The Influence of Fannie and Freddie on Mortgage Loan Terms

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. The Influence of Fannie and Freddie on Mortgage Loan Terms Alex Kaufman 2012-33 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 INFLUENCE OF FANNIE AND FREDDIE ON MORTGAGE LOAN TERMS ALEXKAUFMAN* ABSTRACT. ThispaperusesanovelinstrumentalvariablesapproachtoquantifytheeffectthatGSE purchaseeligibilityhadonequilibriummortgageloantermsintheperiodfrom2003to2007. The technique is designed to eliminate sources of bias that may have affected previous studies. GSE eligibility appears to have lowered interest rates by about 10 basis points, encouraged fixed-rate loansoverARMs,anddiscouragedlow-documentationandbrokeredloans. Thereisnomeasurable effectonloanperformanceorontheprevalenceofcertaintypesof“exotic”mortgages. Theoverall picturesuggeststhatGSEpurchaseshadonlyamodestimpactonloantermsduringthisperiod. JELClassifications: G21,G28,H81,N22. Keywords: Government-SponsoredEnterprises,Mortgages. Date:May7,2012. I thank Ryan Bubb, Hess Chung, Josh Gallin, Claudia Goldin, Adam Guren, Karthik Kalyanaraman, Larry Katz, DavidLaibson,ShaneSherlund,andPaulWillenforvaluablediscussionsandcomments. IamgratefultotheFederal Reserve Bank of Boston for hosting me as I conducted a portion of this research. The analysis and conclusions are my own and do not indicate concurrence by other members of the Federal Reserve research staff or the Board of Governors. *BoardofGovernorsoftheFederalReserveSystem. Email: alex.kaufman@frb.gov. 1

1. INTRODUCTION In 2011 over 75% of all mortgages originated in the United States—over $1 trillion worth— passed through the hands of the Federal National Mortgage Association (Fannie Mae) and the FederalHomeLoanMortgageCorporation(FreddieMac)(InsideMortgageFinance,2012). These institutions,knownastheGovernment-SponsoredEnterprises(GSEs),havetraditionallybeenprivate corporations with a public charter, operating with the implicit backing of the United States government.1 Their mission, as defined by their regulator the Federal Housing Finance Agency (FHFA), is to promote liquidity, affordability, and stability in the U.S. mortgage market. The GSEsaremeanttoaccomplishthesegoalsbypurchasingmortgageloansonthesecondarymarket, whichtheythenpackageintosecuritiesorholdinportfolio. InSeptember2008theGSEs’implicit government backing became explicit when, in the throes of the financial crisis and facing possible bankruptcy, both Fannie and Freddie were placed in conservatorship by FHFA. The cost to taxpayers of their bailout has been estimated at $317 billion so far (Congressional Budget Office, 2011). GiventheGSEs’vastscale,theliabilitytheyrepresenttotaxpayers,andthedecisionsthatmust soon be made about their future, it is crucial to understand how exactly they affect the mortgage markets in which they operate. Unfortunately, modeling GSE activity and estimating its effect is a challenge. Fannie and Freddie are for-profit enterprises bound by a government-mandated mission that is likely at odds with their profit motive (Jaffee, 2009). As such, it is unclear what they maximize. Furthermore, they are large relative to the market. How they affect consumer outcomes, each other, and the rest of the market depends upon details of market structure. For 1TechnicallythetermGovernment-SponsoredEnterprisealsoappliestothetwelveFederalHomeLoanBanks,which aremuchsmallerthanFannieMaeandFreddieMac. Forsimplicityinthispapertheterm“GSE”isusedtoreferonly toFannieandFreddie. 2

instance, Passmore, Sparks, and Ingpen (2002) show that whether or not lower capital costs (due to the implicit government subsidy) are ultimately passed on to borrowers in the form of lower mortgage rates depends crucially on the degree of competition or collusion between Fannie and Freddie, which is theoretically ambiguous.2 The GSEs’ huge market share may also affect their behaviorinotherways. BubbandKaufman(2009),forinstance,explorehowtheGSEs’sizemay allow them to incentivize mortgage originators using a toolbox of strategies to that is unavailable toprivate-labelsecuritizers. Empirical estimation of the GSEs’ impact on outcomes such as interest rates, default rates, and contractstructuresfacesatleastthreeimportantobstacles: selectionbias,externalities,andsorting bias. First, in part due to their government mandate, the loans GSEs buy are not a random subset ofallloans. GSE-purchasedmortgageloansonaveragedifferalongseveraldimensions,including loan size and borrower creditworthiness, from loans purchased by private-label securitizers or left in the portfolio of originating lenders. Such selection must be separated from the true treatment effectofGSEpurchases. Second, even if GSE purchases were indeed random, it would not be sufficient to simply compare mortgages bought by the GSEs with those bought by private securitizers or left in portfolio. GSEs may affect the markets in which they operate by changing equilibrium prices and contract structuresofallloans,notonlythosetheybuy. Inotherwords,eligibilityforGSEpurchasemayinfluenceloancharacteristicsbothforloansthatarepurchasedandthosethat,despitebeingeligible, are not. Because of the potential for such pecuniary externalities, estimates based on comparing 2In the Passmore, Sparks, and Ingpen (2002) model it is even possible that the establishment of the GSEs can raise equilibriuminterestrates. ForthistohappenitmustbethecasethattheGSEsbehavecollusivelyandthattheliquidity of mortgage-backed securities issued by private-label institutions is lowered because the market share of the GSEs cutsintoprivatesecuritizers’economiesofscale. 3

loans purchased by GSEs with loans not purchased will be biased toward zero, even when purchases are randomly assigned. In order to account for such externalities the ideal experiment is instead to compare loans in two similar markets, one in which the GSEs make purchases and one in which they do not, regardless of whether the individual loans being compared are ever bought bytheGSEs. Third, to the extent that GSE purchase eligibility may lead to loan terms that are more (or less) favorable to borrowers, potential borrowers may adjust their loan attributes in order to qualify for (or avoid) categories of loan that the GSEs are likely to buy. Such customer sorting is another potential source of bias. If borrowers sorting into GSE-eligible loans are different from other borrowers, and if those differences influence the features of the loans they receive—for instance, due to preferences or risk-based pricing—then customer sorting will bias estimates of GSE treatment effects. To illustrate this point with a fanciful example, imagine that GSE activity lowers interest rates by 30 basis points, and GSEs follow a government-mandated rule that they will only buy loans made to people who live in red houses. Suppose further that potential borrowers who know this ruleandaresavvyenoughtopainttheirhomesredarealso,onaverage,bettercreditrisks(inaway thatisapparenttoaloanunderwriterbutnottoaneconometricianwithlimiteddata)andsowould naturally receive loans that are cheaper by 15 basis points, regardless of house color. If we were to estimate the effect of GSE intervention on interest rates using the idiosyncrasies of the house colorrule,wewouldincorrectlyfinditis45basispointsbecausewewouldhaveconflatedthetrue treatmenteffectwiththesortingeffect. 4

