Does the Community Reinvestment Act (CRA) Cause Banks to Provide a Subsidy to Some Mortgage Borrowers?

Does theCommunity ReinvestmentAct (CRA)CauseBanks to Provide a Subsidy to Some MortgageBorrowers? (cid:3) Glenn B. Canner ElizabethLaderman SeniorAdvisor Economist Board ofGovernorsoftheFederal Federal ReserveBank ofSan ReserveSystem Francisco Andreas Lehnert WaynePassmore Board ofGovernorsoftheFederal AssistantDirector ReserveSystem Board ofGovernorsoftheFederal Mail Stop93 Reserve System WashingtonD.C. 20551 MailStop 93 (202)452-3325 WashingtonD.C. 20551 Andreas.Lehnert@FRB.GOV (202)452-6432 Wayne.Passmore@FRB.GOV This Version: April2002 LastRevised: April 2,2002 (cid:3) The opinions, analysis and conclusionsof this paper are solely ours and do not necessarily representthose of the Board of Governorsof the Federal Reserve System, the Reserve Bank of SanFranciscoortheirstaffs. WethankMaryDiCarlantonio,JamieIngpen,ErinKlein,JudyPeng andGuillermoPinczukforexcellentresearchassistance. Anyremainingerrorsareourown.

Does theCommunity ReinvestmentAct (CRA)CauseBanks to Provide a Subsidy to SomeMortgageBorrowers? Abstract The Community Reinvestment Act (CRA) encourages lenders to make mortgage loans to certain classes of borrowers. However, the law does not apply to all lenders, and lenders do not necessarily receive credit for all loans made to borrowers of a particular class. Specifically, only commercial banks and savings institutions are subject to the CRA, while mortgage bankers are not. Further, CRA credit is given for loans made to higher-income borrowers who purchase homes in lower-income neighborhoods, but not to other higher-income borrowers. We use thisvariation to test whether or not CRA-affected lenders cut interest rates to CRA-eligible borrowers; in other words, we test for the presence of a regulationdriven subsidy. Our theory suggests that loans made by commercial banks and savings associations (“relationship lenders”) and mortgage companies (“transactionlenders”)willdifferfromoneanotherdependingonborrowerriskandhomeownershipbenefits. Empirically,wefindthatCRA-eligibleloansatCRA-affected institutions do carry lower mortgage spreads compared with other loans at the same institution. However, once we control for risk and benefit effects suggested by our theory, these differences in mortgage spreads become economically and statisticallyinsignificant. Journalof EconomicLiteratureclassificationnumbers: G21,G28,H23, R21 Keywords: CommunityReinvestmentAct, mortgages,bank regulation

1 Introduction The Community Reinvestment Act (CRA) requires the federal banking agencies thatsupervisecommercialbanksandsavingsassociations(“banks”)toencourage suchlenders“tohelpmeetthecreditneedsofthelocalcommunitiesinwhichthey are chartered, consistent with the safe and sound operation of such institutions.” In practice, lenders are judged by their regulators primarily on the number and dollaramountofloansthattheymaketolower-incomeborrowers,orborrowersin lower-income neighborhoods. If lenders satisfy their auditors by attracting such borrowers with lower-than-profitable interest rates, the CRA would be a burden onthebankingsystem.1 We test the effects of the CRA using data on home purchase mortgages made over the period 1995-2000 and variation in the institutions subject to the CRA and the types of borrowers (and neighborhoods) for whom lenders receive CRA credit.2 Thesevariations,however,arestronglycorrelatedwithfactorsinfluencing the nature of the mortgage contract. In our theory section, we develop a model to guide our empirical analysis. Because we embed our analysis in a general model of banking practices, our study has more general applicability. In particular, we find that, as predicted by our theory, institutions offer radically different 1ForstudiesoftheeffectsoftheCRAonbankprofitabilityseeCannerandPassmore(1996), Avery, Bostic and Canner (2000,2002) , Board of Governors of the Federal Reserve System (1993,2000),Meeker and Myers(1996)and Malmquist, Phillips-Patrick, and Rossi (1997). For areviewofthetheoreticalargumentssurroundingCRA, seeCannerandPassmore(1995b). Our paperissimilarininteresttoZinman(2002),whostudiesthesmallbusinessaspectofCRA,while weconcentrateonthemortgageaspect. 2Forastudyoftheextentoflendingtolower-incomeborrowersbyCRA-affectedinstitutions comparedtootherinstitutions,seeCannerandPassmore(1996). 1

services; some institutions offer loans that look like classic capital-market armslength bonds, while other institutions offer loans that feature relationship-based services that,ineffect, increase theborrower’scredit-worthiness. Our paper is of more general interest than as a test of the CRA; in particular, itcontributestotwodifferentbodiesofliterature. First,wecontributetothestudy of the economic effects of laws and regulations using variation in the coverage of laws.3 Our empirical work is closely guided by the theory that we develop; weshowhowsimpleatheoreticalregressionswouldleadtoincorrectconclusions. Second, we contributeto thestudy of the natureand role of financial institutions; in particular, how bank loans differ from bonds and other capital market instruments.4 Theoretical work in this field often presupposes that banks offer distinct services fromcapital markets;herewefind direct evidencethat thisis true. Mortgagesaremadeby awidevarietyoffinancial institutionsto manydifferentborrowertypes. Thegreatest distinctionamonglenders isbetween depository institutions, such as commercial banks and savings associations, on one side and mortgage companies on the other. Banks are subject to a myriad of regulations (includingtheCRA);atthesametime,bankscultivateongoingfinancialrelationships with their borrowers. Mortgage companies are not subject to the same set of regulations as banks, nor do they seek a relationship with borrowers; indeed, becausetheysooftenselltheirmortgages,theycanbeseenasawayforborrowers 3Seee.g.Levitt(1997,1998)AcemogluandAngrist(2001),LehnertandMaki(2002),orPence (2001)whousevariationsacrosstimesorspace;inthispaperweusevariationacrossinstitutions. 4InadditiontoBootandThakor(1997,2000),seealsoBoltonandFreixas(2000)andChemmanurandFulghieri(1994). 2

to access the capital markets. Thus, we refer to banks as relationshiplendersand mortgagecompaniesastransactionlenders. Ourtheoryisbasedontheassumptionthatrelationshiplendershaveaccessto a credit-enhancement technology that effectively increases the borrower’s repayment probability, while transaction lenders simply take the borrower’s apparent credit risk as given and charge the appropriate zero-profit spread. This theory is basedonthecapitalmarketsliteratureaboutfirms’choiceofbankorbondfinancing (see e.g. Boot and Thakor 2000, 1997 and the references therein). The main result of this assumption is that relationship lenders (i.e. banks) have access to a costlytechnologythatmakesborrowersbetteroff;thisistheessenceofthebankborrower relationship.5 For banks to cover the cost of their credit enhancement activitiestheymustchargehighermortgageinterestrates(inturn,someborrowers willbewillingtopay theserates inexchangeforhavingtheircredit enhanced). Thus, even before introducing a law like the CRA, we would observe similar borrowers being charged different mortgage rates depending on whether the borrower contracts with either a bank or a nonbank financial institution. For all but the riskiest borrowers, we show that in a competitive equilibrium, borrowers of thesameapparentcreditriskpayahighermortgagerateatbanksthantheywould at mortgagecompanies. Furthermore,weshowthattheinterestratechargedbybanks(butnottransac- 5Thus we study the distinction between banks and nonbank finance companies in much the samewayCarey,PostandSharpe(1998)do,althoughinaverydifferentmarket. Inaddition,we concentrateontheeffectofasingleregulation. Formoreonthebank-borrowerrelationship,see alsoSharpe(1997). 3

tionlenders)willdepend crucially ontheborrower’sbenefitfrom owningahome (as opposed to renting). If we take this ownership benefit to be purely pecuniary anddrivenbythetaxcode’smortgageinterestdeduction,thenbankswouldcharge lower-incomeborrowerslessthanhigher-incomeborrowers(controllingforcredit quality). In effect, lower-income borrowers have a smaller surplus available to split with banks. Our theory thus suggests that banks will, in general, charge higher mortgage rates than non-banks, but that they will cut mortgage rates to theirlower-income(hencelower-benefit)borrowers. Becausehigh-benefit(i.e.higher-income)borrowersbenefitthemost(controllingforcreditquality)fromthebank’scredit-enhancementtechnology,ourtheory suggeststhat banks willdisproportionatelyservethese borrowers. Lower-income borrowers (again, controlling for credit quality) will prefer to use low-cost nonbankfinancialinstitutions. TheCRA,whichencouragesbankstoserveborrowers of different incomes in proportion to their populations might, as a result, encouragebankstocross-subsidizefromhigher-tolower-incomeborrowerssothattheir borrower population would reflect the larger population. One might then test for such a subsidy by regressing mortgage rates against a standard set of controls and an indicator variable set to unity when a loan is both CRA-eligible and the lender is subject to the CRA. But a negative coefficient on this variable; that is, finding that CRA-affected institutions (i.e. banks) charge CRA-eligible borrowers (i.e. lower-incomeborrowers)lessthan they wouldhavehad theborrowernot been CRA-eligible, or had the borrower instead gone to a transaction lender is evidence consistent with both the view that the CRA is a binding constraint on 4

banksandtheviewthat banksprovidecostlycredit-enhancementservices. Thus, our theory forces us to confront a serious identification issue as we try to discriminate between these two hypotheses. Specifically, it is precisely those institutions with access to the credit enhancement technology (i.e. banks) that are subject to the CRA and it is mostly those borrowers with a relatively low benefit (i.e. lower-income borrowers) to homeownership that are eligible for CRA credit. Ifwecouldobservethebenefittohomeownership,andifthisbenefit did not vary precisely with CRA eligibility, then we could confidently test the hypothesis that the CRA causes banks to subsidize loans to certain borrowers. In such a happy circumstance, we would regress mortgage rates on the standard set of controls and, as additional regressors, borrower credit risk and benefit to homeownership interacted with an indicator variable for relationship lenders. In such a specification, if the coefficient on a variable interacting CRA eligibility with lender type still showed banks cutting rates to CRA-eligible borrowers, we wouldtakethisasevidenceofasubsidy. We take two approaches. First, we make no attempt to control for the borrower’s benefit to homeownership. This provides us with an upper bound on the potential subsidy extended by banks. Second, we make the identifying assumption that the benefit to homeownership is purely pecuniary and driven only by the mortgage interest deduction in the tax code. Under this assumption, higherincome borrowers will benefit the most from homeownership, and this benefit will not be affected by the location of the home. Effectively, banks receive CRA credit for higher-income borrowers only if the borrower purchases a home in a 5

lower-income neighborhood. By restricting our dataset to solely higher-income borrowers,weapproachouridealdataset,because,underouridentifyingassumption, CRA eligibility (i.e. buying a house in a lower-income neighborhood) does notaffectthebenefitofhomeownership. Wethentestwhetherhigher-incomeborrowersreceivelowerinterestratesatbanks,butnottransactionlenders,iftheybuy a house in a lower-income neighborhood. Under our identifying assumption, the existenceoflowerrates wouldbeevidenceofa subsidy. Under our first approach, we find that the upper bound on the potential subsidy is tiny: at the very most, it amounts to less than six basis points. Under our second approach, we can in general go further and reject thehypothesisof a subsidy(although undercertain definitionsof higher-income,our resultsindicatethe presence of a small subsidy). Thus, we conclude that the CRA does not cause banks to extend mortgageloans with substantiallylowermortgagerates to attract CRA eligibleborrowers. Theplanofthispaperisasfollows: section2reviewsthesalientfeaturesofthe CRA, section3 lays outourmodeland associated analysis,section4 describes in somedetail the construction and nature of our data, section 5 presents our results and section 6 briefly concludes. We include further information on the constructionofourfinal datasetandananalysisofhouseholds’propensitytoitemizetheir deductionsin an appendix. Alltablesare at theend ofthepaper. 6

2 The Community Reinvestment Act The CRA was enacted in 1977 and is intended to encourage commercial banks and savings associations to help meet the credit needs of the local communities in which they are chartered. In adopting the CRA, the Congress reaffirmed the principle that depository institutions have an obligation under their charters to serve“theconvenienceandneeds”oftheircommunitiesbyextendingcredittoall partsofthosecommunities.6 TheCRAisdirectedprimarilyatfourfederalsupervisoryagencies–theBoard ofGovernorsoftheFederalReserveSystem,theComptrolleroftheCurrency,the Federal DepositInsuranceCorporation,andtheOfficeofThriftSupervision. The Act calls upon these agencies to (1) use their supervisory authority to encourage each financial institution to help meet local credit needs in a manner consistent with safe and sound operation, (2) assess an institution’s record of meeting the creditneeds ofitsentirecommunity,includinglower-incomeneighborhoods,and (3)considertheinstitution’sCRA performancewhen assessinganapplicationfor a charter, deposit insurance, branch or other deposit facility, office relocation, merger, oracquisition. To enforce theCRA, theregulatoryagencies conductperiodicCRA examinationsofcommercialbanksandsavingsassociationsand,asrequiredbythestatute, evaluate CRA performance during the application process for bank acquisitions, mergers and otheractions. The vaguenessof theaffirmativeresponsibilityplaced 6ForanoverviewofthehistoryoftheCRA,seeGarwoodandSmith(1993). 7

on lenders by the Congress has made it difficult for the regulatory agencies to determine compliance with the CRA. Most institutions receive a rating of satisfactory or better on their CRA performance, and few institutions have had their applicationsformergersoracquisitionsdenied. TheCRAhas,however,prompted institutions to undertake specific actions to enhance their CRA performance before and during the application process; for an overview of these programs and theirprofitability,seeAvery,Bostic,and Canner (2000). In 1995,theagencies beganimplementingarevisedCRA regulationthatuses three distinct performance-based measures: a lending test, an investment test and a service test.7 These tests combine the judgment of CRA examiners with quantitativemeasures of performance, such as the ratio of mortgages extended to lower-income borrowers to all mortgages and the ratio of mortgages extended in lower-income neighborhoods to all mortgages. When adopting the new regulation, the agencies noted that the examination process is inherently subjectiveand requires that performance be measured within the context of (1) a community’s credit needs and (2) thecapabilityof thelender. Thesetwo standardsare referred toas the“performancecontext”. TheCRAlegislationplacesaheavyemphasisontheanalysisofthegeographic distribution of an institution’s lending across its entire community. The current CRAregulationimplementsthislegislativeintentbyclassifyingneighborhoodsin alender’sCRAassessmentareaaslow-,moderate-,middle-,orhigher-income. A 7Inthispaper,wefocusexclusivelyonthelendingtestportionofCRAregulation(see12CFR 228,regulationBB“CommunityReinvestment”;section28containsthelendingtest). 8

low-incomeareaisdefinedasanareawherethemedianfamilyincomeislessthan 50percentofthemedianfamilyincomeforthebroaderarea(suchasametropolitan statistical area or MSA). In a moderate income area, the median family income is at least 50 percent and less than 80 percent of that for the broader area. In a middle-income area, the percentages range from at least 80 percent up to 120 percent. And in a higher-income area, the percentage is at least 120 percent. These income definitions divide the population and the number of census tracts into groups of unequal size, with far fewer people, owner-occupied homes, and census tracts in the lower-income groups.8 We will refer to neighborhoods (or borrowers) with less than 80 percent of the MSA median family income as lower-incomeneighborhoods(orborrowers). ThecurrentCRA regulationalsoextendstheevaluationofabank’slendingto encompass the distribution of loans across low-, moderate-, middle-, and higherincome borrowers, where the income categories follow the same groupings as neighborhoodsbutrelyontheindividual’sincomerelativetothatoftheborrower’s MSA or broader non-metropolitanarea median family income for individualsresiding outside an MSA. Thus, while continuing to place a heavy emphasis on thegeographicdistributionofaninstitution’slending,theagenciesalsofavorably considerloansmadeto lower-incomeindividuals. Oneresultofthisdualapproach(consideringboththeborrower’sownincome as well as the income of the neighborhood in which the borrower seeks to purchase a home) is that loans to higher-income borrowers are favorably considered 8SeeCannerandPassmore(1995b). 9

by regulators for CRA purposes only if the borrower is purchasing a home in a lower-income neighborhood. Indeed, loans to higher-income borrowers not purchasinghomesinlower-incomeneighborhoodsmightbeseenascountingagainst thelenderforCRA purposes,becauseregulatorsusetheratiosofloanstohigherand lower-income borrowers relative to the total as two yardsticks to gauge a lender’sCRA performance. Weexploitthisfact inourempiricalworkbelow. CRAexaminationsconsiderabroadrangeofloanproducts,includingalltypes of residential, consumer, and business loans. Our paper is focused only on home purchase lending, an important component of the lending test, because the data availablepursuant totheHomeMortgageDisclosureAct (HMDA)allowtheempirical investigation of the nature and extent of this type of lending by the mortgageindustrytodifferent neighborhoodsanddifferent borrowersin allMSAs.9 3 A Model of Competitive Banking and Provision of Mortgage Loans Themortgagemarketischaracterizedbyavarietyofinstitutionsthatservedifferent types of borrowers. In our model, we assume that financial institutions have fullinformationaboutallborrowers,butprovidedifferentlevelsofservice(or“relationships”)to borrowers. This view could be easily modified to incorporate the 9TheHMDArequiresfinancialinstitutionswithofficesinmetropolitanareastoprovideinformation on the geographic location of the properties related to the home loans they originate or buy. HMDAalsorequireslenderstodiscloseinformationonthedispositionofhomeloanapplications,thedateoftheloan,andontheraceornationalorigin,gender,andannualincomeofloan applicationsorborrowers. 10

assumptionthatfinancial institutionshaveonlypartialinformationaboutborrowers (thus adding the possibility of adverse selection) and must undertake costly background checks to verify a borrower’s true credit risk. The main conclusions fromthemodelwouldbeunchanged. Bankers often distinguish themselves from other sorts of lenders by saying they develop a relationship with their borrowers. In the mortgage markets, one might pose a stylized contrast between mortgage companies, who focus exclusively on mortgage transactions and with whom the borrower is unlikely to have repeat transaction in the near future, and banks, who develop a long-standing relationshipwith theborrowerand provideotherfinancial services to that borrower over time. Another element of this stylized contrast is that mortgage companies are usually funded by the capital markets and generally sell their mortgages in secondary mortgage markets, whereas banks are funded by both the capital marketsandretaildeposits,andsometimesholdtheirmortgagesintheirportfolios. In addition,banksaresubjecttoextensiveregulationandsupervisionbothforsafety and soundness and for community reinvestment, whereas independent mortgage companies are, for the most part, free of such regulation, particularly as regards communityreinvestment. 3.1 Borrower and Lender Types In ourmodel, the mortgageapplicant has thechoice of borrowingfrom relatively high-costrelationshiplenders(whichweidentifyasbanks)orfromlow-costtransactionlenders(whichweidentifyasindependentmortgagecompanies). However, 11

before determining what typeof institutionfrom which to obtain a mortgage, applicants first need to decide whether they wish to purchase or rent housing. Over all but the shortest horizons, and for all but the lowest-income households, purchasingoffersaclearadvantageoverrenting(seeHendersonandIoannides(1983) and the large literature that followed it). The one exception to this rule is if the homeownerdefaultsandsuffersforeclosure. Althoughthespecificdetailsofforeclosure law vary among states (Pence, 2001), there is little question that renting beatsbeing evicted. Tobegin,weassumethatallfinancialintermediariescanobserveaborrower’s probabilityof repayment( (cid:18) = [ 0 ; 1 ] )withoutincurring anycosts. The borrower’s utilityis: U BUY = 8 > < > : B (cid:0) R withprobability (cid:0) Æ (cid:18) withprobability 1 (cid:0) (cid:18) (1) Here B is the net benefit to the borrower of owning a house (relative to renting), and R istheinterestpaymentonthemortgage. Wenormalizetheutilityofrenting sothat: U RENT = 0 : (2) Theexpectedutilitya borroweroftype (cid:18) is: E (cid:8) U BUY j (cid:18) ; R (cid:9) = (cid:18) ( B + Æ (cid:0) R ) (cid:0) Æ : (3) 12

