizestudentsintoaneducationalinsuranceframingof theISAtoseeifinsurancerelatedattributesincreaseinterestintheprogramanddrawstudentswithprivate informationintotheriskpool. Thesecondquestionexaminestake-upmoderators. Forexample,ISAsaretypicallyofferedbyuniversitiestohelpstudentsmanagetherisksofnon-completion,unemployment,andlowearnings;asweshow below,studentswithhighrisksforthesesituationsarepreciselythestudentswhostandtobenefitthemost 8FormoreinformationontheSAVEIDRplan,see:https://studentaid.gov/announcements-events/save-plan. 13

fromthebuilt-ineducationalinsuranceandaremostlikelytotake-uptheISA.Moregenerally,wemeasure how take-up and the efficacy of an insurance framing varies by student characteristics, specifically those associatedwiththerisksthatISAsinsureagainst. Thethirdquestionaimstounderstandthenatureofadverseselectionandhowpersonalperceptionsof riskaffectstudents’decisionstoparticipateinhypotheticalISAprograms. Ineducationmarkets, adverse selectionariseswhenstudentsuseinformationwhichisnotobservable(orsimplynotpermissibleinalending decision) about their academic ability, possible choice of major, and potential labor market outcomes to sort into an education financing product with insurance features such as an ISA or IDR in a way that iscorrelatedwithrepaymentrates. GiventherichadministrativedataprovidedbytheUniversity,andthe questionsfromoursurveyrelatedtorisktoleranceandincomeexpectations,wecanshowwhichtypesof studentsappeartobeinterestedineducationalinsuranceexante. 3.2 SurveyandExperimentalSettings Thesurveytargetedasampleof8,000studentsenrolledinprogramsthatTheUniversityconsideredoffering possible ISA funding through a third-party provider. The University received a 35% response rate for the survey. Eligible students were undergraduates who were either juniors or seniors (or approximately two years from degree completion) and who have completed a FAFSA form. The University focused on studentsmajoringinpre-licenseandpost-licensenursing,healthinformationmanagement,cybersecurity, orbusinessadministration. PotentialstudyparticipantsreceivedaninvitationfromtheUniversityofferingthemtheopportunityto help the University better understand the student experience with financial aid. The University offered a $10Amazongiftcardstocompensatestudentsfortheirtimeandeffort,informingthemthattheycouldat anytimewithdrawtheirconsenttobeapartofthestudy. TheUniversitybrandedthestudywithitslogo andnameasopposedtotheinstitutionsoftheresearchteamortheISAlender. The survey experiment asked students to choose between a federal student loan with the option of an IDRplanandahypotheticalISA.Monthlyloanpaymentscappedtoafixedshareofanindividual’sincome– orwaivedforverylowincomes–areavailableunderbothoptions,withmarkedlydifferentimplementation and paths for satisfying the loan obligation. In the experiment, students were randomized into two equal groups and, similar to Abraham et al. (2020), the presented options differed in terms of level of detail providedforeachofthechoices. Thefirstgroupwasshowndescriptionsofthestudentloanwithoptionof IDRandtheISA,witharisk-neutralframingthatexplainedthedifferencesinmonthlypayments,general structureoftheloan,thebaselinepaymentterms,andthesourceoffunding. Thetermsofthestudentloan 14

withIDRandtheISAweresettobeactuariallyequivalent. Weexposedthesecondgroup–ourtreatment group–tothesamedescriptionsofthestudentloanwithoptionofIDRandtheISA,butwithanadditional emphasis on the insurance features (nature of the income contingency and maximum repayment term) of thetwofinancingoptions. WerandomizedstudentsintotwodifferentframingsthatcomparedthetermsofahypotheticalISAtoa federalstudentloanwithanoptionofIDR.First,werandomizedhalfoftheparticipantsintothetreatment armthatemphasizedtheISAasaformofinsurancethatprotectsagainstdownsideemploymentrisk,with thestressplacedonfixedrepaymenttermsasopposedtoabalance. Thistreatmentmeasuredtheeffectsof emphasizingthepossibleneedof“educationinsurance"giventhepotentialdownsideriskofcostlyhuman capitalinvestment. Thecontrolgroupconsistedofhalfofthestudents,receivinga“riskneutral”framing which focused on monthly payments and repayment terms. In both the treatment and control groups, studentswereaskedtheirpreferenceandtheintensityoftheirpreferencebetweentheISAandthestudent loanwithIDR.Thissurveyexperimentservesasourprimarycomparisonthroughoutthepaper. Figure1 displaysaflowchartthatdescribestherandomizationinmoredetail. Second,giventhatstudents’riskandtimepreferencesmayshedlightonthenatureofinsurancetakeup,wecreatedasubsequentsurveymodulevaryingthetermsofthestudents’fundingchoice. Ifastudent choseafederalstudentloanwiththeoptionofIDRineitherofthetreatmentarms,weprovidedthemwith alternativeoptionsofISAterms,varyingincomesharepercentageandtermlengthtodoso,whilekeeping theinternalrateofreturn(IRR)constant.ThesealternativeISAoptionsandthestudents’selectionsprovide arevealedpreferencemeasureofdiscountrates,whichcouldbeanimportantmarkerofpotentialadverse selection. Ofcourse,animportantcaveatisthatthesequentialnatureoftheexperimentinducesselectionat eachstage,suchthattheestimatesarenotrepresentativeofthesurveypopulationasawhole. Wedescribe thisingreaterdetailinSection5.4. DetailsoftheresearchdesignandanalysiswerepreregisteredontheSocialScienceRegistry(Balakrishnanetal.,2023). Wediscussourfindingsbelow. 4 Data and Econometric Model 4.1 DataSources For our analysis, we combine data from two unique sources: our primary survey and administrative data fromtheUniversity. Our first data source comes from the survey of students that we conducted in June-July 2022. Out of 15

8,000studentsinvitedtocompletethesurvey,theUniversityreceived2,776responses,a35%participation rate. Thesurveydataincludesdetailedinformationabouthouseholdcomposition,currentlivingsituation, personalincomeaswellashouseholdincome,financialstability9,currenteducationfinancing,andfinancial literacy. Withinthesurvey,werandomizedhowrecipientslearnedaboutahypotheticalISA,framingitas either an insurance option or a risk-neutral option in comparison to a federal student loan with an option of IDR. Our dependent variable is whether they selected the ISA or the federal student loan. Given the importanceofcareerexpectationsonthedesirabilityofinsurance,wealsoaskquestionsoncareer,income, and employment expectations, perceptions on graduation, and risk preferences. Our methodology for all derivedcontrolscanbefoundinAppendixA. Theadministrativedataincludesdetailedinformationaboutsocialanddemographiccharacteristicssuch as age, gender, race/ethnicity, marital status, state of residency, household size, highest education level, employment status, and household income. Along with academic records showing students’ majors, we received information on how many semesters they have completed, attempted and completed number of credits,transfercreditsifany,andcreditsneededtograduate. Thesampleincludes5,465observationsfor 2,776 students – our survey sample – and when available all variables were observed in both June 2022 andJuly2022tomatchthetimeframeofthesurvey’sadministration. Theadministrativedataalsoincludes variables on student finance. The financial aid data include information on students’ academic careers such as their enrollment status (active, dropped, taking a term break, etc.), grade level (undergraduate or graduate), academic standing, receipt of financial aid or award, lifetime federal loan borrowings, and proximitytoaggregatefederalloanlimits. Thefinancialaidsampleincludes7,028observationswith2,776 uniquestudents. Thedataarestructuredasanunbalancedpanelwithannualobservationsofeachstudent forfinancialaidyearsonfilewiththeUniversity,betweenFall2018andFall2022. Thesepropertieshelp ustrackthefinancialaidhistoryandacademicprogressofthestudents. 4.2 EconometricModel ToestimatethedeterminantsforISAtake-up,weestimateEquation4: y =β +β T +β C +γX +ϵ (4) i 0 1 i 2 i i i wherey ,theoutcomevariableofinterest,denotesstudenti’spreferenceforahypotheticalISA(codedas i 1) or a federal student loan (coded as 0). T is a treatment indicator, randomly assigned to respondents, thatrepresentseithertheinsurance-basedframingofanISA(codedas1)orarisk-neutralframing(coded 9SeeAppendixAfordefinition. 16

as 0). C is a vector of the key student characteristics associated with risk and potentially adverse selection. Specifically,weincludethefollowingcharacteristics: currentincomeandemploymentstatus,career expectations including both income and employment expectations after graduation, perceived barriers to graduation,andriskpreferences. X isavectorofdemographiccontrolsincludingage,gender,race/ethnicity,andmaritalstatus. Inthe surveyadministeredtostudents,theUniversityalsocollecteddataonhouseholdsizeandcurrenteducational status. We also include variables that we solicited in our survey before students observed the education insurance framing. These variables include risk preferences, financial stability, measures for income and employmentuncertainty,financingchoicesfortheireducation,andbarrierstograduation. UnlikeinmanyRCTs,whererespondentcharacteristicsandcontrolvariablesareprimarilyincludedto reducebiasorincreaseprecision,wedohaveafirst-orderinterestinthecoefficientsoneveryvariable,as wewishtoassessboththeexistenceandextentofadverseselection. Inotherwords,wetestforeffectsof eachvariableonISApreference. WealsotestforheterogeneouseffectsbyestimatingEquation5: y =β +β T +β T ×H +β H +δC +γ′X +ε (5) i 0 1 i 2 i i 3 i i i i where H is an indicator variable for the dimension of heterogeneity including race/ethnicity, gender, i householdsize, age, maritalstatusandriskaversion. AsinEquation4, T isatreatmentindicator, C isa vectorofthekeystudentcharacteristicsassociatedwithrisk,andX isavectorofdemographicvariables. InSection5.2.1,weadditionallyconsiderdifferencesinISAtake-upinthecontrolandtreatmentgroupsby employmentuncertaintyandincomeuncertainty. 4.3 Balance First,wepresentsummarystatisticsforoursampleandshowcovariatebalanceasevidencethatourrandomizationwassuccessful. Weregresseachofourdemographicvariablesonthetreatmentindicator. Table1 summarizes the characteristics of our sample overall and also checks for balance between treatment and controlgroups. Wedonotseeanyimbalancebetweenourtworandomizedgroupsacrosspre-treatmentcovariatesthatincludeage,sex,race/ethnicity,maritalstatus,householdsize,educationlevel,currentincome andemployment,financialstability,futureincomeandemploymentexpectations,riskaversion,currentfinancing choice, and barriers to graduation. The average student in the survey sample is in their late 30s (37.8),employed,andfinanciallystable. Agreaterfractionofstudentsarewhite(55%)andfemale(71%). Half of the students are married. Most students have an associate’s degree (64%) or a bachelor’s degree 17