ThispaperestimatestheequilibriumtreatmenteffectofGSEinterventiononinterestrates,loan delinquency rates, and mortgage contract features using an instrumental variables regression discontinuitydesignmeanttoaddressselectionbias,sortingbias,andexternalities. Thestrategytakes advantageoftheinteractionoftwofeaturesofthemortgagemarket: theconformingsizelimit,and theubiquityof20%downpayments. Bylaw,theGSEsareonlyallowedtobuyloanssmallerthantheconformingloanlimit,anupper bound that varies from year to year. In 2006 and 2007, for instance, the limit was $417,000 in the continental United States. Loans that exceed the conforming size limit are referred to as jumbo.3 Thispurchaseruleisfairlyrigorouslyobserved: in2007,forinstance,theGSEsbought88%ofall loansinthe$5,000windowjustbelowtheconformingsizelimit,butonly3%ofloansinasimilar windowjustabovethelimit.4 Researchers can potentially overcome two of the three previously mentioned sources of bias— externalities and selection—by exploiting the discontinuity in GSE intervention across the conforming size limit. By comparing loans made in a segment of the market where GSEs dominate (the conforming market) with otherwise similar loans made in a segment of the market where GSEs do not operate (the jumbo market), one can obtain estimates that incorporate pecuniary externalities of GSE purchases on the rest of the market. Also, because the GSE purchase rule is discontinuousandotherrelevantloanfeatures(absentanysortingeffects)varysmoothlywithloan size, bias due to loan selection is not a problem. Loans just above the threshold form a natural comparisongroupforloansjustbelow(see,forexample,DiNardoandLee(2004)). 3The GSEs’ purchase guidelines also include limits on loan-to-value ratios and debt-to-income ratios. While “nonjumbo” would technically be a more accurate term, for simplicity I use the term “conforming” for all loans that are belowtheconformingloanlimit,includingloansthatfailtomeettheseothereligibilitycriteria. Theseomittedcriteria willnotaffecttheregressiondiscontinuityestimation—allthatmattersisthatloansizeispredictiveofwhetheraloan ispurchasedbytheGSEs. 4Thisandotherstatisticscitedintext,unlessotherwisenoted,estimatedusingdatafromLenderProcessingServices (LPS). 5

However, a comparison of loans just above and below the conforming loan limit may still be biasedduetocustomersorting. Indeed,histogramssuchasFigure1suggestthatcustomersbunch just below the conforming loan limit, choosing a larger down payment to avoid getting a jumbo loan. If borrowers that do this are unobservably different from borrowers that don’t, estimates of the GSE treatment effect that use this discontinuity will be contaminated by sorting. Indeed, if sorting on unobservables is similar to sorting on observables (Altonji, Elder, and Taber, 2005) then the evidence is stark: the average credit score of borrowers in the sample who are just below the conforming cutoff is nearly 45 points higher than it is for those just above the cutoff. It thus appearsthatmore-creditworthyborrowersarebetterabletotakeadvantageofconformingloans. Tosimultaneouslyaddressallthreesourcesofbias,thispaperusesaslightlydifferentapproach. Rather than directly compare loans above and below the conforming loan limit, I instrument for whetheraloanis largerorsmallerthanthelimitusing adiscontinuousfunctionofhomeappraisal value. As will be explained in detail in Section 3, certain features of the loan origination process ensurethat,atparticularhomeappraisalvalues,thechancethataborrowergetsaconformingloan jumpssignificantly. Inparticular,abovesomeappraisalvaluesitisimpossibletogetaconforming loan without putting more than 20% down, inducing a jump in the number of jumbo loans at those values. Evidence suggests that these key appraisal values are not salient to either lenders or borrowers,andthereislittleevidenceofmanipulationofappraisalsaroundthesevalues. Thispaperthuscomparespricesandattributesofloansmadetoborrowerswhosehomeshappen to be appraised just below one of these values, with those of borrowers whose homes happen to be appraised just above. I argue that the resulting differences are most plausibly attributed to the different rates at which these borrowers get conforming rather than jumbo loans. Because GSE purchase eligibility is the essential difference between the conforming and jumbo markets, 6

this quasi-random assignment to the conforming loan market allows for a clean estimate of the equilibriumimpactofGSEpurchaseactivitiesonloanattributes. Using this method I find only modest impacts of GSE activity. For a sample of loans originated between 2003 and 2007 I estimate that GSE purchase eligibility lowered interest rates in the conforming market by 8 to 12 basis points, which is slightly smaller than previous estimates of the conforming/jumbospread. Ifindnosignificanteffectonloandefaultorforeclosurerates. GSEactivityappearstohavepromotedfixedratemortgagesoveradjustableratemortgages: Iestimatean increaseof5.3percentagepointsonabaseof61.9percentfixed-rateloans. GSEinterventionalso appearstohavediscouragedlowdocumentationloansandloansboughtthroughabroker. Ifindno effect on the prevalence of contract features such as pre-payment penalties, negative amortization, interest-onlyloans,balloonloans,anddebt-to-incomeratios. This paper joins a growing literature that attempts to measure the impact of GSE intervention on residential mortgage markets. Previous work has largely focused on determining the effect of GSE intervention on contract interest rates. McKenzie (2002) performs a meta-analysis of eight studies that attempt to quantify the size of the conforming/jumbo rate spread, and concludes that the spread has averaged 19 basis points over the years 1996-2000.5 Studies in this literature generallyrunregressionsinwhicha“jumbo”dummyisthecoefficientofinterest,andtheycontrol for observables that may covary with jumbo status. Though extremely useful, such studies are potentiallyvulnerabletoselectionbiasandsortingbias. Laterstudies,suchasPassmore,Sherlund, and Burgess (2005) and Sherlund (2008), yield similar estimates in the 13-24 basis point range whileattemptingtobetteraddresssourcesofbias.6 5StudiesincludeHendershottandShilling(1989);ICFIncorporated(1990);CottermanandPearce(1996);Ambrose, Buttimer, and Thibodeau (2001); Naranjo and Toevs (2001); U.S. Congressional Budget Office (2001); Passmore, Sparks,andIngpen(2002);andPearce(2002) 6Sherlund(2008),forinstance,usesgeographiclocationtocontrolforunobservedborrowercharacteristics. 7