Fortheexpectedutilityofbuyingtoexceedtheutilityofrenting,interestpayments cannot be too high; we can thus derive a schedule of maximuminterest rates that borrowerswouldbewillingto paybefore renting: R m a x ( (cid:18) ) = B + Æ (cid:0) ( Æ = (cid:18) ) : (4) Mortgage companies, in contrast to banks, are assumed to make only transaction or commodity loans; that is, loans that require little contact between the borrowerand lender. Usinga competitivecapitalmarket for theinitialfunding of aloanandthensellingtheloanintothecapitalmarkets,wemodelmortgagecompaniesasacompetitiveindustrywithzeroeconomicprofits. Thus,thetransaction loanrate, R T ,willsatisfy: (cid:18) R T = (cid:26) + c T : (5) where (cid:26) isthecapitalmarket fundingcost fortheloanand c T istheper-mortgage originationcostfortransactionlenders.10 Borrowers can also turn to the banking industry for a mortgage loan. The bankingindustrycanduplicatethebehaviorofmortgagecompaniesoritcan provide relationship loans; that is, loans where the lender undertakes some costly actions that decrease the probability that the borrower defaults. For example, bankersmightencourageandofferfinancialorhomebuyereducationtosomebor- 10Thepurchaserofthemortgagefromthemortgagebankerdiscountsthemortgageratebythe expecteddefaultloss and, for ease of presentation, we assume thata defaultedmortgagehas no value.Inotherwords,theexpectedreturnonthemortgageis (cid:18) R + ( 1 (cid:0) (cid:18) ) 0 . 13

rowers or might provide neighborhood-based loan officers that spend some time becomingfamiliarwiththepeopleandcharacteristicsofaparticulararea. Inaddition,banksmightbeabletouseinformationgleanedfromtheirdepositoryroleto provide valuable services to their borrowers. Each relationship loan will have an origination cost c R > c T and an additional adjustment cost related to the amount ofcreditenhancement undertaken bythelender, (cid:13) . Thebank’sexpectedprofits perdollarofmortgageloan are: (cid:25) (cid:0) (cid:13) ; R R (cid:1) = (cid:18) ( (cid:13) ) R R (cid:0) (cid:13) 2 2 a (cid:0) ( (cid:26) + c R ) ; where (cid:13) 2 [ 0 ; 1 ] ; a > 0 : (6) Here, (cid:18) is thecredit-enhanced probabilityofrepayment, defined as: (cid:18) = (cid:18) + ( 1 (cid:0) (cid:18) ) (cid:13) ; where (cid:18) 2 [ (cid:18) ; 1 ] : (7) 3.2 Competitive Equilibrium Assuminga competitiverelationship lending (i.e. banking)industry, bankers will compete to offer the best feasible contract to each borrower of type (cid:18) . Banks choose the mortgage rate, R R , and the degree of credit enhancement, (cid:13) , to maximizetheborrower’sexpectedutility, V ( (cid:18) ) . Thatis,banksmustsolvetheproblem: V ? R ( (cid:18) ) = m R a R x ;(cid:13) E (cid:8) U BUY j (cid:18) ( (cid:13) ) ; R R (cid:9) ; subject to: (cid:25) (cid:0) (cid:13) ; R R (cid:1) (cid:21) 0 : (8) 14

Here V ? R ( (cid:18) ) is theborrower’s expected utilityfrom takingan optimalrelationship loan. If competition forces the profit per loan to zero, then (from equation 6), we can derive a relationship between the mortgage interest rate at relationship lenders, R R , and thequantityofcredit enhancementundertaken bythelender, (cid:13) : R R ( (cid:13) ) = 1 (cid:18) (cid:18) (cid:13) 2 2 a + (cid:26) + c R (cid:19) : (9) Bysubstitutingthisconditionintothebank’sproblem,equation(8),wecanderive theutility-maximizinglevelofcredit enhancement: (cid:13) ? = a ( 1 (cid:0) (cid:18) ) ( B + Æ ) ; (10) so: (cid:18) ? = (cid:18) + a ( 1 (cid:0) (cid:18) ) 2 ( B + Æ ) : Substituting this result into the borrower’s valuation of the mortgage, we can deriveaclosed-formsolutionfor V ? R ( (cid:18) ) : V ? R ( (cid:18) ) = ( B + Æ ) (cid:18) + a 2 ( B + Æ ) 2 ( 1 (cid:0) (cid:18) ) 2 (cid:0) ( Æ + (cid:26) + c R ) : (11) The value to the borrower of choosing a relationship loan in a competitive banking environment, V ? R ( (cid:18) ) , is quadratic in the borrower’s initial repayment probability. Further, although the quantity of credit enhancement undertaken by the relationship lender is decreasing in (cid:18) , the effective repayment probability, (cid:18) , is non-monotonein (cid:18) . Examples of these policies are shown in figure 1. The top panel of the fig- 15

ure shows how the post-enhancement probability of repayment, (cid:18) , relates to the pre-enhancement probability, (cid:18) . The non-monotone relationship is driven by the amount of relationship investment, (cid:13) , shown in the middle panel. As the borrower’s own credit quality increases, the relationship lender performs less credit enhancement(thatis,investslessintherelationshipwiththeborrower),thusaugmenting the borrower’s credit quality less. These top two panels display graphically the relationshipsderivedin equation (10). Thebottompanel showshow the mortgage interest rate charged by the relationship lender, R R , varies with (cid:18) . Noticethat R R ( (cid:18) ) isdecreasingin (cid:18) ,butthatitisrelativelyflatforlowcredit-quality borrowers(near (cid:18) = 0 ). Next, we compute the value to the borrower of choosing a transaction loan, V ? T ( (cid:18) ) . By substituting R T (from equation5), intotheborrower’s expected utility frompurchasinga house(from equation3),we get: V ? T ( (cid:18) ) = ( B + Æ ) (cid:18) (cid:0) ( Æ + (cid:26) + c T ) : (12) Noticethat V ? T islinearin (cid:18) , withalowerintercept and largerslopethan V ? R . Borrowers with substantial credit risk will prefer to rent rather own a home. We can define a critical level of credit risk, (cid:18) L , where all borrowers with (cid:18) < (cid:18) L willchooseto rent. At the other extreme, some borrowers are almost certain to repay their mortgages even without any credit enhancement, and gain little from credit enhancement. These borrowers will always seek a transaction loan, and we can define 16

Figure1: Optimalpoliciesofrelationshiplenders Relationship Lender Policies 1 Post−enhancement credit−quality 0.5 Pre−enhancement credit−quality 0 0 0.5 1 ) (q ’ q 1 Credit enhancement investment: g 0.5 0 0 0.5 1 ) (q g 0.5 Mortgage rate: R Cost of funds: r 0 0 0.5 1 ) q(R Borrower Credit Quality: q NOTE. Figureshowsoptimalpoliciesofrelationshiplenders(banks)asfunctions ofborrowercreditquality(therepaymentprobability (cid:18) ). Inthisexample,thebenefittohomeownershipis B = 0 : 9 5 ,thecostofdefaultis Æ = 0 : 0 5 ,theorigination costis c R = 0 : 0 1 and thenet costoffundsis (cid:26) = 0 : 1 0 . 17

another critical value, (cid:18) T , where borrowers with (cid:18) > (cid:18) T will not use a bank. By setting V ? R ( (cid:18) T ) = V ? T ( (cid:18) T ) , wecan solveforthislevelofcreditrisk: (cid:18) T = 1 (cid:0) q ( 2 = a B ) c (cid:0) + R Æ (cid:0) c T (cid:1) : (13) If the fixed loan origination costs for the bank and mortgage company are equal ( c R = c T ), the critical level of (cid:18) T is unity; everyone takes out a relationship loanbecausetherelationshiplendercan perfectly replicatethetransactionlender. Because we view relationship loans as usually more expensive than transaction loans (that is, we take c R > c T ) this critical level will be below unity and some high-qualityborrowerswillgo tothemortgagecompanies. Infigures2and3,wegraphtheexpectedutilityofagentsandthemortgageinterestraterespectivelyasafunctionof (cid:18) . Thecriticalpoints (cid:18) L and (cid:18) T aremarked onthegraphs. Notethateventhoughatransactionlendermaybewillingtomake a loan at a significantly lower interest rate than that charged by the relationship lender, someborrowers find it betterto use thecredit enhancement technology of therelationshiplender. 3.3 Varying Homeownership Benefits So far wehavenot addressed thebenefit to homeownership, B . Now assumethat there are two borrower populations: A lower-benefit population (with low values of B ) who make up a proportion (cid:11) of the total and a higher-benefit population (withhighervaluesof B )whomakeuptheremaining 1 (cid:0) (cid:11) . Thetwopopulations 18

Expected Utilities 0 q q 0 L T 1 Repayment Probability: q ) (q *V dna ) (q *V R T Interest Charges 1 Relationship Transaction 0 q q 0 L T 1 Repayment Probability: q Figure2: ExpectedUtility ) (q *R dna ) (q *R R T Relationship Transaction Figure3: InterestRates will be otherwise identical. The key difference between them will be that the benefit tohomeownershipnet of thebenefit torenting willbe lower forthelowerbenefit group. We cannot observe the private benefit to homeownership (as opposed to renting) directly. However, we can assume that this benefit is at least correlated with the borrower’s income, so that high-income agents are more likely to be highbenefit. Thisisasensibleassumptionbecauseatleastpartofthebenefitofhomeownershipisdrivenbyagovernmentpolicythatencourageshomeownership. The U.S. income tax code is also progressive, so that the homeownership subsidy is lower for lower income groups. Brady, Cronin, and Houser (2001) quantify this starkly,showingthattheeffectivemortgageinterestdeductionsubsidyrateismore thantwiceas largeforhouseholdsjointlyearning between $75,000and $100,000 as itisforhouseholdsjointlyearningless than$50,000peryear. We model the different benefits to ownership associated with each benefit 19

groupas differentlevelsofthenet benefit B ofhomeownershipoverrenting. Associated with each level of B will be different values of the critical points (cid:18) L and (cid:18) T . Lower-income households, by virtue of their lower benefit B , will be more likely to rent, all else equal. Thus (cid:18) L will be higher for the lower income group. Bythesametoken,theywillhavelesssurplustosplitwiththerelationshiplender (thegainsfrom credit enhancementwillbelower)and (cid:18) T willshifttotheleft.11 The expected utilities of both types of borrower (high and low benefit) under each type of loan are displayed in figure 4. The associated interest rates are displayed in figure 5. As expected, the transaction lenders do not alter their interest ratesbytypeofborrower. Relationshiplenders,however,lowertheirinterestrates whendealingwithlower-incomeborrowers. Theyarealsoundertakinglesscredit enhancement. EvenwithoutCRAresultinginasubsidy,iftheworldwerelikeourmodel,we would observesimilarborrowers being charged different interest rates depending onwhattypeofinstitutiontheborrowercontractswith. Borrowersatbankswould be charged a higher interest rate, controlling for observable credit quality measures,butwouldbenefitfromthebanks’credit-enhancementtechnology. According to our theory, lower-income borrowers at banks would be charged less than 11If we assume that higher- and lower-incomeborrowershave the same distribution of credit quality (cid:18) , one immediate result of our analysis is that of all the borrowersserviced by relationship lenders, a proportion less than (cid:11) will be low-benefit; and of all the borrowers serviced by transactionlenders,aproportionabove (cid:11) willbelow-benefit. Thusbankswillhavefewerlowerincome(or, more generally,lower-benefit)borrowersthan mortgagecompanies, despite the fact thatbanksperformavaluablecredit-enhancementservice. Further,bankswillhavefewerlowerincomeborrowersthantheoverallborrowerpopulation.Thisresult,whichmayseematoddswith the conventionalwisdom, can beovercomeif we assumethatlower-incomeborrowersare more likelytohavedefaultprobabilities (cid:18) withintherangeservicedbybanks, (cid:18)[ L ; (cid:18) T ] . 20

Expected Utilities 0 0 1 V Interest Rates 1 High−B, Relationship Low−B, Relationship 0.8 High−B, Transaction Low−B, Transaction 0.4 0 q L H q L L q T L q T H 0 Credit Quality: q 1 Figure4: ExpectedUtility R :etaR tseretnI High−B, Relationship Low−B, Relationship High−B, Transaction Low−B, Transaction Figure5: InterestRates higher-income borrowers at banks, holding credit risk constant, because lowerincome borrowers receive fewer benefits from homeownership. Lower-income borrowersatbanksmayormaynotpayhigherratesthansuchborrowersattransaction lenders, depending on their level of credit risk. Thus, a researcher who compares the mortgage rates of lower-income borrowers to higher-income borrowers at banks, or the mortgage rates of lower-incomeborrowers at relationship lenders to those at transaction lenders, has performed the incorrect “natural experiment” to test for a CRA-driven subsidy because he did not correct for (1) the benefit to homeownership and (2) the differing products offered by relationship and transactionlenders. However,ourtheorydoessuggestatleastonenaturalexperimentifwebelieve that the benefits of homeownership are highly correlated with income. This is a natural assumption because the primary pecuniary benefit to homeownership in the United States is probably the mortgage interest deduction; but only relatively 21

higher-incomeborrowersitemizetheirdeductionsand hencebenefit fromthistax break. But mortgages to higher-incomeborrowers receive CRA credit only if the loan is for the purchase of a home in a lower-income neighborhood. Thus, we can focus on higher-income borrowers locating in lower-income neighborhoods as a way to identify the effects of CRA. As a first experiment, if these borrowers (holding credit risk constant) pay lower mortgage rates at banks relative to other higher-income borrowers at banks, then there is evidence of a CRA subsidy. As a second experiment, if these borrowers (holding credit risk constant) pay lower mortgage rates when the mortgage is offered by a relationship lender rather than atransactionlender, thenthereis alsoevidenceofa CRA subsidy. 4 Data ItisdifficulttocomebydatatoempiricallytesttheeffectsofCRA.12 TheHMDA data is the main public source of mortgage borrower information, but lacks information on pricing (includingwhether or not theborrower is paying for private mortgage insurance) and the value of the property securing the mortgage. Here, we merge the HMDA data with two other datasets: first, the Mortgage Interest Rate Survey (MIRS) collected by the Federal Housing Finance Board and, second, data on mortgages with privatemortgage insurance collected by the Federal 12However,thereareafewstudies:Avery,BosticandCanner(2000,2002),BoardofGovernors oftheFederalReserveSystem(1993,2000),CannerandPassmore(1996),Canner,Passmore,and Surette(1996),EvanoffandSegal(1996,1997)andHarvey,Collins,Nigro,andRobinson(2001). 22

FinancialInstitutionsExaminationCouncil(FFIEC).13 HMDA provides information on residential mortgages (including their type: conventional or government-backed), purpose of the loan (home purchase, home improvement, or refinancing), and the amount of the loan. In addition, HMDA includes some information about borrowers (including income, race, ethnicity, and gender), as well as the lender type (commercial bank, savings association or mortgage bank), and location of the property securing the loan (state, MSA, county and census tract). With the property location, the characteristics of the neighborhoodthe property is located in can bedetermined from the 1990 Census ofPopulationandHousing. TheMIRScollectspricinginformationonmortgages, including the contract interest rate, points, the effective interest rate and term to maturity, but only covers conventional home purchase loans.14 It also includes the loan amount, property value and loan-to-value ratio for the mortgage and the locationoftheproperty byzip code. Finally,theFFIEC dataon privatemortgage insurance (PMI) identifies the lender, location of the property by census tract, loan amount and characteristics of the borrower (e.g. race, ethnicity, gender and income)foreach conventionalmortgagebacked by PMI. We briefly describe in this section the procedure used to match records from these three separate datasets. A more complete description of the procedure can befoundintheappendix. Wealsopresentsomestatisticsthatallowanevaluation 13Foranevaluationofanearliertwo-waymatchbetweenHMDAandMIRSdata, see Canner andPassmore(1995a). 14Theeffectiveinterestrateaccountsforboththecontractinterestrate,andpointsandfeespaid bytheborrower. Assuchtheeffectiverateisamoreaccuraterepresentationofthemortgagerate thanthecontractrate. 23

ofthequalityofthematch. Thereisnouniqueborroweridentifier(suchasborrowername,socialsecurity number or street address) in any of these databases nor is there a unique lender identifieracrossallthreedatabases,sowe“statistically”matchedtheloanrecords on conventional home purchase mortgages across the three databases using the following procedure. Broadly speaking, first, we use the date the mortgage was closed from each database to determine the month of mortgage origination. Second, we converted property identifiers in HMDA into zip codes and grouped the HMDA and FFIEC PMI data records by zip codes. Only loans located in MSAs wereconsidered inouranalysis. At this point, we had a set of mortgage loans from each database for each month of the years 1995 to 2000 grouped by zip code. For each of these monthlocationgroups,wematchedrecordsinHMDAandMIRSusingtheloanamount. Records with the least difference in loan amount were matched and kept unless theabsolutedifference wasgreater than$2,000;inthiscase, no matchwas made. The result was a set of statisticallymatched borrower/loan records that contained boththeHMDA and MIRSinformation. ThePMIinformationwasaddedtotheseHMDA/MIRSmatchedrecordsusing anotherstatisticalmatch. Fortherecordsinaparticularmonth-locationgroup,we matched the PMI to the HMDA/MIRS record based on loan amount (using the $2,000 absolute difference again), and borrower race and ethnicity (which had to beexactmatches).15 15Fora completedescriptionofthestatisticalmatchingprocedureseeappendixA.1;fordiag- 24

