The Marginal Effect of Government Mortgage Guarantees on Homeownership

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. The Marginal Effect of Government Mortgage Guarantees on Homeownership Serafin Grundl and You Suk Kim 2019-027 Please cite this paper as: Grundl, Serafin, and You Suk Kim (2019). “The Marginal Effect of Government Mortgage Guarantees on Homeownership,” Finance and Economics Discussion Series 2019-027. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2019.027. 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.

The Marginal Effect of Government Mortgage Guarantees on Homeownership Serafin Grundl and You Suk Kim∗ February 27, 2019 Abstract TheU.S.governmentguaranteesamajorityofresidentialmortgages,whichisoftenjustifiedasameanstopromotehomeownership. Inthispaperweuseproperty-leveldatatoestimate the effect of government mortgage guarantees on homeownership, by exploiting variation of the conforming loan limits (CLLs) along county borders. We find substantial effects on government guarantees, but find no robust effect on homeownership. This finding suggests that governmentguaranteescouldbeconsiderablyreducedwithmodesteffectsonhomeownership, whichisrelevantforhousingfinancereformplansthatproposetoreducethegovernment’sinvolvementinthemortgagemarketbyreducingtheCLLs. ∗FederalReserveBoardofGovernors,serafin.j.grundl@frb.gov,you.kim@frb.gov. WethankMaggieChurchfor outstandingresearchassistance. WethankLaurenLambie-HansenforhelpingustounderstandtheCoreLogicdata. WearegratefultoBrentAmbrose,RonelElul,KrisGerardi,DeekshaGupta,FelipeSeverino,ShaneSherlund,Judit Temesvary, Joe Tracy, and James Vickery for helpful discussions. We are particularly grateful to Scott Frame for extensiveandhelpfulcommentsatvariousstagesoftheproject. WewouldliketothankEdKung,DanRingo,Karen Pence, Neil Bhutta, Raven Molloy and Elliot Anenberg for helpful comments. We also thank seminar participants attheFederalReserveBoard,FreddieMac,KoreaUniversity,theGeorgiaStateUniversity-FRBAtlantaRealEstate Finance Conference, the University of Southern California Dornsife INET Conference, the Federal Reserve System AppliedMicroConference,theFederalReserveResearchScrum,theAsianMeetingoftheEconometricSociety,the KEA-KAEAConference,theUCLA-FRBSFHousingConference,theAREUEA-ASSAConference,theConsumer FinancialProtectionBureauandtheNFAConference2018forhelpfulcomments. Theanalysisandconclusionsset fortharethoseoftheauthorsanddonotindicateconcurrencebyothermembersofthestaff,bytheBoardofGovernors, orbytheFederalReserveSystem. 1

1 Introduction AvastmajorityofresidentialmortgagesintheU.S.areguaranteedbythegovernmentthroughthe Government Sponsored Enterprises (GSEs) Freddie Mac and Fannie Mae, as well as the Federal Housing Administration (FHA). The large presence of the government in mortgage financing is controversial because it exposes taxpayers to the risks of the mortgage market. Indeed, the two GSEs went into conservatorship during the financial crisis in 2008 and received $187 billion from taxpayers.1 Thegovernment’sinvolvementinmortgagefinancingisoftenjustifiedwiththegoalofmaking mortgage credit more available and thereby promoting homeownership. Indeed, the GSEs and the FHA state explicitly that homeownership is one of their goals.2 This raises the question of whether,andbyhowmuchgovernmentmortgageguaranteesincreasehomeownership. Ontheone hand, government guarantees could raise homeownership by providing access to mortgage credit toborrowerswhowouldotherwisenotmeettheunderwritingstandards,orbyloweringtheinterest rates. Ontheotherhand,governmentguaranteescouldhavenoeffectonhomeownershipandonly benefit existing homeowners or new homeowners who would have bought a house even without governmentguarantees. In this paper we estimate the effect of government mortgage guarantees on homeownership in a difference-in-differences analysis. A challenge for estimating this effect is to isolate plausibly exogenous variation in government guarantees. Our strategy is based on geographic variation in thechangesoftheconformingloanlimits(CLLs). TheCLLforacountyisthemaximumloansize that can be guaranteed by the government. The CLLs were increased in 2008, and the increases werelargerincountieswithhighermedianhouseprices. In2011theCLLswerepartiallyreduced, andagainthesereductionswerelargerincountieswithhighmedianhouseprices. The potential problem with using this geographic variation is that the changes were not assigned randomly but were a function of the median house price in a county. We circumvent this problem by constructing a sample of adjacent zip codes that were located in different counties and therefore experienced different CLL changes. We show that prior to the CLL changes these adjacentzipcodeshadsimilaraveragehousepricesandhousepricedistributions,similarlevelsof government guarantees per sale, and similar levels of homeownership. In our analysis we allow for different time trends across different border regions with border-time fixed effects. Therefore weexploitvariationinCLLswithinfairlysmallgeographicareaswithsimilarhousingmarketson 1Sincethenhowever,theGSEshavepaidmorethan$270billionoftheirprofitstotheTreasury. 2For example, the mission of the HUD Office of Housing, which oversees the FHA, includes to “maintain and expand homeownership...” (www.hud.gov/program_offices/housing). Similarly, Freddie Mac states that it makes “homeownershipandrentingmoreaccessibleandaffordable”(http://www.freddiemac.com/about/). ForFannieMae’s commitmenttohomeownershipseehttp://www.fanniemae.com/portal/about-fm/homeownership.html. 2

bothsidesoftheborder. Our research question is highly relevant for the ongoing policy discussion about housing finance reform. After the government bailout of the GSEs in 2008, several proposals were made to reduce the role of the government in mortgage financing. Some of the proposals suggest reducing the government’s role gradually by lowering the CLLs.3 Many reform proposals require a legislativeactandthereforehavearelativelysmallchanceofbeingimplemented. TheCLLshowevercan bechangedwithoutCongresssolelyasanadministrativeact. InparticularrecentlyithasbeendiscussedthatlowerCLLscouldbeimplementedbythenewdirectoroftheFederalHousingFinance Agency(FHFA)whowillbeappointedbythebeginningof2019.4 While reform plans discuss lowering the CLLs, the actual CLLs were increased nationwide in 2017, 2018 and 2019, because their level is linked to house prices.5 These CLL increases were welcomed by some market participants, who argued that they would increase homeownership, whichfurtherhighlightsthepolicyrelevanceofourresearchquestion.6 Despitethepolicyrelevancethereareveryfewrecentempiricalpapersestimatingtheeffectof government mortgage guarantees on homeownership.7 Fieldhouse, Mertens, and Ravn (2018) use aggregatetime-seriesdataandanarrativeapproachtoestimatetheeffectofgovernmentguarantees on various macroeconomic variables, including the national homeownership rate. Our main contribution is to provide causal evidence for the effect of government guarantees on homeownership basedonaquasi-experimentalresearchdesignandproperty-leveldata. To estimate the effect on homeownership we use the CoreLogic real estate database, which provides information about characteristics of houses and their transactions at the property level. This database is particularly suitable for studying homeownership because we are able to track whetherahouseisowner-occupiedovertimeandwhethertheowner-occupancystatuschangedas a result of a transaction. This information is an important advantage compared to mortgage-level data sets. Such data sets sometimes record whether the buyer is a first time home buyer, but there is no information about the owner-occupancy status of the seller. Moreover, we observe not only 3For example, the U.S. Congressional Budget Office lays out different ways to reform the secondary mortgage market,includingreducingtheCLLstothepre2008levels;see“TransitioningtoAlternativeStructuresforHousing Finance” (link). Moreover, Senator Corker’s “Housing Finance Reform and Taxpayer Protection Act” (link) and the American Enterprise Institute’s "Taxpayer Protection Housing Finance Plan" (Wallison, Pinto, Pollock, Lawler, Michel,Oliner,andPeter(2018))proposetoreducethegovernment’srolegraduallybyloweringtheCLLs. 4See https://www.housingwire.com/articles/47283-the-most-powerful-person-in-mortgage-lending-is-about-tobe-replaced. 5ThenationwideCLLwasincreasedfrom$417,000to$484,350inthesethreeyears,anannualincreaseofmore than5percent. 6Forexample,inanofficialstatementtheCaliforniaAssociationofRealtorsstated,"IncreasingtheexistingFannie MaeandFreddieMacconformingloanlimitswillprovidestabilityandcertaintytothehousingmarketandgivetens ofthousandsofCaliforniahomebuyersachanceathomeownership."(link). 7There are some earlier papers such as An, Bostic, Deng, Gabriel, Green, and Tracy (2007), Bostic and Gabriel (2006)andGabrielandRosenthal(2008)thatusedecennialcensusdata. 3

transactionsthatwerefinancedwithamortgagebutalsocashpurchases. A first look at the data shows that the loans that became newly eligible for government guarantees as a result of the CLL increases accounted nationwide for up to 20 percent of the monthly dollar volume of new government-guaranteed loans, and for up to 37 percent in the counties with the largest CLL increases. These statistics suggest that the CLL changes may have had a sizeable effect on government guarantees. Moreover, in the regions where the GSE CLLs were increased in 2008, 39 percent of the purchase loans that became newly eligible for government guarantees in2008(jumbo-conformingloans)weretakenoutbyfirst-timehomebuyers,whichisclosetothe level among conventional conforming loans (42 percent). Nationwide only 32 percent of all conventionalconformingloansaretakenoutbyfirst-timehomebuyers.8 Thesenumberssuggeststhat the effect on homeownership could also be sizeable, although the intention of the CLL changes was not to increase homeownership, but to temporarily support the housing market (Fieldhouse andMertens(2017)). To obtain the effect of government guarantees on homeownership we separately estimate the effect of CLL changes on government guarantees and on homeownership. The effect on government guarantees on homeownership can then be obtained by combining both estimates. We find that CLL changes had a substantial effect on government guarantees. It is not surprising that CLL changeshavesomeeffectonguarantees,butquantifyingthesizeoftheeffectisimportanttoevaluatepolicyproposalsthatchangetheCLLs. OurestimatessuggestthattheCLLincreasesin2008 expanded guarantees for new originations on average by more than $50,000 per sale. This sizable effect is roughly equal to 35 percent of the average guarantee per sale prior to the CLL increases. However, we find no robust effect of the CLL changes on homeownership for either the CLL increases or for the subsequent partial reductions. We then investigate whether the impact of CLL changes differed depending on a measure of credit constraints for a typical borrower, the average loan-to-income ratio in a zip code. We find that the effect on government guarantees was larger in zip codes with higher loan-to-income ratios, but again we find no effect on the homeownership rate. Therefore the additional government guarantees through higher CLLs had no sizeable effect onhomeownership. OurestimatessuggeststhattheCLLchangesaffectedthefinancingchoices,butnothomeownership. Thus, for houses that were affected by the CLL changes, further increases in government guarantees had no effect on homeownership. The increase in government guarantees helped borrowerswhoswitchedtogovernment-backedloansandmayhavehelpedsomeborrowerstoincrease theirloansize,butithadonlyanegligibleeffectonmarginalpotentialhomeowners. An important caveat regarding the scope and implications of our findings is that we can only 8ThesefractionswerecalculatedfromFreddieMacandFannieMaeloanleveldatafromMarch2008toFebruary 2011(link). Afirsttimehomebuyerisanybuyerwhodidnotownahouseinthepreviousthreeyears. 4