Another important strand of the literature has attempted to determine the effect of GSE intervention on the supply of mortgage credit. Ambrose and Thibodeau (2004) uses a structural model to argue that, subsequent to the establishment in 1992 of a set of “Affordable Housing Goals” for the GSEs, the total supply of credit increased slightly more in metropolitan areas with higher proportionsofunderservedborrowers. BosticandGabriel(2006)investigatesthesamesetofhousing goalsbutusestheregulation’sdefinitionofwhatconstitutesa“low-incomeneighborhood”tocompareareasthattheGSEsweresupposedtotargetwithareaswheretheyhadnoparticularmandate, findingnoeffectofGSEtargetingonoutcomessuchashomeownershipratesandvacancyrates. The present paper contributes to this literature in two ways. First, its estimation strategy is designed to eliminate biases that may have affected previous studies. Second, it expands the set of outcomes examined to include contractual forms and features, as well as measures of loan performance. Since the original version of the present paper appeared, Adelino, Schoar, and Severino (2011) has used a related empirical methodology to study a different question: the effect of GSE loan purchases on house prices. The paper finds that being eligible for a conforming loan increases housepricesbyslightlyoveradollarpersquarefoot. Section 2 of this paper presents a brief history of the GSEs and provides background on conforming loan limits. Section 3 describes the estimation strategy in greater detail, while Section 4 discusses the dataset and the econometric specifications used. Section 5 presents results, and Section6concludes. 8

2. BACKGROUND 2.1. History of the GSEs. The Federal National Mortgage Association (Fannie Mae) was establishedin1938asafederalagencyfullycontrolledbytheU.S.government(FannieMae,2010). Its mission was to provide liquidity in the mortgage market by purchasing loans insured by the Federal Housing Administration (FHA). In 1948 that mandate was expanded to include loans insured bytheVeteransAdministration,andbytheearly1950sFannieMaehadgrowntosuchapointthat pressuremountedtotakeitprivate. In1954acompromisewasreachedwherebyFannieprivatized but was still controlled by the government through Treasury ownership of preferred stock. Fannie was also granted special privileges, such as exemption from local taxes, which it maintains to this day. The Housing and Urban Development Act of 1968 took the privatization of Fannie Mae a step farther,splittingitbyspinningoffitsfunctionsbuyingFHA-andVA-insuredloansintothewholly government-controlledGinnieMae,whilepreservingtherestofitsbusinessinthenowsupposedly fully-private Fannie Mae.7 However, Fannie Mae continued to enjoy implicit government backing foritsdebt. In1970thegovernmentcharteredtheFederalHomeLoanMortgageCorporation(FreddieMac) as a private company. Its mission—buying and securitizing mortgages to promote liquidity and stability—was similar to Fannie Mae’s mission, though initially Freddie Mac was only meant to buymortgagesoriginatedbysavingsandloanassociations. Withtimethisdistinctioneroded. Like FannieMae,FreddieMacwasperceivedbymostashavingtheimplicitbackingofthegovernment. 7An often-cited reason for this division is that a 1968 change in public accounting rules made it so that additions to FannieMae’sbalancesheetwouldbetreatedaspublicexpenditures. PrivatizingFannieMaemadefederaldebtappear smaller. 9

In the wake of the the savings and loan crisis, Congress in 1992 passed the Federal Housing Enterprises Financial Safety and Soundness Act, which established the Office of Federal Housing Enterprise Oversight (OFHEO) as the new regulator for the GSEs. The act also expanded the GSEs’ mandate to improve access and affordability for low-income borrowers by creating the AffordableHousingGoalsstudiedinAmbroseandThibodeau(2004)andBosticandGabriel(2006). The rules require the GSEs to buy a certain proportion of their loans from households defined as mid-orlow-income,andfromneighborhoodsdefinedaslow-income. TheGSEs’marketshareballoonedthroughoutthe1990sandearly2000s. Duringthistimeboth institutions expanded their loan purchases and securities issuance, and also began holding more MBS and mortgage loans in portfolio, which they financed by issuing debt.8 Spurred by competition from private-label securitizers, in the mid-2000s the GSEs began expanding their operations into the subprime and Alt-A mortgage markets, which they had traditionally avoided. With the collapse of the housing bubble in mid-2007 the GSEs’ subprime MBS holdings put them at risk of insolvency. The Housing and Economic Recovery Act (HERA) of 2008 replaced the regulator OFHEO with FHFA and granted it the power to place the GSEs in conservatorship, which FHFA did in late 2008, finally making explicit the government’s long-standing implicit backing of GSE debt. SincethentheGSEshavebeenheldinconservatorship,andtheirfutureremainsuncertain. 2.2. ConformingLoanLimits. BylawtheGSEsareonlyallowedtopurchaseloanssmallerthan the conforming loan limit (Federal Housing Finance Agency, 2010). Larger loans are referred to as jumbo. The conforming loan limit varies by both year and location. Prior to 2008 the size limit 8Lehnert, Passmore, and Sherlund (2008) investigates whether the expansion of the GSEs’ portfolios were a major forceaffectingthemortgagerate,andconcludesitwasnot. 10

increased at most once a year, and was constant across all locations within the continental United StatesandPuertoRico.9 In 2008 the passage of HERA retroactively changed the conforming size limits of loans originated after July 1st, 2007, allowing the GSEs to guarantee more loans. Because the act passed in 2008, it is unlikely that the retroactive changing of the conforming limit in some areas affected loans terms at the time of origination.10 Our only variables measured after origination, default and foreclosure, are likely functions of house price appreciation, loan terms, and borrower credit risk, and as such would not be expected to be directly affected by retroactive eligibility for GSE purchase. After HERA it is no longer the case that all continental U.S. locations are treated equally—theActdesignatedasetof“high-cost”countieswithhigherconformingloanlimits. 3. ESTIMATION STRATEGY The estimation strategy in this paper employs a discontinuous function of home appraisal value as an instrument for conforming loan status. Appraisal value is related to conforming status for obvious reasons: more expensive houses are more likely to require mortgage loans larger than the conforming limit. However, the relationship between appraisal value and conforming loan status is not smooth. It is discontinuous because loan-to-value (LTV) ratios of exactly 80 (equivalent to a down payment of 20%) are extremely modal in the U.S. mortgage market. An LTV of 80 is common in part because borrowers are typically required to purchase private mortgage insurance (PMI)forloansabove80LTV.Inaddition,80isconsidered“normal”andmayfunctionasadefault 9Hawaii, Alaska, Guam, and the U.S. Virgin Islands were considered “high-cost areas” and had a conforming limit 50%higherthantherestofthecountry. 10If the law’s passage were anticipated there could be an influence. However, even if passage were anticipated, the exactformulasdeterminingwhichcountieswereaffectedmaynothavebeenanticipated.Ifsuchanticipationdidoccur it would tend to bias the results of this paper toward zero. The data over this period show bunching of loans at the limitsthatwereinforceatthetimeoforigination,butnotattheretroactively-imposedlimits,suggestingthatthelaw changeswerenotanticipated. 11