Althoughtheresultingdatacontainsinformationonadjustable-ratemortgages, 15-year fixed-rate mortgages and 30-year fixed-rate mortgages, we eliminated all observationsthat did not relate to a 30-year fixed-rate mortgage. Such mortgages made up about 90% of our data, and by concentrating solely on a single financial product, we could compare interest rate spreads directly, without having to control for amortization length or the fixed/floating spread. See appendix A.3 for evidencethat mortgagetypeand CRA eligibilityarenotrelated. Our empirical analysiswill feature a full set of lenderand MSA fixed effects; if a particular lender or MSA appears only seldom in our sample, the fixed effect maycompletelyabsorbanyidentificationprovidedbythoseobservations. Further, we want to concentrate on active lenders, as opposed to lenders who make few loans (of the type we study) per year. Thus we impose two conditions on each observation for it to be included in our final dataset. First, we require that each observation come from an institution that comprises at least 0.1 percent of our original database. Second, we require that each observation come from an MSA thatalso comprisesat least 0.1percent ofouroriginaldatabase.16 Theoriginaldatasetcontainedinformationon314,009conventionalhomepurchasemortgagesmadeby2,685differentlendersin304differentMSAs;about76 percentofthemortgagesintheoriginaldatasetweremadebyrelationshiplenders (commercial banks, saving associations and their mortgage affiliates). After apnosticsofthematchquality,seeappendixA.2. 16Further, the survey conducted by the FHLB to construct the MIRS excludes very small lenders;forthisreason,excludingless-activelendersfromourdatasetprobablypurgessomespuriousmatches. 25

plying our two conditions, we are left with a final dataset containing 250,593 mortgages made by 84 different lenders in 144 different MSAs; about 77 percent ofthemortgagesinthisfinal dataset weremadebyrelationshiplenders.17 The resulting dataset appears to closely mirror characteristics of the wider mortgage market. Figure 6 compares the time series of the spread in our data to the Freddie Mac conforming mortgageindex; also, the figure compares the numberofobservationsinourdata(permonth)totheMortgageBanker’sAssociation purchaseindex. ThemeanspreadinourdatatracksthespreadfromFreddieMac’s primary mortgage market survey (PMMS) quite well, with a correlation of about 90 percent. The mean spread in our data, however, consistently exceeds Freddie Mac’s published average, probably because our data include non-conforming and highLTV mortgages;further, we usetheeffective rateon themortgagewhile the index tracks the contract rate and points separately. Meanwhile, the monthly pattern of observations in our data closely mirrors the pattern in the Mortgage Banker’s Association (MBA) purchase index, with a correlation of about 90 percent between the two series. Also, notice that our data also mirror the clear trend intheMBAindexacross years, notjusttheintra-yearseasonalpattern. 17For thepurposesofCRA examination,bankinginstitutionscan chooseto includethe mortgagelendingoftheirmortgageaffiliates. 26

Figure6: Comparisonsbetweenourdatasetandwidermortgagemarketindicators Mortgage Interest Rate Spreads 300 250 200 150 100 r =0.90 50 1995 1996 1997 1998 1999 2000 2001 Date stnioP sisaB Dataset Mean Freddie Mac (a) Spreads Mortgage Activity 8000 6000 4000 2000 r =0.91 0 1995 1996 1997 1998 1999 2000 2001 Date xednI dna snoitavresbO Observations in Dataset MBAA Purchase Index (b)Observations NOTE. Thetoppanel of thefigurecompares themean monthlymortgageinterest ratespreadinourdatatotheFreddieMacmortgagerateindex. Thebottompanel of the figure compares the monthly number of observations in our dataset to the MortgageBanker’s AssociationofAmerica(MBAA)purchaseindex. 27

5 Empirical Specification and Results 5.1 Identification of a Subsidy Identifying Assumptions Webelieve,apriori,thatindependentmortgagecompaniesaremostlikethetransactionlendersinourmodelandthatallotherlenders(commercialbanksand savingsassociations)are mostliketherelationshiplenders inourtheory. Under our theory, a subsidy would result in CRA-eligible borrowers getting lowermortgagespreads at relationshiplendersrelativeto otherborrowersat relationship lenders, holding constant the borrower’s credit risk and the borrower’s benefit to homeownership. Further, our theory predicts that the difference in interestrateschargedthesameborroweratatransactionlenderandatarelationship lenderwilldependonlyontheborrower’sriskcharacteristicsandhomeownership benefit. In an ideal world, in which borrowers are distributed randomly among lenders and the lenders were randomly subject to the CRA, we could identify a subsidy by regressing mortgage spreads on variables related to the borrower’s riskiness, the prevailing cost of funds, MSA and institutional dummies and other variables related to thelocal competitiveenvironment and then interacting lender typeonlywiththosevariablesrelatedtotheborrower’sriskiness,benefittohomeownershipandCRAeligibility. IftheestimatedeffectofCRA eligibilityatCRAaffected institutions were negativein this regression, we would conclude that the CRA was causing affected lenders to attract CRA eligible borrowers with lower 28

mortgage rates than the lenders would normally charge such borrowers, i.e. that the CRA caused banks to subsidize mortgages to eligible borrowers. In the data, though, only relationshiplenders are subject to theCRA; and, further, thebenefit tohomeownershipisprobablycorrelated toaborrower’sCRA-eligibility. We pursue two strategies to identify the effect of CRA. First, we ignore completely the borrower’s benefit to homeownership and instead identify the CRA interaction term as an upper bound on any subsidy. Second, we assume that the benefit to homeownership is largely tax driven.18 Under this assumption, only higher-income borrowers are high-benefit and locating in a lower-income neighborhooddoesnotlowerthatbenefit. Wethenrestrictoursampletohigher-income borrowers only and identify the effect of CRA as the effect of making a loan to a higher-income borrower in a lower-income neighborhood. As a robustness check, we use several different criteria to identify households as higher-income (and hencehighbenefit). Our theory tells us that underthe null hypothesisof no subsidy,thespread on mortgage j , S j , shouldbegivenby: S j = f ( (cid:18) j ) + R j f (cid:21) ( (cid:18) j ) + (cid:13) (cid:21) R j g j + (cid:11) C j + X j B + u j : ( H 0 ) Here (cid:18) j is the credit-worthiness and g j the homeownership benefit of borrower j ; R j = f 0 ; 1 g indicates whether the loan was made by a relationship lender and 18See Brady, Cronin, and Houser (2001) for evidence that higher-income households benefit much more from the mortgage interest deduction in the U.S. tax code than lower or moderate incomehouseholds. 29

C j = f 0 ; 1 g indicateswhethertheloanwasCRA-eligible. Thevector X j contains otherexplanatoryvariables. Thecoefficient (cid:11) onCRA-eligibleloanscapturesthe common reaction to CRA eligibilityby both relationship and transaction lenders. On theotherhand, ifrelationshiplenders activelysolicitCRA eligibleloans with lowermortgagerates,wewouldexpectmortgagespreadstobegivenbythealternatespecification: S j = f ( (cid:18) j ) + R j f (cid:21) ( (cid:18) j ) + (cid:13) (cid:21) R j g j + (cid:11) C j + (cid:11) (cid:21) R j C j + X j B + u j : ( H 1 ) In the presence of a CRA subsidy, (cid:11) (cid:21) would be negative. One approach to discriminating between ( H 0 ) and ( H 1 ) is to estimate the empirical counterpart of ( H 1 )and testwhethertheestimatedcoefficient (cid:11) c (cid:21) = 0 . Becausewehavenogoodwaytodeterminethebenefittohomeownership, g j , for any borrower, we must interpret the estimated coefficient (cid:11) c (cid:21) with care. If we arewillingtoassumethathomeownershipbenefit(andhencethecredit-enhancing investment by relationship lenders) is uncorrelated with CRA-eligibility then we can proceed to test (cid:11) c (cid:21) empirically, despite the lack of a proxy for homeownership benefit. However, it is more likely that homeownership benefit and CRAeligibility are actually negatively correlated. CRA-eligible borrowers are either lower-income, and hence benefit less from the tax code’s subsidy of homeownership, or purchasing homes in lower-incomeneighborhoods, which may contain fewer ownership benefits (e.g. house price appreciation may be more uncertain). This effect will tend to bias (cid:11) c (cid:21) down, possibly leading to a false conclusion that 30

thereisin fact aCRA subsidy. Ontheotherhand,ifwearewillingtoassumethatthehomeownershipbenefit iscompletelytiedtothetaxcode’smortgageinterestdeduction,wecanrestrictour sampletothoselikeliesttoitemizetheirdeductionsandhenceclaimthemortgage interestdeduction. Thusourtheorysuggeststwoidentifyingassumptions: E (cid:8) g j j C j = 1 (cid:9) (cid:20) E (cid:8) g j j C j = 0 (cid:9) ; ( and: I 1 ) E (cid:8) g j j Y j (cid:21) Y (cid:9) = g ( Y j ) ; all C j ( . I 2 ) Here Y j is the borrower’s income and Y is some threshold income above which we assume that that homeowners itemize their deductions. Under the first identifyingassumption,( I 1 ), theestimatedCRA subsidywillbean upperboundonthe true subsidy. Under the second identifying assumption, ( I 2 ), the estimated CRA subsidywillbeconsistent,althoughtheintercepttermforrelationshiplenderswill bebiasedupward, contaminatedby thefunction g ( Y ) . Identifying HigherBenefit Borrowers Despiteouridentifyingassumption,equation( I 2 ), wedo notprimarilyusean absolute income threshold to classify households as being higher-income; instead, weusetheCRA-mandated definitionofhavingan incomeofat least120%ofthe MSA median income.19 Recall from our discussion of the CRA in section 2 that regulatorsusethisdefinitionofhigher-income,andusetheratioofloanstohigher- 19Insection5.4wepresentresultsusingavarietyofabsolutethresholds. 31

incomeborrowersasoneyardstickinevaluatingCRA performance. Thus,households defined as higher-income by our measure will have a range of incomes; in MSAs with lower median incomes, the cutoff will be lower than in MSAs with highermedianincomes. ByusingthisMSA-dependentdefinitionofincomeclass we accord with CRA regulations, but at the cost that some households classified as higher-income will have incomes below those of some households classified as middle income. Figure 7 displays the empirical cumulative distribution functions of borrower income conditional on income class. As shown in the figure, the highest five percent of middle income borrowers have incomes that exceed those of the bottom quarter of higher income borrowers. However, note that all householdsclassified ashigher-incomehaveannualincomesabove$50,000year, and that about 90% have incomes above $60,000. Thus, although there is some overlap between middle- and higher-income borrowers, most higher-income borrowersdo havefairly highincomes. The key issue, though, is whether these higher-income borrowers truly have a higher benefit to homeownership. Because we have identified the benefit to homeownership with the tax code’s subsidy of mortgage interest (the mortgage interest deduction, or MID), we want to be sure that higher-income households (under our MSA-based definition) are actually taking advantage of the MID at a greater rate than lower- or middle-income households. To take advantage of the MID, householdsmustitemizetheirdeductions. Thus we can useinformationon the likelihood of itemizing conditional on income and quantity of mortgage debt to determine if households that we identify as higher-income are in fact using 32

the MID. As explained in appendix B below, we used the Survey of Consumer Finances (SCF) to determine the probability that a household itemizes its deductions. The estimated probabilities, conditional on income and quantity of mortgage debt, are shown in figure 8. Notice that the probability of itemizing rises sharply for households with annual incomes of around $60 to 80,000, and then levelsoff. 5.2 Empirical Specification Althoughourtheoryandidentificationstrategyprovideuswithanempiricalspecification, we present results from two alternate and simpler specifications, so that we have three primary specifications in all. In addition, we use both the standard form of CRA eligibility and an expanded form that treats the two eligibility criteria separately in each of our specifications. In all the specifications we present here, we include fixed effects for both the institution (i.e. the particular bank or mortgagecompanythatmadetheloan)and theMSA. We have a standard set of controls, which we label X i for observation i : In addition, we partition the data into two subsets, so X i = [ W i Z i ] : We will take thesubsetof controls W i to beonly thosecontrols associated withborrowerrisk. A full explanation and variable names are shown in table 1. Sample means of selected variables are in table 2; tables 3, 4 and 5 present sample statistics on the loanspread,borrowerincomeandtheloanamount(mortgagesize)conditionalon lendertypeand borrowerincomeclass(higher, middleorlower). In addition to the standard set of variables, there are four variables of partic- 33

Figure7: EmpiricalCumulativeDistributionsConditionalonIncomeClass 100 95 Higher Lower 50 Middle 25 0 10 50 67 100 250 Income (real, thousands) tnecreP NOTE. Figure gives the empirical CDFs of borrower income, in thousands of real 1996 dollars, conditionalon theborrower’s incomeclass. Incomeclasses are defined relative to the MSA median income, so some middle-income borrowers will have incomes that exceed those of some higher-income borrowers. Table 4 presentssamplestatisticsforborrowerincome. Figure8: ProbabilityofItemizing 1 0.8 0.6 0.4 0.2 0 10 50 67 100 250 Income (real, thousands) ytilibaborP NOTE. Figure gives the probability of itemizing tax deductions conditional on income. The probability is derived from probit estimates on observations from the 1995 and 1998 waves of the Survey of Consumer Finances; see appendix B forfurtherdetails. 34

ular interest. First, CRAELIG is an indicator variable set to unity if the loan is eligibleforCRA credit. Thisvariableiscomprisedoftheunionoftwootherindicatorvariables,LOWMODandIRLT80. Thefirstofthese,LOWMOD,indicates whether the borrower’s neighborhoodis lower income, which makes the loan eligible for CRA credit, regardless of the borrower’s own income. The second of these,IRLT80, indicateswhether theborrower’sown incomeis below80 percent ofthemedianincomeoftheMSA,inwhichcasetheloanalsoiseligibleforCRA credit, regardless of the borrower’s neighborhood. Note thatCRAELIG and its partsdo notdepend onthelendertype,onlytheborrower’scharacteristics. The final variable of interest is RELLEND. This is an indicator variable set to unity if the lender is subject to the CRA. In our theory, we identify these institutionsone-for-onewithrelationshiplenders. We now are ready to describe our three primary specifications. In all of our regressions, the dependent variable S i will be the spread, in percentage points, between the effective rate on loan i and the average prevailing 10-year Treasury rate in the month in which the loan was made. We label the specifications ( A ), 35

( B )and ( C ), correspondingtotheequations: S i = X i B + (cid:11) CRAELIG i + (cid:21) RELLEND i + u i ( A ) S i = X i B + (cid:11) CRAELIG i + (cid:21) RELLEND i ( B ) + (cid:11) (cid:21) CRAELIG i (cid:2) RELLEND S i = W i B W + Z i B Z + (cid:11) i + u i CRAELIG i ( C ) + RELLEND i (cid:2) (cid:2) (cid:21) + W i B W (cid:21) + (cid:11) (cid:21) CRAELIG i (cid:3) + u i : These basic regressions can be thought of as increasing in sophistication. Equation ( A ) is the simplest regression, while equation ( B ) expands upon the basic regression by including an interaction term between lender type and CRA eligibility. If the estimated value for (cid:11) (cid:21) is negative, as discussed previously, one might be tempted to conclude that relationship lenders subsidize loans to CRA eligibleborrowers. Equation( C ),bycontrast,issuggestedbyourtheory. Itisthe empiricalcounterparttoequation( H 1 ). Inourmodel,relationshipandtransaction lendersreacttoaborrower’sapparentcreditriskdifferently,whichwecaptureempirically by interacting the borrower’s apparent risk with lender type. Depending on the identifying assumption ( I 1 or I 2 ) the estimated interaction term can have varyinginterpretations. One immediate generalization that we apply to equations ( A ), ( B ), and ( C ) is to split the definition of CRA eligibility into its parts. Thus, for example, we would estimatethe coefficients of an alternativeversion of equation ( A ), denoted 36

equation( A 0 ): S i = X i B + (cid:11) ( 1 ) LOWMOD i + (cid:11) ( 2 ) IRLT80 i + (cid:21) RELLEND i + u i : ( A 0 ) In the same way, we would estimate the extra coefficients from alternate specificationsofequations( B )and( C )withexpandeddefinitionsofCRAeligibility;we denotethesespecificationsas ( B 0 )and ( C 0 ). 5.3 Results Before turning to the regression results, it is instructive to consider some conditional sample means. Consider again table 3, which gives the means and standard deviations of the spreads on the mortgages in our final dataset conditional on lender type and borrower income category. For comparison’s sake, the average spread between Freddie Mac’s weekly published benchmark mortgage rate and the prevailing ten-year Treasury rate over the same period was about 1.5 percentage points. Notice that transaction lenders charge lower-income borrowers a higher spread than they charge other borrower types, while relationship lenders charge lower-income borrowers a lower spread than they do other borrowers. For other borrower income groups, though, relationship lenders charge a slightly higher spread than do transaction lenders. Without any further investigation, we might conclude that relationship lenders are increasing spreads on non-CRA-eligibleborrowersslightlyinordertoofferCRA-eligibleborrowersattractive,low,spreads. 37