estimate the marginal effect of changing government guarantees, but we cannot speak to the potential effects of abolishing government guarantees entirely.9 We only observe CLL changes at relatively high levels and cannot extrapolate our estimates to a CLL of zero. Many borrowers who are directly affected by the changes we observe are relatively affluent and have higher credit scores. Government guarantees might be most valuable for low and moderate income borrowers whowouldfinditmoredifficulttoaccessmortgagecreditotherwiseandtypicallytakeoutsmaller mortgages. Despite this caveat we argue that our estimates are relevant for housing finance reform plans that propose to reduce the CLLs gradually from their current level. Unlike other housing finance reform proposals such CLL decreases could be implemented administratively through the FHFA, withoutalegislativeact. Inparticular,ithasbeendiscussedwhetherthenewFHFAdirectorcould lowertheCLLs,ashighCLLsarenolongerneededtosupportthehousingmarket. What are the policy implications of our findings? Regarding housing finance reform, our findingssuggestthatloweringtheCLLs,atleasttopre-2008levels,wouldresultinasubstantialreduction of government guarantees and would have, at most, a moderate effect on the homeownership rate. Conversely,thenationwideCLLincreasesin2017,2018and2019willhave,atmost,amoderate impact on the homeownership rate but a sizable impact on government guarantees. This is concerning, because CLL increases can raise house prices (Adelino, Schoar, and Severino (2012), Kung(2014))andtheCLLlevelsarethemselvestiedtohouseprices,whichcouldleadtoapositive feedbackloopthatdestabilizesthehousingmarketbuthasnosizableeffectonhomeownership.10 Ourfindingssuggestthatmoredirectmeasuresthaninterveninginthemortgagemarketwould likely be more effective at achieving the policy goal of higher homeownership. We do not argue however, that there may not be other policy goals, such housing market stability, for government involvementinthemortgagemarket. Literature Few recent papers try to estimate the effect of the government’s involvement on homeownership. An important exception is Fieldhouse, Mertens, and Ravn (2018) who consider various macroeconomic effects of government asset purchases. One of their findings is that expansions of agency mortgage portfolios have increased homeownership. Methodologically, their approachdifferssubstantiallyfromoursandisthereforecomplementary. Theyuseaggregatetime seriesdataandanarrativeapproachwheresomechangesintheagencymortgageholdingsareclas- 9WediscussthelimitationsofouranalysisinmoredetailinSection5.2. 10Our assessment of the 2017, 2018 and 2019 CLL increases stands in stark contrast to the assessment of the CaliforniaAssociationofRealtorsquotedinfootnote6withrespecttotheeffectonthehomeownershiprateandon the stability of the housing market. Adelino, Schoar, and Severino (2012) and Kung (2014) find a sizable effect of CLLchangesonhouseprices. ThiseffectcannotonlyleadtoapositivefeedbackloopofincreasingpricesandCLLs, butalsopartlyexplainsthemoderateeffectonthehomeownershiprateasaffordabilitygainsthroughCLLincreases areoffsetbypriceincreases. 5

sified as unrelated to “short-run cyclical or credit market shocks.” One advantage of this approach compared to ours is that it allows them to study macroeconomic effects. An advantage of our approach is that we use plausible exogenous regional variation in the government’s involvement in the mortgage market, which allows us to estimate effects relative to an adjacent zip code that servesasacontrolgroup. Our paper is also complementary to papers that study the effect of the GSEs on the broader economyandthefinancialsystemwithcalibratedmodels. Jeske,Krueger,andMitman(2013)and GeteandZecchetto(2017)studythedistributionalimpactsofthegovernmentmortgageguarantees intheeconomy. Jeske,Krueger,andMitman(2013)findthatremovinggovernmentguaranteesfor the GSEs would help low-income and low-asset households, increase aggregate welfare by 0.5%. In contrast, Gete and Zecchetto (2017) find that abolishing the GSEs would hurt low- and midincome households, help high-income households. Both papers find that the home ownership rate would decrease.11 Elenev, Landvoigt, and Van Nieuwerburgh (2016) study the effects of phasingouttheGSEsonthemortgage,housingandfinancialmarkets,allowingforrichinteractions betweenthemarkets. In addition, this paper is more broadly related to several strands of the literature. First, it is relatedtopapersthatstudytheeffectsofgovernmentmortgageguaranteesonthemortgagemarket. A large body of work studies how GSE-eligibility affects mortgage interest rates by comparing jumbo and conforming rates. Early work includes Passmore, Sherlund, and Burgess (2005) and Sherlund (2008). More recently, Kaufman (2014) uses a regression discontinuity design around theCLLtoestimatetheeffectofGSE-eligibilityonmortgagecharacteristicssuchasinterestrates. Inaddition,FusterandVickery(2015)studytheeffectsofsecuritizationontheprevalenceoffixedratemortgages,exploitingthefactthatitismoredifficulttosecuritizeajumbomortgageabovethe CLL. Second, this paper is also related to the literature that studies the determinants and consequences of homeownership. Several papers study the effect of the mortgage interest tax deduction on homeownership, including Poterba (1984), Glaeser and Shapiro (2003), Hilber and Turner (2014), and Sommer and Sullivan (2018). Adelino, Schoar, and Severino (2018) study the importance of perceptions of house price risk for homeownership choices. However, relatively few papersstudytheeffectofcreditmarketconditionsonhomeownership. Mostcloselyrelatedtoour study is Bostic and Gabriel (2006), who exploit differences in the definition of lower-income and underserved neighborhoods under the 1992 GSE Act using decennial census data from California andfindthattheGSEmortgagepurchasegoalsonlyhadasmalleffectonhomeownership.12 Fetter 11GeteandZecchetto(2017)findthatitwouldfallfrom68.5%to66.3%. Jeske,Krueger,andMitman(2013)find thatthefractionofhouseholdswhoownatleastasmuchhousingastheyconsumewoulddecreasefrom44%to40%. 12SeealsoAn,Bostic,Deng,Gabriel,Green,andTracy(2007)andGabrielandRosenthal(2008). 6

(2013) uses the mid-century GI-Bill to study the effect of mortgage subsidies on homeownership among veterans. Acolin, Bricker, Calem, and Wachter (2016) and Fuster and Zafar (2016) study the role of borrowing constraints on homeownership using survey data. Caplin, Cororaton, and Tracy (2015) follow FHA borrowers between 2007 and 2009 over time and estimate that, at most, three-fourths of the borrowers will be able to leave the FHA system by selling their house or by refinancingintoanon-FHAloan. Ahighhomeownershiprateisoftenconsidereddesirableduetothepotentialpositiveexternalities of homeownership. DiPasquale and Glaeser (1999) find some evidence that homeowners are “bettercitizens”. AmiorandHalket(2014)studytheinsuranceroleofhomeownership. Homeownership can however also have detrimental effects on the labor market as studied by Blanchflower andOswald(2013)andLaamanen(2017). Third, another related body of work is the literature that studies the effects of credit conditions on the housing market more generally. Many papers study the effects of interest rates on various market outcomes, including mortgage size (DeFusco and Paciorek (2017)), housing market dynamics (Anenberg and Kung 2017), and home buying (Bhutta and Ringo 2017). Moreover, Adelino, Schoar, and Severino (2012) and Kung (2014) study the effects of credit availability on house prices, exploiting an increase in CLLs at different times, and Anenberg, Hizmo, Kung, and Molloy(2016)studytheeffectsofcreditavailabilityonconstructionaswellashousepricesusing adifferentidentificationapproach. The rest of the paper is organized as follows. In Section 2, we explain the CLL changes that weuseforouranalysis. InSection3,wediscussthedata,howwemeasuretreatmentintensityand some summary statistics. In Section 4, we present the main results. In Section 5, we discuss the policyimplicationsandlimitations. InSection6,weconclude. 2 Changes in Conforming Loan Limits TheGSEscanonlypurchasemortgageloansbelowacertainlimitforthemortgageprincipal,called the conforming loan limit (CLL). Similarly, the FHA can only insure loans below a certain loan limit. Loansabovetheselimitsarecalledjumboloansandhavetoeitherstayonthebalancesheets of the lender or be privately securitized. The CLLs therefore limit the government’s involvement inmortgagefinancing. WeexploitregionalchangesofCLLstoestimatetheimpactofgovernment guaranteesonhomeownership. 7

2.1 Timeline of CLL Changes Figure1showsatimelineofthelegislationthatresultedinchangesoftheGSECLLs. Theshaded regionsinthegraphsshowtherangeofCLLs,whichcouldvaryacrosscounties. BeforeMarch2008,theGSECLLsweresetuniformlyat$417,000intheentirecountryexcept for Alaska, Guam, Hawaii, and the U.S. Virgin Islands. In March 2008, the Economic Stimulus Act (ESA) increased the CLLs. Under the ESA, both the GSE CLLs were set to 125 percent of a county’smedianhouseprice,withanuppercapof$729,750andalowerboundof$417,000. In December 2008, the CLLs specified in the ESA were reduced to the lower CLLs specified in the Housing and Economic Recovery Act (HERA). Under the HERA, the CLLs were equal to 115 percent of the county’s median house price and the cap was lowered to $625,500, while the lowerboundremainedunchangedat$417,000. However,onlytwomonthslaterinFebruary2009, theAmericanRecoveryandReinvestmentAct(ARRA)increasedtheCLLsbacktotheESAlevels again. In October 2011, the GSE CLLs were permanently lowered to the lower levels specified in theHERA,whichwerehoweverstillwellabovethepre-ESAlimits. E S A H E R A A R R A E S A Ex pir e d 729,750 625,500 417,000 8 8 9 1 0 0 0 1 0 0 0 0 2 2 2 2 3/ 2/2/ 0/ 0 1 0 1 Date )$( timiL naoL gnimrofnoC CLL = 417,000 CLL = 1.25 x Median House Price (MHP) CLL = 1.15 x MHP Figure1: TimelineofCLLChanges: Thistimelineshowshowtheconformingloanlimitsforthe GSEs were changed by the Economic Stimulus Act (ESA) in 3/2008, the Housing and Economic Recovery Act (HERA) in 12/2008 and the American Recovery and Reinvestment Act in 2/2009. In 10/2011 the GSE CLLs specified in the ESA expired and the lower CLLs specified in HERA wereusedthereafter. 8