choiceformanypeoplewhowouldotherwisechooseadifferentdownpayment. Figure2provides a histogram of the loan-to-value ratios of first-lien mortgage loans, illustrating the importance of 80LTV. To see why the widespread use of 80 LTV induces a discontinuity in the relationship between appraisal value and conforming status, note that the LTV ratio equals the origination amount divided by the appraisal value. In order to have an LTV of 80 while staying under the conforming limit, a home cannot be appraised at more than the conforming limit divided by 0.8. For a conforming limit of $417,000, for instance, this appraisal limit, as I will refer to it, would be $417,000/0.8 = $521,250. Borrowers with homes appraised above $521,250 must choose whether to put 20% or less down and get a jumbo loan, or put greater that 20% down and get a conformingloan—conformingloanswith20%downpaymentsareimpossibleforsuchborrowers. Because of the stickiness of 80 LTV, borrowers whose homes are appraised above this appraisal limit are discontinuously more likely to get a jumbo loan. Figure 3 illustrates the first-stage relationship between appraisal value and jumbo status for the 2006–2007 subsample. So long as borrowers do not sort themselves across the appraisal limit, one can use appraisal value as an instrumentforwhethertheborrowergetsaconformingorjumboloan.11 How easy is it to manipulate appraisal values? Dennis and Pinkowish (2004) provides an overview of the home appraisal process. Independent appraisals are needed because a mortgage lender cannot rely on selling price as a measure of the collateral value of the home. Typically, the lender or mortgage broker contracts a third party to provide an appraisal (Hutto and Lederman, 2003). Borrowers are not allowed to contract appraisers themselves for fear they will shop around 11Estimates using this method can be thought of as the local average treatment effect of GSE intervention on those borrowers who, if moved from an appraisal value just below $521,250 to one just above it, would not respond by raisingtheirdownpaymentsabove20%. 12

for an appraiser willing to inflate the appraisal and thus lower the borrower’s LTV. The appraiser estimates the probable market value of the home by taking into account the neighborhood, the condition of the home, improvements to the home, and recent sale prices of comparable homes in the area. Appraisals usually cost $300-500, and the fee is paid by the borrower when the loan applicationisfiled. The appraisal process is explicitly designed to make it difficult for the borrower to manipulate the appraisal value. However, appraisal manipulation by the lender remains a concern. Anecdotal evidence suggests lenders sometimes leaned on appraisers to inflate values to make loans more attractive for resale on the secondary market.12 Appraisers unwilling to inflate values may have seen a loss of business as a result. Such manipulation may indeed have occurred, but is only relevant for this paper if it occurred across the particular appraisal limit used in the regression discontinuity. If the efforts of lenders to encourage appraisal inflation were less targeted, targeted at another goal, or occurred in small enough numbers, such manipulation would not pose a threat to the empirical strategy. As will be shown in Section 4, there appears to be no bunching around the appraisal limit, suggesting that appraisal values around this limit were not compromised by manipulationbyeitherlendersorborrowers. Borrowerscanmanipulateappraisalvaluesinonelegalway: bybuyingalargerorsmallerhouse. However, this form of manipulation is coarse. It would be difficult for a borrower to inch across the threshold by this means; the appraisal value might change by tens of thousands of dollars, or not at all. So long as our estimate is based on the discontinuity in the local area around the cutoff, we can be reasonably sure borrowers are not using home choice to position themselves just below 12See,forinstance,“InAppraisalShift,LendersGainPowerandCritics,”NewYorkTimes,August18,2009. 13

the threshold. Furthermore, the smooth density function we find around the appraisal limit again suggeststhatthisformofmanipulationisnotaproblem.13 Another potential cause of concern about the estimation strategy is the availability of outside financing that is not observable in the dataset. During the 2003–2007 period it became became tolerated practice to fund down payments with a second-lien mortgage. These so-called “silent seconds” were often 15-LTV (or even 20-LTV) second-lien mortgages on an 80-LTV first-lien mortgage. Because the data do not allow for the linkage of first and second lien mortgages made onagivenproperty,itislikelythatasignificantportionofthe80-LTVloansseeninthedatawere infactsupplementedbyasecond-lienmortgageatthetimeoforigination. However, the invisibility of these second loans does not present a problem for the estimation strategy. Such seconds are the means by which some borrowers managed to stay within the size limitofaconformingloan. Solongasnoteveryborrowerusedsecondloanstostaywithinthesize limit—perhapsbecausesuchsecondswereunavailableorwerealreadymaxedout,ortheborrower was unaware or uninterested in them—then the estimation will provide an unbiased local average treatment effect of GSE purchase activity on those borrowers that would not use seconds in this wayiftheyreceivedanappraisalabovetheappraisallimit. Suchborrowersexistinequalnumbers above and below the appraisal limit, but only above the limit are they more likely to actually get jumboloans. Though appraisal manipulation and silent seconds are unlikely to present problems for the estimation strategy, at least four limitations of the strategy should be mentioned. First, this method is not appropriate for studying the GSEs’ effect on loan terms during the financial crisis itself. 13Somehavenotedatendencyofappraisalstoexactlymatchhomesalevalues.Totheextentappraisalvalueswerenot independentfromsalevalues,thiswouldprovideamechanismbywhichborrowerscouldfinelymanipulateappraisal valuesaroundtheappraisallimit. However,thesmoothnessofthedensityfunctionaroundthelimitsuggeststhatthis mechanismwasnotcommonlyexploited. 14