We present regression coefficients for our six specifications (equations ( A ), ( B ) and ( C ) with both definitions of CRA eligibility) using data from all borrowers in table 6. Table 7 is an adjunct to table 6; it presents the difference in mortgage spreads for all combinations of borrower type (CRA eligible or not) and lender type (relationship or transaction) estimated under specifications ( B ) and ( B 0 ). Figure 9(a) presents the difference in spreads from specification ( C ) conditionalon borrowerand lendertypeand theborrower’sloan-to-valueratio. Tables8and9andfigure9(b)presentthesameconceptsastables6and7and figure9(a),exceptthattheborrowerpoolisrestrictedtohigher-incomeborrowers (borrowerswithincomesat least 120%oftheirMSA’s medianincome). Results From Specifications ( B )and ( B 0 ) Aswediscussedintheprevioussection,specifications( B )and( B 0 )failtocapture the interaction between lender type and the borrower’s credit risk and benefit to homeownership. Nevertheless, these specifications are interesting in their own right,as astartingpointforouranalysis. Notice first from the columns marked ( B ) and ( B 0 ) in tables 6 and 8 that we can reject the hypothesis that there is no interaction effect between lender type and CRA-eligibility, because the estimated interaction coefficients (cid:11) (cid:21) ; (cid:11) ( (cid:21) 1 ) and (cid:11) ( (cid:21) 2 ) arestatisticallydifferentfromzero; moreover,theyarealwaysnegative. This can betaken as evidenceofaCRA subsidy. Tables7and9fleshouttheseresultsabit;theyshowhowspreadschangeconditional on CRA-eligibility status and lender type, relative to non-CRA-eligible 38

borrowers at transaction lenders. Notice that CRA-eligible borrowers at relationshiplenders generally pay thelargest spreads and thatnon-CRA-eligibleborrowersattransactionlendersthelowest. Loansatrelationshiplendersgenerallycarry higherspreads forall borrowertypes; also, loans to CRA-eligibleborrowers generally carry higher spreads at both lender types. We refer to the first difference as the relationship premium and the second as the CRA premium. The estimated interactionterms (cid:11) (cid:21) ; (cid:11) ( (cid:21) 1 ) and (cid:11) ( (cid:21) 2 ) canbedecomposedintovariationsinthesepremiums, as shown in the bottom right-hand corners of tables 7 and 9. Overall, these results suggest a CRA subsidy of around seven basis points; although for borrowers who are both lower-income and purchasing houses in lower-income neighborhoods,thissubsidycouldriseashighas14basispoints. Thisisasignificant amount, equivalent in credit risk terms to dropping the loan’s loan-to-value ratiofrom over100%to under80%. Results From Specifications ( C )and( C 0 ) According to our theory there is an important interaction between lender type and the borrower’s homeownership benefit and credit quality. Higher-risk and higher-benefit borrowers benefit more from the services provided by relationship lenders. If this were true, we would expect that the relationship premium would varysignificantlybyborrowercredit risk. Consider again table 6 and compare the estimatedinteraction terms under the pairs of columns ( B ), ( C ) and ( B 0 ), ( C 0 ). As discussed previously, the results under specifications ( B ) and ( B 0 ) might lead one to conclude that there is a CRA 39

subsidy of about seven or eight basis points. Under the specifications suggested bytheory,however,theestimatedinteractiontermsfallbyatleasthalfinabsolute value. UsingthesimplestdefinitionofCRAeligibility,CRAELIG,theestimated interaction effect drops from more than seven basis points, to just over one basis point. The results displayed in tables 6 and 7 use the entire sample of borrowers. In this sample, identifyingassumption ( I 1 ) seems more appropriate; in other words, the unobserved variation in homeownership benefit is negatively correlated with CRA eligibility. Thus the estimated interaction term (cid:11) (cid:21) is an upper bound on the actual subsidy (if any). Even this upper bound is quite small. This is strong evidenceagainstaCRA subsidy. Table 8 presents the same results as table 6 when we restrict the dataset to containhigher-incomeborrowersonly. Inthiscase,CRAeligibilityisdetermined purely by the median income of the neighborhood where the borrower purchases his house. Our identifying assumption here, ( I 2 ), holds that the benefit to homeownershipdoes notvaryby neighborhoodincome. Under this identifying assumption, the interaction term (cid:11) (cid:21) provides a consistent estimate of any CRA subsidy. However, the intercept term (cid:21) will be biased upward by the effect of the higher benefit on mortgage rates at relationship lenders. Comparing the columns marked ( B 0 ) and ( C 0 ) in table 8, we see that in fact the estimated intercept term, (cid:21) b , does increase significantly under the alternate specification. At the same time, the estimated intercept and interaction terms on 40

CRAeligibility, (cid:11) b and (cid:11) c (cid:21) fall(inabsolutevalue). Infact,undercolumn( C 0 ),CRA eligibilitydoesnothaveastatisticallysignificanteffectonmortgagespreadsatall. Thuswecan reject thehypothesisofaCRA subsidy. Ourtheoryalsomakespredictionsaboutdifferencesinmortgagespreadsconditional on lender type and borrower risk. In particular, from equation ( H 0 ), the difference between the spread paid by two similar borrowers at a bank and at a transactionlenderisgivenby: E (cid:8) S i j (cid:18) ; RELLEND i = 1 ; X (cid:9) (cid:0) E (cid:8) S j j (cid:18) ; RELLEND j = 0 ; X (cid:9) = f (cid:21) ( (cid:18) ) + (cid:13) : (14) Theory predicts that the spread difference between relationship and transaction lenders conditional on credit-worthiness, (cid:18) , should be increasing in (cid:18) . See figure 3 for example. Indeed, for very high-risk borrowers, it is entirely possible that borrowers will pay less at relationship lenders than at transaction lenders; as (cid:18) goesto zero, thespread attransactionlenders goes toinfinity. Infigures9(a)and9(b)wepresenttheempiricalcounterpartsofequation(14). Weusethecoefficientsestimatedunderspecification( C )intable6tocomputethe estimateddifferenceinspreadspaidbythesameborroweratarelationshiplender versusatatransactionlender. Weholdallborrowerqualitiesexceptloan-to-value (LTV) fixed. Higher values of LTV are associated with greater credit risk, and thuslowervaluesof (cid:18) . Notice that low-LTV borrowers (presumably the safest) pay higher mortgage rates at relationship lenders while the highest-LTV borrowers (presumably the 41

riskiest) pay lower spreads. This fits well with the theory developed in section 3; relatively safe borrowers pay more at relationship lenders than at transaction lenders while the very riskiest borrowers pay less. Indeed, the highest LTV borrowerspaylargespreadsattransactionlendersbutnotatrelationshiplenders. The figurealso includesa separatelineforCRA-eligibleborrowers;thislineis nearly indistinguishablefromthelinefornon-CRA-eligibleborrowers,emphasizingthat therelativeimportanceofCRA-eligibilityistiny. Figure 9: Estimated difference in spreads between transaction and relationship lenders,by loan-to-valueratio. Spread: Relationship − Transaction Lender 40 0 −40 −80 −120 70 75 80 85 90 95 100 105 stnioP sisaB Spread: Relationship − Transaction Lender (Higher−Income Sample) 40 0 −40 −80 Not CRA Eligible CRA Eligible −120 70 75 80 85 90 95 100 105 Loan−to−Value (Percent) (a)Allborrowers stnioP sisaB Not CRA Eligible CRA Eligible Loan−to−Value (Percent) (b) Higher-incomeborrowersonly 5.4 Results Using Absolute Income Thresholds As we have discussed, until now we have classified borrowers as higher-income based on the ratio of their income to the median income in the MSA. We used thisapproachbecauseCRA regulatorsuseit;weassumethatlendersareasaware ofthehigher-incomethresholddefined bytheirregulatorastheyareofthelower- 42

incomethreshold. However,onecanimagineseveralreasonsforusinganabsolute incomethresholdto define ahouseholdas higher-income(and thushigh-benefit). First, it fits better with our identifying assumption ( I 2 ); second, we can compute, roughlyspeaking,theprobabilitythataborroweritemizes(andhencereceivesthe MID)conditiononhisorherincome;third,byrestrictingthesampletoborrowers with high incomes, we might be able to control for any unobserved variation in trueborrowerquality. Clearly, some of our controls are endogenous; moreover, the choice of lender type might depend on unobserved borrower-level characteristics. Until now, we have maintained the assumption that, even if there were unobserved variation in borrowercreditquality,thisvariationwasnotsystematicacrosslendertypes. Itis possible, though, that our estimate of (cid:11) c (cid:21) could be biased if lenders observe more about borrowers than we do, and if this unobserved component of credit quality varies systematically with lender type. For example, if among those households with the same apparent credit quality, the relatively lower-quality borrowers got loans from banks while the relatively higher-quality borrowers got loans from transaction lenders, we would tend to overstate the true difference in mortgage spreads. Thisendogenousunobservedheterogeneityinborrowertypeswouldnot, by itself, affect our estimate of the CRA subsidy. However, if this self-selection weregreateramongborrowerseligibleforCRAcredit,ourestimatesoftheofthe CRA subsidywouldbebiasedtowardszero, understatingthetruesubsidy. One way of controlling for unobserved differences in credit quality is to restrict attentionto householdswith highincomes. Amonghouseholdswith annual 43

incomes greater than $100,000,for example, there is likely to beless unobserved heterogeneity than among the general borrower population. Recall also that we include a full set of institution and MSA fixed effects. Thus we do not have to control for idiosyncraticvariation in borrowerqualityacross lenders, only within lenders. Notethatsinceweconcentrateonrelativelyhigh-incomeborrowers,nonewill be eligible for CRA credit purely on the basis of the income test alone (in other words, none will have incomes below 80% of the MSA median income). From examiningfigure8,notethattheprobabilityofitemizing(andhenceclaimingthe MID) rises sharply at annual incomes of between $60 and 80 thousand. Thus we chose a series of income thresholds to bracket this critical region; in particular, we estimated our model after restricting the dataset to only those borrowers with annualincomesabove$60,80, 100or120thousand. Table 10 presents sample statistics for each of the restricted samples. Notice that as borrower incomegrows, the average loan-to-valueand loan-to-incomeratios actually fall slightly,while theloan spread rises slightly. Borrowers’ propensity to use relationship lenders is essentially flat, while the proportion buying housesin lower-incomeareas falls by abouta thirdfrom thelowestto thehighest income threshold. The dataset formed by eliminating those households with annualincomesbelow$60thousandcontainsmorethan4,500loansmadetoborrowers purchasing homes in lower-income neighborhoods, while the dataset formed by eliminating those households with annual incomes below $120,000 contains only about 450 such loans. Thus although we may be controlling for borrower 44

heterogeneitybyincreasingthethreshold,wedosoat thecostofstatisticalprecision. In table 11 we present results from ourprimary specification of interest, ( C 0 ), for each threshold income threshold. The estimated coefficient on the interaction term is always negative, but, again, relatively small. For the dataset formed by eliminating all borrowers with incomes below $100,000, the point estimate indicates a subsidy of about 5.5 basis points. This is larger than the estimate from our other definition of higher-income (in table 8), but still quitesmall in absolute terms. Moreover, it is still smallerthan the potential subsidythat we estimated in thecompletedataset(from table6). 6 Conclusion We studied the effect of the Community Reinvestment Act (CRA) on mortgage lending. We developed a theory of the services provided by banks to borrowers based on their ongoingrelationship with the borrower. In our model, banks compete against firms that provide only one-time financial transactions, in this case mortgages,and thatdo notdeveloparelationshipwiththeborrower. Using this theory, we showed that, controlling for credit risk, borrowers at banksareexpectedtopaymorethanborrowersfromnonbankinstitutionsbecause banks provide a costly extra service (identified as credit enhancement in our theory). Thus, the spread a borrower pays on a mortgage, we showed, depended on theborrower’screditqualityandbenefittohomeownership. Theseresultssuggest 45

that an empirical finding that banks charge higher mortgage rates than nonbanks, but provide lower rates to CRA-eligible (lower-income) borrowers is consistent withboththeviewthatCRAisataxandtheviewthatbanksproviderelationships thathavevaluetotheirborrowers. Our theory thus forced us to confront a serious identification issue as we attemptedto discriminatebetween thesetwo hypotheses. We took two approaches: (1)Weassumedthatthebenefittohomeownershipwasnegativelycorrelatedwith CRA eligibility; and (2) We assumed that this benefit was highest and constant forhigher-incomeborrowersandnotaffectedbythemedianincomeoftheneighborhood. Ourresultsunderassumption(1)can betakenas anupperboundonthe potentialmortgageinterestsubsidyprovokedbytheCRA;whileourresultsunder assumption (2) can be taken as direct evidence of the presence or absence of a subsidy. Underassumption(1),wefoundthatthelargestpotentialsubsidyprovidedby banks in order to attract CRA eligibleloans was tiny, less than six basis points in allspecifications. Under assumption (2), we restricted our sample to higher-income borrowers only; we used both the CRA criteria and a variety of absolute income thresholds to classify borrowers as higher-income. We then compared the mortgage interestratespaidbythesehigher-incomeborrowersatrelationshiplendersandat transactionlenderswhentheydidordidnotbuyahouseinalower-incomeneighborhood. Those higher-income borrowers who bought homes in lower-income neighborhoods using mortgages from banks did not, in general, pay a different 46

mortgageratethanotherhigher-incomeborrowersatbanks. Attheverymost,the potentialCRA-linkedsubsidywaslessthansixbasispoints. Thus,oursecondtest didnotfind compellingevidenceofasubsidy. Our findings are consistent with the view that the CRA does not cause banks toextend mortgageloanswithsubstantiallylowermortgagerates to attract CRAeligible borrowers. However, it may still be the case that the CRA forced banks toinstitutecostlyspeciallendingprogramsorotherwisepayafixedinvestmentin order to make loans to CRA-eligible borrowers. Beyond the effects on mortgage rates thispaperdoesnot addresstheoverallcosts orbenefitsoftheCRA. 47

A Constructing the Base Dataset Ourfinaldataset, asdiscussed in the maintext, contains information onindividual mortgageloans,borrowersandlendersdrawnfromthreeprimarydatasets: HMDA,MIRSand a PMI dataset. Further, we dropped all loans that were not 30-year and fixed-rate, and those loans made by lenders or in MSAs for whom we had relatively few observations. In this section we describe in some detail the precise nature of the statistical matching procedure usedtoconstruct thebasedataset, andtheeffectofourrestrictions onthefinal dataset. A.1 The Statistical Match Our first major step was to match the loan application register (the “LAR”) from the HomeMortgageDisclosureActrecordstotherecordsfromtheFederalHousingFinance Board’s(FHFB)MonthlyInterestRateSurvey(the“MIRS”).Oursecondmajorstepwas tomatchtheserecordstothePrivateMortgageInsurer(the“PMI”)records. Timing ofthe LAR-MIRSmatch Ideally, both theLARandtheMIRSwouldcontain specificloan identifiers (suchasborrowernameorpropertystreetaddress)thatwouldallowprecisematches. Suchidentifiers are not present in either data file. Similarly, it would be desirable if both files contained precisedates withthesamedateconcept (thatis,thesamedefinitionofthedayonwhich the mortgage is complete). Instead, MIRS dates are monthly, and the LAR and MIRS featureslightlydifferentdateconcepts. TheMIRSrecordscontaina“cycledate”variable (coded, e.g., as 9601, 9602, etc for January 1996, February 1996 and so on) rather than an “action date” variable as in the LAR records.20 Information from the institution conducting the MIRS(the FHFB)led us to believe that observations are drawn mostly from the end of the cycle date month (the official documentation says the loans must close in thelastfiveworkingdaysofthemonth). Thus,wematchedMIRSrecordsfromonecycle date to LAR records from the middle of the cycle date month to the middle of the next month. Summary ofprocedure Our procedure can be divided into several parts. First, we prepared the LAR records, then we prepared the MIRS records, then we matched these two datasets. After that, we matchedtothePMIrecords. 20Inthecaseofloanoriginations,theactiondateistheclosingdateoftheloan. Thedategives thedayandmonth. 48

In producing the LAR dataset to prepare to match, we extracted only those HMDA LARrecords that: 1. Wereoriginations 2. Wereforhomepurchases 3. Wereconventional 4. Hadavalidgeocode (state,countyandcensustract) 5. HadavalidMSA(no“non-MSA”records) We then calculated the cycle date for the LAR records (using the month and day of the LAR action date) and determined the ZIP codes for the LAR records. Note that LAR recordsmaysometimesbematchedtomorethanoneZIPcode. We then turned to the MIRS dataset. We discarded only those MIRS records with invalidormissingMSAcodes. Wethenmatchedthetwodatabases: 1. WesortedeachdatabasebyMSA,stateandcycledatetoproduceasetofpotential matches(“couplets”) 2. IftheZIPcodesinacoupletwerenotanexactmatch,wediscarded thecouplet 3. Fortheremainingcouplets, wecalculated thedifference inloanamounts 4. We then sorted the couplets into groups by MSA, state, and ZIP code (the “geocode”)andthecycledate 5. In each geocode-month group, we chose the firstcouplet (that is, the couplet with thesmallestdifference inloanamount)asapotential matchcouplet 6. The next couplet in each group (the one with the second smallest loan amount difference)wasexaminedtoseeifeithersidehadpreviously beenmatched;ifsoit wasdiscarded, ifnot,itwasthenextpotential matchcouplet 7. Allpotentialmatchcoupletswithloanamountdifferencesgreaterthan$2,000were discarded. AttheendofthisprocedurewewereleftwithasetofLAR-MIRSmatchcouplets. Inone sense,thisdatabasealreadycontainsalmostallofthevariablesofinterest. However,many of these mortgages will carry private mortgage insurance. PMIrates are (broadly speaking)setbystateinsurancecommissionsandPMIitselfaffectstheeffectiveinterestrateon themortgagebecausethepaymentsarerolledintothemortgage’sAPR.Abouteightcompanies provide PMI insurance; these companies voluntarily submit a HMDA-like record totheFFIEC,whichwerefertoasthePMIdatabase.21 Weaugmented ourLARS-MIRS coupletmatchdatasetwithafurtherstatisticalmatchagainstthePMIdatabase(thelender nameissuppressed inthePMIdatabasebythePMIcompanies, sowecannotdoanexact 21By2000,amergerhadreducedthenumberofPMIcompaniestoseven. 49

match). BecausethePMIdatabase closelyfollowstheHMDAdatabaseinform,wewere abletomatchonseveralidentifying criteria. To prepare the PMI database for matching, we discarded all records without a valid geocode (MSA, state, county and census tract), that were attached to loans for purposes other than the purchase ofa1-4 family home, orwerefor anaction other than loan origination. To match the resulting PMI database with our LARS-MIRS matched couplets we: 1. SortedthePMIdatabase bygeocode (state, MSA,county andcensus tract)toproducethesamegeocodecategoriesasintheLARS-MIRSdatabase;thisgaveusour setofpotential triplets 2. We then compared the “race or national origin” (RONO) codes from the LARS data and the PMI data; these had to match exactly, producing a winnowed set of potential triplets 3. For each of these potential triplets we calculate the absolute difference in three variablesintheLARSandPMIdata: (a) Loanamount (b) Borrowerincome (c) Actiondate 4. We discarded all potential triplets in which the loan amount or borrower income differedbymorethan$2,000orinwhichtheactiondatedifferedbymorethan150 days 5. Toproduce theAmatch: Wethensorted bygeocode, loanamountdifference, borrowerincomeamountdifference and action date difference; for each geocode group we picked the first triplet asan actual triplet, wethen checked subsequent triplets tosee ifeither side hadalreadymatched,ifsothenthattripletwasdiscarded,ifnot,itbecameanactual triplet 6. Toproduce theBmatch: WethencollectedallunmatchedcoupletsandPMIrecordsandtriedtomatchthem again using slightly different criteria; in particular, we tightened the loan amount differenceto$1,000butwidenedtheincomeamountdifference to$10,000(onthe theorythatdifferingdefinitions of“income”mightbeaffectingourmatch). Intheend,wesimplydiscarded thefewextramatchesgenerated underthe“B”matching procedure. 50