The intention of the ESA, the HERA and the ARRA was not to increase homeownership, but to support the housing market during the crisis (Fieldhouse and Mertens (2017)). In this paper we do not aim to evaluate whether the CLL changes achieved their intended goal, but instead use the policychangestoestimatetheeffectonhomeownership. Indeed,forouranalysisitisadvantageous that the intention of the legislation was not to increase homeownership, which makes it less likely thattheextentbywhichacountybenefitedfromtheCLLincreasesisrelatedtounobservablesthat arerelatedtohomeownership. A first look at the data suggests that the effect on homeownership could have been sizable. In the regions where the GSE CLLs were increased in 2008, 39 percent of the purchase loans that became newly eligible for government guarantees in 2008 (jumbo-conforming loans) were taken out by first-time home buyers. This share is close to the share for conventional conforming loans (42 percent). Nationwide only 32 percent of all conventional conforming loans are taken out by first-timehomebuyers.13 WewouldliketonotethattheselegislationschangednotonlyGSECLLsbutalsoFHACLLs. As shown in Figure 8 in the Appendix, the ESA increased FHA CLLs to the same level as GSE CLLs in March 2008, but the FHA CLLs were lower than the GSE CLLs before the ESA. In January 2014 the FHA CLLs were decreased to the levels specified by the HERA, which are the same as the GSE CLLs under the HERA. However, the GSE CLLs were already decreased in October 2011. To measure government guarantees we do not distinguish whether a loan is guaranteed by the GSEs or insured by the FHA because the government would be exposed to the creditriskoftheloaneitherway. During 2008, at the time of the CLL increase, the jumbo-conforming spread, i.e. the interest rate gap between loans just above and just below the conforming loan limit, was around 80-100 basis points.14 In subsequent years the jumbo-conforming spread narrowed as the housing market calmeddownandprivatejumboloansbecamemoreeasilyavailable. In2011whentheGSECLLs were reduced, the gap had narrowed to around 30-40 basis points and in more recent years the spread even turned negative at some times. As the spread was positive in 2008 and in 2011 we should expect both CLL changes to have an effect, but because the spread was larger in 2008 a givenCLLchangemayhavehadalargereffect. Jumbo-Conforming Share Figure 2 demonstrates the impact of the CLL changes on the portfolio of government guaranteed loans nationwide. The increase in CLLs made loans between 13ThesefractionswerecalculatedfromFreddieMacandFannieMaeloanleveldatafromMarch2008toFebruary 2011(link). Afirsttimehomebuyerisanybuyerwhodidnotownahouseinthepreviousthreeyears.Thereforethese buyers may not be “true” first time buyers, but for our analysis it is only important that they were not homeowners beforetheypurchasedthehouse. 14Seethetimeseriesforthe“adjustedspread”inFigure2inthisCoreLogicInsightsBlogpost(link). 9

the pre-ESA and post-ESA limits eligible for the GSEs and FHA insurance. Such newly eligible loans are commonly referred to as “jumbo-conforming” loans.15 The graph plots the share of jumbo-conforming loans among new purchase loan originations over time. The figure shows the jumbo-conforming share if it is measured as the fraction of the total loan count (count measure) andifitismeasuredastheshareofthetotalcreditextended(dollarmeasure). It shows that the share of jumbo-conforming loans increased after the CLLs were increased in March 2008. Due to the temporary decrease in the CLLs in early 2009 before the ARRA was passed, the share decreased slightly around that time. Eventually, the jumbo conforming share reached a level close to 10 percent using the loan count measure and levels close to 20 percent using the dollar measure. This difference arises because jumbo-conforming loans are larger than conventionalconformingloans. Thus,thefiguresuggeststhattheincreaseinCLLspotentiallyled toasubstantialincreaseingovernmentguarantees.16 After the GSE CLLs were lowered in October 2011 the jumbo-conforming shares decreased substantially. ThereductionofFHACLLsinJanuary2014hadonlyamodesteffectonthejumboconforming share because prior to that change only a small share of FHA loans (2 to 3 percent) was between the ESA and the HERA limits and some of these borrowers responded to the CLL reductionbydecreasingtheloansizeinordertobewithinthenewlimits. 15BecausetheGSEsandtheFHAhaddifferentpre-andpost-ESAlimitsforacounty,aloanthatwouldbeclassified asjumbo-conformingbytheFHAmightstillbeaconformingloan. Forexample,consideracountywhoseFHAlimit increasedfrom$362,790to$729,750andwhoseGSElimitincreasedfrom$417,000to$729,000. AnFHAloanof $400,000wouldbeajumbo-conformingloan,butaGSEloanwiththesamesizewouldstillbeaconformingloan. 16IfwefocusoncountiesthatweremostaffectedbytheESAthejumbo-conformingshareincreasedevenmore.For exampleincountieswheretheGSECLLswereraisedtotheceilingof$729,750thejumbo-conformingsharereached 23percentusingthecountmeasureand37percentusingthedollarmeasure. 10

20 15 10 5 0 )%( erahS 2007m32008m3 2011m10 2014m1 Month Dollar Amount Loan Count Figure 2: The Share of Jumbo-Conforming Loans among New Originations Guaranteed by theGovernment. Thisfiguredisplaystheshareofjumbo-conformingloansamongpurchaseloans originated in each month that are eventually securitized by the GSEs or insured by the FHA. The vertical gray line denotes March 2008 when the ESA increased the CLLs. Source: Black Knight McDashdata. For our analysis, we do not distinguish whether a loan is guaranteed by the GSEs or insured by the FHA because the government would be exposed to the credit risk of the loan either way. However, we use the GSE CLLs rather than the FHA CLLs to define our treatment intensity measure, which captures how much a house was affected by the CLL changes. Moreover we focus on the effect of the GSE CLL reduction in 2011 rather than the effect of the FHA CLL reduction in 2014. We made this choice because as shown in Figure 2 the reduction of GSE CLLs had a much largereffect. Nevertheless,wealsoobtainedestimatesusingtheFHACLLstodefinethetreatment intensityasarobustnesscheck. 2.2 Variation of CLL Changes Across Counties The extent to which the CLLs were raised or lowered varied across counties. This is illustrated in Figure 3, where the GSE CLLs are plotted as a function of a county’s median house price, before and after the ESA. The CLLs prior to March 2008 are shown by the red lines and the post-ESA CLLsbythebluelines. 11

y = 1.25x 729,750 417,000 0 250,000 500,000 750,000 1,000,000 Median House Price ($) )$( timiL naoL gnimrofnoC Pre−ESA Post−ESA Figure 3: Conforming Loan Limit Changes for GSEs through the Economic Stimulus Act (ESA). This figure describes how a county’s GSE CLLs were determined before and after the ESA. The red lines represent the old CLLs before the ESA, and the blue lines represent the new CLLsaftertheESA. Themajorityofcounties,where125percentofthemedianhousepricedidnotexceed$417,000, didnotexperienceanyincreaseintheGSECLLs,sotheredandbluelinescoincide. However,the CLL increased, in so called high-cost counties, where 125 percent of the median house price did exceed$417,000. Theincreasewaslargerforcountieswithhighermedianhouseprices,butasthe CLLswerecappedat$729,750themaximumincreasewas$312,750($729,750minus$417,000). In our empirical analysis, we exploit the regional differences in CLL increases to estimate the effect of government guarantees on homeownership. However, naively using this cross-county variation could be problematic if counties with lower median house prices are not a valid control groupforthehighpricecountieswithlargerCLLchanges. Asweexplaininmoredetailbelow,we circumventthisproblembyfocusingonadjacentzipcodesalongacountyborderthatexperienced differentCLLchanges. WeshowthatpriortotheCLLchangestheseadjacentzipcodeshadsimilar housepricelevels,housepricedistributions,governmentguaranteesandhomeownershiprates. 12

3 Data, Treatment Intensity and Summary Statistics 3.1 Data HomeownershipData ThemaindatasetweusetoestimatetheeffectofCLLchangesonhomeownershipistheCoreLogicRealEstateData(CoreLogicdata,henceforth). Thisdatasetprovides multiple files that contain different types of information. For this paper, we use the file with information about individual house transactions (the deeds file) and the file with information about characteristicsofindividualhouses(thetaxfile). The deeds file provides detailed information about individual house transactions such as the date of the house sale, mortgage characteristics associated with the sale and whether a buyer is an owner-occupant.17 Importantvariablesfromthetaxfilearewhetherahouseisowner-occupiedand the assessed value of the house by tax authorities. Information about whether a house is owneroccupied is crucial for studying homeownership. We need to observe the owner-occupancy status of a house before and after its sale to see whether a house sale leads to a net increase or decrease in homeownership. We obtain information about the owner-occupancy status of a house initially fromthetaxfileandthenupdatethisinformationusingthetransactionfileiftheowner-occupancy status changed as the result of a sale. Thus, this data set allows us to measure homeownership at thehouselevel: whetherahouseisowner-occupiedornot.18 This information is an important advantage compared to typical mortgage data sets. Such data sets sometimes record whether the buyer is a first time home buyer, but there is no information about whether the seller is an owner-occupant. Moreover, we observe not only transactions that werefinancedwithmortgagesbutalsocashpurchases,whichcouldalsoleadtoachangeinowner occupancystatus. Another important variable is the assessed value of a house by tax authorities. This variable is important for predicting the loan size necessary to purchase a house. Many previous papers on relatedtopicsusedappraisalvaluesorlistprices,whichareonlyavailableforhousesthatareonthe market.19 Moreover,theassessedvaluealsoallowsustocontrolforpotentialdifferentialtrendsfor 17CoreLogic constructs the variable indicating whether a buyer is an owner-occupant by comparing the buyer’s mailingaddresswiththepropertyaddress. WewouldlikethankLaurenLambie-Hansonforhelpingustounderstand howthisisdone. 18ThisdefinitionofhomeownershipissimilartothedefinitionofhomeownershipusedbytheU.S.CensusBureau thatistheratioofowner-occupiedhousingunitsandtotaloccupiedhousingunits.TheonlydifferencebetweenourdefinitionandtheCensusdefinitionisthedenominator. Becausewecannotdistinguishoccupiedandunoccupiedhouses, ourdenominatorincludesmorehouses. Forexample,housesusedasvacationhomesareincludedinourdenominator, whereastheyareexcludedintheCensusdenominator. Ourdefinitionofhomeownershipisthereforelikelytounderstatethehomeownershiprateslightly,comparedwiththeCensusdefinition. SeethefollowinglinkformoreinformationaboutthedefinitionofhomeownershipusedbytheCensus: https://www.census.gov/housing/hvs/definitions.pdf. 19For example, Adelino, Schoar, and Severino (2012) use the appraisal value in predicting whether a house will benefitfromanincreaseinCLLs. Kung(2014)usesthelistpriceofahouseonthemarketforasimilarpurpose. 13

different segments of the housing market. A potential problem with this variable is that in some counties assessed values tend to be far below market prices. Therefore we adjust the assessed values by exploiting data from sold properties where purchase price data is available. We discuss thisadjustmentinmoredetailbelow. Government Guarantee Data To estimate the effect of CLL changes on the amount of government guarantees we also use the CoreLogic data. An important limitation is that the data does not allowustoobservedirectlywhetheraloancarriesagovernmentguarantee. Insteadweassumethat a loan carries a government guarantee if it is eligible for a guarantee. To evaluate whether this is areasonableapproximationweusetheBlackKnightMcDashmortgagedataset,whichassembles data from several large mortgage servicers.20 This data allows us to see whether a loan carries a government guarantee, either through the GSEs or the FHA, which is not recorded in the CoreLogicdata. IntheBlackKnightMcDashdata,91.4percentofloansthatareeligibleforgovernment guaranteesareindeedguaranteedbythegovernment. SampleSelection Weselectthesubsampleforouranalysisasfollows. Wekeeponlyresidential properties such as single-family houses or condos. However, we exclude apartments. Throughout the paper, we will refer to all properties in our sample, including condos, as “houses”. We also drophousesthatwentthroughforeclosureduringthesampleperiod. 3.2 Treatment Intensity We measure the treatment intensity at the house level by calculating how much the CLL change increasesordecreasesthefractionofthehousevaluethatcanbefinancedwithaconformingGSE loan. Formallywedefinehousei’streatmentintensityT asfollows: i (cid:110) (cid:111) (cid:110) (cid:111) min 0.8V,CLLGSE −min 0.8V,CLLGSE i c(i),post i c(i),pre T = , (1) i V i where V refers to house i’s adjusted value assessed for tax purposes prior to the beginning of i our sample period.21 CLLGSE and CLLGSE refer to the GSE CLL before and after the CLL c(i),pre c(i),post changes for house i’s county c(i), respectively. For a CLL increase T measures the additional i proportionofV thatcanbefinancedwithaGSEloan,assumingaborrowermakesadownpayment i of 20 percent. Analogously, for the partial CLL decrease in 2011 T measures the reduction of the i 20Thedatacontainsinformationonmorethan175millionmortgagesandhomeequityloans. 21FortheanalysisoftheCLLincreasesin2008weusetheassessedvaluein2006andfortheanalysisoftheCLL reductionsin2011and2014weusetheassessedvaluein2010. NotethatintheCoreLogicdatatheassessedvalues areavailableregardlessofwhetherahousewassold. 14