Fromlate2007onwardtherewasacollapseinthejumboloanmarket. Thoughthisitselfsuggests that the GSEs may have played an important role ensuring access to credit during the crisis, the tiny number of jumbo loans in the 2008–2011 period eliminates the control group necessary for the estimation strategy. In effect, there is no longer a first-stage relationship between appraisal value and jumbo status because there are, to a first approximation, no longer jumbo loans. This paperthereforefocusesontheperiod2003–2007,andestimatestheeffectsofGSEactivityduring non-crisistimes. Second,allestimatesapplytoborrowerstakingloansneartheconformingloanlimit. Despitethe factthatthesampleperiodof2003–2007sawanunprecedentedextensionoflargemortgageloans to poorer borrowers, it is still the case that most borrowers taking loans close to the conforming limit were relatively affluent. Therefore this estimation strategy is not able to address the question ofwhateffectGSEinterventionsmayhavehadontheloantermsoflessaffluentborrowers. Third,thisstrategyisill-suitedtoestimatingtheGSEs’effectonaccesstomortgagecredit. The continuity that we see in the loan density function across the appraisal limit suggests that there is little GSE effect on credit availability, at least for more affluent borrowers in the non-crisis 2003– 2007 period. However, developing a formal test of this proposition would necessitate adapting a density discontinuity estimation approach such as McCrary (2008) for use in an instrumental variables framework. Such an exercise might be of little use in any event, as GSE credit access effectsmightbeexpectedmoststronglyforlessaffluentborrowersorduringcrises. Lastly, these estimates cannot be interpreted as more general estimates of the effects of loan securitization. Though the proportion of conforming loans displays a discontinuity around the appraisallimit,thesecuritizationrateitselfdoesnotdisplayadiscontinuity(thoughitdoeschange 15

slope). The results should instead be interpreted as the effects on price, contract structure, and defaultofbeinginasegmentofthemarketeligibleforpurchasebytheGSEs. 4. DATA AND SPECIFICATIONS 4.1. Data. ThedatausedinthispapercomefromLenderProcessingServicesAppliedAnalytics, Inc. (LPS).14 These are loan-level data collected through the cooperation of mortgage servicers, including the ten largest servicers in the United States.15 The data cover over half of outstanding mortgages in the United States and contain more than 32 million active loans. Key variables includeoriginationamount,homeappraisalamount,loanterms,securitizationstatus,andmonthly paymentperformance. The analysis sample contains first-lien, non-FHA non-VA insured mortgage loans backed by owner-occupied, single-family homes and originated between the years 2003 to 2007. To be included in the sample, both the origination amount and the appraisal value must be $1,000,000 or less. Table 1 provides summary statistics for this sample of approximately 14.9 million mortgage loans. Thenumbersforthefullsamplearebroadlyconsistentwithstatisticsfoundinstudiesusing other data sources.16 The rightmost columns provide averages for loans that fall within a $5000 band on either side of their appraisal limit. This provides a base rate against which the size of the regressionestimatescanbejudged.17 14ThesedataareoftenreferredtobythenameMcDash. LenderProcessingServicesacquiredMcDashAnalyticsin November2008. 15Mortgage servicers fulfill a role similar to building superintendents: they collect payments from borrowers and pursue accounts thatare delinquent. A mortgage loan’s servicing rightsare often sold separatelyfrom rights to that loan’sstreamofpayments.AllthemortgagesintheLPSdatasetwereeitheroriginatedbyoneitsparticipatingservicers orhavehadtheirservicingrightssoldtooneoftheseservicers. 16Direct comparisons with other studies are difficult because of variation in sample selection. Mayer, Pence, and Sherlund(2009),forinstance,coversthesametimeperiodbutfocusesmoreontheAlt-Aandsubprimemarketsthan thepresentstudydoes. 17Becausethisbaserateiscalculatedusingloansneartheappraisallimit,thevastmajorityofwhichareconforming, thisrateshouldbeinterpretedastheratethatexistswithGSEintervention,whilethisrateminustheregressionpoint estimateyieldstheratethatwouldexistintheabsenceoftheGSEs. 16

Figure 1 presents a histogram of loan frequency by origination amount for the continental U.S. in the years 2006 and 2007.18 Visual inspection confirms that there is an atom of borrowers positioned just below the conforming size limit of $417,000. The figure also displays evidence of rounding. Dollaramountsendingineven$5,000,$10,000,and$50,000incrementsaremorecommon than other amounts. The presence of rounding makes formal analysis of the discontinuity (as in McCrary (2008)) unreliable. However, because $417,000 falls between tick marks (where we would expect to find a smooth density despite rounding), and because the density there is larger than in any other bin, the atom is very likely not an artifact of rounding. It appears that some borrowersarebunchingjustbelowthelimitinordertoavoidjumboloans. Bunching below the limit can only create bias if borrowers below the limit are different from borrowersabovethelimit. LPSdatacontainlimitedinformationaboutborrowercharacteristics,but they do contain one important measure: credit (FICO) score. Taking our 2006–2007 continental U.S. sample, the average FICO score of borrowers in the $5000 bin just below the conforming limit of $417,000 is 740.9, while the average FICO of borrowers in the $5000 bin just above is only 696.5. This swing of nearly 45 FICO points represents a very sizable drop-off in credit quality. ThoughitispossibletoexplicitlycontrolforobservablessuchasFICOscore,thissorting on observables suggests there may be sorting on unobservables as well. This motivates the use of aninstrumentalvariablesspecificationbasedonappraisalvalue. Figure 4 presents a histogram of loan frequency by appraisal value for the same sample. Again thereisevidenceofrounding,thistimemakingitdifficulttovisuallydeterminewhetherthereisan atom. Figure 5 provides a close-up of the area around the $521,250 cutoff, which confirms there 18Because the conforming loan limit varies by year and location, histograms of the full sample are not easily interpretable. However,the2006-07continentalU.S.subsamplehasasingleconforminglimit($417,000)andsoiseasily visuallyinterpreted.Othersubsamplesexhibitnearlyidenticalbehavioraroundtheirrespectiveconforminglimits.For the sake of interpretability all figures use the 2006–2007 continental U.S. subsample, while all regression estimates usethefull2003–2007sample. 17

is no evidence of abnormal bunching. The average FICO score of borrowers in the $5000 bin just belowthecutoffis719.6,whiletheaverageFICOscoreofborrowersinthebinjustaboveis719.3. Itthusappearsthatappraisalvalueisnotmeaningfullycompromisedbyborrowersorting,andisa validrunningvariableforourregressiondiscontinuityanalysis. 4.2. Specification. The instrumental variables regression discontinuity specification used in this paper fits a flexible polynomial on either side of the appraisal cutoff and measures the size of the discontinuity using a dummy variable taking value 1 for observations below the cutoff. The first-stagespecificationis: (1) X = α +α Z + f(APP)+g(APP)∗Z +α S +υ i 0 1 i i i i 2 i i Where X is anindicator for whetherthe loan originationamount is under theconforming limit, i f(·)andg(·)are7th-orderpolynomialfunctionsofappraisalamount,Z isanindicatorforwhether i the appraisal amount is under the appraisal limit, and S is a vector of control variables including i refinance status, dummies for FICO score in 5-point bins, and over 600,000 dummies for every zip code/month of origination combination in the dataset, allowing us to control for local market conditions extremely flexibly.19 Although the appraisal limit varies by year and location, all data ispooledbyre-centeringthedatasuchthat,foreachyearandlocation,therelevantappraisallimit is equal to zero. This allows the full 2003–2007 sample to be run in a single regression. Table 19These variables were chosen because they are all pre-treatment variables with respect to home appraisal. Other variables, such as loan-to-value ratio, or whether the loan is fixed- or adjustable-rate, are omitted because they are determinedpost-treatment.However,includingthesevariablesdoesnotmeaningfullychangetheresults.Additionally, dummiesareincludedforwhethertheappraisalvalueisanexactmultipleof$5000or$1000inordertoaccountfor anypotentialreportingeffectsrelatedtotheroundingseeninthedata. 18