A.2 Evaluating Match Quality One might expect that the differences in loan amount would be zero; by allowing nonzerodifferences, wemaybeadmitting spurious matches. However,boththePMIandthe HMDAreporting guidelines callforloan amountstoberounded tothenearest thousand; potentially, if the loan amount is close to the rounding midpoint, different institutions might report the loan amount differently. Further, the MIRSdata come from avoluntary survey of institutions, who may truncate rather than round, or who may have a slightly different idea of what constitutes the loan balance, e.g. excluding certain fees rolled into theloanbalance. Asmentionedinthetext,weexcludeallobservationsassociatedwithsmallerorinactive lending institutions. We have several reasons for doing so; in particular, because we includeafullsetoflenderandMSAfixedeffects,suchobservations willlikelyhavelittle effectonourcoefficientestimatesanyway,andless-activelendersmaybespreading their fixed costs over fewer loans. However, although we do not know the list of institutions polled by the FHFB in conducting the MIRS, we know that they are relatively large and well-established lenders. By excluding loans from smaller, less-active lenders, we are purging ourdatasetofspurious matches. At each step in the matching procedure, we produced diagnostic variables that can be used to evaluate the quality of the match. In table A.1 we present summary statistics and the empirical CDF of the loan differences from each step of the match. As shown in the table, only about 36% of HMDA-PMI matches had a difference of exactly zero; however, 90% had a loan difference of $500 or less and only about four percent had a loandifferencegreaterthan$1,000(theHMDAroundingpoint). WeallowedtheHMDA- PMIloandifferencetobealittlelarger,becausewewerealsomatchingonraceornational original (which had to be an exact match) and borrower income (which generally had to bewithin$2,000ofeach;seesectionA.1formoreinformation). A.3 Other Types of Mortgages Inthispaper wehave exclusively used fixed-rate mortgages witha30-year amortization; in reality, households use a variety of amortization horizons and rate structures. Among the most popular are 15-year fixed-rate mortgages and certain types of adjustable-rate mortgages(ARMs). Thirty-year fixed-rate mortgages have the considerable advantage of being relatively homogeneous products; the main source of variation among them is the effective rate on the mortgage. However, if CRA-eligible borrowers are steered towards these other typesofmortgages,wemightbemissingtherelevantvariationinmortgageterms. Thisis doubly trueifonelendertyperather thananother favors ARMsor15-year mortgages for lower-incomeborrowers. 51

TableA.1: Statisticalmatch diagnostic: Loanamountdifferences HMDAMatch MIRS PMI Thousands Mean 0 : 2 3 6 3 0 : 1 6 3 2 Median 0 : 1 6 0 0 0 Std. Dev. 0 : 3 2 0 0 0 : 3 6 9 5 Percent Diff = 0 100% 75% 50% HMDA−MIRS Match HMDA−PMI Match 25% 36 84 0 0 1 2 Thousands NOTE. Tableand figuregivestatisticsonthedistributionofthedifferenceinloan amountsfrom theHMDA-MIRSandtheHMDA-PMImatch. Because the FHFB ask about these alternate mortgage products, we can test this proposition directly. Although we exclude ARMs and 15-year mortgages from our final dataset, we can check their relative frequency in the base dataset by lender and borrower type. Table A.2 shows the percent of observations in our base dataset that are 30-yearfixed-ratemortgagesconditionalonborrowertype,neighborhood typeandlender type. The table has several interesting features: First, in all combinations of borrower, neighborhood and lender type, at most 12% of mortgages are not 30-year FRMs. Second, transaction lenders are more likely than relationship lenders to use 30-year FRMs. Third,withineachlendertype,thereisnoappreciabledifferenceinthetreatmentofCRAeligibleborrowers. Thuswearereassured thatwearenotmissinganimportantfeatureof thetreatmentofCRA-eligibleborrowersbystudying 30-year FRMsexclusively. 52

TableA.2: Distributionof30-year, fixed-ratemortgagesin dataset IncomeClass Lender Type Borrower Neighborhood Transaction Relationship Total —Percent— Lower Lower 9 8 : 0 9 9 0 : 5 8 9 1 : 7 1 ? Lower Higher 9 7 : 7 9 9 0 : 0 4 9 1 : 5 2 ? Middle Lower 9 8 : 0 1 8 8 : 5 7 9 0 : 5 3 ? Middle Higher 9 7 : 6 7 8 9 : 3 5 9 1 : 2 6 Higher Lower 9 9 : 8 6 8 8 : 9 8 9 1 : 4 5 ? Higher Higher 9 7 : 6 5 8 8 : 8 7 9 1 : 0 6 Total 9 7 : 7 3 8 9 : 3 1 9 1 : 2 3 CRA-EligibleBorrowers 9 8 : 0 1 8 9 : 9 4 Non CRA-EligibleBorrowers 9 7 : 6 5 8 9 : 0 5 NOTE. Table givespercent of loans from base dataset that are 30-year, fixed-rate byborrowerandneighborhoodincomeclass. Borrowerandneighborhoodincome classes are defined relative to the MSA median income; lower income borrowers andneighborhoodsaredefinedashavingincomeslessthan80%oftheMSA’smedianincome. Middleincomeborrowershaveincomesbetween80%and 120%of the MSA median and high-incomeborrowers have incomes greater than or equal to 120% of the MSA median. All lower-income borrowers, regardless of neighborhood and all borrowers who purchase homes in lower-incomeneighborhoods, regardlessofincome, areeligibleforCRA credit. ? : Loans eligibleforCRA credit. 53

A.4 Results Without PMI Aswediscussed insectionsA.1andA.2,theflagvariableforprivatemortgageinsurance comes from a separate database, and hence a separate matching procedure. As with the primarymatch(betweentheHMDAandtheMIRSrecords), thissecondary matchisalso subject to potential error. Asa robustness check, wepresent our results without the PMI variable. Althoughintheprimaryregressions, thePMIvariableispositiveandsignificant (asexpected), eliminating ithaslittleeffectontheparameters ofinterest. TableA.3: CompleteRegressionResultsforAll BorrowerTypes(withoutPMI) EstimatedCoefficients ByRegressionSpecification Variable A A 0 B B 0 C C 0 Relationship Lender: RELLEND (cid:21) ( 0 0 0: 0: 7 0 2 4 ) ( 0 0 0: 0: 7 0 2 4 ) ( 0 0 0: 0: 8 0 8 4 ) ( 0 0 0: 0: 9 0 0 4 ) ( 0 0 3: 0: 4 1 8 4 ) ( 0 0 3: 0: 5 1 0 4 ) CRAEligible: CRAELIG (cid:11) ( 0 0 0: 0: 5 0 1 2 ) ( 0 0 1: 0: 0 0 8 4 ) ( 0 0 0: 0: 7 0 6 4 ) Interaction term: CRAELIG (cid:2) RELLEND (cid:11) (cid:21) (cid:0) ( 0 0 0: 0: 7 0 2 5 ) (cid:0) ( 0 0 0: 0: 1 0 8 5 ) CRAEligible: LOWMOD (cid:11) (1) ( 0 0 0: 0: 1 0 2 3 ) ( 0 0 0: 0: 6 0 3 8 ) ( 0 0 0: 0: 4 0 4 8 ) CRAEligible: IRLT80 (cid:11) (2) ( 0 0 0: 0: 5 0 1 2 ) ( 0 0 1: 0: 1 0 4 5 ) ( 0 0 0: 0: 8 0 1 5 ) Interaction term: LOWMOD (cid:2) RELLEND continuedonnextpage 54

TableA.3(continuedfrompreviouspage) EstimatedCoefficients ByRegression Specification (withoutPMI) Variable (cid:11) A A 0 B B 0 C C 0 (1) (cid:21) (cid:0) ( 0 0 0: 0: 6 0 2 8 ) (cid:0) ( 0 0 0: 0: 3 0 6 8 ) Interaction term: IRLT80 (cid:2) RELLEND (cid:11) (2) (cid:21) (cid:0) ( 0 0 0: 0: 7 0 8 5 ) (cid:0) ( 0 0 0: 0: 2 0 2 6 ) RiskVariables Loan-to-value ratio 7 8 9 9 (cid:23) 5 1 1 6 (cid:21) (cid:20) (cid:20) (cid:20) (cid:20) (cid:23) (cid:23) (cid:23) (cid:23) < < < < 8 9 9 9 1 1 6 9 (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 (cid:23) 0: 0: 0: 0: 0: 0: 0: 0: 2 0 3 0 6 0 4 0 4 2 4 3 1 3 7 4 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 0: 0: 0: 0: 2 0 3 0 6 0 4 0 3 2 5 3 2 3 7 4 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 0: 0: 0: 0: 2 0 3 0 6 0 4 0 4 2 5 3 1 3 8 4 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 0: 0: 0: 0: 2 0 3 0 6 0 4 0 3 2 5 3 2 3 9 4 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 1: 0: 2: 0: 0 0 7 0 7 0 7 1 3 5 8 6 0 5 0 1 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 1: 0: 2: 0: 0 0 8 0 7 0 6 1 2 5 0 6 1 6 9 1 ) ) ) ) 99 (cid:0) ( 0 0 0: 0: 4 0 7 8 ) (cid:0) ( 0 0 0: 0: 4 0 6 8 ) (cid:0) ( 0 0 0: 0: 4 0 3 8 ) (cid:0) ( 0 0 0: 0: 4 0 1 8 ) ( 1 0 3: 0: 9 2 6 ) ( 1 0 2: 0: 4 2 6 ) Loan-to-income ratio ‘ ‘ 2 = 1 0 0 (cid:0) ( ( 0 0 0 0 0: 0: 4: 0: 6 0 6 1 ‘ 7 1 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 4: 0: 6 0 7 1 8 1 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 4: 0: 6 0 7 1 8 1 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 4: 0: 6 0 7 1 9 1 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 2: 0: 3 0 1 1 2 2 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 2: 0: 3 0 2 1 4 2 ) ) FANFRED (cid:0) ( 0 0 0: 0: 0 0 6 3 ) (cid:0) ( 0 0 0: 0: 0 0 6 3 ) (cid:0) ( 0 0 0: 0: 0 0 7 3 ) (cid:0) ( 0 0 0: 0: 0 0 7 3 ) (cid:0) ( 0 0 1: 0: 3 0 6 8 ) (cid:0) ( 0 0 1: 0: 3 0 6 8 ) SOLDOTH (cid:0) ( 0 0 0: 0: 9 0 5 4 ) (cid:0) ( 0 0 0: 0: 9 0 6 4 ) (cid:0) ( 0 0 0: 0: 9 0 1 4 ) (cid:0) ( 0 0 0: 0: 9 0 2 4 ) ( 0 0 0: 0: 3 1 6 0 ) ( 0 0 0: 0: 3 1 6 0 ) CONFORMSIZE continuedonnextpage 55

TableA.3(continuedfrompreviouspage) EstimatedCoefficients ByRegression Specification (withoutPMI) Variable (cid:0) ( 0 0 A 1: 0: 7 0 9 4 ) (cid:0) ( 0 0 A 1: 0: 0 7 0 9 4 ) (cid:0) ( 0 0 B 1: 0: 7 0 8 4 ) (cid:0) ( 0 0 B 1: 0: 0 7 0 8 4 ) ( 0 0 C 0: 0: 0 0 1 9 ) (cid:0) ( 0 0 C 0: 0: 0 0 0 0 9 ) PMI n.a. n.a. n.a. n.a. n.a. n.a. PCTVAC ( 0 0 0: 0: 1 0 3 5 ) ( 0 0 0: 0: 1 0 6 5 ) ( 0 0 0: 0: 1 0 3 5 ) ( 0 0 0: 0: 1 0 6 5 ) ( 0 0 0: 0: 3 1 3 1 ) ( 0 0 0: 0: 3 1 4 1 ) RiskVariableInteraction Terms Loan-to-value ratio 7 8 9 9 (cid:23) 5 1 1 6 (cid:21) (cid:20) (cid:20) (cid:20) (cid:20) (cid:23) (cid:23) (cid:23) (cid:23) < < < < 8 9 9 9 1 1 6 9 (cid:23) (cid:0) (cid:0) (cid:0) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 1: 0: 2: 0: 2 0 5 0 3 0 5 1 3 6 1 7 7 6 9 2 ) ) ) ) (cid:0) (cid:0) (cid:0) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 1: 0: 2: 0: 2 0 5 0 3 0 5 1 3 6 1 7 7 6 6 2 ) ) ) ) 99 (cid:0) ( 1 0 4: 0: 0 2 3 7 ) (cid:0) ( 1 0 4: 0: 0 2 1 7 ) Loan-to-income ratio ‘ ‘ 2 = 1 0 0 ‘ (cid:0) ( ( 0 0 0 0 0: 0: 8: 0: 5 0 0 3 8 2 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 8: 0: 5 0 1 3 9 2 ) ) FANFRED ( 0 0 1: 0: 4 0 4 9 ) ( 0 0 1: 0: 4 0 3 9 ) SOLDOTH (cid:0) ( 0 0 1: 0: 6 1 8 1 ) (cid:0) ( 0 0 1: 0: 6 1 9 1 ) CONFORMSIZE (cid:0) ( 0 0 1: 0: 9 1 4 0 ) (cid:0) ( 0 0 1: 0: 9 1 3 0 ) PMI n.a. n.a. continuedonnextpage 56

TableA.3(continuedfrompreviouspage) EstimatedCoefficients ByRegressionSpecification (withoutPMI) Variable A A 0 B B 0 C C 0 PCTVAC (cid:0) ( 0 0 0: 0: 2 1 4 2 ) (cid:0) ( 0 0 0: 0: 2 1 1 2 ) OtherControlVariables Intercept ( 2 0 4: 0: 5 1 ) ( 2 0 4: 0: 5 1 ) ( 2 0 4: 0: 4 1 ) ( 2 0 4: 0: 4 1 ) ( 2 0 2: 0: 4 1 ) ( 2 0 2: 0: 4 1 ) CLOANS (cid:0) ( 0 0 0: 0: 7 0 1 7 ) (cid:0) ( 0 0 0: 0: 7 0 5 7 ) (cid:0) ( 0 0 0: 0: 7 0 2 7 ) (cid:0) ( 0 0 0: 0: 7 0 5 7 ) (cid:0) ( 0 0 0: 0: 6 0 8 7 ) (cid:0) ( 0 0 0: 0: 7 0 1 7 ) LENDERS ( 0 0 1: 0: 6 2 5 4 ) ( 0 0 1: 0: 7 2 6 4 ) ( 0 0 1: 0: 6 2 7 4 ) ( 0 0 1: 0: 7 2 7 4 ) ( 0 0 1: 0: 4 2 9 4 ) ( 0 0 1: 0: 6 2 0 4 ) YearDummies YR95 (cid:0) ( 0 0 5: 0: 0 0 7 5 ) (cid:0) ( 0 0 5: 0: 0 0 8 5 ) (cid:0) ( 0 0 5: 0: 0 0 6 5 ) (cid:0) ( 0 0 5: 0: 0 0 7 5 ) (cid:0) ( 0 0 5: 0: 1 0 1 5 ) (cid:0) ( 0 0 5: 0: 1 0 2 5 ) YR96 (cid:0) ( 0 0 6: 0: 6 0 2 5 ) (cid:0) ( 0 0 6: 0: 6 0 3 5 ) (cid:0) ( 0 0 6: 0: 6 0 1 5 ) (cid:0) ( 0 0 6: 0: 6 0 2 5 ) (cid:0) ( 0 0 6: 0: 6 0 2 5 ) (cid:0) ( 0 0 6: 0: 6 0 3 5 ) YR97 (cid:0) ( 0 0 6: 0: 6 0 5 5 ) (cid:0) ( 0 0 6: 0: 6 0 6 5 ) (cid:0) ( 0 0 6: 0: 6 0 5 5 ) (cid:0) ( 0 0 6: 0: 6 0 6 5 ) (cid:0) ( 0 0 6: 0: 6 0 9 5 ) (cid:0) ( 0 0 6: 0: 7 0 0 5 ) YR98 (cid:0) ( 0 0 3: 0: 6 0 8 4 ) (cid:0) ( 0 0 3: 0: 6 0 8 4 ) (cid:0) ( 0 0 3: 0: 6 0 7 4 ) (cid:0) ( 0 0 3: 0: 6 0 8 4 ) (cid:0) ( 0 0 3: 0: 7 0 4 4 ) (cid:0) ( 0 0 3: 0: 7 0 5 4 ) YR99 (cid:0) ( 0 0 3: 0: 6 0 6 3 ) (cid:0) ( 0 0 3: 0: 6 0 6 3 ) (cid:0) ( 0 0 3: 0: 6 0 6 3 ) (cid:0) ( 0 0 3: 0: 6 0 6 3 ) (cid:0) ( 0 0 3: 0: 6 0 7 3 ) (cid:0) ( 0 0 3: 0: 6 0 7 3 ) POSTRUSSIA ( 0 0 2: 0: 3 0 8 2 ) ( 0 0 2: 0: 3 0 8 2 ) ( 0 0 2: 0: 3 0 8 2 ) ( 0 0 2: 0: 3 0 8 2 ) ( 0 0 2: 0: 3 0 7 2 ) ( 0 0 2: 0: 3 0 7 2 ) F-Value 102.08 102.20 101.57 101.57 95.78 95.96 Pr > F R 2 < 0 0 0: 3: 0 2 1 < 0 0 0: 3: 0 2 1 < 0 0 0: 3: 0 3 1 < 0 0 0: 3: 0 3 1 < 0 0 0: 3: 0 4 1 < 0 0 0: 3: 0 4 1 SampleSize 250,593 NOTE. Table gives selected regression coefficients and standard errors (in parentheses) from OLS regressions of the spread on a mortgage against the specified variables; see discussion in text and equations ( A ), ( B ) and ( C ) for more details. In these regressions, the PMI variable was not used. All regressions contained a full set of MSA and lending institution dummies; the F-statistic from a test of the hypothesis that all dummies were jointlyequaltozeroisshown. 57