(20%).Forhalfofthestudents,theirannualtotalpersonalincomeisbetween$40,000and$90,000whereas almostone-thirdofthestudentshaveanannualtotalhouseholdincomethatis$120,000ormore. Students inoursamplebelievethattheyarehighlylikelytobeemployedfull-timewithin5yearsofgraduationandto earnanincomearoundtheaverageoftheindividualmiddle-incomespectrumat$80,000.Amongbarriersto graduation,theyseeworkresponsibilities(79%),familyconstraints(58%),andfinancialconstraints(41%) as the main reasons for not getting as much out of their educational experience as they would like to. In termsofenrollmentbalance,Columns3and4showthatthestudentsrandomizedintothetreatmentgroup areoverallverysimilartothoseinthecontrolgroup,confirmingthatweexecutedtherandomassignment acrossframingnarrativescorrectly. 5 Results 5.1 ImpactofInsuranceFramingonTake-up First, weexplorewhetherframinganISAasaformofeducationalinsuranceandputtingemphasisonits protection against downside labor market risk is effective in increasing interest in the program. Table 2 summarizesourfindings. Wepresentvariousspecificationinthefollowingway: Column(1)capturesthe baselineeffectoftreatmentwithnoadditionalcontrols.Column(2)addsadditionalcontrolssuchascurrent personal income, employment status, and financial stability (not reported) to the control set. Column (3) adds controls for the expectations for post-graduation career plans (not reported; available upon request), income, and employment. Column (4) adds risk preferences. In Column (5), we incorporate the current financing choice of the student. Model (6) adds controls for barriers to graduation. Finally, Column (7) additionallycontrolsfordemographicvariables–age,gender,race/ethnicity,maritalstatus,householdsize, andeducationalstatus(wereportsignificantcoefficientsonly). Column(7)reflectsthespecificationlisted in our preregistration. Except for Column (1), all specifications include fixed effects for students’ major. Table2containscorrespondingcolumnsforeachofourproposedmodelspecifications. OurresultsindicatethattheinsuranceframingincreasesthelikelihoodoftakingupanISA,compared toafederalloanwiththeoptionofIDR,byroughly10percentagepoints. Thiscoefficientisrobusttoour modelchoiceandstatisticallysignificant.Givenameanof23percentagepointsISAtake-upforthecontrol group,thiseffectcorrespondstoa43%relativeincreaseinthetake-uprateforISAsinthetreatmentgroup. The large baselineeffect is striking, andemphasizes the value that individualsplace on the ISA’sbuilt-in insurance against adverse employment/income outcomes and the shorter maximum repayment term. Our treatment effects for the insurance framing are comparable in magnitude to Abraham et al. (2020) and is 18

stableacrossobservabledimensionsanddemographiccharacteristics. 10 5.2 HeterogeneousTreatmentEffects Variousdemographicgroupsorindividualswithdifferentriskpreferencesmayhavedisparatesusceptibilitiestodefaultriskandthereforemaybeaffecteddifferentlybytheframingofafinancingcontract. Wetest for heterogeneous treatment effects along several dimensions including race/ethnicity, gender, household size, age, marital status and risk aversion. We use the Coefficient of Relative Risk Aversion (CRRA), a measure for risk preferences following a similar method to Kimball et al. (2008). We focus on students’ characteristicswherewewouldexpecttoseeadifferentialimpactexante. Butwedonotfindanystatisticallysignificantheterogeneoustreatmenteffectsunderthecontextofanyofthesecharacteristics. Table3 presentstheheterogeneoustreatmenteffects. Forallsubsamplesweinvestigate,weobservethattheinsuranceframingisthedrivingforcebehindtheISAtake-up.11 Onemighthaveexpectedthatsomeaspectsof adverse selection (e.g. future income risk) could lead to heterogeneous impacts; however, we do not find anyevidenceofadverseselectioninthiscontext. One possible explanation for the lack of differential effects may be related to the imperfect nature of theriskmetricsandsmallsubgroupcellsizesthatmayleaveuswithlittlepowertopreciselyestimateany differentialeffect(Balakrishnanetal.,2023). Additionalanalysis,includingvisualevidenceandmachinelearningmethods,ispresentedinthenextsection. 5.2.1 AdditionalAnalysis We further investigate the heterogeneous treatment effects based on future expectations and uncertainty over employment and income to uncover adverse selection, if any. Figure 2 looks at the percentage of students choosing an ISA given degrees of employment uncertainty and their treatment group. One can observethattreatmentincreasesthepercentageofstudentschoosingISAineverycategoryofemployment uncertainty. For example, among the people who are very certain that they would get a job in the future, 22%ofstudentsinthecontrolgroupinthiscategorychoseISAwhereasthenumberis32%forthetreated students. We investigate whether these percentages are significantly different from each other within and acrosscategories. Theconfidenceintervalsdrawnonthegraphanswerthatquestion. Weseethatthepercentageofstudents choosing ISA is significantly higher in the treatment group than the control group for both those 10Wepreregisteredcorrectionsformultiplecomparisonsusingthefalsediscoveryrate. However,sinceourdependentvariableis adummyvariableindicatingISAtake-up,wedidnotneedtousethiscorrection. 11See Table A1 to Table A6 in Appendix B for detailed results. In addition to the variables in Appendix B, we considered heterogeneouseffectsbasedonincomeuncertaintyandincomeuncertainty, butsimilarlydidnotfindanystatisticallysignificant differences;weomitthoseresultsforbrevityandfocusontheanalysisin5.2.1forthesetwovariables. 19

whoareverycertainandthosewhoaremoderatelyuncertainoftheiremploymentprospects. Onecanalso conclude that the percentage of students choosing the hypothetical ISA for treatment and control groups separatelyacrosscategoriesarenotstatisticallydifferentfromeachotheratthe5%significancelevel.12 Yet thereappearstobesomeevidencethatthosewhohaveveryuncertainemploymentforecastshaveaslightly higher preference for the ISA. That may be why the treatment effect in that subgroup is not statistically significant. Thereforeitcouldbethataninsurance"message"foragroupthatdoesnothavemuchconcern aboutemploymentuncertaintyiswherethereispotentialforbehaviorchange. Figure 3 carries out a comparable analysis for (high) income uncertainty. Similarly to what we observeforemploymentuncertainty,thetreatmentincreasesthepercentageofstudentschoosingISAinevery categoryofuncertaintyoverearningahigherincomeinthefuture. Thegapissmallestforthosewiththe highestlevelofuncertainty.Forpeoplewhoarevery/moderatelycertainabouttheirincomeprospectsbeing high,thetreatmentsignificantlyincreasesthepercentageofstudentstakingupthehypotheticalISA.13 Interestingly,theISApreferenceinthecontrolgroupshowsconsiderableheterogeneity,rising15percentage pointsbetweentheverycertainandveryuncertaincategories(withthedifferencebeingborderlinesignificant). Thetreatmenteffectforthetwogroupsisalmostthesame,soitisthedifferenceinthebaselinethat explainsthesignificanceofthetreatmenteffect(orlackthereof). Figure2andFigure3revealthattheresultsaremostclearfor(high)incomeexpectations,likelybecause thiscategorydoesnothavetheaddedcomplexityoffull-time/part-timeexpectations.Aswouldbeexpected, thereisastrongerpreferencefortheISAcontractamongthosewithveryuncertainbeliefsaboutpositive futureincomeandemploymentprospects. Amongthemostuncertainrespondents,thereisnodiscernible treatment effect. However, as the certainty expressed by respondents increases, the framing of the ISA contractbecomesmuchmoreimportant. Itistheserespondentswithrelativelycertain/securebeliefsabout thefuturethatappeartobedrivingthebaselineresults. One possible mechanism driving this result is the salience of downside risk among different types of respondents.Thosewhoinitiallyexpresssignificantuncertaintyabouttheirfutureprospectsbeingfavorable arealreadypronetoimaginingnegativeoutcomes,andhencetheydidnotneedtheextraframingtoview an ISA in terms of its insurance properties. In other words, for those with considerable uncertainty over highfutureincomes,perhapsnoadditionalmessagingaboutinsuranceisnecessarybecausethataspectof theISAisalreadysalienttothem. Forthosewithlittleuncertainty,though,theprimingfromthetreatment mayhaveagreateropportunitytoaffectsalience.Oneshouldalso,onceagain,notethesmallersamplesize 12Wenotethatinthetreatmentgroup,thepercentageofstudentschoosinganISAisstatisticallydifferentatthe10%significance levelbetweenthosewhoaremoderatelyuncertainandverycertain,andmoderatelyuncertainandveryuncertainabouttheirfuture employment. 13Onecanalsonotethatforthecontrolgroup,thepercentageofstudentschoosingISAisstatisticallydifferentatthe10%significancelevelbetweenthosewhoareverycertainandveryuncertainabouttheirfutureincomebeinghigh. 20

–103studentsinthe“Veryuncertainofhighincome”category–andthefactthattheuncertaintymeasures inthesurveyaredefinedbasedonLikertscales,whichcaptureriskinlimitedways. Asarobustnesscheckandtoincorporatetheotherdimensionsofrisk,wealsoemploythedata-driven approach by Athey and Imbens (2016) to estimate treatment effect heterogeneity. We investigate several risk factors including perceived barriers to graduation, current financing choices, risk aversion, future income/employmentexpectations,anddemographics14 ascandidatesforsourcesofheterogeneity. Thesuggestedmethodologybuildsonregressiontreemethodsfromtheprediction-basedmachinelearningliterature (Breimanetal.,1984;Breiman,2001). Thesampleisfirstdividedintotwoparts: trainingandestimation data. We use the training data to partition the population according to covariates, and then we use the estimation data to obtain treatment effects for each subpopulation. This approach provides an advantage forcaseswheretherearemanycovariatesbylettingthedatatelltheresearcherwhichrelevantsubgroups oneshouldlooktoforheterogeneity. Byseparatingthedatasetsusedtoselectthemodelstructureandto estimate,AtheyandImbens(2016)mitigatethepossiblebiasinmachinelearningmethodswherespurious correlationsbetweencovariatesandoutcomesaffecttheselectedmodel. Theyarguethatalthoughthereducedsamplesizeineachstepthroughpartitionleadstolossofprecision, theestimatesareunbiasedfor everysubpopulation. In this exercise, we simulate with different cut-off points in the sample to get the training and the estimationdata.Ingeneral,allsimulationsreturnhousinginsecurityandotherfinancialconstraints,collegerelatedrisks,workresponsibilities,havinggovernmentversusprivatestudentloans,andriskaversionasthe mainsourcesofheterogeneity. Wepresentevidencethatmanyofthesesubsampleshavesignificanteffects; however,whenwetestthesymmetryofthetreatmenteffects,wefindfewsignificantformsofheterogeneous treatmenteffects. FigureA1inAppendixCsummarizesourfindingswhenweuse40%ofoursampleasthetrainingdata andtherestasourestimationsample. 5.3 AdverseSelectioninTake-up Adverse selection is a key concern to the stability of any insurance market. In particular, preference for a financing contract that depends on a prediction of future income is difficult to navigate for a lender. If the degree of selection is too large, the market can quickly unravel. Historical examples include dental insurance,certainannuitymarkets,andevenlifetimeairfareproducts(Einavetal.,2023). ThereareseveraldimensionsalongwhichadverseselectioncouldpresentitselfinanISAcontext.First, 14Thedemographicsincludeage,gender,maritalstatus,raceandethnicity,householdsize,andeducationalattainment. 21