shareofV thatcanbefinancedwithaGSEloan. Weusehousei’svalueassessedfortaxpurposes i priortothebeginningofoursampleperiodtomeasureV,soT isunchangedthroughoutoursample i i period. Asassessedvaluesareclosetoactualpricesinsomecountiesbutnotinothers,werescale the assessed values by multiplying it with the median ratio of purchase price to assessed value for eachcounty. OurmainestimatesarebasedonthechangesinGSECLLstocalculatethetreatmentintensity. As a robustness checks we also obtain estimates using the changes in FHA CLLs. We also report robustness checks with a 10 percent downpayment to calculate T, because such loans are eligible i forGSEguaranteesiftheborrowerhasmortgageinsurance. 3.3 Sample of Border Zip Codes UsingthevariationinCLLchangesacrosscountiesispotentiallyproblematicbecausethechanges were not assigned randomly but were a function of the median house price in a county. To circumventthisproblemweassembleasampleofadjacentzipcodesthatarelocatedintwodifferent counties where the zip code to the left and to the right of their common border experienced differentchangesofGSECLLs. In2008theCLLsforGSEloanswereincreasedinsocalled“high-cost” counties, but remained constant elsewhere. Similarly, the GSE CLLs were partially decreased in high-costcountiesin2011andremainedunchangedelsewhere. Our sample includes borders between high-cost counties where the GSE CLLs were increased and adjacent counties where the GSE CLLs remained unchanged. In addition, we also exploit variation of CLL changes within the region of high-cost counties by including borders where the CLL changes on both sides of the county border were different. We exclude border regions where theCLLsdifferedbylessthan$50,000toguaranteeaminimumwithin-bordervariationofCLLs. Figure 4 shows the zip codes in our sample on a map. Naturally, both coasts, and especially California, account for a sizable part of the sample, because they account for a large share of high-costcounties. For our analysis we focus on houses with assessed values in 2006 between $500,000 and $1,000,000. We argue that this segment of the housing market was most affected by the CLL changes. The idea is that houses with values below $500,000 were not much affected by the CLL increases because even with the prior CLLs of $417,000 they could have been financed with a conformingloan. Forhousesabove$1,000,000theCLLincreaseslikelyalsoplayedasmallerrole because the CLL increases represent a smaller fraction of the house value. Moreover, unless the down payment is unusually large, such houses cannot be financed with a conforming loan even aftertheCLLincreases.22 22In the next subsection in Figure 5b we show that government guarantees have increased for houses between $500,000and$1,000,000asaresultoftheCLLincreasesin2008. IntheAppendix,Figure9weshowthatthisisnot 15

Figure 4: Map of Border Zip Codes This map shows the zip codes along county borders we use in our analysis. Adjacent zip codes to both sides of the county border experienced CLL changes throughtheESAthatdifferedbyatleast$50,000. SummaryStatisticsPriortoCLLIncrease Table1showssummarystatisticsfromMarch2007 prior to the CLL increase. In our econometric analysis we exploit variation in treatment intensity, which varies continuously at the house level. To present summary statistics and graphs however, we divide the houses by zip code into two groups. The left column shows the zip code for each borderthatexperiencedthesmallerCLLchangeandtherightcolumnshowsthezipcodewiththe higher CLL change. We refer to these two groups as “lower CLL” and “higher CLL” zip codes, respectively. BeforetheCLLswereincreasedin2008 theyweresetat$417,000nationwide. AftertheCLL increase they increased by $70,371 to $487,371 for the zip codes in the left column. For the zip codesintherightcolumntheCLLsincreasedby$277,264to$694,264. The average adjusted assessed house value in 2006 is similar in both groups of zip codes. The averageadjustedassessedvaluein“lowerCLL”zipcodeswas$672,033comparedto$683,110in thecaseforhousesbelow$500,000.Thissuggeststhatthissegmentofthehousingmarketwasindeednotsubstantially affectedbytheCLLchanges. 16

“higher CLL” zip codes. The indicator variables for house value bins ranging from $500,000 to $1,000,000showthatnotonlywastheaverageadjustedassessedvaluesimilarbutthehousevalue distributionwassimilaraswell. Table1: BorderSampleSummaryStatistics-PriortoCLLIncrease. Thistablepresentssummary statistics from March2007 beforethe ESA increasedthe CLLs. Column (1) containsborder zip codes where the CLL either remained constant or increased only slightly, whereas column (2) showsadjacentzipcodeswheretheCLLwasincreasedsubstantially. (1) (2) LowerCLL HigherCLL Pre-TreatmentCLL($) 417,000 417,000 Post-TreatmentCLL($) 487,371 694,264 ChangeinCLL($) 70,371 277,264 AdjustedAssessedHouseValuein2006($) 672,033 683,110 ShareofHouseValue∈[$500K,$600K) 0.383 0.351 ShareofHouseValue∈[$600K,$700K) 0.254 0.248 ShareofHouseValue∈[$700K,$800K) 0.170 0.182 ShareofHouseValue∈[$800K,$900K) 0.116 0.129 ShareofHouseValue∈[$900K,$1000K) 0.077 0.090 ShareofHouseswithT >0 0.660 0.903 i AvgT 0.064 0.157 i AvgT forHouseValue∈[$500K,$600K) 0.025 0.040 i AvgT forHouseValue∈[$600K,$700K) 0.086 0.154 i AvgT forHouseValue∈[$700K,$800K) 0.096 0.233 i AvgT forHouseValue∈[$800K,$900K) 0.089 0.282 i AvgT forHouseValue∈[$900K,$1000K) 0.077 0.293 i ShareofOwner-occupiedHouses 0.837 0.829 ProbabilityofHouseSale 0.022 0.021 NumberofHouses 247,752 323,231 Next, consider the treatment intensity for both groups. The average treatment intensity in zip codeswithsmallerCLLincreaseswas0.064comparedto0.157inzipcodeswithlargerincreases. In other words, in the lower-CLL zip codes an additional 6.4 percent of the assessed house value canbefinancedwithaGSEloanduetotheCLLincreases,comparedto15.7percentintheadjacent higher-CLL zip codes. Moreover, 90.3 percent of the houses in the right column have a positive treatmentintensitycomparedtoonly66.0percentforthezipcodesintheleftcolumn. Notice that the treatment intensity is higher for houses with higher assessed values. This is becauselessexpensivehousescouldbefinancedalmostentirelywithaGSEloanevenpriortothe CLLincreases. Foreachhousevaluebin,howevertheaveragetreatmentintensityisalsohigherin zipcodeswithlargeCLLchanges. Inourdifference-in-differencesanalysisweincludeinteraction 17

terms between house value bins and quarters, which absorb the variation in treatment intensity acrossdifferentvaluebins. Therefore,ouranalysisusesmainlyvariationtreatmentintensitywithin valuebinsacrosszipcodes. Theshareofowner-occupiedhousesisverysimilarinbothgroupswith83percentinzipcodes wheretheCLLchangesweresmallcomparedto83.8percentinzipcodeswheretheCLLschanged more. Thisnumberishigherthanthenationalhomeownershiprateof64.4percentfortworeasons. Themainreasonisthatweincludesingle-familyhomesandcondosinouranalysisbutweexclude apartments.23 Lastly, the probability that a house was sold in a quarter was 0.8 percent in the zip codes with smallchangesand0.7percentinthezipcodeswithlargerchanges. 3.4 Graphs: CLLs, Guarantees, and Homeownership Figure 5 presents graphs showing how the ESA in March 2008 affected CLLs, government guarantees,andthehomeownershiprate. ThegraphsshowtwoseparatelinesforzipcodesthatexperiencedthelargerCLLincreases(“HigherCLL”)withinaborderandthezipcodeswithsmallerincreases(“LowerCLL”).ThezipcodesaregroupedinthesamewayasinTable1. Figure5ashows the CLLs, which increased from $417,000 to $487,371 in zip codes with small CLL changes, and from$417,000to$694,264inadjacentzipcodeswithlargechanges. Next consider Figure 5b, which shows the average government guarantee for houses that were sold in the same two zip code groups. The government guarantee of a house is equal to the mortgage principal if the house was bought with government-backed loan, and zero otherwise. Prior to the CLL increases average government guarantees are almost identical in both zip code groups and change in a parallel fashion. In both groups government guarantees increased substantially in 2007 as the private securitization market collapsed. After the CLL increase, however, the average guarantees diverge, and guarantees in zip codes with larger CLL increases are about $20,000 higher. Figure 9 in the Appendix shows government guarantees for houses below $500,000 for comparison. There is no divergence in government guarantees for these houses because they are notdirectlyaffectedbytheCLLchanges. Thisisreassuringasitsuggeststhattheadjustedassessed valuefortaxpurposesallowsustoselecthousesthatareaffectedbythepolicychanges. 23Ifapartmentsareincluded,thehomeownershiprateinoursubsampleisclosetothenationalrateinthedata. 18

700 in $1000 600 500 400 2007q2 2008q2 2009q2 2010q2 Higher CLL Lower CLL (a)CLLs 220 in $1000 200 180 160 140 2007q2 2008q2 2009q2 2010q2 Higher CLL Lower CLL (b)GovernmentGuarantees -.001 -.0015 -.002 -.0025 -.003 2007q2 2008q2 2009q2 2010q2 Higher CLL Lower CLL (c)HomeownershipTransitions Figure 5: Adjacent Zip Codes with Small (“Lower CLL”) and Large (“Higher CLL”) CLL Changes. These figures show zip codes where the CLLs were increased substantially (blue line) and for adjacent zip codes where the CLLs were either increased less or remained unchanged (red line). Source: Authors’ calculations based on data from CoreLogic, Inc., CoreLogic Real Estate Data. 19

Lastly, consider the homeownership transitions shown in Figure 5c. The variable plotted on this graph captures changes in homeownership and has a value of 1 for house sales from investors toowner-occupants,asthesetransactionscreateone additionalhomeowner;avalueof-1forsales from owner-occupants to investors; and zero otherwise.24 The graph shows that average homeownership transitions in both groups change in a parallel fashion – both before and after the CLL change. Taken together, the graphs in Figure 5 suggest that the CLL increases had an effect on governmentguarantees,butnovisibleeffectonthehomeownershiprate. Thusincreasedgovernment guaranteeshadnoapparenteffectonhomeownership. Inourdifference-in-differencesanalysiswe investigate this further by controlling for various factors that could drive the patterns in the raw averages. 4 Main Analysis 4.1 Specifications Weestimatedifference-in-differencesregressionsofthefollowingform: y =β Post ×T (2) i,q 0 q i +β T +X β +ξ +µ +ε . 1 i i,q x zip border×q i,q y =∑β 1[q=q(cid:48)]×T (3) i,q 0,q i q(cid:48) +β T +X β +ξ +µ +ε . 1 i i,q x zip border×q i,q The unit of analysis is at the level of a house (i) and quarter (q) pair. The outcome variable y is either the amount of government guarantees, or a variable capturing changes of the owneri,q occupancy status of the house. T is our treatment intensity measure that varies at the house level. i The vector X contains interaction terms between bins for the adjusted assessed value, interacted i,q with quarter fixed effects. The bin size for the house value is $100,000. These interaction terms are meant to control for different trends across different segments of the housing market. Thus, 24IntheAppendixwealsoshowthegraphforthelevelofhomeownershipinFigure10. 20