B Probability of Itemizing Tax Deductions We used the 1995 and 1998 waves of the SCF to calculate the probability of itemizing tax deductions conditional on outstanding mortgage debt and income. We computed the (weighted) sample probabilities and estimated a probit model of itemization. Table B.1 presentsweightedsamplestatistics;notethatweeliminatedhouseholdsinwhichthehead wasyoungerthan25orolderthan61. Inaddition, weeliminated households withannual reported totalincomesthatwerenegative, zerooraboveonemilliondollars (inreal1996 dollars). In our remaining sample, about one third of all households reported itemizing their deductions. However, this conceals significant variation. We split the sample into three parts conditional on the household’s ratio of mortgage debt to income. First, those households with zero mortgage debt (and hence a ratio of zero) and then, among those households withpositivemortgagedebt, thoseaboveandbelowthemedianratioof1.03. Asshown in table B.1, among those with zero debt, the itemization rate is about 12 percent, while among those households with positive mortgage debt, the itemization rate is around50percent. In table B.2 we present the (weighted) tax deduction itemization rate conditional on income range and mortgage debt. Although the itemization rate rises with income for allhouseholds, itrises sharpest forthose withpositive mortgage debt. Beyond $60 or80 thousandtheitemizationratelargelylevelsoffamonghouseholdswithmortgages,whileit continues toriseamongthosewithout. About80%ofhouseholds withpositivemortgage debt and incomes greater than $80,000 itemize their deductions (although this rate does bouncearoundabit). A parametric approach is to estimate a probit model of the probability of itemizing. Because we are particularly interested in how this probability reacts to income, we include ageneral specification ofincome (a setof dummies aswellas themore traditional log income). The coefficient estimates are shown in B.3. Figure 8 presents the results graphically; inthefigureallvariablesexceptincomearesettotheirconditionalmeansfor each income range. The results indicate that, as in the non-parametric approach, above about$60–80thousandperyearofincome,theprobabilityofitemizinglargelylevelsoff. 58

TableB.1: SCF SampleStatistics MortgageDebt toIncomeRatio Median Zero Below Above Variable MortgageDebta ( 4 4 3 7 : : 1 6 6 9 ) ( 9 8 9 4 : : 3 5 3 4 ) Income a ( 3 4 5 3 : : 4 8 0 7 ) ( 8 9 4 3 : : 2 8 8 8 ) ( 5 4 4 0 : : 1 4 5 8 ) Ageb ( 4 1 0 0 : : 7 4 1 4 ) 4 ( 4 8 : : 7 7 3 0 ) 4 ( 1 9 : : 2 1 5 2 ) —Percent— Itemize 1 2 : 4 2 5 3 : 3 3 5 2 : 3 2 Second Mortgage 1 : 8 1 2 : 3 7 1 2 : 7 2 HighSchool Degree 5 6 : 0 4 5 4 : 8 4 5 2 : 8 7 CollegeDegree 2 4 : 5 9 3 4 : 6 4 3 8 : 8 8 1995SCF 4 9 : 3 1 4 8 : 3 6 4 7 : 0 8 Observations 2,649 1,489 1,486 NOTE. Table gives weighted sample statisticsfrom the combined 1995 and 1998 waves of the Survey of Consumer Finances. Sample is broken into three parts conditionalonthemortgage-to-incomeratio: Thosewithzeromortgagedebt(and hence a ratio of zero) and, among thosewith positivemortgagedebt, thoseabove and thosebelowthemedian of1.03. aInthousandsofreal1996dollars. bAgeofhouseholdhead. 59

TableB.2: ConditionalProbabilityofItemizingDeductions MortgageDebt toIncomeRatio Median Zero Below Above IncomeRange Income (cid:20) 20............... 0 : 0 2 0 : 0 0 0 : 1 4 20 < Income (cid:20) 40.......... 0 : 0 9 0 : 2 5 0 : 3 7 40 < Income (cid:20) 60.......... 0 : 2 1 0 : 4 4 0 : 5 2 60 < Income (cid:20) 80.......... 0 : 2 8 0 : 6 6 0 : 7 9 80 < Income (cid:20) 100......... 0 : 4 0 0 : 7 4 0 : 8 4 100 < Income (cid:20) 120 ....... 0 : 5 4 0 : 6 1 0 : 7 7 Income > 120.............. 0 : 6 1 0 : 8 7 0 : 8 1 NOTE. Table gives weighted incidence of tax itemization conditional on income range and the ratio of mortgagedebt to income; data are from the 1995 and 1998 wavesoftheSCF. 60

TableB.3: ProbitResults ofTax ItemizationModel EstimatedCoefficient Variable IncomeClassDummyVariables 20 < Income (cid:20) 40 ........................... ( 0 0 1: 1: 7 2 ) 40 < Income (cid:20) 60 ........................... ( 0 0 3: 1: 2 6 ) 60 < Income (cid:20) 80 ........................... ( 0 0 6: 1: 1 9 ) 80 < Income (cid:20) 100.......................... ( 0 0 6: 2: 8 2 ) 100 < Income (cid:20) 120......................... ( 0 0 3: 2: 7 6 ) Income > 120 ............................... ( 0 0 4: 3: 7 0 ) Logincome.................................. ( 0 0 6: 1: 6 1 ) ZeroMortgageDebt.......................... (cid:0) ( 0 0 7: 0: 3 5 ) SecondMortgagePresent..................... ( 0 0 3: 0: 7 7 ) (MortgageDebtLevel)/(100,000).............. ( 0 0 1: 0: 1 4 ) Age......................................... ( 0 0 0: 0: 3 2 ) Age 2 ........................................ (cid:0) ( 0 0 0: 0: 2 2 ) HighestDiplomaAchieved HighSchoolDegree.......................... (cid:0) ( 0 0 1: 0: 6 7 ) continuedonnextpage 61

TableB.3(continuedfrompreviouspage) EstimatedCoefficient Variable CollegeDegree .............................. (cid:0) ( 0 0 2: 0: 3 8 ) Observationfrom1995SCF................... ( 0 0 0: 0: 3 4 ) NOTE. Table gives coefficient estimates and standard errors for a probit regression. The dependent variable wasanindicator settounityifthehousehold headitemizedhisorher deductions intheprevioustaxyear. 62

References Acemoglu, D. and J. D. Angrist (2001). Consequences of employment protection? The case of the Americans with DisabilitiesAct. Journal of Political Economy109(5),915–57. Avery,R., R. Bostic, and G. B. Canner (2002). AssessingtheCRA’s necessity and efficiency.Manuscript,Federal Reserve Board, WashingtonDC. Avery, R. B., R. W. Bostic, and G. B. Canner (2000). CRA special lending programs.FederalReserveBulletin86(11), 711–31. Board of Governors of the Federal Reserve System (1993). Report to the Congress on Community Development Lending by DepositoryInstitutions. Washington,DC: Federal Reserve Board. Board of Governors of the Federal Reserve System (2000).Report on the PerformanceandProfitabilityofCRA-related Lending.Washington,DC: Federal Reserve Board. Bolton, P. and X. Freixas (2000). Equity, bonds, and bank debt: Capital structureandfinancialmarket equilibriumunderasymmetricinformation.Journalof PoliticalEconomy108(2),324–51. Boot,A.W.A.andA.V.Thakor(1997).Financialsystemarchitecture.Review of FinancialStudies10(3), 693–733. Boot, A. W. A. and A. V. Thakor (2000). Can relationship banking survive competition? The JournalofFinance55(2), 679–713. Brady, P., J.-A. Cronin, and S. Houser (2001). Regional differences in the utilizationofthemortgageinterestdeduction.OfficeofTaxAnalysisWorking Paper88, U.S. DepartmentoftheTreasury. Canner, G.B.andW.Passmore(1995a).Creditriskandtheprovisionofmortgages to lower-income and minority homebuyers. Federal Reserve Bulletin81(11), 989–1016. Canner, G. B. and W. Passmore (1995b). Home purchase lending in lowincomeneighborhoodsandtolow-incomeborrowers.FederalReserveBulletin81(2), 71–103. Canner, G. B. and W. Passmore (1996). The relative profitability of commercial banks active in lending in lower-income neighborhoods and to lowerincomeborrowers.InProceedingsofthe32ndAnnualConferenceonBank Structureand Competition.Federal ReserveBank ofChicago. 63

Canner, G. B., W. Passmore, and B. Surette (1996). Distribution of credit risk among providers of mortgages to lower-incomeand minority homebuyers. FederalReserveBulletin82(12), 1077–1101. Carey,M.,M.Post, andS. A.Sharpe(1998). Doescorporatelendingbybanks and finance companies differ? Evidence on specialization in private debt contracting.The JournalofFinance53(3), 845–78. Chemmanur, T. J. and P. Fulghieri (1994). Reputation, renegotiation, and the choice between bank loans and publicly traded debt. Review of Financial Studies7(3),475–506. Evanoff, D. D. and L. M. Segal (1996). CRA and fair lending regulations: Resulting trends in mortgage lending. Federal Reserve Bank of Chicago EconomicPerspectives20(6), 19–46. Evanoff,D. D.andL. M.Segal(1997).Strategicresponsestobankregulation: Evidence from HMDA data. Journal of FinancialServices Research 11(1- 2), 69–93. Garwood, G. L. and D. S. Smith (1993). The community reinvestment act: Evolutionand current issues.FederalReserveBulletin79(4),251–67. Harvey,K. D., M. C. Collins,P. Nigro, and B. Robinson (2001). Disparitiesin mortgage lending, bank performance, economic influence, and regulatory oversight.Journalof RealEstateFinanceandEconomics23, 379–410. Henderson, J. V. and Y. M. Ioannides (1983). A model of housing tenure choice. AmericanEconomicReview73(1), 98–113. Lehnert, A. and D. M. Maki (2002). Consumption, debt, and portfolio choice: Testing the effects of bankruptcy law. Manuscript, Federal Reserve Board ofGovernors,Washington,DC. Levitt,S.D.(1997).Usingelectoralcyclesinpolicehiringtoestimatetheeffect ofpoliceoncrime. AmericanEconomicReview87(3),270–290. Levitt,S.D.(1998).Juvenilecrimeandpunishment.JournalofPoliticalEconomy106(6), 1156–1185. Malmquist,D.,F.Phillips-Patrick,andC.Rossi(1997).Theeconomicsoflowincomemortgagelending.JournalofFinancialServicesResearch11(1-2), 169–88. 64

Meeker, L. and F. Myers (1996). Community Reinvestment Act Lending: Is it profitable? In Financial Industry Perspectives 1996, pp. 13–35. Kansas City,MO:Federal ReserveBank ofKansas City. Pence, K.(2001).Foreclosingonopportunity? Statelawsandmortgagecredit. Manuscript,Federal ReserveBoard, WashingtonDC. Sharpe, S. A. (1997). The effect of consumer switching costs on prices: A theory and its application to the bank deposit market. Review of Industrial Organization12(1),79–94. Zinman, J. (2002). The causes and real effects of credit constraints: Evidence from the community reinvestment act. Manuscript, M.I.T. Department of Economics,CambridgeMA. 65

Table1: ControlVariableNames and Definitions Name Definition RiskVariables w LTVBTW7581 Loan tovalueratio ( (cid:23) ) between75% and 80% w LTVBTW8191 Loan tovalueratio ( (cid:23) ) between81% and 90% w LTVBTW9196 Loan tovalueratio ( (cid:23) ) between91% and 95% w LTVBTW9699 Loan tovalueratio ( (cid:23) ) between96% and 98% w LTVGTEQ99 Loan tovalueratio ( (cid:23) ) greaterthanorequal to99% w LOANINC Loan toincomeratio (alsodenoted ‘ ) w LOANINCSQ Loan toincomeratio squared ( ‘ 2 ) w FANFRED Loan soldtoFannieMaeorFreddie Mac w SOLDOTH Loan soldtoanothersecuritizer w CONFORMSIZE Loan sizeis withintheconformingloanlimit w PMI Loan carries privatemortgageinsurance w PCTVAC Syntheticvariablecombininginformationonpercent ofdwellingsvacant andpercent boardedup. OtherControlVariables z CLOANS Number of conventional home purchase loans made in the same calendar year and in the same census tract, divided by the number of 1-4 family units in thetract z LENDERS Numberofdifferentlendersmakingloansinthetract in the calendar year, divided by the number of 1-4 familyunitsinthetract z YRxx Year dummies z NAMExxx Institutiondummies z MSAxxx MSA dummies NOTE. Tablegivesnames,definitions,andcategoriesofthestandardcontrolvariables used in the regressions. A designation of w indicates that the variable is associatedwithaborrower’screditrisk,whileadesignationof z indicatesthatthe variableisassociated withotherfactors affecting mortgagespreads. 66

Table2: SampleMeansofSelected Variables LenderType Variable All Relationship Transaction Mortgagespread ( 1 0 : : 8 5 5 1 ) ( 1 0 : : 8 5 5 2 ) ( 1 0 : : 8 4 7 9 ) LoanAmounta 1 ( 2 6 6 6 : : 4 0 3 9 ) 1 ( 2 6 4 5 : : 2 1 2 3 ) 1 ( 3 6 4 8 : : 0 7 1 2 ) Loan-to-valueratio( (cid:23) ) ( 8 1 2 4 : : 0 5 5 1 ) ( 8 1 1 4 : : 9 7 3 8 ) ( 8 1 2 3 : : 4 5 7 4 ) Loan-to-incomeratio( ‘ ) ( 2 1 : : 1 2 5 4 ) ( 2 0 : : 1 8 4 7 ) ( 2 2 : : 0 0 6 6 ) —Percent— FANFRED 7 0 : 5 6 7 : 1 8 2 : 4 SOLDOTH 1 1 : 8 1 1 : 5 1 2 : 7 CONFORMSIZE 9 3 : 8 9 4 : 3 9 1 : 9 PMI 3 6 : 3 3 9 : 2 2 6 : 2 PCTVAC 3 : 1 3 : 3 2 : 5 CRAELIG 2 8 : 0 2 9 : 5 2 2 : 6 LOWMOD 8 : 1 8 : 6 6 : 1 IRLT80 2 3 : 9 2 5 : 4 1 8 : 8 Observations 250,593 193,827 56,766 NOTE. Table gives means and standard deviations (in parentheses) of selected variables for the indicated subsets of the data. Relationship lenders are defined as commercial banks and savings institutions; transaction lenders are defined as independentmortgagebankers. aInthousandsofreal,1996,dollars. 67

Table 3: Mean and Standard Deviation of Mortgage Spread Conditional On BorrowerIncomeand LenderType Lender Type Relationship Transaction All Lower-IncomeBorrowers: Income < 80% ofMSA Median Spread Mean 1 : 7 8 1 8 1 : 8 9 6 3 1 : 8 0 2 3 Spread Std. Dev. ( 0 : 5 9 5 0 ) ( 0 : 5 1 4 2 ) ( 0 : 5 8 3 0 ) Observations 49,174 10,700 59,874 Medium-IncomeBorrowers: 80% (cid:20) Income < 120%ofMSA Median Spread Mean 1 : 8 5 8 4 1 : 8 5 6 7 1 : 8 5 8 0 Spread Std. Dev. ( 0 : 4 9 3 7 ) ( 0 : 4 8 6 1 ) ( 0 : 4 9 2 0 ) Observations 55,198 16,086 71,284 Higher-IncomeBorrowers: Income (cid:21) 120% ofMSAMedian Spread Mean 1 : 8 7 8 8 1 : 8 6 8 0 1 : 8 7 6 1 Spread Std. Dev. ( 0 : 4 8 6 1 ) ( 0 : 4 8 5 0 ) ( 0 : 4 8 5 8 ) Observations 89,455 29,980 119,435 AllIncomeCategories Spread Mean 1 : 8 4 8 4 1 : 8 7 0 2 1 : 8 5 3 3 Spread Std. Dev. ( 0 : 5 1 9 4 ) ( 0 : 4 9 1 1 ) ( 0 : 5 1 3 3 ) Observations 193,827 56,766 250,593 NOTE. Table gives means and standard deviations of the spreads (effective rate minus prevailing ten-year Treasury rate) on mortgages conditional on borrower income and lender type, as well as the number of observations in each cell. Note thatloanstolower-incomeborrowersarealwayseligibleforCRA credit; loansto other types of borrowers are only eligible for CRA credit if the purchased home isin alower-incomeneighborhood. 68

Table4: Borrowerincomeconditionalonclass Borrower incomeclass Mean Median All borrowerincomeclasses .......................... ( 6 5 5 4 : : 2 0 1 7 ) 5 6 : 3 1 Lower: income (cid:20) 80% ofMSAmedian)................ 2 ( 9 8 : : 5 1 8 9 ) 2 9 : 5 9 Middle: 80% < income < 120%ofMSA median....... 4 ( 9 9 : : 1 5 1 0 ) 4 8 : 0 4 Higher: income (cid:21) 120% ofMSAmedian .............. ( 9 6 2 7 : : 6 0 9 9 ) 8 0 : 0 0 NOTE. Tablegivesmean and medianincomes(in thousandsofreal 1996dollars) ofborrowersinthedatasetbyincomeclass;standarddeviationsareinparentheses. Table5: Loan AmountConditionalon Lenderand BorrowerType LenderType Borrowerincomeclass All Relationship Transaction All............................... 1 ( 2 6 6 6 : : 4 0 3 9 ) 1 ( 2 6 4 5 : : 2 1 2 3 ) 1 ( 3 6 4 8 : : 0 7 1 2 ) Lower............................ ( 7 3 6 1 : : 3 6 9 5 ) ( 7 3 5 0 : : 3 9 0 8 ) ( 8 3 1 4 : : 4 1 0 1 ) Middle........................... 1 ( 1 4 0 0 : : 5 4 3 7 ) 1 ( 0 4 9 0 : : 8 4 8 3 ) 1 ( 1 4 2 0 : : 7 5 8 5 ) Higher ........................... 1 ( 6 7 1 1 : : 0 2 1 9 ) 1 ( 5 7 9 0 : : 9 2 5 5 ) 1 ( 6 7 4 4 : : 1 2 8 0 ) NOTE. Table gives means and standard deviations of the mortgage amount in thousands of real 1996 dollars conditional on lender type and borrower income class. 69