thecurrentfinancialconditionsofindividuals(highincome,stableemployment,etc.)and/orfutureemployment/incomeexpectationsmayplayasignificantroleinwhetherrespondentswishtopledgeapercentage of future income in exchange for downside risk protection. These student characteristics may not be observabletothelenderormaybeobservablebutnotpermissibletouseinloandecisioning(consistentwith, for example, non-discrimination laws). If these characteristics are correlated with loan repayment rates, then adverse selection is present in the market. Contrary to our expectations, we do not find that current income or employment status is predictive of ISA takeup in our student sample, many of whom report income/employmentconcurrentwiththeirstudies. Similarly, individuals may have private information about their future income/employment prospects, irrespective of their current income/employment, which could substantively influence the decision to buy insuranceandislikelytobecorrelatedwitheventualrepaymentrates.Astudentmaybelievetheyarelikely tohaveahighincomeforanynumberofreasons. Forinstance,theymayhaveaninternshipwithavirtually guaranteed job offer, or their degree may open up a promotion in their current job. In such cases, this privateinformationwouldleadtothesestudentsselectingoutoftheinsurancemarketandoptingtorepaya studentloanasquicklyaspossibleattheleastamountoftotalcost. Conversely,studentswhobelievethey areunlikelytohaveincomesabovetheISAincomethresholdwouldbedisproportionatelylikelytooptfor insurance,asitwouldrendertheirtuitionbilleffectivelyfree. Suchstudentsmayincludethoseanticipating beingoutoftheworkforce,workingpart-time,takinglow-payingjobswithhighnonpecuniarybenefits,or others. Themeasurementofadverseselectionisthekeydeterminantofwhetheragiveninsurancemarket isviable,andisthusanimportantaspectofourcurrentinvestigation. To examine the potential for adverse selection, we asked students to rate the likelihood (based on a 5-point Likert scale) that they would be employed 5 years after graduating. We presented a comparable questionaroundexpectationsoffutureincomebeinglow,moderate,orhigh. Ofparticularnote,atthetime studentsreportedtheirbeliefsaboutfutureincome/employment,theywereunawareofthehypotheticaleducationfinancingchoicestheywouldbepresentedlaterinourstudy.Aswiththecurrentincome/employment variables,uncertaintyoverfutureemploymentprospectsdoesnotappeartodriveISApreferences,indicatingasomewhatsurprisinglackofadverseselectionalongthisdimension.However,higheruncertaintyover futureincomebeinghighdoesappeartohaveaneconomicallylargeandmarginallysignificantimpacton studentchoiceoffinancing. Employment and wage risk conditional on degree completion are not the only risks for providers of education financing. Another important dimension where adverse selection might emerge relates to the likelihood that students do not complete their college degree. Students who believe they are less likely to complete based on private information should be more likely to wish to insure against future negative 22

outcomes. Weseesomeevidenceofthisinthe"college-relatedrisks"variableinTable2. Wealsoinvestigateotherriskstoemployment/income,includinghousingsecurity,familyhealthrisks, and current work responsibilities/risks. We find no evidence that these risk factors seem to be driving take-up. Wedofindthatcertainstudentcharacteristicsarestronglyassociatedwithtake-up. Race,marital status,andthespecificwayinwhichstudentsarealreadyfinancingtheireducationallhavesomepowerin predictingtake-up.15 Namely,studentswhohavepreviouslyfinancedtheireducationwithfederalstudent loansareconsiderablymorelikelytochooseastudentloanwiththeoptionofIDR. Overall, there is less evidence suggestive of adverse selection across a variety of variables than we mighthaveimaginedexante. Futureemploymentuncertainty,forexample,doesnotseemtobeassociated with higher ISA take-up. However, we do find strong suggestive evidence of adverse selection based on uncertainty around high future incomes. Given that income uncertainty is the focal point for educational insurancedecisions,thisfindingisanimportantcontributionofourstudy. 5.4 ChangeinPreferencesinResponsetoChangesinPrice Another dimension we investigate is whether students’ preferences change when offered alternative ISA contracts with different income share percentages and term lengths, while keeping the offers actuarially equivalent(comparableinternalrateofreturn). Asmentionedinsection3.2,studentswhochoseafederal student loan with an option of IDR against the original ISA contract (2.3% share for 10 years) in either of the treatment arms were asked if they would switch to an ISA contract where they would pay smaller installments over a longer period (2.2% share for 12 years). Those who still chose the federal student loan option in this group were then asked to choose between the same federal student loan and another actuarially equivalent ISA contract where they would pay larger installments over a shorter period (2.8% shareforsevenyears). Separately,studentswhochosetheoriginalISAcontractagainstthefederalstudent loan were then presented with both the alternative ISA options at the same time. Figure 4 and Figure 5 depict the flow of the follow-up questions on alternative ISA options depending on the student’s initial choice. Outofthe2001studentswhoinitiallychosethefederalstudentloan,365(18%ofthesample)switchto thelongertermISAcontract. Forthosewhostillpreferredthefederalstudentloanovertheoriginalandthe longer-termISAs,285outof1636(17%)wouldnowswitchtotheactuariallyequivalentISAcontractwitha shortertermandhigherincomeshare.Ontheotherhand,among775studentswhooriginallychosetheISA, 162(21%)wouldstillchoosetheoriginalISA,137(18%)switchtothelongertermISAcontractwherethey 15Wereiteratethattherecanbeadverseselectiononvariables–suchascurrentemployment/incomeandmostdemographics– thatare"observable"tothelenderbutmaynotlegallybeusedindecision-making. Consequently,suchvariablescanbeasourceof adverseselectioneventhoughtheyare"observable." 23

payalowershareoftheirincomeoveralongertimeperiod,while475(61%)optfortheshortertermISA withhigherincomeshare. WhileacomparableshareofstudentsswitchtothelongertermISAcompared with the original student loan choosers, a much higher share (61%) of original ISA choosers switch to a shorter term ISA compared with the original student loan choosers (17%). Several factors may explain thisresult. Ourstudentsampleismorerepresentativeofworkingadults(re)enteringhighereducationmidcareer, as the average age of our sample is 38 and most students are employed while attending college. Comparedwithtraditional-ageundergraduatestudents,theydon’tperceiveasignificantemploymentrisk, as we see in Table 2. Moreover, as these students are further along in their age-earnings profile than traditional-agestudents,theymayreasonablyexpecttoearnmorethanmostgraduatesinsubsequentyears. They may thus prefer a shorter-term ISA to a longer-term income-contingent repayment contract during high-incomeyears. Table 4 considers the treatment effect on the switching behavior, separately for original student loan choosers (Panels A and B) and original ISA choosers (Panels C and D). Panel A shows that treatment increasesthelikelihoodofswitchingfromstudentloantoISAwithalongertermby5.1percentagepoints (33%ofcontrolmean)inourpreferredmodel. Theeffectisrobustacrossallspecificationsandstatistically significant. In Panel B, however, we do not observe a significant treatment effect when we consider the likelihood of switching from federal student loan to the shorter term ISA contract. Given that treated studentswereexposedtoourinsuranceframing,partofwhichemphasizedthemaximumrepaymentterm for a student loan with IDR (20 years) compared with the ISA (12 years), and picked the student loan anyway,itmakessensethatthetreatedstudentsweremuchmorelikelytoselectthelongertermhypothetical ISAovertheoriginalhypotheticalISAcomparedwiththecontrolgroup. Ourfindingsareconsistentwith Mumford(2020). Ontheotherhand,weseecontraryresultsforthosewhochosetheISAcontractinthefirstround. Panel D shows that treatment increases the likelihood of switching to a shorter-term contract significantly for original ISA choosers – by 8.6 percentage points (15% of control mean) in our preferred model. We do notfindasimilareffectforswitchingtoalonger-termISAfromtheoriginalISAinPanelC.Incontrastto ourresultsfortheoriginalstudentloanchoosers,amongoriginalISAchoosersthetreatmentgroupismore likely to switch to a shorter-term ISA after consistently being exposed to our insurance framing of lower maximumrepaymenttermforthehypotheticalISA(12years)comparedwiththestudentloanwithIDR(20 to25years)andpotentiallyselectingtheISAinthefirstroundinpartbecauseofthatfactor. Whengiven the opportunity, in follow-up questions to pick an actuarially equivalent shorter term ISA, these students aremorelikelytodosothanthecontrolgroup. 24

5.5 ProposedStudentLoanCancellation Finally,inaseparatefollow-upquestiontoallstudentsinoursurvey,wetestedwhetherrespondentswould preferthestudentloanwithIDRifthestudentloanoptionincorporatedapossibilityofstudentloancancellation. In the question, we informed respondents of a 20% chance that the federal government would forgivethe$10,000theywereborrowingviathefederalstudentloanwiththeoptionofIDR.Studentscould thenrespondbysayingtheywereeither: (a)muchmorelikelytofavorastudentloan,(b)somewhatmore likelytofavorastudentloan,or(c)thatthenewinformationhadnoeffectontheirpreferenceforastudent loan. Foreaseofinterpretation,wecreateanindicatorvariablethattakesthevalueof1ifastudentexpresses anyadditionalinterestinstudentloans(thefirsttwooptions)andthevalueofzeroifthestudentexpresses nochangeinpreference. Table5showsresultsfromthisquestion,withmodelspecificationscomparableto thoseinTable2,apartfromthedependentvariablebeingincreasedpreferenceforastudentloaninsteadof ISAtake-up. InPanelA,weconsiderallstudentswhoparticipatedintheexperimentandhighlightwhetherastudent choseanISAinthefirstroundasanexplanatoryvariable. Wefindthatstudentswhooriginallychosethe ISAare7.7percentagepointslesslikelytoexpressanincreasedpreferenceforstudentloanswhenpresented withthechanceofstudentloanforgiveness,comparedwithstudentswhochosethestudentloanwithIDR. Essentially,studentswhochosetheISAinthefirstroundappeartoreinforcetheiroriginalreluctancetotake out a student loan. Taken with the other results in our study, this is consistent with the idea that students choosing the ISA are committed to the ISA or especially loan averse, even if there is a chance that the federalgovernmentwillcancelanyloanstheymaytakeout. InPanelB,weconsiderthosestudentswhochosealoaninthefirstroundandtowhomwesubsequently offeredthesmaller-sharebutlonger-termhypotheticalISAinthesecondround.Weshowthatstudentswho switched from the student loan with IDR to the longer-term, smaller-share ISA are 6.4 percentage points morelikely(inourpreferredspecification)topreferthestudentloanwithIDRifitcomeswitha20%chance ofloanforgiveness. ThesestudentsaremarginallyattachedtotheISA(sincetheychangedtheirchoiceto anISAinthesecondround)andmaybeeasilyinducedtoswitchbetweenthetwofinancingchoicesgiven relativelysmallchangestoterms.Thiseffectiscomparableinsizeanddirectionforthegroupwhoswitched totheshorter-termISA(PanelC);furtherevidenceforourhypothesisofmarginalattachment. Finally,inPanelC,weonlyconsiderstudentswhostillchosethestudentloanintheprevioustworounds andtowhomweofferedahigher-sharebutshorter-termISAinthethirdround. Wefindthatthelikelihood ofstudentloanforgivenessinducesasignificantincreaseinexpressedpreferenceforstudentloansby8.9 25

percentagepoints,similartothepreviousgroup.Again,thisresultisevidencethatthosestudentsmarginally attached to an ISA (being switchers to the ISA in the third round) are willing to switch back to a loan if giventheprospectofloanforgiveness. Ourresultsshowthatoffersofstudentloanforgivenessdolittletochangethestudentloanpreferenceof studentswhoinitiallypreferredthehypotheticalISA.However,forstudentswhodidswitchtheirpreference duetoouralternativeoffersofincomeshareandrepaymentterm,thepotentialforstudentloanforgiveness has a large effect on respondents’ preferences changing back to a student loan with IDR. These and our previous results provide suggestive evidence that offering a lower income share that is longer-term may drivemoreadverseselectiontotheISA,thatstudentloanforgivenesscouldpotentiallydrawsomeofthat adverseselectionbackintothefederalloanprogram. 6 Conclusion Income contingent education financing is designed to lessen the burden of repaying federal student loans withanoptionofIDRbytyingthemonthlypaymentstotheborrower’sincome. Students’preferences,the riskfactorsthatmayaffecttheirtake-upandhencethepotentialadverseselectionineducationalinsurance marketsareofinterestforbothpolicymakersandresearchers. Weaddressthesequestionsbyrandomizing themessagingofhypotheticalfinancingcontractsfor2,766studentsatalarge,non-profituniversitytounderstandhowtheframingoffinancialcontracts,alongwithstudents’characteristicsandfutureexpectations, caninfluencedecisions. The results from this study have key implications for how student financing contracts should be designed, and they offer crucial insights into the behavioral responses of students. First, student choices dependonthepresentationofcontractterms. Thesinglebiggestresultfromourinterventionisthedegree towhichinformationandframingimpactsstudents’preferencesforhowtofinancetheireducation.Thefact thatemphasizingtheinsuranceaspectsofanISAleadstoaroughly43%increaseinhypotheticaltake-up underscoreshowimportantmessagingisforbothprivatemarketparticipantslikelendersandservicers,and fortheDepartmentofEducation. Ourstudyclearlyshowsthatprovidingmoredetailsaboutthenuancesoffinancialcontracts,andhow repaymentswouldbeaffectedbyemploymentandfinancialcircumstances,enablesstudentstomakemore informed choices about their financing decisions. The distinguishing feature of the hypothetical ISA is itsproportionalpaymentstructurewithbuilt-ininsurancefeatures. Specifically, theISAcontractrequires repayments proportional to income while providing contingencies in periods of low income; and it caps themaximumamounttorepayment. Studentsdoperceivedownsideincomeprotectionriskasaninsurance 26