our estimates of the treatment effect use mainly variation of T within a segment of the housing i market across different zip codes. We also include zip code fixed effects ξ . Lastly, in our zip main specification we also include fixed effects for each combination of a border region and a quarter µ .25 These fixed effects capture any unobserved differential trends for different border×q border regions. Controlling for such differential trends is important as the housing crisis and the subsequent recovery affected different regions differently. By including these fixed effects we exploit variation in CLLs within a fairly small geographic areas with similar housing markets on bothsidesoftheborder. The main coefficients of interest are β and β , respectively. The difference between the two 0 0,q specifications is that in Equation (2) we estimate a single coefficient of interest that combines all quarterspriortotheCLLincreaseandallquartersaftertheCLLincrease,whereasinEquation(3) weestimateaseparatecoefficientforeachquarter. The time window for the analysis of the CLL increases runs from 2007:Q2 to 2011:Q1. The time window for the analysis of the CLL reductions runs from 2010:Q4 to 2014:Q3. We extend the sample window three years after the CLL changes, because home purchase decisions take considerabletime. There were large changes in the housing market during our sample period, especially during the financial crisis. For example, the homeownership rate declined nationwide, and the private securitizationmarketcollapsed. However,nationwidechangesandevenchangesspecifictoaborder area are differenced out in our specification. With the difference-in-differences specification, we estimate differential trends for houses with higher T within a border area. Nationwide or even i borderareaspecificchangesshouldthereforenotaffectourestimates. 4.2 Effect on Government Guarantees First, we investigate whether, and by how much, the higher loan limits increased government guarantees. Weconsideronlypurchaseloans,becauserefinancingloansarenotdirectlyassociated with changes of the homeowner and can therefore not have a direct impact on the homeownership rate.26 Thegoalofthisanalysisisnottostudywhethertherewereanyeffectsongovernmentguaranteesornot,buttoquantifythesizeoftheeffect. Quantifyingtheeffectsizeiscrucialtoevaluatethe trade-off between government guarantees and home ownership. This trade-off arises for example inpolicyproposalsthataimtolowergovernmentguaranteesbyloweringtheCLLs,butalsointhe 25We also show some estimates without µ in Tables 2 and 4. In this case we include quarter fixed effects border×q instead. 26Refinancingloanscouldhoweverhaveaneffectonthehomeownershiprateifrefinancinghelpstroubledhomeownerstokeeptheirhouse. 21

currentpolicythattiestheCLLstohousepricesandthereforeleadstoregularCLLincreases. Ouroutcomevariableistheamountofgovernmentguaranteesforeachloan. Formally, GovAmt =1{Loani isguaranteedbythegovernment}×LoanSize i i where LoanSize refers to the size of loan i. This outcome variable measures changes in the exi tensive and intensive margins, so it captures the possibilities that the number of guaranteed loans increasesbutalsothatborrowersincreasetheloansizeduetothehigherloanlimits. The sample for the analysis consists of all houses that were sold in a quarter, so our analysis capturestheeffectongovernmentguaranteesconditionalonasale. CLLchangescouldalsoaffect guarantees through their effect on the number of sales. In Section 4.5 we investigate the effect of CLLchangesonsalesandfindnosubstantialeffect. Thereforetheeffectongovernmentguarantees conditionalonasaleisclosetotheeffectunconditionaleffect. In the CoreLogic data we do not observe directly whether a loan is guaranteed by the government. Instead, we assume that loans, which are eligible for government guarantees, are indeed guaranteed by the government. We assume that any fixed-rate mortgage under the CLL is eligible for a government guarantee. This assumption is a reasonable approximation. In the Black Knight McDash data, which contains information about whether a loan carries a government guarantee, 91.4 percent of loans that are eligible for government guarantees are indeed guaranteed by the government. We have also obtained estimates of the effect on government guarantees using Black KnightMcDashdirectlyandobtainedsimilarresults. HerewereporttheestimatesusingtheCore- Logic data because it makes our estimates more comparable to our homeownership estimates that usethesamedata. Inparticularthisallowsustocalculateourmeasureoftreatmentintensityusing thesamevariablefortheassessedhousevalue. 22

Table 2: Effect on Government Guarantees. This table shows estimates from a difference-indifferences as in Equation (2). The dependent variable is the loan size guaranteed by the government (GovAmt). The table shows β ×T rather than β , where the average treatment intensity i 0 i 0 T is the average in the counties with high CLLs. The magnitude of the estimates is expressed in i $1,000. Standarderrorsareclusteredatthezipcodelevel. Source: Authors’calculationsbasedon datafromCoreLogic,Inc.,CoreLogicRealEstateData. CLLIncrease PartialCLLReduction (1) (2) (3) (4) T -32.9∗∗∗ -32.9∗∗∗ 3.3 3.3 i (7.5) (7.5) (2.3) (2.3) Post=1×T 57.1∗∗∗ 57.1∗∗∗ -10.2∗∗∗ -10.2∗∗∗ i (8.1) (8.2) (2.8) (2.8) QtrFE Y N Y N ZipcodeFE Y Y Y Y HouseValueBinxQtr Y Y Y Y BorderxQtrFE N Y N Y N.Obs. 64,744 64,744 41,752 41,752 Adj. R2 0.152 0.151 0.169 0.167 Table2showsestimatesfromEquation(2). Tomaketheestimateseasiertointerpretwereport β ×T, where T is the average treatment intensity in zip codes with larger CLL changes. Our 0 i i estimatessuggestthatfortheaveragehouseinthezipcodeswithlargerCLLchangesgovernment guarantees increased by $57,000 as a result of the CLL increases. This sizable effect corresponds tomorethan35percentoftheaveragegovernmentguaranteespriortotheCLLincrease. Similarly,thepartialCLLreductionsin2011reducedgovernmentguaranteesby$10,200. The impact of the CLL reductions was smaller because the CLLs were not reduced all the way to their levels before the ESA, and therefore T was smaller in magnitude. The estimates are identical i whetherweincludeborder-quarterfixedeffects(columns(2)and(4))ornot(columns(1)and(3)). Note that these findings cannot be driven by a chaning composition of houses that are sold as we controlforhousevaluebinsinteractedwithquarterdummies. Figure 6 shows estimates from the difference-in-differences specification given by Equation (3) for the CLL increase in 2008 in panel (a), and the CLL reduction in 2011 in panel (b). The introduction of the ESA in 2008 and the reduction of the GSE limits in 2011 are demarcated with vertical lines. We plot the product of the coefficient point estimates and the average treatment intensity β ×T, where the average treatment intensity T is the average in the counties with 0,ym i i highCLLs. 23

Figure 6: Effect on Government Guarantees. This figure plots estimated difference-indifferences coefficients from the regression given by Equation (3). The dependent variable is the loansizeguaranteedbythegovernment(GovAmt). Themarkershowsβ ×T,whereT istheavi 0,q i i eragetreatmentintensityinthezipcodeswithlargeCLLchanges. Theshadedareashowsthe90% confidence interval of each estimate. The magnitude of the estimates is expressed in $1,000. The regression contains year-quarter fixed effects, zip code fixed effects, border-quarter fixed effects, and the additional control variables described in the main text. Standard errors are clustered at the zipcodelevel. Source: Authors’calculationsbasedondatafromCoreLogic,Inc.,CoreLogicReal EstateData. 100 in $1000 20 in $1000 10 50 0 0 -10 -50 -20 2007q2 2008q2 2009q2 2010q2 2010q4 2011q4 2012q4 2013q4 (a)CLLIncrease (b)PartialCLLReduction In panel (a) we see that the increase of the loan limits in 2008 led to an increase of average guarantees by approximately $50,000 within one or two quarters. In panel (b) we see that the reductionoftheGSElimitsin2011loweredgovernmentguaranteeswithinoneortwoquartersby approximately$10,000forahousewiththeaveragetreatmentintensityinthehighCLLgroup. Robustness: DifferentTreatmentIntensityMeasure InFigure11weusethemeasureoftreatment intensity based on the FHA CLL. Recall that the FHA CLLs were increased simultaneously with the GSE CLLs in 2008, but the FHA CLLs were reduced later in 2014 rather than 2011. As theFHACLLswerenotchangedin2011itisperhapsquestionabletousetheFHAbasedtreatment intensity measure for the reduction of GSE CLLs in 2011. However, for completeness we show both, the estimates for the CLL increase in 2008, and the estimates for the partial CLL reduction in2011inFigure11. We obtain very similar estimates of around $50,000 for the CLL increase in 2008. For the CLL reduction in 2011 the estimates look similar, with effects of $5,000 to $10,000 for most of thesamplewindow,buttheeffectdisappearstowardtheendofthesamplewindowin2014. Tothe extent that the estimates differ from the baseline estimates we believe that the baseline estimates aremorerelevantaswepreviouslyargued. 24

InFigure13wecalculatethemeasureoftreatmentintensitybyusinga90percentloan-to-value ratioratherthan80percentasinthebaselineestimates,becauseborrowerswhotakeoutmortgage insurance are eligible for GSE guarantees with a 10 percent downpayment. The estimates for the CLL increase in 2008 are similar to the baseline, but for the CLL reduction in 2011 we estimate that government guarantees decreased by approximately $20,000 rather than by $10,000 as in the baselineestimates. AlthoughtheestimatedeffectsizeisdifferentfortheCLLreduction,thefinding ofalargeeffectongovernmentguaranteesisqualitativelyrobust. 4.3 Effect on Homeownership Transitions Next,weconsidertheeffectonhomeownership. AsshowninTable3,weconstructavariablethat takesavalueof1ifahousetransitionsfromnon-owner-occupiedtoowner-occupied,avalueof-1 ifittransitionsfromowner-occupiedtonon-owner-occupied,andzerootherwise. Table3: TransitionMatrixofOwnerOccupancyStatus Buyer OwnerOccupied NotOwnerOccupied OwnerOccupied 0 -1 Seller NotOwnerOccupied +1 0 Thus,inthiscase  1 ifhousei’sstatustransitionsfrominvestor-ownedtoowner-occupied    y i,q = 0 ifhousei’sstatusdoesnotchangeasaresultofasaleorisnotsold (4)    −1 ifhousei’sstatustransitionsfromowner-occupiedtoinvestor-owned Notethaty =0ineachofthethreefollowingcases: (i)atransitionfromanowner-occupantseller i,q to an owner-occupant buyer, (ii) a transition from a non-owner-occupant seller to a non-owneroccupant buyer, and (iii) the house is not sold. Thus, we estimate the effect on homeownership usingallthehousesinoursample,regardlessofwhethertheyweresoldornot. 25

Table4: EffectonHomeownershipTransitions. Thistableshowsestimatesfromthedifferencein-differences specification in Equation (2). The dependent variable takes a value of 1 if a house transitionsfromnon-owner-occupiedtoowner-occupied,avalueof-1ifittransitionsfromowneroccupied to non-owner-occupied, and zero otherwise. The table shows β ×T, where T is the 0 i i average treatment intensity in the zip codes with large CLL changes. The magnitude of the estimates is expressed in percentage points. The regression contains quarter fixed effects, zip code fixed effects, and the additional control variables described in the main text. Standard errors are clustered at the zip code level. Source: Authors’ calculations based on data from CoreLogic, Inc., CoreLogicRealEstateData. CLLIncrease PartialCLLReduction (1) (2) (3) (4) T 0.026 0.024 0.019∗∗∗ -0.004 i (0.016) (0.015) (0.007) (0.006) Post=1×T -0.021 -0.018 -0.022∗∗∗ 0.008∗ i (0.014) (0.013) (0.007) (0.005) QtrFE Y N Y N ZipcodeFE Y Y Y Y HouseValueBinxQtr Y Y Y Y BorderxQtrFE N Y N Y N.Obs. 9,930,192 9,930,192 4,592,992 4,592,931 Adj. R2 0.001 0.001 0.001 0.001 Table 4 shows estimates from the specification in Equation (2) measured in percentage points. For the CLL increase, we find no statistically significant effect and the point estimates imply that the CLL increase led to a small reduction in home ownership. For the partial CLL reduction, we find a small negative statistically significant effect on home ownership in column (3). However, the sign flips once we include border-quarter fixed effects in column (4), and we obtain a small positive effect. Overall the estimates in Table 4 show little evidence for a substantial effect of the CLLchangesonhomeownership. Figure 7 shows estimates using Equation (3). This specification includes border-quarter fixed effects. There appears to be no positive effect of the CLL increases in 2008 and no negative effect oftheCLLreductionsin2011. 26