2 provides a summary of the applicable conforming limits and appraisal limits for all years and locationsinthesample. Thesecond-stagespecificationis: (2) Y = β +β Xˆ +h(APP)+k(APP)∗Z +β S +(cid:15) i 0 1 i i i i 2 i i Where Y is an outcome, such as interest rate, and Xˆ is the predicted value from the first stage. i i TheeffectonoutcomeY ofgettingaloanintheconformingmarketasopposedtothejumbomari ket is estimated by the coefficient β . The estimate can be thought of as a local average treatment 1 effect of GSE activity onthose borrowers who would not respond to a slightlyhigher appraisal by increasingtheirdownpaymentabove20%inordertostayintheconformingmarket. Many of the outcome variables (Y) used in this study are binary, suggesting a probit or logit i specification. However, the size of the dataset (nearly 15 million observations) coupled with the number of independent variables (over 600,000) makes such an estimation impractical. For this reasonalinearprobabilitymodelisusedinstead. 5. RESULTS As a first step, Figure 3 confirms that there is power in the first stage by presenting a scatterplot of percent conforming against appraisal value for the continental U.S. in 2006 and 2007. Visual inspection shows a clear discontinuity at the appraisal limit of $521,250. Virtually all borrowers with homes appraised at $521,000 end up with conforming loans, whereas borrowers with homes appraisedat$521,500arediscontinuouslymorelikelytogetjumboloans. Table3showstheresults 19

of a formal first-stage regression using the full sample. There is a discontinuity of 8.8 percentage points,significantatthe1%level,inwhetherornottheborrowergetsaconformingloan. Tables4and5presenttheregressionresults. Eachcoefficientinthetablesrepresentsaseparate instrumentalvariablesregression,eachusingappraisalvalueastherunningvariableandincluding thecompletesetofcontrolvariables. TheestimateinTable4ofa10-basispointjumbo/conforming spread is about half the size of many estimates in the literature (McKenzie, 2002). If previous estimates suffered from customer sorting (specifically, more-creditworthy borrowers choosing conforming loans over jumbo loans) this would tend to bias those estimates upwards. However, the disparitycouldalsobeduetootherfactors,suchasthedifferenceinsampleperiod. While conforming status appears to push basic interest rates down, the estimate of its effect on introductory ARM teaser rates is positive 4.6 basis points. Why might teaser rates move in the opposite direction from other rates? One possibility is that lower teaser rates are associated with contractsthataremoreexpensiveinotherways. BubbandKaufman(2011)showsthatinasample of credit card contracts, for-profit investor-owned credit card issuers were more likely to offer low teaser rates but high interest rates and penalties later on, while cards issued by credit unions have higher teaser rates but lower charges otherwise. Seen in that light, higher teaser rates and lower baseratesmaybeanaturalpairing. LoanseligibleforGSEpurchaseappeartoenterdefaultandforeclosureatthesamerateasother loans—neitherestimateissignificant. AnegativeeffectofGSEinterventionondefaultwouldhave been slightly more in line with prior work. Both Elul (2009) and Krainer and Laderman (2009) 20

compare the delinquency outcomes of GSE-securitized loans and privately securitized loans, attempting to control for relevant risk characteristics, and conclude that GSE-securitized loans generally perform better. However these studies look at realized securitization status, not purchase eligibility,anddonotattempttoaccountforsortingbias. Note that the interest rate effect, in the absence of any significant loan performance effect, suggests that the price difference is not simply due to less risky borrowers receiving a discount. It suggests instead that the price difference is a true effect of GSEs passing on the implicit governmentsubsidytoborrowers. Table 5 examines the GSE effect on a number of mortgage contract features. There appears to be no effect on the prevalence of a number of “exotic” contract features: pre-payment penalties, interest-onlyloans,loansallowingnegativeamortization,andloanswithballoonpaymentsallhave pointestimatesindistinguishablefromzero. However,thereisaGSEeffectonatleastthreeaspects of the contract. The conforming market appears to favor fixed-rate mortgages over adjustable-rate mortgages: the prevalence of adjustable-rate mortgages is estimated to drop by 5.3 percentage points. This result is consistent with Green and Wachter (2005), and suggests the GSEs may play aroleinallowingborrowerstoavoidinterestraterisk. The results further show that GSE activity lowers the prevalence of brokered loans by 4.9 percentagepoints,andoflowdocumentationloansby7.8percentagepoints. Bothlowdocumentation andtheuseofbrokershasbeenassociatedwithpoorloanperformanceduringthecrisis. However, it appears that the drops in low documentation and brokerage induced by GSE activity are not enoughtohavehadanaffectondefaultorforeclosure. 21

6. CONCLUSION ThispapercontributestotheliteratureonGSEinterventioninthemortgagemarketintwoways. First,itemploysanoveleconometricstrategydesignedtoproduceestimatesfreeofselectionbias, sortingbias,andexternalities. Second,itexpandsthesetofoutcomesexaminedbyincludingcontract features and measures of loan performance. For borrowers with loans near the conforming limit, during the 2003–2007 period, GSE activity lowered interest rates by 8 to 12 basis points, while modestly decreasing the prevalence of adjustable-rate mortgages, low documentation loans, and loans originated through a broker. Effects on contract structure are mixed. There is no measurable effect on loan performance. As the post-conservatorship future of Fannie and Freddie is debated,thissetofeffectsshouldbeweighedagainstthecostofgovernmentsupportoftheGSEs, aswellasthepotentialtoachievesuchoutcomesthroughothermeans. REFERENCES ADELINO, M., A. SCHOAR, AND F. SEVERINO (2011): “Credit Supply and House Prices: EvidencefromMortgageMarketSegmentation,”Unpublishedmanuscript. ALTONJI, J., T. ELDER, AND C. TABER (2005): “Selection on Observed and Unobserved Variables: AssessingtheEffectivenessofCatholicSchools,”JournalofPoliticalEconomy,113(1). AMBROSE, B., R. BUTTIMER, AND T. THIBODEAU (2001): “A New Spin on the Jumbo/Conforming Loan Rate Differential,” Journal of Real Estate Finance and Economics, 23. AMBROSE, B., AND T. THIBODEAU (2004): “HavetheGSEAffordableHousingGoalsIncreased theSupplyofMortgageCredit?,”RegionalScienceandUrbanEconomics,34. 22

