Liquidity Problems and Early Payment Default Among Subprime Mortgages

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Liquidity Problems and Early Payment Default Among Subprime Mortgages Nathan B. Anderson and Jane K. Dokko 2011-09 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Liquidity Problems and Early Payment Default Among Subprime Mortgages Nathan B. Anderson and Jane K. Dokko ∗ November 22, 2010 Abstract The lack of property tax escrow accounts among subprime mortgages causes borrowers to make large lump-sum tax payments that reduce liquidity. Different property tax collection datesacrossstatesandcountiescreateexogenousvariationinthetimebetweenloanorigination and the first property tax due date, affording the opportunity to estimate the causal effect of loan-level exposure to liquidity reductions on mortgage default. We find that a nine-month delay in owing property taxes reduces the probability of first-year default by about 4 percent, oraboutone-thirdoftheeffectofareductioninequityfrom10%tonegative20%. ∗WewouldliketothankJoshuaMiller,ColinMotley,andMichaelMulhallforinvaluableresearchassistance. We aregratefultoRandyReback,FernandoFerreira,HuiShan,AndreasLehnert,DavidAlbouy,RobertKaestner,seminar participants at the Federal Reserve Board, the University of Illinois-Chicago, the Harris School at the University of Chicago, Cornell University, University of California-Santa Cruz, and conference participants at the National Tax Association annual conference and the EEA and AEA annual meetings for providing helpful comments. The views expressed in this paper are those of the authors and do not reflect the opinions of the Federal Reserve Board or the FederalReserveSystem. 1

1 Introduction Highratesofearlypaymentdefault(EPD)amongsubprimemortgages,whichiswhenaborrower defaults in the first year of mortgage origination, triggered large financial losses among many subprimelendersandcontributedtothelargestfinancialcrisissincetheGreatDepression(Mayer, Pence&Sherlund(2009)). Theliteratureonmortgagedefaultpresentstworeasonswhyborrowers default: illiquiditythatlimitsahousehold’sabilitytomakemortgagepaymentsandnegativeequity that leaves households unwilling to pay even though they may be able. In this paper, we provide evidencethatliquidityconstraintsamongsubprimeborrowerscontributedgreatlytothehighEPD rates. While establishing and quantifying the relative importance of illiquidity is a topic of vigorous debate, prior work is limited by the inability to observe exogenous, loan-level differences in liquidity.1 Acommonstrategyfoundintheliteratureistouseproxiesforliquiditydifferencesamong borrowers, such as the unemployment or credit card delinquency rate in a borrower’s county, the divorce rate in a borrower’s state, or credit card utilization rates (Elul, Souleles, Chomsisengphet & Glennon (2010)).2 The problem with this approach is that aggregate proxies like unemployment rates and credit card delinquency rates are often correlated with unobserved determinants of default, such as borrower-specific default costs or expectations of capital gains. For example, borrowers in high unemployment counties may be more likely to default because they expect lower future capital gains than borrowers in counties with low unemployment. Credit card utilization rates face a similar limitation: borrowers with higher discount rates may also have higher credit card utilization and thus value future capital gains less, which may lead to default, irrespective of borrowers’ illiquidity. Thus, the endogeneity of the available proxy variables for between-loan differences in borrowers’ illiquidity prevents the identification of the causal effect of loan-level liquiditydifferencesonmortgagedefault.3 We use local property tax due dates to observe plausibly exogenous and anticipated loan-level reductions in borrowers’ liquidity. The prolonged absence of property tax escrow accounts in the subprimemortgagemarketensuresthatpropertytaxduedatesrepresentlargefinancialobligations for these borrowers.4 In 2007, among housing units with mortgages, the median annual property tax payment was $2,099, which was 140% of the median monthly housing cost and 2.9% of the medianannualhouseholdincome.5 AsdiscussedinCabral&Hoxby(2010),propertytaxbillsare verysalienttohomeownerswithoutescrowaccountsandlargeenoughthattheymusteithersaveor increasecreditcardborrowingtopaythesebills. Theperiodsimmediatelyfollowingapropertytax 1SeeDeng,Quigley&vanOrder(2000),Bajari,Chu&Park(2008),Foote,Gerardi&Willen(2008),Experian- OliverWyman(2009),Ghent&Kudlyak(2009),Mayeretal.