Figure 7: Effect on Homeownership Transitions. This figure plots estimated difference-indifferences coefficients with the regression given by Equation (3). The dependent variable takes a value of 1 if a house transitions from non-owner-occupied to owner-occupied, a value of -1 if it transitions from owner-occupied to non-owner-occupied, and zero otherwise. The vertical axis is measured in percentage points. The marker shows β ×T, where T is the average treatment 0,q i i intensity in the zip codes with large CLL changes. The shaded area shows the 90% confidence interval of each estimate. The regression contains quarter fixed effects, zip code fixed effects, and the additional control variables described in the main text. Standard errors are clustered at the zip code level. Source: Authors’ calculations based on data from CoreLogic, Inc., CoreLogic Real EstateData. in pp .0006 in pp .0005 .0004 0 .0002 -.0005 0 -.001 -.0002 -.0015 -.0004 2007q2 2008q2 2009q2 2010q2 2010q4 2011q4 2012q4 2013q4 (a)CLLIncrease (b)PartialCLLReduction Robustness: DifferentTreatmentIntensityMeasure InFigure12weusethemeasureoftreatment intensity based on the FHA CLL. The estimates are similar to the main specification as the point estimates typically have the “wrong” sign and are not statistically significant. In Figure 14 we show the estimated effect on homeownership with a 90 percent loan-to-value ratio. As in our baseline estimates we find no statistically significant effects and the point estimates mostly have the“wrong”sign. Overallourestimatesappeartobequalitativelyrobusttochangesinhowthetreatmentintensity measure is calculated. We consistently find substantial effects on government guarantees, but not onhomeownership. Dependingonthetreatmentintensitymeasurehowever,themagnitudesofthe estimated effects on government guarantees varies somewhat, especially for the CLL reduction in 2011. 4.4 Heterogeneous Effects So far we have shown that the effects of CLL changes are substantial for government guarantees but no robust effects for homeownership. One possibility why we do not find any effect on home- 27

ownershipisbecauseCLLchangesonlyhaveaneffectinregionswherepotentialhomebuyersare more credit-constrained. To investigate this hypothesis we use the average loan-to-income ratio at the zip code level as a measure of credit constraints. We measure the loan-to-income ratio prior to the CLL changes — in 2007 for the CLL increase, and in 2010 for the partial CLL reduction.27 Wethenestimatethefollowing“triple-diff”regression: y =β Post ×T +β Post ×T ×LTI i,q 0 q i 1 q i zip +β T +β Post ×LTI +β T ×LTI +X β +ξ +µ +ε . (5) 2 i 3 q zip 4 i zip i,q x zip border×q i,q Here our main coefficient of interest is β , which captures how much the effect changes with the 1 zipcode-levelloan-to-incomeratioLTI . zip First, we investigate whether there are larger effects on government guarantees in zip codes where loans are larger relative to incomes in Table 5. Indeed we find that the effect is larger in zip codes with higher loan-to-income ratios for both the CLL increase and the CLL reduction. The estimates imply that there are positive effects for all except about 10 percent of zip codes with the lowest loan-to-income ratios, but the effects are substantially larger for zip codes with high loan-to-incomeratios. In Table 6 we look for heterogeneous effects for homeownership transitions by interacting the treatment variable with the loan-to-income ratio. The interaction terms are not statistically significant at conventional levels. For the CLL increase the estimated sign suggests that zip codes with largerloan-to-incomeratioshavesmallerpositiveorevennegativeeffectsonhomeownership. This is inconsistent with the estimates in Table 5 and with basic economic theory predicting that there shouldbelargerpositiveeffectinregionswithtightercreditconstraints. Thus,theseestimatessuggestthattherewasnosubstantialeffectonhomeownership—noteveninzipcodeswithrelatively highloan-to-incomeratios. 27We calculate the average zipcode-level loan-to-income ratio based on the Home Mortgage Disclosure Act data. We only use purchase loans for one-to-four family housing to construct this variable. In the sample for the CLL increase, themeanloan-to-incomeratiois2.75withastandarddeviationof0.31. The10thand90thpercentilesare 2.30and3.03,respectively. InthesampleforthepartialCLLreduction,themeanloan-to-incomeratiois2.92witha standarddeviationof0.50. The10thand90thpercentilesare2.13and3.43,respectively. 28

Table 5: Heterogeneous Effects on Government Guarantees. This table shows estimates from a difference-in-differences as in Equation (5). The dependent variable is the loan size guaranteed by the government (GovAmt). The table shows the product of the estimated coefficients and the i average treatment intensity T in the zip codes with large CLL changes. Here we also include i interaction terms with the loan-to-income ratio at the zip code level. The regression contains quarter fixed effects, zip code fixed effects, and the additional control variables described in the main text. Standard errors are clustered at the zip code level. Source: Authors’ calculations based ondatafromCoreLogic,Inc.,CoreLogicRealEstateData. CLLIncrease PartialCLLReduction (1) (2) (3) (4) T 152.6∗∗∗ 152.6∗∗∗ -17.5∗∗ -17.5∗∗ i (38.7) (38.7) (7.9) (7.9) Post=1×T -192.1∗∗∗ -192.1∗∗∗ 26.1∗∗∗ 26.1∗∗∗ i (38.2) (38.2) (9.2) (9.2) T ×Loan-to-IncomeRatio -64.1∗∗∗ -64.1∗∗∗ 7.5∗∗∗ 7.5∗∗∗ i (13.2) (13.2) (2.8) (2.9) Post=1×Loan-to-IncomeRatio 0.0 0.0 -0.2∗∗∗ -0.2∗∗∗ (0.1) (0.1) (0.1) (0.1) Post=1×T ×Loan-to-IncomeRatio 86.0∗∗∗ 86.0∗∗∗ -13.0∗∗∗ -13.0∗∗∗ i (13.0) (13.0) (3.3) (3.3) QtrFE Y N Y N ZipcodeFE Y Y Y Y HouseValueBinxQtr Y Y Y Y BorderxQtrFE N Y N Y N.Obs. 64,744 64,744 37,156 37,156 Adj. R2 0.156 0.154 0.168 0.166 4.5 Effect on Sales In Table 7 we estimate the effect of the CLL changes on the probability that a house is sold measuredinpercentagepoints. EvenifhigherCLLshavenosubstantialeffectonhomeownershipthey might be beneficial if they increase the turnover of houses, which results in a better allocation of housestohouseholdsandmayincreasegeographicmobility. Thesignsofourcoefficientestimates are consistent with this hypothesis, but with the exception of column (4) the estimates are not statistically significant. The economic magnitude of the coefficients is also moderate. For example the estimates in columns (2) and (4) suggest that the effect on the average house in a county with largeCLLchangesisapproximately1-2percentoftheaveragesaleprobability(seeTable1). 29

Table 6: Heterogeneous Effects on Homeownership Transitions. This table shows estimates from a difference-in-differences as in Equation (5). The dependent variable is given by Equation (4). The table shows the product of the estimated coefficients and the average treatment intensity T in the zip codes with large CLL changes. Here we also include interaction terms with the loani to-income ratio at the zip code level. The magnitude of the estimates is expressed in percentage points. Theregressioncontainsyear-quarterfixedeffects,zipcodefixedeffects,andtheadditional control variables described in the main text. Standard errors are clustered at the zip code level. Source: Authors’calculationsbasedondatafromCoreLogic,Inc.,CoreLogicRealEstateData. CLLIncrease PartialCLLReduction (1) (2) (3) (4) T 0.156 0.190 0.004 -0.054 i (0.137) (0.126) (0.040) (0.037) Post=1×T 0.089 0.043 -0.037 0.040 i (0.091) (0.078) (0.039) (0.030) T ×Loan-to-IncomeRatio -0.042 -0.059 0.005 0.017 i (0.049) (0.045) (0.014) (0.013) Post=1×Loan-to-IncomeRatio 0.001∗∗∗ 0.000 -0.002∗∗∗ -0.000 (0.000) (0.000) (0.001) (0.000) Post=1×T ×Loan-to-IncomeRatio -0.044 -0.022 0.006 -0.011 i (0.032) (0.028) (0.013) (0.010) QtrFE Y N Y N ZipcodeFE Y Y Y Y HouseValueBinxQtr Y Y Y Y BorderxQtrFE N Y N Y N.Obs. 9,930,192 9,930,192 4,143,616 4,143,555 Adj. R2 0.001 0.001 0.001 0.001 4.6 Summary of Findings In summary, we find that the CLL changes had a substantial effect on government guarantees. Jumbo-conforming loans that became newly eligible for government guarantees as a result of the CLLincreaseaccountednationwideforupto20percentoftheGSEportfolioindollarterms. Our estimatesusingthesampleofadjacentborderzipcodessuggestthatfortheaveragehouselocated on the side of the border where the CLLs increased more government guarantees increased by about$50,000,or35percentofitsmeanwhentheCLLswereincreasedin2008anddecreasedby about$10,000whentheCLLswereloweredin2011. Despitethissizableeffectongovernmentguaranteeswefindnosignificanteffectonhomeownership. Our estimates are typically not statistically significant, and the point estimates sometimes suggestthatincreasedguaranteesareassociatedwithlowerhomeownership. 30