BOSTIC, R., AND S. GABRIEL(2006): “DotheGSEMattertoLow-IncomeHousingMarkets? An Assessment of the Effects of the GSE Loan Purchase Goals on California Housing Outcomes,” JournalofUrbanEconomics,59. BUBB, R., ANDA. KAUFMAN(2009): “SecuritizationandMoralHazard: EvidencefromaLender CutoffRule,”FederalReserveBankofBostonPublicPolicyDiscussionPaperNo.09-5. (2011): “ConsumerBiasesandFirmOwnership,”NewYorkUniversityLawandEconomicsWorkingPaper11-6. CONGRESSIONAL BUDGET OFFICE (2011): “The Budgetary Cost of Fannie Mae and Freddie MacandOptionsfortheFutureFederalRoleintheSecondaryMortgageMarket,”. COTTERMAN, R., AND J. PEARCE (1996): “TheEffectoftheFederalNationalMortgageAssociationandtheFederalHomeLoanMortgageCorporationonConventionalFixed-RateMortgage Yields,” Studies on Privatizing Fannie Mae and Freddie Mac, U.S. Department of Housing and UrbanDevelopment. DENNIS, M., AND T. PINKOWISH (2004): Residential Mortgage Lending: Principles and Practices,5thEdition.ThomsonSouth-Western. DINARDO, J., AND D. LEE (2004): “Economic Impacts of New Unionization on Private Sector Employers: 1984-2001,”QuarterlyJournalofEconomics,119. ELUL, R. (2009): “Securitization and Mortgage Default: Reputation vs. Adverse Selection,” FederalReserveBankofPhiladelphiaWorkingPaper09-21. FANNIE MAE (2010): “AboutFannieMae,”http://www.fanniemae.com/. FEDERAL HOUSING FINANCE AGENCY (2010): “Conforming Loan Limit,” http://www.fhfa.gov/. 23

GREEN, R., AND S. WACHTER (2005): “The American Mortgage in Historical and International Context,”JournalofEconomicPerspectives,19(4). HENDERSHOTT, P., AND J. SHILLING (1989): “The Impact of the Agencies on Conventional Fixed-RateMortgageYields,”JournalofRealEstateFinanceandEconomics,2. HUTTO, G., AND J. LEDERMAN (2003): Handbook of Mortgage Lending. Mortgage Bankers AssociationofAmerica. ICF INCORPORATED(1990): “EffectsoftheConformingLoanLimitonMortgageMarkets,”U.S. DepartmentofHousingandUrbanDevelopment. INSIDE MORTGAGE FINANCE (2012): MortgageMarketStatisticalAnnual. JAFFEE, D. (2009): “The Future Role of Fanny Mae and Freedie Mac in the U.S. Mortgage Market,”Unpublishedmanuscript. KRAINER, J., AND E. LADERMAN (2009): “Mortgage Loan Securitization and Relative Loan Performance,”FederalReserveBankofSanFranciscoWorkingPaper09-22. LEHNERT, A., W. PASSMORE, AND S. SHERLUND (2008): “GSEs, Mortgage Rates, and SecondaryMarketActivities,”JournalofRealEstateFinanceandEconomics,36(3). MAYER, C., K. PENCE, AND S. SHERLUND (2009): “The rise in mortgage defaults,” Journal of EconomicPerspectives,23(1),27–50. MCCRARY, J. (2008): “Manipulation of the running variable in the regression discontinuity design: Adensitytest,”JournalofEconometrics,142(2),698–714. MCKENZIE, J.(2002): “AReconsiderationoftheJumbo/Non-JumboMortgageRateDifferential,” JournalofRealEstateFinanceandEconomics,25. NARANJO, A., AND A. TOEVS(2001): “TheEffectsofPurchasesofMortgagesandSecuritization by Government Sponsored Enterprises on Mortgage Yield Spreads and Volatility,” Journal of 24

RealEstateFinanceandEconomics,25. PASSMORE, W., S. SHERLUND, AND G. BURGESS (2005): “The Effect of Government- SponsoredEnterprisesonMortgageRates,”RealEstateEconomics,33. PASSMORE, W., R. SPARKS, AND J. INGPEN (2002): “GSEs,MortgageRates,andtheLong-Run EffectsofMortgageSecuritization,”JournalofRealEstateFinanceandEconomics,25. PEARCE, J. (2002): “ConformingLoanDifferentials: 1992-1999,”Unpublishedmanuscript. SHERLUND, S. (2008): “The Jumbo-Conforming Spread: A Semiparametric Approach,” Federal ReserveBoardFinanceandEconomicsDiscussionSeries2008-01. U.S. CONGRESSIONAL BUDGET OFFICE(2001): “InterestRateDifferentialsBetweenJumboand ConformingMortgages,”U.S.CongressionalBudgetOffice. 25

APPENDIX ytisneD 20. 510. 10. 500. 0 Histogram of Origination Amount Continental US 2006−2007 0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 Origination Amount (in $1000s) FIGURE 1. Histogram of loan origination amounts for 2006-07 continental U.S. subsample. Theverticallineisthe$417,000conformingsizelimit. ytisneD 2. 51. 1. 50. 0 Histogram of Loan−To−Value Ratios Continental US 2006−2007 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 Loan−To−Value Ratio FIGURE 2. Histogramofloan-to-valueratiosforthe2006-07continentalU.S.subsample. 26

gnimrofnoC tnecreP 1 8. 6. 4. Percent in Conforming Market by Appraisal Amount Continental US 2006−2007 450 500 550 600 Appraisal Amount (in $1000s) FIGURE 3. Proportion of loans smaller than the conforming limit, by home appraisal amount, for 2006-07 continental U.S. subsample. The vertical line is the $521,250“appraisalsizelimit”equaltotheconforminglimitdividedby0.8. ytisneD 10. 800. 600. 400. 200. 0 Histogram of Appraisal Amount Continental US 2006−2007 0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 Appraisal Amount (in $1000s) FIGURE 4. Histogram of home appraisal amounts for 2006-07 continental U.S. subsample. The vertical line is the $521,250 “appraisal size limit” equal to the conforminglimitdividedby0.8. 27

ytisneD 40−e0.4 40−e0.3 40−e0.2 40−e0.1 0 Histogram of Appraisal Amount: Detail Continental US 2006−2007 0 0 0 0 0 0 0 0 0 0 0 0 5 5 5 5 5 5 5 5 5 5 5 5 2 2 2 2 2 2 2 2 2 2 2 2 4 5 6 7 8 9 0 1 2 3 4 5 1 1 1 1 1 1 2 2 2 2 2 2 5 5 5 5 5 5 5 5 5 5 5 5 Appraisal Amount ($) FIGURE 5. Detailofhistogramofhomeappraisalamountsfor2006-07continental U.S.subsample. Theverticallineisthe$521,250“appraisalsizelimit”equaltothe conforminglimitdividedby0.8. 28