Table6: CompleteRegressionResults forAllBorrowerTypes EstimatedCoefficients ByRegressionSpecification Variable A A 0 B B 0 C C 0 Relationship Lender: RELLEND (cid:21) ( 0 0 0: 0: 7 0 0 4 ) ( 0 0 : : 0 0 7 0 0 4 ) ( 0 0 0: 0: 8 0 5 4 ) ( 0 0 0: 0: 8 0 8 4 ) ( 0 0 3: 0: 4 1 7 4 ) ( 0 0 3: 0: 4 1 9 4 ) CRAEligible: CRAELIG (cid:11) ( 0 0 0: 0: 5 0 1 2 ) ( 0 0 1: 0: 0 0 8 4 ) ( 0 0 0: 0: 7 0 6 4 ) Interaction term: CRAELIG (cid:2) RELLEND (cid:11) (cid:21) (cid:0) ( 0 0 0: 0: 7 0 2 5 ) (cid:0) ( 0 0 0: 0: 1 0 8 5 ) CRAEligible: LOWMOD (cid:11) (1) ( 0 0 : : 0 0 1 0 2 3 ) ( 0 0 0: 0: 6 0 3 8 ) ( 0 0 0: 0: 4 0 5 8 ) CRAEligible: IRLT80 (cid:11) (2) ( 0 0 : : 0 0 5 0 1 2 ) ( 0 0 1: 0: 1 0 4 5 ) ( 0 0 0: 0: 8 0 1 5 ) Interaction term: LOWMOD (cid:2) RELLEND (cid:11) (1) (cid:21) (cid:0) ( 0 0 0: 0: 6 0 2 8 ) (cid:0) ( 0 0 0: 0: 3 0 6 8 ) Interaction term: IRLT80 (cid:2) RELLEND (cid:11) (2) (cid:21) (cid:0) ( 0 0 0: 0: 7 0 8 5 ) (cid:0) ( 0 0 0: 0: 2 0 2 6 ) continuedonnextpage 70

Table6(continuedfrompreviouspage) EstimatedCoefficients ByRegressionSpecification Variable A A 0 B B 0 C C 0 RiskVariables Loan-to-value ratio 7 8 9 9 (cid:23) 5 1 1 6 (cid:21) (cid:20) (cid:20) (cid:20) (cid:20) (cid:23) (cid:23) (cid:23) (cid:23) < < < < 8 9 9 9 1 1 6 9 (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 (cid:23) 0: 0: 0: 0: 0: 0: 0: 0: 2 0 3 0 5 0 4 0 4 2 0 3 6 3 2 4 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 0: 0: 0: 0: 2 0 3 0 5 0 4 0 4 2 1 3 7 3 3 4 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 0: 0: 0: 0: 2 0 3 0 5 0 4 0 4 2 0 3 7 3 3 4 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 0: 0: 0: 0: 2 0 3 0 5 0 4 0 3 2 1 3 8 3 5 4 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 1: 0: 2: 0: 0 0 7 0 6 0 6 1 3 5 3 6 5 6 5 1 ) ) ) ) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 1: 0: 2: 0: 0 0 7 0 6 0 6 1 2 5 5 6 7 6 4 1 ) ) ) ) 99 (cid:0) ( 0 0 0: 0: 4 0 9 8 ) (cid:0) ( 0 0 0: 0: 4 0 8 8 ) (cid:0) ( 0 0 0: 0: 4 0 6 8 ) (cid:0) ( 0 0 0: 0: 4 0 3 8 ) ( 1 0 2: 0: 4 2 6 ) ( 1 0 2: 0: 4 2 6 ) Loan-to-income ratio ‘ ‘ 2 = 1 0 0 (cid:0) ( ( 0 0 0 0 0: 0: 4: 0: 6 0 6 1 ‘ 7 1 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 4: 0: 6 0 7 1 8 1 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 4: 0: 6 0 7 1 8 1 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 4: 0: 6 0 7 1 9 1 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 2: 0: 3 0 1 1 2 2 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 2: 0: 3 0 2 1 4 2 ) ) FANFRED (cid:0) ( 0 0 0: 0: 0 0 7 3 ) (cid:0) ( 0 0 0: 0: 0 0 7 3 ) (cid:0) ( 0 0 0: 0: 0 0 8 3 ) (cid:0) ( 0 0 0: 0: 0 0 8 3 ) (cid:0) ( 0 0 1: 0: 3 0 8 8 ) (cid:0) ( 0 0 1: 0: 3 0 7 8 ) SOLDOTH (cid:0) ( 0 0 0: 0: 9 0 5 4 ) (cid:0) ( 0 0 0: 0: 9 0 6 4 ) (cid:0) ( 0 0 0: 0: 9 0 2 4 ) (cid:0) ( 0 0 0: 0: 9 0 2 4 ) ( 0 0 0: 0: 3 1 4 0 ) ( 0 0 0: 0: 3 1 4 0 ) CONFORMSIZE (cid:0) ( 0 0 1: 0: 8 0 0 4 ) (cid:0) ( 0 0 1: 0: 8 0 0 4 ) (cid:0) ( 0 0 1: 0: 7 0 9 4 ) (cid:0) ( 0 0 1: 0: 7 0 9 4 ) (cid:0) ( 0 0 0: 0: 0 0 0 9 ) (cid:0) ( 0 0 0: 0: 0 0 1 9 ) PMI ( 0 0 0: 0: 1 0 0 2 ) ( 0 0 0: 0: 1 0 0 2 ) ( 0 0 0: 0: 1 0 0 2 ) ( 0 0 0: 0: 1 0 0 2 ) ( 0 0 0: 0: 1 0 5 4 ) ( 0 0 0: 0: 1 0 5 4 ) PCTVAC ( 0 0 0: 0: 1 0 3 5 ) ( 0 0 0: 0: 1 0 6 5 ) ( 0 0 0: 0: 1 0 3 5 ) ( 0 0 0: 0: 1 0 6 5 ) ( 0 0 0: 0: 3 1 4 1 ) ( 0 0 0: 0: 3 1 5 1 ) continuedonnextpage 71

Table6(continuedfrompreviouspage) EstimatedCoefficients ByRegressionSpecification Variable A A 0 B B 0 C C 0 RiskVariableInteraction Terms Loan-to-value ratio 7 8 9 9 (cid:23) 5 1 1 6 (cid:21) (cid:20) (cid:20) (cid:20) (cid:20) (cid:23) (cid:23) (cid:23) (cid:23) < < < < 8 9 9 9 1 1 6 9 (cid:23) (cid:0) (cid:0) (cid:0) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 1: 0: 2: 0: 2 0 5 0 3 0 6 1 2 6 2 7 9 6 1 2 ) ) ) ) (cid:0) (cid:0) (cid:0) (cid:0) ( ( ( ( 0 0 0 0 0 0 0 0 0: 0: 0: 0: 1: 0: 2: 0: 2 0 5 0 3 0 5 1 3 6 2 7 9 6 8 2 ) ) ) ) 99 (cid:0) ( 1 0 4: 0: 0 2 4 7 ) (cid:0) ( 1 0 4: 0: 0 2 2 7 ) Loan-to-income ratio ‘ ‘ 2 = 1 0 0 ‘ (cid:0) ( ( 0 0 0 0 0: 0: 8: 0: 5 0 1 3 9 2 ) ) (cid:0) ( ( 0 0 0 0 0: 0: 8: 0: 5 0 1 3 9 2 ) ) FANFRED ( 0 0 1: 0: 4 0 3 9 ) ( 0 0 1: 0: 4 0 2 9 ) SOLDOTH (cid:0) ( 0 0 1: 0: 6 1 7 1 ) (cid:0) ( 0 0 1: 0: 6 1 8 1 ) CONFORMSIZE (cid:0) ( 0 0 1: 0: 9 1 4 0 ) (cid:0) ( 0 0 1: 0: 9 1 3 0 ) PMI ( 0 0 0: 0: 0 0 2 5 ) (cid:0) ( 0 0 0: 0: 0 0 2 5 ) PCTVAC (cid:0) ( 0 0 0: 0: 2 1 4 2 ) (cid:0) ( 0 0 0: 0: 2 1 1 2 ) OtherControlVariables Intercept ( 2 0 4: 0: 5 1 ) ( 2 0 4: 0: 5 1 ) ( 2 0 4: 0: 4 1 ) ( 2 0 4: 0: 4 1 ) ( 2 0 2: 0: 4 1 ) ( 2 0 2: 0: 4 1 ) CLOANS (cid:0) ( 0 0 0: 0: 7 0 3 7 ) (cid:0) ( 0 0 0: 0: 7 0 6 7 ) (cid:0) ( 0 0 0: 0: 7 0 4 7 ) (cid:0) ( 0 0 0: 0: 7 0 7 7 ) (cid:0) ( 0 0 0: 0: 7 0 1 7 ) (cid:0) ( 0 0 0: 0: 7 0 4 7 ) LENDERS ( 0 0 1: 0: 7 2 0 4 ) ( 0 0 1: 0: 8 2 1 4 ) ( 0 0 1: 0: 7 2 1 4 ) ( 0 0 1: 0: 8 2 2 4 ) ( 0 0 1: 0: 5 2 7 4 ) ( 0 0 1: 0: 6 2 8 4 ) continuedonnextpage 72

Table6(continuedfrompreviouspage) EstimatedCoefficients ByRegressionSpecification Variable A A 0 B B 0 C C 0 YearDummies YR95 (cid:0) ( 0 0 5: 0: 0 0 7 5 ) (cid:0) ( 0 0 5: 0: 0 0 8 5 ) (cid:0) ( 0 0 5: 0: 0 0 6 5 ) (cid:0) ( 0 0 5: 0: 0 0 7 5 ) (cid:0) ( 0 0 5: 0: 1 0 1 5 ) (cid:0) ( 0 0 5: 0: 1 0 2 5 ) YR96 (cid:0) ( 0 0 6: 0: 6 0 1 5 ) (cid:0) ( 0 0 6: 0: 6 0 2 5 ) (cid:0) ( 0 0 6: 0: 6 0 1 5 ) (cid:0) ( 0 0 6: 0: 6 0 2 5 ) (cid:0) ( 0 0 6: 0: 6 0 1 5 ) (cid:0) ( 0 0 6: 0: 6 0 3 5 ) YR97 (cid:0) ( 0 0 6: 0: 6 0 5 5 ) (cid:0) ( 0 0 6: 0: 6 0 6 5 ) (cid:0) ( 0 0 6: 0: 6 0 5 5 ) (cid:0) ( 0 0 6: 0: 6 0 5 5 ) (cid:0) ( 0 0 6: 0: 6 0 9 5 ) (cid:0) ( 0 0 6: 0: 7 0 0 5 ) YR98 (cid:0) ( 0 0 3: 0: 6 0 8 4 ) (cid:0) ( 0 0 3: 0: 6 0 8 4 ) (cid:0) ( 0 0 3: 0: 6 0 8 4 ) (cid:0) ( 0 0 3: 0: 6 0 8 4 ) (cid:0) ( 0 0 3: 0: 7 0 4 4 ) (cid:0) ( 0 0 3: 0: 7 0 5 4 ) YR99 (cid:0) ( 0 0 3: 0: 6 0 6 3 ) (cid:0) ( 0 0 3: 0: 6 0 6 3 ) (cid:0) ( 0 0 3: 0: 6 0 6 3 ) (cid:0) ( 0 0 3: 0: 6 0 6 3 ) (cid:0) ( 0 0 3: 0: 6 0 7 3 ) (cid:0) ( 0 0 3: 0: 6 0 7 3 ) POSTRUSSIA ( 0 0 2: 0: 3 0 8 2 ) ( 0 0 2: 0: 3 0 8 2 ) ( 0 0 2: 0: 3 0 8 2 ) ( 0 0 2: 0: 3 0 8 2 ) ( 0 0 2: 0: 3 0 8 2 ) ( 0 0 2: 0: 3 0 8 2 ) F-Value 102.08 102.20 101.57 101.57 95.78 95.96 Pr > F R 2 < 0 0 0: 3: 0 2 1 < 0 0 0: 3: 0 2 1 < 0 0 0: 3: 0 3 1 < 0 0 0: 3: 0 3 1 < 0 0 0: 3: 0 4 1 < 0 0 0: 3: 0 4 1 SampleSize 250,593 NOTE. Table gives selected regression coefficients and standard errors (in parentheses) from OLS regressions of the spread on a mortgage against the specified variables; see discussion in text and equations ( A ), ( B ) and ( C ) for more details. All regressions contained a full set of MSA and lending institutiondummies; the F-statisticfromatestofthehypothesisthatalldummieswerejointlyequaltozero isshown. 73

Table7: EffectiveInteractionofLenderand BorrowerTypes fromTable6 LenderType Specification Relationship Transaction RelationshipPremium CRA eligible ( B ) CRAELIG=1 ( B 0 ) 1 2 : 2 1 0 : 8 1 : 4 LOWMOD=1 ( B 0 ) 9 : 0 6 : 3 2 : 7 IRLT80=1 1 2 : 5 1 1 : 4 1 : 1 Not CRA eligible ( B ) CRAELIG=0 ( B 0 ) 8 : 6 0 8 : 6 LOWMOD=0 ( B 0 ) 8 : 9 0 8 : 9 IRLT80=0 8 : 9 0 8 : 9 CRA Premium 3 : 6 0 : 1 3 : 6 1 0 : 8 6 : 3 1 1 : 4 (cid:0) 7 : 2 (cid:0) 6 : 2 (cid:0) 7 : 8 NOTE. Table gives the difference in mortgage spreads (in basis points) of each combination of CRA-eligibility status and lender type relative to non-CRA-eligible borrowers at transaction lenders (the excluded category in the regressions). 74

Table8: CompleteRegressionResults ForHigherIncomeBorrowers Only RegressionSpecification Variable A 0 B 0 C 0 RelationshipLender: RELLEND (cid:21) ( : : 0 0 8 0 9 4 0 9 6 8 ) ( : : 0 0 9 0 0 5 2 0 4 0 ) ( : : 3 0 0 2 0 2 3 7 9 8 ) CRA Eligibility:LOWMOD (cid:11) ( 1 ) (cid:0) ( : : 0 0 1 0 0 6 9 0 8 1 ) ( : : 0 0 1 1 6 2 3 5 4 4 ) ( : : 0 0 0 1 5 2 4 5 7 5 ) InteractionTerm: LOWMOD (cid:2) RELLEND (cid:11) ( (cid:21) 1 ) (cid:0) ( : : 0 0 3 1 5 4 1 1 2 5 ) (cid:0) ( : : 0 0 1 1 1 4 6 2 1 0 ) RiskVariables Loan-to-valueratio 7 8 9 9 (cid:23) 5 1 1 6 (cid:21) (cid:20) (cid:20) (cid:20) (cid:20) 9 (cid:23) (cid:23) (cid:23) (cid:23) 9 < < < < 8 9 9 9 1 1 6 8 (cid:23) ( ( ( ( ( : : : : : : : : : : 0 0 0 0 1 0 1 0 4 0 0 0 8 0 2 0 7 0 1 1 3 3 1 4 4 3 8 8 3 8 1 2 4 0 5 9 6 1 2 3 1 8 6 4 4 0 9 7 2 3 ) ) ) ) ) ( ( ( ( ( : : : : : : : : : : 0 0 0 0 1 0 1 0 4 0 0 0 8 0 2 0 7 0 1 1 3 3 1 4 4 3 8 8 4 8 0 2 4 0 4 9 9 1 1 3 9 8 3 4 7 0 5 7 6 4 ) ) ) ) ) ( ( ( ( 1 ( : : : : : : : : : : 0 0 1 0 2 0 4 0 6 0 2 0 0 0 0 0 2 2 8 3 0 6 7 7 3 7 5 0 3 7 3 7 2 7 4 4 9 6 1 6 8 7 1 8 1 2 3 6 7 0 ) ) ) ) ) Loan-to-incomeratio ‘ ‘ (cid:0) ( : : 3 0 2 0 2 7 5 7 ) (cid:0) ( : : 3 0 2 0 2 7 3 6 ) (cid:0) ( : : 2 0 3 1 4 6 8 1 ) continuedonnextpage 75

Table8(continuedfrompreviouspage) RegressionSpecification Variable ‘ 2 ( : : 0 0 A 5 0 0 8 1 1 9 ) ( : : 0 0 B 5 0 0 8 1 1 9 ) ( : : C 0 0 4 0 0 1 4 9 0 ) FANFRED (cid:0) ( : : 0 0 7 0 9 3 7 9 ) (cid:0) ( : : 0 0 7 0 9 3 9 9 ) (cid:0) ( : : 1 0 6 1 0 0 0 5 ) SOLDOTH ( : : 0 0 6 0 2 5 3 5 ) ( : : 0 0 6 0 2 5 4 5 ) ( : : 0 0 4 1 7 2 9 3 ) CONFORMSIZE (cid:0) ( : : 0 0 8 0 0 4 4 7 ) (cid:0) ( : : 0 0 8 0 0 4 3 7 ) ( : : 0 0 0 1 3 0 0 8 ) PMI ( : : 0 0 0 0 8 2 9 8 5 9 ) ( : : 0 0 0 0 8 2 9 8 8 9 ) ( : : 0 0 3 0 0 5 5 9 8 3 ) PCTVAC ( : : 0 0 1 0 5 7 0 0 5 1 ) ( : : 0 0 1 0 4 7 9 0 7 1 ) ( : : 0 0 3 1 3 4 9 5 1 5 ) RiskVariableInteractionTerms Loan-to-valueratio 7 8 9 9 (cid:23) 5 1 1 6 (cid:21) (cid:20) (cid:20) (cid:20) (cid:20) 9 (cid:23) (cid:23) (cid:23) (cid:23) 9 < < < < 8 9 9 9 1 1 6 9 (cid:23) (cid:0) (cid:0) (cid:0) (cid:0) (cid:0) ( ( ( ( 1 ( : : : : : : : : : : 0 0 0 0 1 0 2 0 6 0 2 0 3 0 0 0 9 2 7 4 1 7 3 9 7 8 5 2 3 2 1 7 1 0 9 6 1 4 9 9 5 1 4 5 6 5 5 4 4 5 ) ) ) ) ) Loan-to-incomeratio ‘ ‘ 2 ‘ (cid:0) ( ( : : : : 1 0 0 0 1 1 2 0 2 8 0 4 2 0 7 5 6 9 1 0 ) ) continuedonnextpage 76