Table 7: Effect on Sales. This table shows estimates from a difference-in-differences as in Equation (2). The dependent variable is an indicator variable that is equal to one if the house was sold duringthequarter. Thetableshowsβ ×T,wheretheaveragetreatmentintensityT istheaverage 0 i i inthecountieswithhighCLLs. Themagnitudeoftheestimatesisexpressedinpercentagepoints. Theregressioncontainsyear-quarterfixedeffects,zipcodefixedeffects,andtheadditionalcontrol variables described in the main text. Standard errors are clustered at the zip code level. Source: Authors’calculationsbasedondatafromCoreLogic,Inc.,CoreLogicRealEstateData. CLLIncrease PartialCLLReduction (1) (2) (3) (4) T 0.015 0.048 0.017 0.027∗ i (0.040) (0.039) (0.016) (0.016) Post=1×T 0.053 0.008 -0.010 -0.022∗ i (0.032) (0.029) (0.011) (0.012) QtrFE Y N Y N ZipcodeFE Y Y Y Y HouseValueBinxQtr Y Y Y Y BorderxQtrFE N Y N Y N.Obs. 9,930,192 9,930,192 4,592,992 4,592,937 Adj. R2 0.004 0.004 0.006 0.006 We also find that the effect on government guarantees was larger in zip codes with high loanto-income ratios. This finding suggests that government guarantees are more important in regions where house prices are high relative to incomes, because borrowers in these regions might not qualifyforaloanthatisnotguaranteedbythegovernment. Wealsoinvestigatewhethertheeffect onhomeownershipvarieddependingontheloan-to-incomeratio,butagainwefindnoeffect. This finding suggests that the CLL changes affected the financing choices, but not homeownership. Thus, for houses that were affected by the CLL changes, further increases in government guarantees had no effect on homeownership. The increase in government guarantees helped borrowerswhoswitchedtogovernment-backedloansandmayhavehelpedsomeborrowerstoincrease theirloansize,butithadonlyanegligibleeffectonmarginalpotentialhomeowners.28 Is This Finding Unsurprising? The intention of the CLL increase in 2008 was not to expand homeownership but to support the housing market. Moreover, in 2008 the housing market was in 28OurfindingsdonotcontradictAdelino,Schoar,andSeverino(2012)andKung(2014)whofindthatCLLchanges have substantial effects on house prices. Indeed, increasing house prices can offset the effect of CLL increases on creditavailability,becausealargerloanisrequiredtopurchasethesamehouse. Thiscouldinturnexplainwhywesee alimitedeffectonhomeownership. 31

turmoilandpotentialbuyersmaythereforehavebeenreluctanttobuy. Nevertheless,wearguethat our finding is by no means obvious a priori. Indeed, in the regions where the GSE CLLs were increased in 2008, 39 percent of the purchase loans that became newly eligible for government guarantees in 2008 (jumbo-conforming loans) were taken out by first-time home buyers. This share is close to the share for conventional conforming loans (42 percent). Nationwide only 32 percent of all conventional conforming loans are taken out by first-time home buyers.29 These numbers suggests that the effect on homeownership could be sizable. Moreover, general housing market conditions in 2008 are unlikely to explain our findings because they are differenced out, and we do find a substantial effect of the CLL increases on government guarantees. In addition, weobtainevensmallerestimatesfortheeffectonhomeownershipusingtheCLLchangesin2011 whenthehousingmarketwascalmer. In addition we argue that even if our findings are qualitatively unsurprising for some, policy decisions should be based on quantitative estimates. For example, our estimates can guide policy makers as they help to project the expected reduction in government guarantees if CLLs are reduced. 5 Policy Implications and Limitations 5.1 Policy Implications In this section, we discuss the policy implications of our findings. The estimated effects of CLL changes inform two current policy issues. First, the GSE CLLs were increased nationwide in the pastthreeyears. Second,severalhousingreformproposalssuggestthattheGSEscouldbephased outbygraduallyloweringtheCLLs. Recent CLL Increases We first discuss the recent CLL increase. Between 2017 and 2019, the CLLs for the GSEs and the FHA increased from $417,000 to $484,350 outside of the high cost areas, or more than five percent per year. Our findings suggest that such increases are likely to increasegovernmentguarantees,butwilllikelyatmosthaveamodesteffectonthehomeownership rate. ThereasonfortheCLLincreaseisthattheCLLsaretiedtohousepricesandhousepriceshave increased in recent years. Tying CLLs to house prices is problematic if CLL increases themselves contributetohousepriceincreases(Adelino,Schoar,andSeverino(2012)andKung(2014)),which 29ThesefractionswerecalculatedfromFreddieMacandFannieMaeloanleveldatafromMarch2008toFebruary 2011(link). Afirsttimehomebuyerisanybuyerwhodidnotownahouseinthepreviousthreeyears.Thereforethese buyers may not be “true” first time buyers, but for our analysis it is only important that they were not homeowners beforetheypurchasedthehouse. 32

thencouldleadtofurtherCLLincreases,andsoforth. Thiscouldresultinapositivefeedbackloop that leads to continually increasing house prices and increased government guarantees, while the homeownership rate would be largely unaffected. Moreover, the increase in house prices that is drivenbyCLLincreasescoulddestabilizethehousingmarketinthelongrun. Our assessment stands in stark contrast to commentary by the California Association of Realtors(C.A.R.),whichcommentedtheCLLincreaseasfollows30: "C.A.R.applaudstheFHFAforrecognizingCalifornia’scontinuinghomepriceincreasesover thelastfewyearsandraisingmaximumconformingloanlimits,"saidC.A.R.PresidentSteveWhite. "IncreasingtheexistingFannieMaeandFreddieMacconformingloanlimitswillprovidestability andcertaintytothehousingmarketandgivetensofthousandsofCaliforniahomebuyersachance athomeownership."31 HousingFinanceReform Next,weturntotheongoingdebateonhousingfinancereform. This debate revolves around two broad issues: (1) how the government should be involved in the mortgagemarket,and(2)thescopeofthegovernment’sinvolvement. Thehowissueisconcernedwith questions like which kinds of mortgage contracts the government should favor (e.g. 30 year fixed rate mortgage), or whether the mortgage insurance model by the FHA is preferable to the GSE model. The scope issue is related to this paper, because one way to determine the scope of the government’sinvolvementisbyadjustingtheCLLs. Indeed, some of the reform plans propose to phase out the GSEs by gradually lowering the CLLs, for example, the reform proposals by the American Enterprise Institute (Wallison, Pinto, Pollock, Lawler, Michel, Oliner, and Peter, 2018) or the “Housing Finance Reform and Taxpayer Protection Act” by Senator Corker.32 Moreover, the Congregational Budget Office discussed a reductionof CLLsin“Transitioningto AlternativeStructuresforHousing Finance”(CBO,2014). Manyreformproposalsrequirealegislativeactandthereforehavearelativelysmallchanceofbeingimplemented. TheCLLshowevercanbechangedwithoutCongresssolelyasanadministrative act. Inparticular,recentlyithasbeendiscussedthatlowerCLLscouldbeimplementedbythenew director of the FHFA, who will be appointed by the beginning of 2019, as the high CLLs are no longer needed to support the housing market.33 Our findings suggest that decreasing the CLLs to thepre-ESAlevelswouldlikelyreducegovernmentguaranteessubstantiallybutwouldlikelyhave onlyamodesteffectonthehomeownershiprate. 30It should be noted that realtors stand to gain from increased house prices, because their they typically earn a percentageofthehousepricefromhousetransactions. 31See https://www.prnewswire.com/news-releases/california-realtors-commend-fhfa-for-raising-fannie-mae-andfreddie-mac-conforming-loan-limits-300563003.html 32Forthecompletetextofthebill,seehttps://www.congress.gov/bill/113th-congress/senate-bill/1217. 33See https://www.housingwire.com/articles/47283-the-most-powerful-person-in-mortgage-lending-is-about-tobe-replaced. 33

We conclude this section with three remarks regarding the implications of our findings for housing finance reform. First, even though we argue that the CLLs could be lowered from current levelswithoutsubstantiallyaffectingthehomeownershiprate,theremaybeotherpolicygoalsthat justify the high CLLs. Second, besides lowering the CLLs there may be other policies to better align the governmentwith the goal of increasinghomeownership. For example, ifthe government guarantees were restricted to purchase loans, government guarantees could be lowered, arguably without substantial impacts on the homeownership rate. Third, there may be more direct ways to achievethegoalofincreasedhomeownershipthantointerveneinthemortgagemarket. 5.2 Limitations Thereareseveralimportantlimitationsofouranalysis. First,ouranalysisdoesnotanswerthequestionofwhatwouldhappenifgovernmentmortgage guarantees were eliminated entirely. We only observe a change in CLLs at relatively high levels, whichallowsustoestimatethemarginaleffectofchangesingovernmentguarantees. Ourfindings suggest that this change does not affect marginal homeowners for the most part. It is entirely possible that reducing the CLL to zero would affect more low- and moderate-income households, whicharemorelikelytohavedifficultiesinobtainingcreditintheprivatemarket,andaretherefore marginal homeowners.34 Reducing the CLLs to zero could therefore have a substantial effect on thehomeownershiprate. Ourpaperisthereforecomplementarytotheoreticalpapersthatsimulate counterfactuals in which government guarantees are entirely eliminated such as Jeske, Krueger, and Mitman (2013), Elenev, Landvoigt, and Van Nieuwerburgh (2016) and Gete and Zecchetto (2017). Second, our analysis does not take into account some of the effects that may be present if the CLLs would be lowered in the whole country rather than only in some counties. For example it may be the case that banks are able to absorb only a certain amount of mortgages on their balancesheetsandadditionalmortgageswouldhavetobeprivatelysecuritized. Theremayalsobe macroeconomic effects of a nationwide reduction in the CLLs that are not present for the regional reduction we observe. Our paper is therefore complementary to Fieldhouse, Mertens, and Ravn (2018),whoestimatethemacroeconomiceffectsofmortgageassetpurchasesbythegovernment. Lastly, in 2013 the GSEs started to shift part of their credit risk to private investors, which reduced government guarantees (Finkelstein, Strzodka, and Vickery (2018)). These credit risk transfer programs are structured such that the GSEs bear the “first loss” in a mortgage pool — a tranche of about 0.5 percent. The tranches from about 0.5 to about 4.0 percent are sold to private investors, and the “catastrophic risk” above 4.0 percent is borne by the GSEs again. Due to these 34ForinstanceBhuttaandRingo(2017)findthatareductioninFHAmortgageinsurancepremiumshadasizeable effectonhome-buyingamongpotentialbuyerswhorelyonFHAloans. 34

programs, changes in the loan volume that is sold to the GSEs are not equivalent to changes in governmentguarantees. Theseprogramshighlightthatreformproposalsthatsuggestloweringthe CLLsarenottheonlywaytolowergovernmentguarantees. 6 Conclusion The U.S. government guarantees a majority of mortgages, which is often justified as a means to promotehomeownership. Inthispaper,weestimatetheeffectbyusingadifference-in-differences design, with detailed property-level data, that exploits changes of the conforming loan limits (CLLs) along county borders. We find a sizable effect of CLLs on government guarantees but no robust effect on homeownership. Thus, government guarantees could be considerably reduced, withverymodesteffectsonthehomeownershiprate. Ourfindingisparticularlyrelevantforrecent housing finance reform plans that propose to gradually reduce the government’s involvement in the mortgage market by reducing the CLLs. Our findings also suggest, that to achieve the policy goal of raising homeownership the government should use more direct instruments than high conformingloanlimits. References ACOLIN, A., J. BRICKER, P. CALEM, AND S. WACHTER (2016): “Borrowing Constraints and Homeownership,”AmericanEconomicReview,106(5),625–29. ADELINO, M., A. SCHOAR, AND F. SEVERINO (2012): “Credit Supply and House Prices: Evidence from Mortgage Market Segmentation,” Discussion paper, National Bureau of Economic Research. (2018): “Perception of House Price Risk and Homeownership,” Discussion paper, NationalBureauofEconomicResearch. AMIOR, M., AND J. HALKET (2014): “Do Households Use Home-ownership to Insure Themselves? EvidenceacrossUSCities,”QuantitativeEconomics,5(3),631–674. AN, X., R. W. BOSTIC, Y. DENG, S. A. GABRIEL, R. K. GREEN, AND J. TRACY (2007): “GSE Loan Purchases, the FHA, and Housing Outcomes in Targeted, Low-Income Neighborhoods,” Brookings-WhartonPapersonUrbanAffairs,pp.205–256. 35