whilealsocreatinganincomesmoothingeffect. Ourstudyshowsthatinformationaboutthesefeatureshas astrongeffectonstudentpreference. Second,uncertaintyaboutfutureearningpotentialcontributestoadverseselectionintheISAinsurance market. Theevidencewefindisconsistentwithaconnectionbetweenthetake-upofISAsanduncertainty aboutearningahigherincomeinthefuture. Whilecurrentincomeandfutureunemploymentstatusarenot relatedtotake-up,incomeuncertaintymatters. Otherrisks,withtheexceptionofcollegecompletionrisks, donotseemtopredictISAtake-up. Third,contractchoicemaydependonfamiliaritywithexistingfinancialcontracts: Asignificantbarrier thatISAproviders(oranyalternativeformoffinancinghighereducation)encounteristhatthestudentloan system is well-known, even if not well-liked, by many students. Indeed, our survey finds that students currently financing their education with federal student loans are significantly less likely to prefer ISAs. Giventhepervasivenessofparticipationinthefederalstudentloanmarket,thisdescriptiveresultillustrates theuphillclimbthatanypotentialcompetitortothetraditionalstudentloanmodelfaces. Thisisespecially true given the recent expansion of favorable repayment terms via the SAVE IDR plan and the continuing possibility of loan forgiveness being expanded to more federal student loan borrowers. Indeed, the UniversitylargelyabandonedtheirpotentialISAofferingshortlyaftertheBidenadministration’sloweringof mandatorymonthlypayments. Finally, there are indications that students prefer different financial contracts when they are presented with different versions of monthly payments and contract lengths. Specifically, students who chose the federalstudentloanwiththeoptionofIDRchangedtheirpreferencetoISAwhentheyweregivenalonger contractwithalowerincomesharepercentage. Incontrast, studentswhochosetheoriginalISAcontract preferred a shorter term contract with larger monthly payments. Conventional wisdom suggests that borrowersareaversetothelong-timehorizonforforgivenessofferedbypreviousIDRplans. Thisproposition is one of the justifications both for system-wide student loan forgiveness and also the recent redesign of IDR. Our results cast some doubt on this assertion, as we find that hypothetical borrowers who currently borrow in the student loan system prefer lower payments stretched out over a longer time horizon. This resultislikelymoreusefulforthedesignoffuturegovernmentrepaymentprogramsthanforISAs,asthe governmentisinauniquepositiontobeabletoforegorepaymentoveralongerperiodoftime. Ourstudyexamineswhyeducationalinsuranceremainsunderdevelopedorabsentdespitetheincome and employment risk to student loan borrowers. We randomized students into a framing that presented a riskneutralchoicebetweenastudentloanwithIDRoranISA,oraframingthatemphasizedtheinsurance attributesofanincomeshareagreement. Wefindsignificantevidenceforadverseselectionamongstudents whohadthehighestuncertaintyabouthigherfutureincomes.WealsofindthatthemarginalISAparticipant 27

prefers a plan with a lower share and higher term; perhaps suggesting that those participants have higher discount rates or present biased preferences. However, we do find that this group will switch back when offeredthepossibilityofstudentloanforgiveness. Theseresultsshowthatwhilethereissignificantdemandforeducationalinsurance, adverseselection among students with higher expected income uncertainty or present biased discount rates may make it difficult for traditional firms to provide such insurance in the private market. Furthermore, we find that offers of student loan cancellation may increase the amount of adverse selection present in federal loan programs.Thesefindingsareimportantforourunderstandingofthenatureofinsuranceinhighereducation marketsandhowpolicymakerscouldhelpstudentsmakemoreefficientdecisions. 28

References Abraham, K. G., Filiz-Ozbay, E., Ozbay, E. Y., & Turner, L. J. (2020). Framing effects, earnings expectations, and the design of student loan repayment schemes. Journal of Public Economics, 183, 104067. Akerlof,G.A.(1970).Themarketfor"lemons":Qualityuncertaintyandthemarketmechanism.TheQuarterlyJournalofEconomics,84(3),488–500. Akers,B.(2021).Makingcollegepay:Aneconomistexplainshowtomakeasmartbetonhighereducation. Currency. Arcidiacono,P.,Hotz,V.J.,&Kang,S.(2012).Modelingcollegemajorchoicesusingelicitedmeasuresof expectationsandcounterfactuals.JournalofEconometrics,166(1),3–16. Athey,S.,&Imbens,G.(2016).Recursivepartitioningforheterogeneouscausaleffects.Proceedingsofthe NationalAcademyofSciences,113(27),7353–7360. Baker, R., Bettinger, E., Jacob, B., & Marinescu, I. (2018). The effect of labor market information on communitycollegestudentsâmajorchoice.EconomicsofEducationReview,65,18–30. Balakrishnan,S.,Bettinger,E.,Kofoed,M.S.,Ritter,D.,Webber,D.,Aksu,E.,&Hartley,J.S.(2023).Pre- AnalysisPlanfor:AnEvaluationofIncomeShareAgreements:TheEffectofEducationInsurance FramingandtheNatureofAdverseSelection.https://doi.org/10.1257/rct.8936-1 .2 Balakrishnan, S., & Cynamon, B. Z. (2018). Borrowing arrangements and returns to college education. AvailableatSSRN3263639. Bettinger,E.P.,Long,B.T.,Oreopoulos,P.,&Sanbonmatsu,L.(2012).Theroleofapplicationassistance andinformationincollegedecisions:Resultsfromtheh&rblockfafsaexperiment.TheQuarterly JournalofEconomics,127(3),1205–1242. Black,S.E.,Denning,J.T.,Dettling,L.J.,Goodman,S.,&Turner,L.J.(Forthcoming).Takingittothe limit: Effects of increased student loan availability on attainment, earnings, and financial wellbeing.AmericanEconomicReview. Bodie, Z. (1994). On the risk of stocks in the long run. Capital Markets eJournal. https://api .semanticscholar.org/CorpusID:153602556 29

Bowen,W.,Chingos,M.,&McPherson,M.(2009).Crossingthefinishline:Completingcollegeatamerica’spublicuniversities.PrincetonUniversityPress. Breiman,L.(2001).Randomforests.Machinelearning,45,5–32. Breiman,L.,Friedman,J.,Stone,C.J.,&Olshen,R.A.(1984).Classificationandregressiontrees.CRC press. Broady,K.,&Hershbein,B.(2020).Majordecisions:Whatgraduatesearnovertheirlifetimes.Brookings Institution Papers. https://www.brookings.edu/articles/major-decisions -what-graduates-earn-over-their-lifetimes/ Brodeur,A.,Cook,N.,Hartley,J.,&Heyes,A.(2022).Dopre-registrationandpre-analysisplansreduce p-hackingandpublicationbias?AvailableatSSRN. Chakrabarti,R.,Gorton,N.,&Lovenheim,M.F.(2020).Stateinvestmentinhighereducation:Effectson human capital formation, student debt, and long-term financial outcomes of students (Working PaperNo.27885).NationalBureauofEconomicResearch. Chapman,B.,&Dearden,L.(2017).Conceptualandempiricalissuesforalternativestudentloandesigns: The significance of loan repayment burdens for the united states. The ANNALS of the American AcademyofPoliticalandSocialScience,671(1),249–268. Chapman, B., & Lounkaew, K. (2015). An analysis of stafford loan repayment burdens. Economics of EducationReview,45,89–102. Chetty,R.,&Szeidl,A.(2007).Consumptioncommitmentsandriskpreferences.TheQuarterlyJournalof Economics,122(2),831–877. Conlon,J.J.(2021).Majormalfunction:Afieldexpirementcorrectingundergradaute’sbeliefsaboutsalaries. JournalofHumanResources,56(3),922–939. Cox,J.C.,Kreisman,D.,&Dynarski,S.(2020).Designedtofail:Effectsofthedefaultoptionandinformationcomplexityonstudentloanrepayment.JournalofPublicEconomics,192,104298. Denning,J.T.,&Jones,T.R.(2021).Maxedout?theeffectoflargerstudentloanlimitsonborrowingand educationoutcomes.JournalofHumanResources,56(4),1113–1140. 30

Eide,E.R.,Hilmer,M.J.,&Showalter,M.H.(2016).Isitwhereyougoorwhatyoustudy?therelative influence of college selectivity and college major on earnings. Contemporary Economic Policy, 34(1),37–46. Einav,L.,Finklestein,A.,&Fisman,R.(2023).Riskybusiness:Whyinsurancemarketsfailandwhattodo aboutit.YaleUniversityPress. Friedman, M. (1955). The role of government in education. Economics and the public interest, 2(2), 85– 107. Goyal,A.,&Welch,I.(2007).AComprehensiveLookatTheEmpiricalPerformanceofEquityPremium Prediction.TheReviewofFinancialStudies,21(4),1455–1508. Guiso, L., & Paiella, M. (2008). Risk Aversion, Wealth, and Background Risk. Journal of the European EconomicAssociation,6(6),1109–1150. Hendricks, L., & Leukhina, O. (2018). The return to college: Selection and dropout risk. International EconomicReview,59(3),1077–1102. Herbst,D.(2023).Theimpactofincome-drivenrepaymentonstudentborroweroutcomes.AmericanEconomicJournal:AppliedEconomics,15(1),1–25. Herbst,D.,&Hendren,N.(2021).Opportunityunraveled:Privateinformationandthemissingmarketsfor financinghumancapital.NBERWorkingPaperSeries,29214. Herbst,D.,Palacios,M.,&Yannelis,C.(2022).Equityandincentivesinhouseholdfinance.SSRN,(4398141), 1–47. Karamcheva,N.,Perry,J.,&Yannelis,C.(2020).Income-drivenrepaymentplansforstudentloans:Budgetarycostsandpolicyoptions.CongressionalBudgetOfficeReport. Kimball, M. S., Sahm, C. R., & Shapiro, M. D. (2008). Imputing risk tolerance from survey responses. JournaloftheAmericanStatisticalAssociation,103(483),1028–1038. Kirkeboen,L.J.,Leuven,E.,&Mogstad,M.(2016).Fieldofstudy,earnings,andself-selection.TheQuarterlyJournalofEconomics,131(3),1057–1111. Kofoed,M.S.(2017).Toapplyornottoapply:Fafsacompletionandfinancialaidgaps.ResearchinHigher Education,58,1–39. 31