(2009)andVandell(1995)forasummary. 2Fay,Hurst&White(2002)andKeys(2010)observeindividual-leveladverseeventsinastudiesaboutbankruptcy. 3In addition to endogeneity, using aggregate proxies prevents the estimation of how the average borrower is affectedbyilliquidityandassigningaggregateproxiestoindividualloansintroducesclassicalmeasurementerror,which attenuatesourunderstandingofhowilliquiditycontributestomortgagedefault. Othersourcesofloan-levelliquidity reductions, such as interest rate resets, typically occur two to three years after origination and thus cannot provide informationonthecausesofEPD. 4Industry estimates suggest that prior to 2007 only about 25% of subprime loans had escrow accounts (National MortgageNewsMortgageWireArchive,March7,2005). 5These figures likely underestimate the annual financial obligation among subprime mortgage holders, but, in states with semi-annual installments, may overstate the size of an individual tax bill. Source: U.S. Census Bureau, 2007AmericanCommunitySurvey. 2

due date are thus associated with a decrease in cash-on-hand or an increase in debt commitments. Survey evidence from 2006 suggests that property tax liabilities were the proximate cause of as muchas12%ofsubprimemortgagedelinquencies.6 By combining loan-level information on the payment status of subprime mortgages with administrativedataonpropertytaxcollectiondates,weobserveexactlywhenanindividualborrower faces a property tax due date.7 Differences in property tax collection dates across states, counties, cities, and school districts make the timing of the first property tax due date relative to when a mortgageisoriginatedplausiblyorthogonaltounobserveddeterminantsofnegativeequity,default costs, expectations of capital gains, and the size of property tax bills. This natural experiment allowsustoobserveotherwise-identicalborrowersthatdifferonlyinthetimingoftheirfirstproperty taxduedaterelativetowhenamortgageisoriginated. The exogenous variation in the timing of the first due date allows us to identify variation in borrowers’ duration of exposure to a reduction in liquidity, which we use to estimate the effect of reduced liquidity on mortgage delinquency and default outcomes. For example, one year after origination, some borrowers have been “treated early” and paid their identical property tax bill 11 months ago, while otherwise similar borrowers have been “treated late” and paid their identical propertytaxbillonly1monthago. Unlessborrowersareabletoquicklyincreasecash-on-handor decrease debt commitments, in the first year after origination borrowers treated early are exposed to up to 11 more months of reduced liquidity than borrowers with late due dates. Our estimation strategythusreliesontheidentifyingassumption,whichisconsistentwithourdata,thatborrowers originating loans near and far away from property tax due dates are observably and unobservably similar. Using data on subprime loans originated between 2000 and 2007 for home purchases, we find that an approximately nine month delay in owing property taxes reduces the probability of EPD by about 3 to 4 percent. This estimate suggests that the effect of a nine month delay in owing property taxes is about one-third as large as the effect of a transition from 10% equity to 20% negative equity. If the reduction in liquidity due to property taxes were, on average, immediate and brief, we would expect “early treated” and “late treated” loans to experience similar rates of EPD (i.e., default within the first year of mortgage origination). We therefore infer that the likely mechanismfortheestimatedeffectoccursthroughthepersistenceoftheliquidityreductionasthe first property tax bill increases borrowers’ sensitivity to income or expenditure shocks after the due date. That we also find an 11 to 16 percent greater likelihood of making up missed mortgage payments(i.e. “curing”)whenpropertytaxesaredelayedcorroboratesthisinterpretation. 