ANENBERG, E., A. HIZMO, E. KUNG, AND R. MOLLOY(2016): “TheEffectofMortgageCredit AvailabilityonHousePricesandConstruction: EvidencefromaFrontierEstimationApproach,” Discussionpaper,BoardofGovernorsoftheFederalReserveSystem. ANENBERG, E., AND E. KUNG (2017): “InterestRatesandHousingMarketDynamicsinaHousingSearchModel,”Discussionpaper,BoardofGovernorsoftheFederalReserveSystem. BHUTTA, N., AND D. R. RINGO(2017): “TheEffectofInterestRatesonHomeBuying: Evidence from a Discontinuity in Mortgage Insurance Premiums,” Discussion paper, Board of Governors oftheFederalReserveSystem. BLANCHFLOWER, D. G., AND A. J. OSWALD (2013): “Does High Home-ownership Impair the LaborMarket?,”Discussionpaper,NationalBureauofEconomicResearch. BOSTIC, R. W., AND S. A. GABRIEL (2006): “Do the GSEs Matter to Low-Income Housing Markets? AnAssessmentoftheEffectsoftheGSELoanPurchaseGoalsonCaliforniaHousing Outcomes,”JournalofUrbanEconomics,59(3),458–475. CAPLIN, A., A. CORORATON, AND J. TRACY (2015): “Is the FHA Creating Sustainable Homeownership?,”RealEstateEconomics,43(4),957–992. CBO (2014): “TransitioningtoAlternativeStructuresforHousingFinance,”Discussionpaper. DEFUSCO, A. A., AND A. PACIOREK (2017): “TheInterestRateElasticityofMortgageDemand: Evidence from Bunching at the Conforming Loan Limit,” American Economic Journal: EconomicPolicy,9(1),210–240. DIPASQUALE, D., AND E. L. GLAESER(1999): “IncentivesandSocialCapital: AreHomeowners BetterCitizens?,”JournalofUrbanEconomics,45(2),354–384. ELENEV, V., T. LANDVOIGT, AND S. VAN NIEUWERBURGH (2016): “Phasing Out the GSEs,” JournalofMonetaryEconomics,81(C),111–132. FETTER, D. K. (2013): “How Do Mortgage Subsidies Affect Home Ownership? Evidence from theMid-CenturyGIBills,”AmericanEconomicJournal: EconomicPolicy,5(2),111–147. FIELDHOUSE, A. J., AND K. MERTENS (2017): “A Narrative Analysis of Mortgage Asset PurchasesbyFederalAgencies,”Discussionpaper,NationalBureauofEconomicResearch. FIELDHOUSE, A. J., K. MERTENS, AND M. O. RAVN (2018): “The Macroeconomic Effects of Government Asset Purchases: Evidence from Postwar US Housing Credit Policy,” The QuarterlyJournalofEconomics,133(3),1503–1560. 36

FINKELSTEIN, D., A. STRZODKA, AND J. I. VICKERY(2018): “CreditRiskTransferanddefacto GSEReform,”Discussionpaper,FederalReserveBankofNewYork. FUSTER, A., AND J. VICKERY (2015): “Securitization and the Fixed-Rate Mortgage,” Review of FinancialStudies,28(1),176–211. FUSTER, A., AND B. ZAFAR (2016): “To Buy or Not to Buy: Consumer Constraints in the HousingMarket,”AmericanEconomicReview,106(5),636–40. GABRIEL, S. A., AND S. ROSENTHAL (2008): “The GSEs, CRA and Homeownership in TargetedUnderservedNeighborhoods,”inConferenceonBuiltEnvironment: Access,Finance,and Policy. GETE, P., AND F. ZECCHETTO (2017): “DistributionalImplicationsofGovernmentGuaranteesin MortgageMarkets,”TheReviewofFinancialStudies,p.hhx083. GLAESER, E. L., AND J. M. SHAPIRO (2003): “The Benefits of the Home Mortgage Interest Deduction,”TaxpolicyandtheEconomy,17,37–82. HILBER, C. A., AND T. M. TURNER (2014): “The Mortgage Interest Deduction and Its Impact onHomeownershipDecisions,”ReviewofEconomicsandStatistics,96(4),618–637. JESKE, K., D. KRUEGER, AND K. MITMAN (2013): “Housing,MortgageBailoutGuaranteesand theMacroEconomy,”JournalofMonetaryEconomics,60(8),917–935. KAUFMAN, A. (2014): “The Influence of Fannie and Freddie on Mortgage Loan Terms,” Real EstateEconomics,42(2),472–496. KUNG, E. (2014): “The Effect of Credit Availability on House Prices: Evidence from the Economic Stimulus Act of 2008,” Discussion paper, Working Paper, University of California-Los Angeles. LAAMANEN, J.-P. (2017): “Home-ownership and the Labour Market: Evidence from Rental HousingMarketDeregulation,”LabourEconomics. PASSMORE, S. W., S. M. SHERLUND, AND G. BURGESS (2005): “The Effect of Housing Government-SponsoredEnterprisesonMortgageRates,”Discussionpaper. POTERBA, J. M. (1984): “Tax Subsidies to Owner-Occupied Housing: An Asset-Market Approach,”TheQuarterlyJournalofEconomics,99(4),729–752. SHERLUND, S. M. (2008): “The Jumbo-Conforming Spread: A Semiparametric Approach,” Discussionpaper. 37

SOMMER, K., AND P. SULLIVAN(2018): “ImplicationsofUSTaxPolicyforHousePrices,Rents, andHomeownership,”TheAmericanEconomicReview. WALLISON, P. J., E. J. PINTO, A. J. POLLOCK, P. LAWLER, N. MICHEL, S. D. OLINER, AND T. PETER (2018): “The Taxpayer Protection Housing Finance Plan: Gradually Winding Down FannieMaeandFreddieMacandImprovingtheFHA,”Discussionpaper. 38

A Additional Figures E S A H E R A A R R A E S A Ex pir e d 729,750 625,500 362,790 271,050 200,160 8 8 9 4 0 0 0 1 0 0 0 0 2 2 2 2 3/ 2/2/ 1/ 0 1 0 0 Date )$( timiL naoL gnimrofnoC CLL = 417,000 CLL = 0.95 x Median House Price (MHP) CLL = 1.25 x MHP CLL = 1.15 x MHP Figure 8: Timeline of FHA CLL Changes. This timeline shows how the conforming loan limits for the FHA were changed by the Economic Stimulus Act (ESA) in 3/2008, the Housing and Economic Recovery Act (HERA) in 12/2008 and the American Recovery and Reinvestment Act in 2/2009. In 1/2014 the CLLs specified in ESA expired and the lower CLLs specified in HERA wereusedthereafter. 39

200 in $1000 180 160 140 120 100 2007q2 2008q2 2009q2 2010q2 Higher CLL Lower CLL Figure9: GSEGuaranteesBelow$500,000. Thisfigureshowstheaveragegovernmentguarantee forzipcodeswheretheCLLswereincreasedalotandforadjacentzipcodeswheretheCLLswere increased less. Unlike Figure 5b this graph shows houses with assessed values below $500,000. Source: Authors’calculationsbasedondatafromCoreLogic,Inc.,CoreLogicRealEstateData. .84 .83 .82 .81 .8 .79 2007q2 2008q2 2009q2 2010q2 Higher CLL Lower CLL Figure 10: Homeownership Rate. This graph shows the same two groups of adjacent zip codes with different CLL changes as Figure 5c. However, unlike Figure 5c it shows the level of homeownershipratherthanchangesofthehomeownershiprate. Source: Authors’calculationsbasedon datafromCoreLogic,Inc.,CoreLogicRealEstateData. 40

Figure 11: Effect on Government Guarantees - FHA CLL Treatment Intensity. This figure plots estimated difference-in-differences coefficients from the regression given by Equation (3). Unlike our baseline specification we use the FHA CLLs rather than the GSE CLLs to calculate the treatment intentisity measure T for these estimates. The dependent variable is the loan size i guaranteed by the government (GovAmt). The marker shows β ×T, where T is the average i 0,q i i treatment intensity in the zip codes with large CLL changes. The shaded area shows the 90% confidence interval of each estimate. The magnitude of the estimates is expressed in $1,000. The regression contains year-quarter fixed effects, zip code fixed effects, border-quarter fixed effects, and the additional control variables described in the main text. Standard errors are clustered at the zipcodelevel. Source: Authors’calculationsbasedondatafromCoreLogic,Inc.,CoreLogicReal EstateData. in $1000 10 in $1000 100 0 50 0 -10 -50 -20 2007q2 2008q2 2009q2 2010q2 2010q4 2011q4 2012q4 2013q4 (a)CLLIncrease (b)PartialCLLReduction 41

Figure 12: Effect on Homeownership - FHA CLL Treatment Intensity. This figure plots estimated difference-in-differences coefficients with the regression given by Equation (3). Unlike our baseline specification we use the FHA CLLs rather than the GSE CLLs to calculate the treatment intentisity measure T for these estimates. The dependent variable takes a value of 1 if a house i transitionsfromnon-owner-occupiedtoowner-occupied,avalueof-1ifittransitionsfromowneroccupied to non-owner-occupied, and zero otherwise. The vertical axis is measured in percentage points. The marker shows β ×T, where T is the average treatment intensity in the zip codes 0,q i i withlargeCLLchanges. Theshadedareashowsthe90%confidenceintervalofeachestimate. The regressioncontainsquarterfixedeffects,zipcodefixedeffects,andtheadditionalcontrolvariables described in the main text. Standard errors are clustered at the zip code level. Source: Authors’ calculationsbasedondatafromCoreLogic,Inc.,CoreLogicRealEstateData. in pp in pp .001 .0006 .0004 0 .0002 0 -.001 -.0002 -.002 -.0004 2007q2 2008q2 2009q2 2010q2 2010q4 2011q4 2012q4 2013q4 (a)CLLIncrease (b)PartialCLLReduction 42

Figure 13: Effect on Government Guarantees - 90 Percent LTV. This figure plots estimated difference-in-differencescoefficientsfromtheregressiongivenbyEquation(3). Herewecalculate the measure of treatment intensity by using a 90 percent loan-to-value ratio rather than 80 percent asinthebaselineestimates. Thedependentvariableistheloansizeguaranteedbythegovernment (GovAmt). Themarkershowsβ ×T,whereT istheaveragetreatmentintensityinthezipcodes i 0,q i i with large CLL changes. The shaded area shows the 90% confidence interval of each estimate. The magnitude of the estimates is expressed in $1,000. The regression contains year-quarter fixed effects, zip code fixed effects, border-quarter fixed effects, and the additional control variables described in the main text. Standard errors are clustered at the zip code level. Source: Authors’ calculationsbasedondatafromCoreLogic,Inc.,CoreLogicRealEstateData. 100 in $1000 20 in $1000 50 0 0 -20 -50 -40 2007q2 2008q2 2009q2 2010q2 2010q4 2011q4 2012q4 2013q4 (a)CLLIncrease (b)PartialCLLReduction 43

Figure14: EffectonHomeownership-90PercentLTV.Thisfigureplotsestimateddifference-indifferences coefficients with the regression given by Equation (3). Here we calculate the measure of treatment intensity by using a 90 percent loan-to-value ratio rather than 80 percent as in the baseline estimates. The dependent variable takes a value of 1 if a house transitions from nonowner-occupied to owner-occupied, a value of -1 if it transitions from owner-occupied to nonowner-occupied, and zero otherwise. The vertical axis is measured in percentage points. The markershowsβ ×T,whereT istheaveragetreatmentintensityinthezipcodeswithlargeCLL 0,q i i changes. The shaded area shows the 90% confidence interval of each estimate. The regression containsquarterfixedeffects,zipcodefixedeffects,andtheadditionalcontrolvariablesdescribed in the main text. Standard errors are clustered at the zip code level. Source: Authors’ calculations basedondatafromCoreLogic,Inc.,CoreLogicRealEstateData. .001 in pp .001 in pp .0005 .0005 0 -.0005 0 -.001 -.0015 -.0005 2007q2 2008q2 2009q2 2010q2 2010q4 2011q4 2012q4 2013q4 (a)CLLIncrease (b)PartialCLLReduction 44