Levine,P.B.,&Ritter,D.(2023).Theracialwealthgap,financialaid,andcollegeaccess.JournalofPolicy AnalysisandManagement,forthcoming. Lovenheim, M. F., & Smith, J. (2022). Returns to different postsecondary investments: Institution type, academicprograms,andcredentials.HanbookofEconomicsofEducation,6,187–318. Markle,G.(2015).Factorsinfluencingpersistenceamongnontraditionaluniversitystudents.AdultEducationQuarterly,65(3),267–285. Marx,B.M.,&Turner,L.J.(2019).Studentloannudges:Experimentalevidenceonborrowingandeducationalattainment.AmericanEconomicJournal:EconomicPolicy,11(2),108–141. Merton, R. C. (1971). Optimum consumption and portfolio rules in a continuous-time model. Journal of EconomicTheory,3(4),373–413. Mezza,A.,Ringo,D.,Sherlund,S.,&Sommer,K.(2020).Studentloansandhomeownership.Journalof LaborEconomics,38(1),215–260. Mezza, A., & Sommer, K. (2015). A trillion dollar question: What predicts student loan delinquencies? JournalofStudentFinancialAid,46(3). Mueller,H.,&Yannelis,C.(2022).Increasingenrollmentinincome-drivenstudentloanrepaymentplans: Evidencefromthenavientfieldexperiment.TheJournalofFinance,77(1),367–402. Mumford,K.J.(2020).Studentselectionintoanincomeshareagreement.Proceedings.AnnualConference onTaxationandMinutesoftheAnnualMeetingoftheNationalTaxAssociation,113,1–42. Ritter,D.,&Webber,D.A.(2019).Modernincome-shareagreementsinpostsecondaryeducation:Features, theory,applications.FRBofPhiladelphiaPaymentCardsCenterDiscussionPaper,(19-6). Rothstein,J.(2023).Thelostgeneration?labormarketoutcomesforpostgreatrecessionentrants.Journal ofHumanResources,58(5),1452–1479. Stange, K. M. (2012). An empirical investigation of the option value of college enrollment. American EconomicJournal:AppliedEconomics,4(1),49–84. Thakor, A. V. (2020). Fintech and banking: What do we know? Journal of Financial Intermediation, 41, 100833. 32

Webber,D.A.(2016).Arecollegecostsworthit?howability,major,anddebtaffectthereturnstoschooling. EconomicsofEducationReview,53,296–310. Webber,D.A.(2017).Statedivestmentandtuitionatpublicinstitutions.EconomicsofEducationReview, 60,1–4. Webber, D. A. (2022). Decomposing changes in higher education return on investment over time. FEDS Notes. https://www.federalreserve.gov/econres/notes/feds-notes/ decomposing-changes-in-higher-education-return-on-investment-over -time-20220713.html Zaber,M.A.,Steiner,E.D.,Gebremariam,H.,Spears,A.,Tariq,Z.H.,&Carman,K.G.(2023).Income share agreements: Market structure, communication, and equity implications of a student loan alternative.RANDResearchReport,1–35. Zweifel,P.,&Manning,W.G.(2000).Moralhazardandconsumerincentivesinhealthcare.Handbookof HealthEconomics,409–459. 33

7 Tables Table1: SummaryStatisticsandBalance (1) (2) (3) (4) N Overallmean Controlmean Treatment (overall) (std.dev) (std.dev) dummy Currentincome 2776 70235.951 69244.060 1985.211 (37762.513) (37023.692) (1433.248) Currentlyemployed 2315 0.834 0.839 -0.011 (0.372) (0.367) (0.014) Veryuncertainofhighfutureincome 2599 0.040 0.040 -0.002 (0.195) (0.197) (0.008) Veryuncertainoffutureemployment 2739 0.128 0.125 0.006 (0.334) (0.331) (0.013) Riskaversion(CCRA) 2776 3.854 3.848 0.012 (1.095) (1.108) (0.042) Financiallystable 2151 0.775 0.772 0.006 (0.418) (0.420) (0.016) Currentfinancingchoice Cash(selfandfromfamilyandfriends) 1416 0.510 0.500 0.021 (0.500) (0.500) (0.019) Institutionalandfederal/stategrants 916 0.330 0.328 0.003 (0.470) (0.470) (0.018) Governmentandprivatestudentloans 1205 0.434 0.437 -0.006 (0.496) (0.496) (0.019) Other 777 0.280 0.275 0.010 (0.449) (0.447) (0.017) Barrierstograduation Housinginsecurity 210 0.076 0.080 -0.009 (0.264) (0.271) (0.010) Difficultydoingcollege-levelwork 266 0.096 0.091 0.010 (0.294) (0.287) (0.011) Workresponsibilities 2193 0.790 0.790 -0.001 (0.407) (0.407) (0.015) Familyconstraints 1595 0.575 0.581 -0.013 (0.494) (0.494) (0.019) Disabilityorhealthconcerns(physicalormental) 576 0.207 0.201 0.013 (0.406) (0.401) (0.015) Financialconstraints 1124 0.405 0.400 0.011 (0.491) (0.490) (0.019) Internetconnectivityissues 257 0.093 0.088 0.009 (0.290) (0.283) (0.011) Issueswithacademicsupportservices 260 0.094 0.102 -0.016 (0.291) (0.302) (0.011) Difficultybeingatransferstudent 59 0.021 0.024 -0.005 (0.144) (0.152) (0.005) Other 123 0.044 0.045 -0.001 (0.206) (0.207) (0.008) Race/Ethnicity African-American 196 0.071 0.069 0.003 (0.256) (0.254) (0.010) Hispanic 417 0.150 0.143 0.014 (0.357) (0.350) (0.014) Asian 200 0.072 0.068 0.007 (0.259) (0.253) (0.010) Otherrace/ethnicity 1666 0.600 0.614 -0.028 (0.490) (0.487) (0.019) Noethnicityreported 297 0.107 0.105 0.004 (0.309) (0.307) (0.012) Highestdegreeearned Highschooldegreeorless 409 0.147 0.138 0.020 (0.355) (0.345) (0.013) Somecollegedegree 2313 0.833 0.843 -0.020 (0.373) (0.364) (0.014) Graduatedegree 46 0.017 0.017 -0.001 (0.128) (0.130) (0.005) Age 2776 37.778 37.917 -0.278 (9.188) (9.237) (0.349) Female 1970 0.710 0.715 -0.011 (0.454) (0.452) (0.017) Married 1399 0.504 0.511 -0.014 (0.500) (0.500) (0.019) Householdsize 2776 2.932 2.952 -0.042 (1.783) (1.717) (0.068) Notes. Foreverycategoricalvariableandbinaryvariable,N showsthenumberofstudentswho fallintothatcategory. Foreveryothervariable, N showsthenumberofstudentsforwhomthe dataisavailable. Fordetailsonderivedvariablesincludingfutureemploymentuncertainty,future uncertaintyofhighincome,andfinancialstability,seeAppendixA. Theminimumageis19whereasthemaximumis71. Theminimumandmaximumvaluesforthe householdsizeare1and15,respectively. Acontinuouscurrentincomevariableiscreatedfroma 12-categorycategoricalvariable(lessthan$20,000,$20,000-$29,999,...,$110,000-$119,999, $120,000ormore)wherethemiddlepointofeachcategoryisassignedtostudentswhochosethat categoryastheircurrentincomevalue.Theminimumcurrentincomeis$10,000whereasthemaximumis$150,000. 34

Table2: DeterminantsofISATake-up (1) (2) (3) (4) (5) (6) (7) Treatment 0.099*** 0.099*** 0.104*** 0.105*** 0.103*** 0.103*** 0.102*** (0.017) (0.017) (0.018) (0.018) (0.018) (0.018) (0.018) Currentincome -0.007 0.000 0.000 0.001 -0.001 -0.003 (0.014) (0.015) (0.015) (0.015) (0.015) (0.016) Currentlyemployed 0.020 0.023 0.022 0.027 0.025 0.033 (0.031) (0.032) (0.032) (0.032) (0.032) (0.033) Veryuncertainofhighfutureincome† 0.091* 0.091* 0.083* 0.078 0.075 (0.049) (0.049) (0.049) (0.049) (0.049) Veryuncertainoffutureemployment† 0.015 0.015 0.012 0.010 0.014 (0.028) (0.028) (0.028) (0.028) (0.028) Riskaversion(CCRA) -0.003 -0.001 -0.001 -0.000 (0.008) (0.008) (0.008) (0.008) Currentfinancingchoice Cash(selfandfromfamilyandfriends) 0.018 0.016 0.016 (0.022) (0.022) (0.022) Institutionalandfederal/stategrants 0.030 0.029 0.033 (0.020) (0.020) (0.021) Governmentandprivatestudentloans -0.069***-0.072*** -0.074*** (0.022) (0.022) (0.022) Other 0.011 0.009 0.008 (0.021) (0.021) (0.021) Barrierstograduation Housingsecurityandotherfinancialrisks -0.016 -0.009 (0.016) (0.016) College-relatedrisks 0.031* 0.030* (0.017) (0.016) Familyandhealthrisks 0.013 0.013 (0.016) (0.016) Currentworkresponsibilitiesandotherrisks 0.039 0.036 (0.024) (0.025) Race/Ethnicity African-American -0.095** (0.043) Hispanic -0.001 (0.038) Asian -0.012 (0.046) Otherrace/ethnicity -0.041 (0.032) Married 0.034* (0.020) ControlMean 0.230 0.230 0.230 0.230 0.230 0.230 0.230 ControlStd 0.421 0.421 0.421 0.421 0.421 0.421 0.421 Totalnumberofobservations 2776 2776 2599 2599 2599 2599 2594 AdjustedR-Squared 0.012 0.011 0.013 0.013 0.019 0.020 0.022 Notes. VariabledefinitionsfollowthePre-AnalysisPlan;seeAppendixAfordetailsonourderivedvariables. TheregressionmodelinallcolumnsfollowsfromEquation3inthePre-AnalysisPlan. ThedependentvariableistheISA take-up,whichtakesthevalueof1ifthestudentchosethehypotheticalISAoverafederalstudentloanwithIDR,0otherwise.Foreachindependentvariable,wereportthecoefficientandthestandarderrorsinparentheses.Thecoefficientof interest,thetreatmenteffectontheISAtake-up,isreportedinthefirstrow.ExceptforModel1,allmodelsincludefixed effectsforthestudent’smajor.Thepreferredregressionmodel(Column7)alsocontrolsfor(notreportedorstatistically significant)age,gender,householdsize,highestdegreeearned,careerexpectations,andfinancialstability. Weuselogtransformationforcurrentincome. †Uncertaintymeasuresforfutureemploymentandincomearedefinedasindicatorvariablestakingthevalueof1ifthe studentisveryuncertainoftheirfutureemploymentorofearningahighincome,and0otherwise. SeeAppendixAfor detailsontheseandotherderivedvariables. Forriskaversion,weusetheCoefficientofRelativeRiskAversion(CRRA),ameasureforriskpreferencesfollowinga similarmethodtoKimballetal.(2008).Barrierstograduation-originally10variables-arereducedto4maincategories bymaximumlikelihoodfactoranalysis. Wereportthecoefficientsforthesetofstandardizedfactorestimates. Collegerelatedrisksincludedifficultydoingcollegework,beingatransferstudent,andlackofsupportfromadvisor/academic supportservices. Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 35