Table8(continuedfrompreviouspage) RegressionSpecification Variable A 0 B 0 C 0 FANFRED ( : : 0 0 8 1 8 1 9 1 7 8 ) SOLDOTH ( : : 0 0 0 1 1 3 9 7 8 6 ) CONFORMSIZE (cid:0) ( : : 1 0 0 1 3 1 1 8 0 4 ) PMI (cid:0) ( : : 0 0 2 0 2 6 0 7 3 7 ) PCTVAC (cid:0) ( : : 0 0 2 1 2 6 4 2 7 3 ) OtherControlVariables Intercept 2 ( : : 5 0 6 1 8 1 5 8 6 9 ) 2 ( : : 5 0 6 1 7 1 4 9 7 0 ) 2 ( : : 4 0 1 2 4 1 8 4 7 9 ) CLOANS (cid:0) ( : : 0 0 3 0 3 8 2 4 7 5 ) (cid:0) ( : : 0 0 3 0 3 8 3 4 6 5 ) (cid:0) ( : : 0 0 3 0 3 8 7 3 1 8 ) LENDERS ( : : 0 0 8 3 4 0 2 6 3 0 ) ( : : 0 0 8 3 4 0 8 6 9 0 ) ( : : 0 0 8 3 1 0 9 3 6 2 ) YR95 (cid:0) ( : : 4 0 5 0 5 7 7 7 8 2 ) (cid:0) ( : : 4 0 5 0 5 7 7 7 0 2 ) (cid:0) ( : : 4 0 5 0 9 7 4 6 1 5 ) YR96 (cid:0) ( : : 5 0 7 0 5 6 6 4 8 2 ) (cid:0) ( : : 5 0 7 0 5 6 5 4 5 2 ) ( (cid:0) (cid:0) : : 5 0 7 0 8 6 7 3 8 7 ) YR97 (cid:0) ( : : 6 0 3 0 0 6 1 8 9 3 ) (cid:0) ( : : 6 0 3 0 0 6 1 8 4 3 ) (cid:0) ( : : 6 0 3 0 4 6 0 7 1 8 ) YR98 (cid:0) ( : : 3 0 5 0 7 4 4 5 4 8 ) (cid:0) ( : : 3 0 5 0 7 4 3 5 7 8 ) (cid:0) ( : : 3 0 6 0 1 4 5 5 4 5 ) YR99 (cid:0) ( : : 3 0 6 0 1 4 9 1 4 1 ) (cid:0) ( : : 3 0 6 0 1 4 8 1 5 1 ) (cid:0) ( : : 3 0 6 0 3 4 7 0 6 8 ) POSTRUSSIA ( : : 2 0 2 0 0 2 3 8 0 9 ) ( : : 2 0 2 0 0 2 2 8 9 9 ) ( : : 2 0 2 0 0 2 2 8 6 6 ) continuedonnextpage 77

Table8(continuedfrompreviouspage) RegressionSpecification Variable A 0 B 0 C 0 F-Value 44.52 44.53 44.18 Pr > F R 2 < 0 0 : : 0 3 0 5 0 1 < 0 0 : : 0 3 0 5 0 1 < 0 0 : : 0 3 0 6 0 1 SampleSize 119,435 NOTE. Table gives selected regression coefficients and standard errors (in parentheses) from OLS regressions of the spread on a mortgage against the specified variables; see discussion in text and equation ( A 0 ) for more details. Here only higher-income borrowers (borrowers with incomes above 120% of the MSA median income) are included in the regression; CRA eligibility then stems purely from the borrower’s neighborhood; if the borrower purchases a home in a lowerincome neighborhood (whenLOWMOD = 1 ), the mortgage is eligible for CRA credit. All regressions contained a full set of MSA and lending institution dummies; the F-statistic from a test of the hypothesis that all dummies were jointly equalto zero is shown. 78

Table9: EffectiveInteraction ofLender andBorrowerTypes fromTable8 LenderType Specification Relationship Transaction RelationshipPremium CRA eligible ( B 0 ) LOWMOD 7 : 2 1 : 7 5 : 5 Not CRA eligible ( B 0 ) LOWMOD 9 : 0 0 9 : 0 CRA Premium (cid:0) 1 : 8 1 : 7 (cid:0) 3 : 5 NOTE. Table gives the difference in mortgage spreads (in basis points) of each combination of CRA-eligibility status and lender type relative to non-CRAeligible borrowers at transaction lenders (the excluded category in the regressions). Herethedatasetislimitedtohigh-incomeborrowers,definedasborrowers withincomesgreaterthan orequalto 120%oftheMSAmedian income. 79

Table10: Restricted SampleStatisticsUsingIncomeThresholds IncomeThreshold(thousands) Variable 60 80 100 120 MortgageSpread ( 1 0 : : 8 4 7 9 ) ( 1 0 : : 8 4 8 9 ) ( 1 0 : : 9 4 0 9 ) ( 1 0 : : 9 4 1 8 ) Loan Amount 1 ( 7 7 3 3 : : 8 4 0 5 ) 2 ( 0 8 0 2 : : 1 7 9 4 ) 2 ( 2 9 5 3 : : 0 8 8 8 ) ( 2 1 4 0 8 4 : : 6 7 2 2 ) Loan-to-ValueRatio ( 8 1 1 3 : : 3 7 8 1 ) ( 7 1 9 3 : : 9 7 0 9 ) ( 7 1 8 3 : : 4 8 9 6 ) ( 7 1 7 3 : : 4 7 3 2 ) Loan-to-IncomeRatio ( 1 0 : : 8 6 7 4 ) ( 1 0 : : 7 6 2 2 ) ( 1 0 : : 6 6 0 3 ) ( 1 0 : : 5 6 0 3 ) —Percent— RelationshipLender 7 4 : 9 3 7 4 : 4 3 7 4 : 1 9 7 4 : 0 3 LOWMOD 4 : 0 0 3 : 3 8 2 : 9 8 2 : 7 8 FANFRED 7 0 : 8 9 6 4 : 4 4 5 6 : 3 4 4 8 : 8 1 SOLDOTH 1 2 : 4 3 1 7 : 7 8 2 4 : 2 4 3 0 : 1 3 CONFORMSIZE 8 6 : 6 2 7 7 : 0 0 6 5 : 6 4 5 5 : 0 6 PMI 3 0 : 9 6 2 3 : 7 2 1 7 : 3 6 1 2 : 8 7 Observations 114,280 60,363 32,190 17,810 NOTE. Tablegivessamplestatisticsforsubsamplesofourprimarydataset;households were included only if their reported annual income exceeded the indicated threshold(in thousandsofreal 1996dollars). 80

Table11: ResultsUsingAbsoluteIncomeThresholds IncomeThreshold(thousands) Variable 60 80 100 120 RelationshipLender: RELLEND (cid:21) ( 0 0 : : 3 0 1 2 7 3 ) ( 0 0 : : 3 0 4 2 9 9 ) ( 0 0 : : 2 0 5 3 6 7 ) ( 0 0 : : 1 0 1 4 4 8 ) CRA Eligibility:LOWMOD (cid:11) ( 1 ) ( 0 0 : : 0 0 0 1 4 3 ) ( 0 0 : : 0 0 3 1 1 9 ) ( 0 0 : : 0 0 5 2 0 7 ) ( 0 0 : : 0 0 1 3 5 6 ) InteractionTerm: LOWMOD (cid:2) RELLEND (cid:11) ( (cid:21) 1 ) (cid:0) ( 0 0 : : 0 0 0 1 9 5 ) (cid:0) ( 0 0 : : 0 0 3 2 3 1 ) (cid:0) ( 0 0 : : 0 0 5 3 2 0 ) (cid:0) ( 0 0 : : 0 0 1 4 0 2 ) RiskVariables Loan-to-valueratio 7 8 9 9 (cid:23) 5 1 1 6 (cid:21) (cid:20) (cid:20) (cid:20) (cid:20) 9 (cid:23) (cid:23) (cid:23) (cid:23) 9 < < < < 8 9 9 9 1 1 6 8 (cid:23) ( ( ( ( ( 0 0 0 0 0 0 0 0 1 0 : : : : : : : : : : 0 0 1 0 2 0 3 0 6 0 2 0 1 0 1 0 9 2 3 3 4 7 5 8 4 8 4 2 3 9 ) ) ) ) ) ( ( ( ( ( 0 0 0 0 0 0 0 0 1 0 : : : : : : : : : : 0 0 1 0 2 0 4 0 8 0 1 0 1 1 3 1 4 3 0 5 9 9 2 1 2 1 9 7 1 0 ) ) ) ) ) ( ( ( ( ( 0 0 0 0 0 0 0 0 1 0 : : : : : : : : : : 0 0 1 0 2 0 4 0 7 0 0 1 1 1 4 1 0 5 5 7 5 2 2 4 1 5 0 9 3 5 ) ) ) ) ) ( ( ( ( ( 0 0 0 0 0 0 0 0 1 0 : : : : : : : : : : 0 0 1 0 2 0 6 1 8 1 2 1 1 1 7 2 5 1 2 1 4 5 3 9 1 0 8 1 8 1 ) ) ) ) ) Loan-to-incomeratio ‘ ‘ (cid:0) ( 0 0 : : 2 0 3 1 8 6 ) (cid:0) ( 0 0 : : 2 0 0 2 1 0 ) (cid:0) ( 0 0 : : 2 0 4 2 0 6 ) (cid:0) ( 0 0 : : 2 0 8 3 6 6 ) continuedonnextpage 81

Table11(continuedfrompreviouspage) IncomeThreshold(thousands) Variable 60 80 100 120 ‘ 2 ( 0 0 : : 0 0 4 0 3 4 ) ( 0 0 : : 0 0 3 0 4 5 ) ( 0 0 : : 0 0 4 0 5 7 ) ( 0 0 : : 0 0 5 1 9 0 ) FANFRED (cid:0) ( 0 0 : : 1 0 6 1 0 1 ) (cid:0) ( 0 0 : : 1 0 7 1 3 4 ) (cid:0) ( 0 0 : : 1 0 6 2 3 0 ) (cid:0) ( 0 0 : : 1 0 5 2 2 7 ) SOLDOTH ( 0 0 : : 0 0 4 1 9 2 ) ( 0 0 : : 0 0 2 1 6 5 ) ( 0 0 : : 0 0 4 1 3 8 ) ( 0 0 : : 0 0 2 2 3 2 ) CONFORMSIZE ( 0 0 : : 0 0 0 1 4 1 ) (cid:0) ( 0 0 : : 0 0 1 1 7 3 ) (cid:0) ( 0 0 : : 0 0 3 1 0 8 ) (cid:0) ( 0 0 : : 0 0 6 2 7 4 ) PMI ( 0 0 : : 0 0 2 0 8 6 ) ( 0 0 : : 0 0 3 0 1 9 ) ( 0 0 : : 0 0 3 1 7 4 ) ( 0 0 : : 0 0 5 2 9 0 ) PCTVAC ( 0 0 : : 0 0 3 1 6 6 ) ( 0 0 : : 0 0 2 2 5 0 ) ( 0 0 : : 0 0 1 2 0 6 ) (cid:0) ( 0 0 : : 0 0 2 3 0 3 ) RiskVariableInteractionTerms Loan-to-valueratio 7 8 9 9 (cid:23) 5 1 1 6 (cid:21) (cid:20) (cid:20) (cid:20) (cid:20) 9 (cid:23) (cid:23) (cid:23) (cid:23) 9 < < < < 8 9 9 9 1 1 6 9 (cid:23) (cid:0) (cid:0) (cid:0) (cid:0) (cid:0) ( ( ( ( ( 0 0 0 0 0 0 0 0 1 0 : : : : : : : : : : 0 0 0 0 1 0 2 0 5 0 2 0 3 0 0 0 7 2 8 4 5 8 5 9 9 9 2 4 8 6 ) ) ) ) ) (cid:0) (cid:0) (cid:0) (cid:0) (cid:0) ( ( ( ( ( 0 0 0 0 0 0 0 0 1 0 : : : : : : : : : : 0 0 0 0 1 0 3 0 6 0 1 1 2 1 0 1 3 4 2 6 8 0 4 2 7 2 4 1 5 3 ) ) ) ) ) (cid:0) (cid:0) (cid:0) (cid:0) ( ( ( ( ( 0 0 0 0 0 0 0 0 1 0 : : : : : : : : : : 0 0 0 0 0 0 2 0 5 0 0 1 1 1 9 1 7 6 5 9 8 3 2 7 6 8 5 5 7 4 ) ) ) ) ) (cid:0) (cid:0) (cid:0) (cid:0) (cid:0) ( ( ( ( ( 0 0 0 0 0 0 0 0 1 0 : : : : : : : : : : 0 0 0 0 0 0 5 1 8 1 1 1 0 2 9 2 1 2 0 3 1 7 2 3 2 5 2 0 4 9 ) ) ) ) ) Loan-to-incomeratio ‘ ‘ 2 ‘ (cid:0) ( ( 0 0 0 0 : : : : 1 0 0 0 1 1 2 0 9 8 2 5 ) ) (cid:0) ( ( 0 0 0 0 : : : : 2 0 0 0 0 2 5 0 4 3 1 6 ) ) (cid:0) ( ( 0 0 0 0 : : : : 1 0 0 0 2 3 3 0 5 0 6 8 ) ) ( ( 0 0 0 0 : : : : 0 0 0 0 1 4 0 1 2 1 5 1 ) ) continuedonnextpage 82

Table11(continuedfrompreviouspage) IncomeThreshold(thousands) Variable 60 80 100 120 FANFRED ( 0 0 : : 0 0 8 1 6 2 ) ( 0 0 : : 0 0 7 1 6 5 ) ( 0 0 : : 0 0 5 2 5 1 ) ( 0 0 : : 0 0 4 2 1 8 ) SOLDOTH ( 0 0 : : 0 0 0 1 5 4 ) ( 0 0 : : 0 0 1 1 1 7 ) (cid:0) ( 0 0 : : 0 0 2 2 7 0 ) (cid:0) ( 0 0 : : 0 0 0 2 2 5 ) CONFORMSIZE (cid:0) ( 0 0 : : 1 0 0 1 1 2 ) (cid:0) ( 0 0 : : 0 0 8 1 8 5 ) (cid:0) ( 0 0 : : 0 0 8 1 3 9 ) (cid:0) ( 0 0 : : 0 0 4 2 7 6 ) PMI (cid:0) ( 0 0 : : 0 0 2 0 1 7 ) (cid:0) ( 0 0 : : 0 0 3 1 3 0 ) (cid:0) ( 0 0 : : 0 0 4 1 0 6 ) (cid:0) ( 0 0 : : 0 0 6 2 7 3 ) PCTVAC (cid:0) ( 0 0 : : 0 0 1 1 6 7 ) (cid:0) ( 0 0 : : 0 0 1 2 0 3 ) ( 0 0 : : 0 0 1 2 7 9 ) ( 0 0 : : 0 0 5 3 1 7 ) OtherControlVariables Intercept ( 2 0 : : 4 0 0 2 5 2 ) ( 2 0 : : 3 0 8 2 5 7 ) ( 2 0 : : 4 0 2 3 2 5 ) ( 2 0 : : 4 0 4 4 7 5 ) CLOANS (cid:0) ( 0 0 : : 0 0 3 0 3 9 ) (cid:0) ( 0 0 : : 0 0 2 1 2 2 ) (cid:0) ( 0 0 : : 0 0 2 1 4 5 ) (cid:0) ( 0 0 : : 0 0 2 2 0 0 ) LENDERS ( 0 0 : : 0 0 8 3 8 1 ) ( 0 0 : : 0 0 7 4 6 3 ) ( 0 0 : : 0 0 9 5 4 4 ) ( 0 0 : : 0 0 7 6 0 5 ) YR95 (cid:0) ( 0 0 : : 4 0 5 0 4 8 ) (cid:0) ( 0 0 : : 4 0 6 1 4 1 ) (cid:0) ( 0 0 : : 4 0 8 1 1 5 ) (cid:0) ( 0 0 : : 5 0 0 2 2 1 ) YR96 (cid:0) ( 0 0 : : 5 0 7 0 7 7 ) (cid:0) ( 0 0 : : 5 0 8 0 0 9 ) (cid:0) ( 0 0 : : 5 0 8 1 1 3 ) (cid:0) ( 0 0 : : 5 0 9 1 2 7 ) YR97 (cid:0) ( 0 0 : : 6 0 3 0 3 7 ) (cid:0) ( 0 0 : : 6 0 4 1 3 0 ) (cid:0) ( 0 0 : : 6 0 6 1 4 3 ) (cid:0) ( 0 0 : : 6 0 8 1 4 8 ) YR98 (cid:0) ( 0 0 : : 3 0 5 0 8 5 ) (cid:0) ( 0 0 : : 3 0 4 0 9 6 ) (cid:0) ( 0 0 : : 3 0 4 0 6 9 ) (cid:0) ( 0 0 : : 3 0 4 1 0 1 ) YR99 (cid:0) ( 0 0 : : 3 0 6 0 3 4 ) (cid:0) ( 0 0 : : 3 0 7 0 3 6 ) (cid:0) ( 0 0 : : 3 0 7 0 8 8 ) (cid:0) ( 0 0 : : 3 0 9 1 6 0 ) POSTRUSSIA ( 0 0 : : 2 0 2 0 1 3 ) ( 0 0 : : 2 0 1 0 5 4 ) ( 0 0 : : 2 0 1 0 3 5 ) ( 0 0 : : 2 0 1 0 5 7 ) F-Value 42.84 25.61 15.62 9.86 continuedonnextpage 83

Table11(continuedfrompreviouspage) IncomeThreshold(thousands) Variable 60 80 100 120 Pr > F R 2 < 0 : : 0 3 0 6 1 < 0 : : 0 3 0 8 1 < 0 : : 0 3 0 8 1 < 0 : : 0 4 0 0 1 SampleSize 114,280 60,360 32,190 17,810 NOTE. Table gives selected regression coefficients and standard errors (in parentheses) from OLS regressions of the spread on a mortgage against the specified variables; see discussion in text and equation ( C 0 ) for more details. Here only thoseborrowerswithreportedannualincomes(inreal1996dollars)exceedingthe indicated thresholds were included in the regression; CRA eligibility then stems purely from the borrower’s neighborhood; if the borrower purchases a home in a lower-income neighborhood (when LOWMOD = 1 ), the mortgage is eligible for CRA credit. All regressions contained a full set of MSA and lending institution dummies; the F-statistic from a test of the hypothesis that all dummies were jointlyequal tozero is shown. 84