TABLE 1. SummaryStatistics FullSample NearAppraisalLimit Mean S.D. Obs. Mean S.D. Obs. OriginationAmount($) 212,322 129,932 14,941,284 303,385 88,241 162,235 AppraisalValue($) 308,559 191,472 14,941,284 458,768 50,650 162,235 Jumbo .095 .294 14,941,284 .089 .285 162,235 FICOScore 711.6 61.9 12,733,244 722.3 56.0 139,257 Loan-to-ValueRatio 72.0 16.9 14,815,612 65.9 16.9 161,282 InterestRate(%) 6.25 1.38 14,284,352 6.01 1.35 153,771 AdjustableRateMortgage .279 .448 14,812,239 .354 .478 160,722 ARMTeaserRate(%) 5.34 2.26 4,116,418 4.88 2.15 55,110 Pre-PaymentPenalty .133 .340 14,593,905 .152 .359 159,565 Interest-OnlyAllowed .126 .332 14,941,284 .175 .380 162,235 NegativeAmortizationAllowed .050 .219 14,941,284 .066 .248 162,235 Balloon .009 .092 14,941,283 .009 .096 162,235 Brokered .310 .462 9,866,479 .327 .485 106,208 LoworNoDocumentation .320 .466 8,117,111 .379 .485 87,858 Debt-to-IncomeRatio 34.9 13.4 10,033,173 35.2 12.7 112,091 61+DayDefault .107 .309 14,941,284 .098 .297 162,235 Foreclosure .071 .258 14,941,284 .065 .246 162,235 Notes: Sample of first-lien, non-FHA insured, non-VA insured loans made to borrowers with owner-occupied single-family residences between the years 2003 and 2007. The sample contains only loans with both origination amount and appraisal value $1,000,000 or less. Near Appraisal Limit contains the subset of loans that fall in the $5000 band on either side of their ownappraisallimit. InterestRatedefinedascontractinterestrateforfixed-ratemortgageloans, and as post-teaser margin plus index for adjustable rate mortgage loans. Index value taken at time of origination. 61+ Day Default and Foreclosure equal to 1 if loan ever attains that status withina36-monthwindowfollowingorigination. 29

TABLE 2. ConformingLoanLimitsandAppraisalLimits Standardareas High-costareas ConformingLimit AppraisalLimit ConformingLimit AppraisalLimit 2003 $322,700 $403,375 $484,050 $605,063 2004 $333,700 $417,125 $500,550 $625,688 2005 $359,650 $449,563 $539,475 $674,344 2006 $417,000 $521,250 $625,500 $781,875 2007 $417,000 $521,250 $625,500 $781,875 Notes: High-cost areas are defined during the sample period as Alaska, Hawaii, Guam, and the U.S. Virgin Islands. The standard limit applies to the continental U.S. and Puerto Rico. During the sample period the high-cost limit is always 50% largerthanthestandardlimit. Appraisallimit isdefinedastheapplicableconforminglimitdividedby0.8. TABLE 3. FirstStage BelowConformingLimit AboveAppraisalLimit(α) -0.088*** s.e. (.001) BaseRate 0.969 N 14,941,284 Notes: First stage regression of conforming status on a dummy indicating whether a loan is above the appraisal limit. Controlsincludea7th-orderpolynomialoneitherside of the appraisal limit, dummy variables for every combination of zip code and origination month, as well as refinance status and FICO score in 5-point bins. Base Rate is the sample average in the $5000 band below the appraisal limit. Standarderrorsinparentheses. ***,**,and*denotestatisticalsignificanceatthe1%,5%,and10%levels,respectively. 30

TABLE 4. PriceandPerformance InterestRate ARMTeaserRate 61+DayDefault Foreclosure (basispoints) (basispoints) β -10.07*** 4.59*** 0.008 0.007 s.e. (1.04) (0.82) (0.009) (0.009) BaseRate 600.94 487.89 0.098 0.065 N 14,284,352 4,116,418 14,941,284 14,941,284 Notes: Eachcellisaninstrumentalvariablesregressionofthedependentvariableon conforming status, instrumenting for conforming status with appraisal value. Controls include a 7th-order polynomial on either side of the appraisal limit, dummy variables for every combination of zip code and origination month, as well as refinance status and FICO score in 5-point bins. Interest Rate defined as contract interest rate for fixed-rate mortgage loans, and as post-teaser margin plus index for adjustable rate mortgage loans. Index value taken at time of origination. 61+ Day DefaultandForeclosureequalto1ifloaneverattainsthatstatuswithina36-month window following origination. Base Rate is the sample average in the $5000 band on either side of the appraisal limit. Standard errors in parentheses. ***, **, and * denotestatisticalsignificanceatthe1%,5%,and10%levels,respectively. TABLE 5. ContractFeatures AdjustableRate Pre-PaymentPenalty InterestOnly NegativeAmortization β -0.053*** -0.014 0.003 0.008 s.e. (0.009) (0.009) (0.009) (0.009) BaseRate 0.354 0.152 0.175 0.066 N 14,812,239 14,593,905 14,941,284 14,941,284 Balloon Brokered LowDocumentation DTIRatio β 0.003 -0.049*** -0.078** 2.633 s.e. (0.009) (0.012) (0.014) (1.713) BaseRate 0.009 0.327 0.379 35.196 N 14,941,283 9,866,479 8,117,111 10,033,173 Notes: Each cell is an instrumental variables regression of the dependent variable on conforming status, instrumenting for conforming status with appraisal value. Controls include a 7th-order polynomialoneithersideoftheappraisallimit,dummyvariablesforeverycombinationofzipcodeand origination month, as well as refinance status and FICO score in 5-point bins. Low Documentation includes no documentation loans. Base Rate is the sample average in the $5000 band on either side of the appraisal limit. Standard errors in parentheses. Sample sizes vary due to missing data for some dependent variables. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels,respectively. 31