Table3: HeterogeneousTreatmentEffectsonHypotheticalISATake-Up (1) (2) (3) (4) (5) (6) (7) (8) (9) Blackor Noresponsefor White African-American Hispanic race/ethnicity Female HouseholdSize BelowMedianAge Married RiskAversion × × × × × × × × × Treated Treated Treated Treated Treated Treated Treated Treated Treated ISAtake-up -0.016 0.019 0.058 -0.073 0.002 -0.002 -0.020 0.021 0.019 (0.056) (0.080) (0.069) (0.075) (0.039) (0.010) (0.035) (0.035) (0.016) CurrentIncome Y Y Y Y Y Y Y Y Y CurrentEmployment Y Y Y Y Y Y Y Y Y FinancialStability Y Y Y Y Y Y Y Y Y CareerExpectations Y Y Y Y Y Y Y Y Y FutureIncomeUncertainty Y Y Y Y Y Y Y Y Y FutureEmploymentUncertainty Y Y Y Y Y Y Y Y Y RiskAversion Y Y Y Y Y Y Y Y Y CurrentFinancingChoice Y Y Y Y Y Y Y Y Y BarrierstoGraduation Y Y Y Y Y Y Y Y Y Race/Ethnicity Y Y Y Y Y Y Y Y Y MaritalStatus Y Y Y Y Y Y Y Y Y 55.37% 7.06% 15.02% 10.70% 70.97% 50% 50.40% Samplecomposition White African-American Hispanic Noresponse Female — below-medianage married — Notes.Weonlyreporttheestimatesbasedonourpreferredmodel,Column7inTable2.FullestimatesfromothermodelscanbefoundinAppendixB. VariabledefinitionsfollowthePre-AnalysisPlan;seeAppendixAfordetailsonourderivedvariables.TheregressionmodelfollowsfromEquation3inthePre-AnalysisPlan.We reporttheheterogeneoustreatmenteffectontheISAtake-upforeachdimensionreportedinthecolumnandthestandarderrorsinparentheses.Theregressionmodelalsocontrols for(notreported)age,gender,householdsize,andhighestdegreeearned,aswellasfixedeffectsforstudent’smajor.Weuselog-transformationforcurrentincome. Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 36

Table4: ResponsestoActuariallyEquivalentISAOfferswithDifferentTerms/Shares (1) (2) (3) (4) (5) (6) (7) PanelA:SwitchingfromStudentLoantoISAwithLongerTerm(2.2%sharefor12years) Treatment 0.057*** 0.060*** 0.051*** 0.052*** 0.051*** 0.051*** 0.051*** (0.017) (0.017) (0.018) (0.018) (0.018) (0.018) (0.018) ControlMean 0.156 0.156 0.156 0.156 0.156 0.156 0.156 N 2001 2001 1878 1878 1878 1878 1874 PanelB:SwitchingfromStudentLoantoISAwithShorterTerm(2.8%sharefor7years) Treatment -0.017 -0.019 -0.015 -0.014 -0.016 -0.016 -0.013 (0.019) (0.019) (0.020) (0.020) (0.020) (0.020) (0.020) ControlMean 0.182 0.182 0.182 0.182 0.182 0.182 0.182 N 1636 1636 1543 1543 1543 1543 1539 PanelC:SwitchingfromOriginalISAtoISAwithLongerTerm(2.2%sharefor12years) Treatment -0.035 -0.032 -0.022 -0.025 -0.022 -0.022 -0.032 (0.028) (0.028) (0.029) (0.029) (0.029) (0.029) (0.029) ControlMean 0.197 0.197 0.197 0.197 0.197 0.197 0.197 N 775 775 721 721 721 721 720 PanelD:SwitchingfromOriginalISAtoISAwithShorterTerm(2.8%sharefor7years) Treatment 0.088** 0.085** 0.083** 0.085** 0.083** 0.081** 0.085** (0.036) (0.036) (0.037) (0.037) (0.037) (0.037) (0.037) ControlMean 0.561 0.561 0.561 0.561 0.561 0.561 0.561 N 775 775 721 721 721 721 720 CurrentIncome N Y Y Y Y Y Y CurrentEmployment N Y Y Y Y Y Y FinancialStability N Y Y Y Y Y Y CareerExpectations N N Y Y Y Y Y FutureIncomeUncertainty N N Y Y Y Y Y FutureEmploymentUncertainty N N Y Y Y Y Y RiskAversion N N N Y Y Y Y CurrentFinancingChoice N N N N Y Y Y BarrierstoGraduation N N N N N Y Y Race/Ethnicity N N N N N N Y MaritalStatus N N N N N N Y Notes.StudentswhochoseafederalstudentloanwithIDRagainsttheoriginalISAcontractinthefirstround(ineitherofthearms)werepresented withanactuariallyequivalenthypotheticalISAcontractwheretheywouldpaysmallerinstallmentsoveralongerperiod(2.2%sharefor12years)in thesecondround(PanelA).Thosewhostillchosethefederalstudentloanw/IDRoverthelongertermISAcontractinthesecondroundwerethen giventheoptiontochoosebetweenthesamestudentloanw/IDRandanactuariallyequivalenthypotheticalISAcontractwheretheywouldpaylarger installmentsoverashorterperiod(2.8%sharefor7years)(PanelB).Separately,studentswhochosetheoriginalISAcontract(2.3%sharefor10 years)againstafederalstudentloanwithIDR(ineitherofthearms)werepresentedbothalternative,actuariallyequivalentISAcontractsdescribed aboveatthesametime.OurresultsforswitchingpreferencetothelongertermISAarepresentedinPanelC,andresultsforswitchingpreferenceto theshortertermISAin(PanelD).Forfurtherdetailsonhowweaskedthesefollow-upquestionstothestudents,pleaseseeFigure4andFigure5. VariabledefinitionsfollowthePre-AnalysisPlan. TheregressionmodelinallcolumnsisgivenbyEquation1inthePre-AnalysisPlan,withthe exceptionofthedependentvariable. Thedependentvariabletakesthevalueof1ifthestudentswitchedtothealternativecontractofferedinthe particularround,0otherwise.Weonlyreportthecoefficientofinterest,thetreatmenteffectonthelikelihoodofswitchingtooneofthealternative ISAs.ExceptforModel1,allmodelsincludefixedeffectsforthestudent’smajor.Thepreferredregressionmodel(Model7)alsocontrolsfor(not reported)age,gender,householdsize,andhighestdegreeearned.Weuselog-transformationforcurrentincome.Nisthesamplesize. Thesampleisthe2,001studentswhochosethestudentloanw/IDRovertheoriginalISAofferinroundoneforPanelA;1,636studentswhochose thefederalstudentloanw/IDRoverinthefirstroundandhaven’tswitchedtothelongertermISAinroundtwoforPanelB;and775studentswho chosetheoriginalISAoverthefederalstudentloanw/IDRforPanelsCandD. Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 37

Table5: PreferenceforStudentLoanw/IDRif20%ChanceofLoanForgiveness (1) (2) (3) (4) (5) (6) (7) PanelA:Sample=AllRespondents(FirstRound) ChoseISA=1 -0.084*** -0.083*** -0.081*** -0.081*** -0.078*** -0.078*** -0.078*** (0.021) (0.021) (0.022) (0.022) (0.022) (0.022) (0.022) N 2776 2776 2599 2599 2599 2599 2594 PanelB:Sample=ChoseStudentLoanw/IDRinFirstRound(SecondRound) SwitchedtoLonger-TermISA=1 0.069** 0.062** 0.069** 0.068** 0.067** 0.064** 0.063** (0.028) (0.028) (0.029) (0.029) (0.029) (0.029) (0.029) N 2001 2001 1878 1878 1878 1878 1874 PanelC:Sample=ChoseStudentLoanw/IDRinSecondRound(ThirdRound) SwitchedtoShorter-TermISA=1 0.096*** 0.095*** 0.079** 0.078** 0.082** 0.085*** 0.088*** (0.031) (0.032) (0.032) (0.032) (0.033) (0.033) (0.033) N 1636 1636 1543 1543 1543 1543 1539 CurrentIncome N Y Y Y Y Y Y CurrentEmployment N Y Y Y Y Y Y FinancialStability N Y Y Y Y Y Y CareerExpectations N N Y Y Y Y Y FutureIncomeUncertainty N N Y Y Y Y Y FutureEmploymentUncertainty N N Y Y Y Y Y RiskAversion N N N Y Y Y Y CurrentFinancingChoice N N N N Y Y Y BarrierstoGraduation N N N N N Y Y Race/Ethnicity N N N N N N Y MaritalStatus N N N N N N Y Notes. Allstudents,regardlessoftheirchoicebetweenafederalstudentloanw/IDRandthehypotheticalISAinthecoreexperiment(round1) wereaskedwhethera20%chancethatthe$10,000theyborrowfromthefederalgovernmentviathestudentloanmaybeforgiveninthefuture wouldchangetheirpreferenceforthestudentloanw/IDR. PanelAshowshowchoosingtheoriginalISAinthecoreexperimentisassociatedwiththelikelihoodofpreferringthehypotheticalISAforthe fullsample(round1).PanelBrepeatsthesameexerciseforthestudentswhoinitiallychosethestudentloanbutwerethenofferedanactuariallyequivalenthypotheticalISAwithalowerincomeshareandalongerterm(round2). Finally,PanelCrepeatsthesameexerciseforthestudents whochosethestudentloaninround2andwerethenofferedanactuarially-equivalenthypotheticalISAwithahigherincomeshareandshorter terminround3. VariabledefinitionsfollowthePre-AnalysisPlan; seeAppendixAfordetailsonourderivedvariables. Theregressionmodelinallcolumns followsfromEquation3inthePre-AnalysisPlan. TheregressionmodelscontrolforallvariablesshowninTable2intherespectivemodels. Thedependentvariableisadiscretevariabletakingthevalueof1ifastudentreportedbeingmoreormuchmorelikelytoselectthefederal studentloanifloanforgivenessisavailableand0iftheyreportednochangeintheirpreference. Weonlyreportthecoefficientofinterest,the associationbetweenchoosinganISAandthechangeinpreferenceforafederalstudentloanwithIDRwhenloanforgivenessispossible.Weuse log-transformationforcurrentincome.Nisthesamplesize. Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 38

8 Figures Notes: Thesurveypresentedhypotheticalfinancingoptionstostudentsthroughtwoframings. The presentedoptionsdifferedintermsoflevelofdetailprovidedforeachofthechoices. Halfofthe studentswererandomizedintoa"riskneutral"framingofthestudentloanwithoptionofIDRand theISAthatexplainedthedifferencesinmonthlypayments,generalstructureoftheloan,thebaseline paymentterms,andthesourceoffunding.ThetermsofthestudentloanwithIDRandtheISAwere settobeactuariallyequivalent.Weexposedtheotherhalfofstudents–ourtreatmentgroup–tothe samedescriptionsofthestudentloanwithoptionofIDRandtheISA,butwithanadditionalemphasis ontheinsurancefeatures(natureoftheincomecontingencyandmaximumrepaymentterm)ofthetwo financingoptions.PleaseseeAppendixDforthedetailsofthesurvey. Figure1: CoreExperiment 39

Notes: ThefigureinvestigatestheshareofstudentschoosingthehypotheticalISAgiventheirreportedemploymentuncertaintyandtheirtreatment status.Forexample,amongrespondentswhoareverycertainthatthey’dbeemployedinthefuture,22%ofcontrolstudentsinthiscategorychosethe hypotheticalISAwhereastheshareis32%forthetreatedstudents. TreatmentincreasestheshareofstudentschoosingtheISAineverycategoryof employmentuncertainty.Verticallinesshowtheconfidenceintervals.ThesharesofstudentschoosingtheISAaresignificantlyhigherinthetreatment groupthanthecontrolgroupforthosewhoareverycertainorwhoarearemoderatelyuncertainoftheirfutureemploymentprospects. Onecanalso concludethattheshareofstudentschoosingtheISAfortreatmentandcontrolgroupsseparatelyacrosscategoriesarenotstatisticallydifferentfromeach otheratthe5%significancelevel.SeeAppendixAfordetailsonourderivationoftheemploymentuncertaintycategories. Figure2: HypotheticalISATake-upbyTreatmentandEmploymentUncertainty 40

Notes: ThefigureinvestigatestheshareofstudentschoosingthehypotheticalISAgiventheirreportedincomeuncertaintyandtheirtreatmentstatus. Forexample,amongthepeoplewhoareverycertainthatthey’dearnahighincomeinthefuture,21%ofcontrolstudentsinthiscategorychosethe hypotheticalISAwhereastheshareis32%forthetreatedstudents. TreatmentincreasestheshareofstudentschoosingthehypotheticalISAinevery categoryofincomeuncertainty. Thegapisthesmallestforthosewithhighestuncertaintyoverhighfutureincome. Verticallinesshowtheconfidence intervals. Forstudentswhoareveryandmoderatelycertainabouttheirhighincomeprospects,treatmentsignificantlyincreasestheshareofstudents takingupthehypotheticalISA.SeeAppendixAfordetailsonourderivationoftheincomeuncertaintycategories. Figure3: HypotheticalISATake-upbyTreatmentand(High)IncomeUncertainty 41

Notes:StudentswhochoseafederalstudentloanwithanoptionofIDRagainsttheoriginalISAcontract(2.3%sharefor10years)ineitherof thecoreexperimentarmsinthefirstroundwereaskediftheywouldswitchtoanISAcontractwheretheywouldpaysmallerinstallmentsover alongerperiod(2.2%sharefor12years).Thosewhostillchosethefederalstudentloanoptioninthesecondroundwerethenaskedtochoose betweenthesamefederalstudentloanandanotheractuariallyequivalentISAcontractwheretheywouldpaylargerinstallmentsoverashorter period(2.8%sharefor7years).PleaseseeAppendixDforthedetailsofthesurvey. Figure4: TestingforPreferencesoverActuariallyEquivalentISAswithAlternativeTerms- FederalStudentLoanw/IDRChoosersinCoreExperiment 42

Notes:StudentswhochosetheoriginalISAcontract(2.3%sharefor10years)againstthefederalstudentloanwithanoptionofIDRineither ofthetreatmentarmsinthefirstroundwerepresentedwithtwootherISAcontractsatthesametimeinthesecondround.Thestudentswere askedtochoosebetweenthetwoactuariallyequivalentISAcontractswheretheywouldpaysmallerinstallmentsoveralongerperiod(ISA option#1:2.2%sharefor12years)orlargerinstallmentsoverashorterperiod(ISAoption#2:2.8%sharefor7years).PleaseseeAppendix Dforthedetailsofthesurvey. Figure5: TestingforPreferencesoverActuariallyEquivalentISAswithAlternativeTerms- ISAChoosersinCoreExperiment 43

Appendix A DerivedVariableDefinitions FutureEmploymentUncertainty Forcertaintyoffutureemploymentprospects,weusethefollowingconventiontocategorizerespondents:16 • Category1(VeryCertainofBeingEmployed):Peoplewhoanswered"Likely"tobefullyemployedand"Unlikely"tobeunemployed • Category 2 (Moderately Certain of Being Employed): People who answered “Unlikely” to be unemployed and “Likely” or “Neither likely nor unlikely”tobeemployedpart-time • Category3(ModeratelyUncertainofBeingEmployed):Peoplewhoanswered“Neitherlikelynorunlikely”tobeunemployed • Category4(VeryUncertainofBeingEmployed):Peoplewhoanswered“Likely”tobeunemployed Ourdummyof"Veryuncertainofbeingemployed"takesthevalueof1ifthestudentisinthelastcategory,and0otherwise. FutureIncomeUncertainty Inasimilarmanner,thedegreeofcertaintyoverfutureincome/earningspotentialiscategorizedasfollows: • Category1(VeryCertainofEarningHigherIncome):Peoplewhoanswered“Likely”toearnmorethanUS$75,000and“Unlikely”toearnlessthan US$35,000 • Category2(ModeratelyCertainofEarningHigherIncome):Peoplewhoanswered"Unlikely"toearnlessthanUS$35,000and“Likely”or“Neither likelynorunlikely”toearnbetweenUS$55,001andUS$75,000 • Category3(ModeratelyUncertainofEarningHigherIncome):Peoplewhoanswered"Likely"toearnbetweenUS$35,001andUS$55,000 • Category4(VeryUncertainofEarningHigherIncome):Peoplewhoanswered“Likely”or“Neitherlikelynorunlikely”toearnlessthanUS$35,000 Ourdummyof"Veryuncertainofhighincome"takesthevalueof1ifthestudentisinthelastcategory,and0otherwise. FinancialStability Thefinancialstabilityisassessedusingthefollowingquestion,followingtheFederalReserveBoard’sSurveyofHouseholdEconomicsandDecisionmaking: Supposethatyourhouseholdhasanemergencyexpensethatcosts$400.Basedonyourhousehold’scurrentfinancialsituation,howwouldyour householdpayforthisexpense?Ifyourhouseholdwouldusemorethanonemethodtocoverthisexpense,pleaseselectallthatapply. • 1.Useacreditcardandpayitoffinfullatthenextstatement • 2.Borrowmoneyfromafriendorfamilymember • 3.Useapaydayloan,informalloan,depositadvance,oroverdraft • 4.Usemoneycurrentlyinachecking/savingsaccountorusecash • 5.Usemoneyfromabankloanorlineofcredit • 6.Useacreditcardandpayitoffovertime • 7.Sellsomething • 8.Wouldn’thaveaccesstoanyoftheaboveoptions Studentsare"financiallystable"iftheychoseoption1oroption4(inotherwords,cancoverthe$400expensewithcashorcashequivalent)and"not financiallystable"iftheydidnot. 16Itis,ofcourse,possibletocreatemorecategoriesgiventhehighlevelofgranularityofthedata. Thecategoriesherearedecidedinamannerthatthe certaintymonotonicallygoesfromhighertolowerandthatthereareenoughobservationstomakeameaningfulinferenceineachcategory. 44

Appendix B HeterogeneousTreatmentEffects-FullOutput Thisappendixincludessupplementaryanalysis,withTablesA1-A6presentingthefullresultsofheterogeneoustreatmenteffectsforeachdimensionthatis summarizedinTable3.Wereportallcoefficients,inadditiontotheinteractionterm.Thesetablesconfirmthattheinsuranceframingistheprimarydriving forcebehindtheISAtake-up. TableA1: TreatmentEffectonHypotheticalISATake-Up-byRace (1) (2) (3) (4) (5) (6) (7) Treatment 0.092* 0.094** 0.106** 0.107** 0.110** 0.112** 0.109** (0.048) (0.048) (0.051) (0.051) (0.051) (0.051) (0.051) White 0.026 0.022 0.015 0.015 0.023 0.028 0.080* (0.036) (0.036) (0.037) (0.037) (0.037) (0.037) (0.045) BlackorAfricanAmerican -0.020 -0.019 -0.065 -0.064 -0.065 -0.061 0.002 (0.051) (0.052) (0.051) (0.051) (0.051) (0.051) (0.057) Hispanic 0.014 0.011 0.011 0.011 0.015 0.018 0.076 (0.044) (0.044) (0.046) (0.046) (0.045) (0.045) (0.051) Noresponse 0.066 0.066 0.063 0.063 0.062 0.069 0.144** (0.049) (0.049) (0.051) (0.051) (0.051) (0.051) (0.058) Treatment×White 0.001 -0.002 -0.007 -0.008 -0.015 -0.018 -0.016 (0.053) (0.053) (0.056) (0.056) (0.056) (0.056) (0.056) Treatment×BlackorAfricanAmerican 0.040 0.037 0.026 0.025 0.026 0.023 0.019 (0.078) (0.078) (0.079) (0.079) (0.080) (0.080) (0.080) Treatment×Hispanic 0.049 0.048 0.059 0.059 0.055 0.053 0.059 (0.065) (0.065) (0.069) (0.069) (0.069) (0.069) (0.069) Treatment×Noresponse -0.035 -0.041 -0.073 -0.074 -0.075 -0.077 -0.073 (0.072) (0.072) (0.074) (0.075) (0.075) (0.075) (0.075) N 2776 2776 2599 2599 2599 2599 2594 VariabledefinitionsfollowthePre-AnalysisPlan;seeAppendixAfordetailsonourderivedvariables. Theregressionmodelinallcolumnsfollowsfrom Equation3inthePre-AnalysisPlan. TheregressionmodelscontrolforallvariablesshowninTable2intherespectivemodels. Weuselog-transformation forcurrentincome. N isthesamplesize. Outof2,776respondents,1,537areWhite,196areBlackorAfricanAmerican,417areHispanic,329identify withotherrace/ethnicitycategories,whereas297declinedtoanswer.Otherrace/ethnicitycategoriesincludeAsian(200),AmericanIndianorAlaskaNative (29),NativeHawaiianorOtherPacificIslander(22)andmulti-racial(78). Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 45

TableA2: TreatmentEffectonHypotheticalISATake-Up-byGender (1) (2) (3) (4) (5) (6) (7) Treatment 0.102*** 0.100*** 0.098*** 0.098*** 0.100*** 0.102*** 0.100*** (0.032) (0.032) (0.032) (0.032) (0.032) (0.032) (0.033) Female -0.011 -0.016 -0.013 -0.012 -0.001 -0.002 -0.004 (0.025) (0.028) (0.029) (0.029) (0.029) (0.029) (0.030) Treatment×Female -0.004 -0.001 0.009 0.009 0.005 0.001 0.002 (0.038) (0.038) (0.039) (0.039) (0.038) (0.039) (0.039) N 2776 2776 2599 2599 2599 2599 2594 VariabledefinitionsfollowthePre-AnalysisPlan;seeAppendixAfordetailsonourderivedvariables. Theregressionmodelinallcolumnsfollowsfrom Equation3inthePre-AnalysisPlan. TheregressionmodelscontrolforallvariablesshowninTable2intherespectivemodels. Weuselog-transformation forcurrentincome.Nisthesamplesize. Thecoefficientfortheinteractiontermineverymodelcapturestheheterogeneoustreatmenteffects. Outof2,776respondents,1,970arefemale. Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 46

TableA3: TreatmentEffectonHypotheticalISATake-Up-byHouseholdSize (1) (2) (3) (4) (5) (6) (7) Treatment 0.100*** 0.100*** 0.109*** 0.109*** 0.111*** 0.111*** 0.107*** (0.033) (0.033) (0.034) (0.034) (0.034) (0.034) (0.034) Householdsize 0.004 0.005 0.004 0.004 0.005 0.003 0.001 (0.007) (0.007) (0.007) (0.007) (0.007) (0.007) (0.007) Treatment×Householdsize -0.000 -0.000 -0.002 -0.002 -0.002 -0.003 -0.002 (0.010) (0.010) (0.010) (0.010) (0.010) (0.010) (0.010) N 2776 2776 2599 2599 2599 2599 2594 VariabledefinitionsfollowthePre-AnalysisPlan;seeAppendixAfordetailsonourderivedvariables. Theregressionmodelinallcolumnsfollowsfrom Equation3inthePre-AnalysisPlan. TheregressionmodelscontrolforallvariablesshowninTable2intherespectivemodels. Weuselog-transformation forcurrentincome.Nisthesamplesize. Thecoefficientfortheinteractiontermineverymodelcapturestheheterogeneoustreatmenteffects. Themeanhouseholdsizeinoursampleis2.93withaminimumof1andamaximumof15. Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 47

TableA4: TreatmentEffectonHypotheticalISATake-Up-byAge(MedianSplit) (1) (2) (3) (4) (5) (6) (7) Treatment 0.114*** 0.114*** 0.118*** 0.118*** 0.115*** 0.114*** 0.112*** (0.024) (0.024) (0.025) (0.025) (0.025) (0.025) (0.025) Belowmedianage 0.031 0.030 0.023 0.022 0.015 0.015 0.024 (0.023) (0.023) (0.023) (0.023) (0.024) (0.024) (0.034) Treatment×Belowmedianage -0.031 -0.031 -0.027 -0.027 -0.024 -0.023 -0.020 (0.034) (0.034) (0.035) (0.035) (0.035) (0.035) (0.035) N 2776 2776 2599 2599 2599 2599 2594 VariabledefinitionsfollowthePre-AnalysisPlan;seeAppendixAfordetailsonourderivedvariables. Theregressionmodelinallcolumnsfollowsfrom Equation3inthePre-AnalysisPlan. TheregressionmodelscontrolforallvariablesshowninTable2intherespectivemodels. Weuselog-transformation forcurrentincome.Nisthesamplesize. Thecoefficientfortheinteractiontermineverymodelcapturestheheterogeneoustreatmenteffects. Themedianageinoursampleis37. Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 48

TableA5: TreatmentEffectonHypotheticalISATake-Up-byMaritalStatus (1) (2) (3) (4) (5) (6) (7) Treatment 0.093*** 0.094*** 0.099*** 0.099*** 0.099*** 0.096*** 0.091*** (0.024) (0.024) (0.024) (0.024) (0.025) (0.025) (0.025) Married 0.029 0.030 0.032 0.032 0.035 0.032 0.023 (0.023) (0.023) (0.023) (0.023) (0.023) (0.023) (0.025) Treatment×Married 0.013 0.011 0.011 0.011 0.010 0.015 0.021 (0.034) (0.034) (0.035) (0.035) (0.035) (0.035) (0.035) N 2776 2776 2599 2599 2599 2599 2594 VariabledefinitionsfollowthePre-AnalysisPlan;seeAppendixAfordetailsonourderivedvariables. Theregressionmodelinallcolumnsfollowsfrom Equation3inthePre-AnalysisPlan. TheregressionmodelscontrolforallvariablesshowninTable2intherespectivemodels. Weuselog-transformation forcurrentincome.Nisthesamplesize. Thecoefficientfortheinteractiontermineverymodelcapturestheheterogeneoustreatmenteffects. Outof2,776respondents,1,399aremarried. Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 49

TableA6: TreatmentEffectonHypotheticalISATake-Up-byRiskAversion (1) (2) (3) (4) (5) (6) (7) Treatment 0.058 0.058 0.041 0.041 0.030 0.032 0.027 (0.062) (0.062) (0.064) (0.064) (0.064) (0.064) (0.064) Riskaversion -0.012 -0.011 -0.011 -0.011 -0.010 -0.010 -0.010 (0.010) (0.010) (0.010) (0.010) (0.010) (0.010) (0.010) Treatment×Riskaversion 0.011 0.011 0.016 0.016 0.019 0.018 0.019 (0.015) (0.015) (0.016) (0.016) (0.016) (0.016) (0.016) N 2776 2776 2599 2599 2599 2599 2594 VariabledefinitionsfollowthePre-AnalysisPlan;seeAppendixAfordetailsonourderivedvariables. Theregressionmodelinallcolumnsfollowsfrom Equation3inthePre-AnalysisPlan. TheregressionmodelscontrolforallvariablesshowninTable2intherespectivemodels. Weuselog-transformation forcurrentincome.Nisthesamplesize. Thecoefficientfortheinteractiontermineverymodelcapturestheheterogeneoustreatmenteffects. Outof2,776respondents,1,303(46.94%)arecategorizedas"VeryRiskAverse",571(20.57%)as"ModeratelyRiskAverse",441(15.89%)as"Willingto TakeRisk"and461(16.61%)as"VeryWillingtoTakeRisk". Standarderrorsinparentheses;*p<0.1,**p<0.05,***p<0.01 50

Appendix C RobustnessCheck: AData-DrivenApproachtoUnderstandingAdverseSelection Asarobustnesscheckandtoincorporateeverydimensionofrisk,weemploythedata-drivenapproachbyAtheyandImbens(2016)toestimatetreatment effect heterogeneity. We investigate several risk factors including perceived barriers to graduation, current financing choices, risk aversion and future expectationsaswellasdemographics17ascandidatesforsourcesofheterogeneity. Thesuggestedmethodologybuildsonregressiontreemethodsfromthe prediction-basedmachinelearningliterature(Breimanetal.,1984;Breiman,2001).Thesampleisfirstdividedintotwoparts:trainingandestimationdata. Weusethetrainingdatatocreatepartitionofthepopulationaccordingtocovariates,andthenweusetheestimationdatatoestimatetreatmenteffectsforeach subpopulation. Theapproachprovidesanadvantageforcaseswheretherearemanycovariatesbylettingthedatatelltheresearchertherelevantsubgroups whereoneshouldlookforheterogeneity.Byseparatingthedatasetsusedtoselectthemodelstructureandtoestimate,AtheyandImbens(2016)mitigatethe possiblebiasinmachinelearningmethodswherespuriouscorrelationsbetweencovariatesandoutcomesaffecttheselectedmodel.Theyarguethatalthough thereducedsamplesizeineachstepthroughpartitionleadstolossofprecision,theestimatesareunbiasedforeverysubpopulation. Inthisexercise,wesimulatewithdifferentcut-offpointsinthesampletogetthetrainingandtheestimationdata. FigureA1summarizesourfindings whenweuse40%ofoursampleasthetrainingdataandtherestasourestimationsample. Thetrainingdataidentifieshousingsecurityandotherfinancial risksasthefirstsourceofheterogeneity. Theseconditerationonthesubsamplesbasedonhousingsecurityandotherfinancialrisksreturnriskaversion andcollege-relatedrisksassecondsourcesofheterogeneity. Havingenoughsamplesizeinoneofthebranches,wecarrythethirditerationtoidentify employmentuncertaintyasanothersourceofheterogeneity. Foreachsubsample,wereportthetreatmenteffectsandstandarderrorsinbracketsbasedon theestimationdata.Althoughweshowsignificanttreatmenteffectsforeachsubsample,thestatisticalsignificanceonlyexistsforoneofthebrancheswithin subsample,i.e.,whentheindexforhousingsecurityandotherfinancialrisksislessthan0.69butnottheotherwayaround.Thislackofsymmetryprevents ustohaveconclusiveevidenceonheterogeneityandadverseselection. Wedo,however,presentsuggestiveevidenceforpotentialriskfactorsthatleadto significantheterogeneoustreatmenteffectswhichmayberevealedinlargersampleswithmorepower. 17Thedemographicsincludeage,sex,raceandethnicity,householdsizeandeducationalattainment. 51

Notes: WeusetheCoefficientofRelativeRiskAversion(CRRA),ameasureforriskpreferencesfollowingasimilar methodasKimball,Sahm,andShapiro(2008). Theindicesforhousingsecurityandotherfinancialrisksandcollegerelatedrisksarecreatedviafactoranalysisbasedonstudents’answerstosurveyquestionsonbarrierstograduation.The partition,andhencethep-values,thatdeterminesthesubsampleswheretreatmentheterogeneityexistsarebasedonthe trainingdata.Thenumbersinparenthesesattheleavesindicatehowmanydatapointsinthetrainingdatabelongtoeach leaf. Alltreatmenteffectsareestimatedusingtheestimationdataforeachsubsamplewiththestandarderrorsshownin bracketsandsamplesizeinparentheses. FigureA1: AtheyandImbens(2016)regressiontree 52

Appendix D CoreExperiment-DescriptionoftheRandomizationandTreatment Thesurveypresentedhypotheticalfinancingoptionstostudentsviatwoarmsinwhatweconsiderthe"core"experiment.Halfofthestudentswererandomized intoa"riskneutral"framing.ThiscontrolarmprovidedstudentswithlimitedinformationonafederalstudentloanwiththeoptionofIDRandahypothetical ISA;thisinformationincludedabasicdescriptionofmonthlypayments,repaymentterm,anddownsideprotectionsofthetwofinancingoptions.Theother halfofstudentswererandomlyselectedintothetreatmentarmthatemphasizededucationalinsurancefeaturesthatwouldprotectthestudentagainstdownside employmentandincomerisks,includingprovidingadditionalinformationandmoredetailedexamplesonrepaymentcaps,incomethresholdsfornopayments due,andobligationsatisfactionconditionsforeachfinancingmethod.Theexactwordingofthechoicespresentedtostudentsisshownbelow.Theadditional insuranceemphasisforthetreatmentarmishighlightedinboldtextforthereader(butwasnothighlightedinanywayforsurveyrespondentsinthetreatment group). FigureA2: PresentationofHypotheticalFinancingOptionsintheCoreExperiment: TreatmentandControlGroups 53

AlternativeOptions-StudentLoanChoosersinCoreExperiment Insubsequentrounds,wetestedstudentpreferencesoveralternativecontracttermsforthehypotheticalISA.Weshouldnotethatwewillnotbeabletomake causalinferenceswiththeseprice/termvariations,asthesamplewhowasofferedthealternative,actuariallyequivalent,contractswasendogenoustothecore experiment. Theexactwordingofbothalternative,actuariallyequivalent,hypotheticalISAsisshownbelow. Asbefore,theadditionalinsuranceemphasis forthetreatmentarmishighlightedinboldtextforthereader(butwasnothighlightedforsurveyrespondentsinthetreatmentgroup).Thedifferencesinthe incomesharesandthecontractlengthsbetweentheISAalternativesareunderlinedforthereader(butwerenotunderlinedforsurveyrespondents). Inthesecondround,studentswhochoseafederalstudentloanwithanoptionofIDRagainsttheoriginalISAcontract(2.3%sharefor10years)ineither ofthetreatmentarmsinthefirstroundwereaskediftheywouldswitchtoanISAcontractwheretheywouldpaysmallerinstallmentsoveralongerperiod (2.2%sharefor12years). FigureA3: PresentationofHypotheticalActuariallyEquivalentISAswithAlternativeTerms- OriginalStudentLoanw/IDRChoosersinCoreExperiment- Alternative(Option3): ISAw/LongerTerm(2.2%Sharefor12Years) 54

Inthethirdround,thosewhostillchosethefederalstudentloanoptioninthesecondroundwerethenaskedtochoosebetweentheoriginalfederal studentloanw/IDRandanother,actuariallyequivalent,ISAcontractwheretheywouldpaylargerinstallmentsoverashorterperiod(2.8%sharefor7years). FigureA4: PresentationofHypotheticalActuariallyEquivalentISAswithAlternativeTerms- StudentLoanw/IDRChoosersinSecondRound- Alternative(Option4): ISAw/ShorterTerm(2.8%Sharefor7Years) 55

AlternativeOptions-ISAChoosersinCoreExperiment Ontheotherhand,thestudentswhochosetheoriginalhypotheticalISAcontractoverthefederalstudentloanwiththeoptionofIDRinthecoreexperiment werealsoofferedthealternativeISAcontracts. TheywerepresentedwithallthreeISAoptions(Option2,Option3,andOption4)withdifferentpricing andpaymenttermsdescribedaboveatthesametime,asshownbelow. Asbefore,theadditionalinsuranceemphasisforthetreatmentarmishighlightedin boldtextforthereader(butwasnothighlightedforsurveyrespondentsinthetreatmentgroup).Thedifferencesintheincomesharesandthecontractlengths betweentheISAalternativesareunderlinedforthereader(butwerenotunderlinedforsurveyrespondents). FigureA5: PresentationofHypotheticalActuariallyEquivalentISAswithAlternativeTerms- InitialISAChoosersinCoreExperiment- Alternatives: ISAw/LongerTerm(2.2%Sharefor12Years)orISAw/ShorterTerm(2.8%Sharefor7Years) 56