2 Property Tax Due Dates, Liquidity, and Default To understand how less liquidity ensues following a property tax due date, we begin with an explanation of the property tax remittance process in the United States. The property tax remittance processintheUnitedStatescreatestwotypesofborrowers: thoseforwhompropertytaxduedates produce no reduction in liquidity because they have escrow accounts and those for whom prop- 6SeeTable4in: PartnershipLessonsandResults: ThreeYearFinalReport,p. 31HomeOwnershipPreservation Initiative(July17,2006)atwww.nhschicago.org/downloads/82HOPI3YearReport_Jul17-06.pdf. 7Knowledgeofwhichborrowersexperiencealiquidityreductionallowsustoavoidthemeasurementerrorassociatedwithaggregateproxiesandtoestimatetheeffectsofliquidityreductionsfortheaverageborrower. 3

erty tax due dates produce a reduction in liquidity because they do not have escrow accounts. A borrower and his lender jointly select one of two processes for remitting property taxes: an equal portionofthetotaltaxpaymenteachmonthalongwiththemortgagepaymentoralumpsumproperty taxpayment on or beforea tax due date.8 In thecase where property taxesare included in the monthly mortgage payment through an escrow account, a property tax due date requires no action and produces no post-due date liquidity reduction. This is because an escrow account, which is a bank account set up by a lender (or servicer) in which he deposits monthly payments collected from the borrower, spreads out the borrower’s tax payments over time. A borrower’s monthly escrow payments are fixed throughout any single year.9 Thus, for escrowed borrowers, a property tax due date is not associated with a post-due date increase in financial obligations nor any direct remittancetothelocalgovernment. For borrowers without escrow accounts, which includes upwards of 75% of subprime borrowers, a property tax due date requires action and has effects on their liquidity, even if these bills are well anticipated.10 A non-escrowed borrower must decide whether or not to remit a lump sum tax payment to the government. Both actions, either paying the tax bill on time or not paying on time, reduce the non-escrowed borrower’s liquidity via two mechanisms: less cash-on-hand and an increase in debt commitments.11 If a borrower elects not to pay their property tax bill on time, he increases his debt commitments by incurring tax delinquency penalties and a typical annual interest rate of 18%. If the bill remains unpaid the local government possesses the first lien on the homeandhastherighttotakeownershipoftheproperty.12 Anon-escrowedborrowerchoosingtopaythepropertytaxbillontimehasatleasttwotypesof optionstofinancehispayment,bothofwhichresultinareductioninliquidity(albeitwithdifferent implications for the magnitude and duration of the reduction). First, if the borrower has adequate cash-on-hand to pay the property tax bill by cash or check, the property tax payment reduces post-due date liquidity by reducing cash-on-hand. For some borrowers, it may take many months to restore their cash-on-hand to their pre-due date levels. Second, a borrower with inadequate cash-on-handmaychoosetobecomedelinquentonthemortgage,stoppayingnon-mortgagebills, such as utility or credit card bills, or increase their borrowing (or some combination of these three actions). Each of these three actions leads to greater debt commitments via increased borrowing and possibly higher borrowing costs, which may take many months to undo. Even if borrowers optimally select the action that least increases their debt commitments, their liquidity is lower after the property tax due date. For example, the permanent income hypothesis predicts that, to smooth consumption, borrowing increases (savings falls) in the period of an anticipated increase infinancialobligations. Ifmortgagedelinquencyrepresentstheleastexpensiveborrowingvehicle, householdsmayelecttobecomedelinquenttocovertheirfinancialobligations.13 8Notethataborrower’stotalannualpropertytaxpaymentdoesnotdependontheremittanceprocess. 9Aborrowerensuresanadequate“cushion”,i.e.,enoughmoneyintheaccounttopaythetaxbillontheduedate, overthecourseoftheyearbymakinganinitialdepositintotheescrowaccountatclosing(Anderson&Dokko(2009)). 10MortgageServicingBullentin(MSB),March7,2005. 11Althoughthisassumesthatthetaxbillislargeenoughthatitisimpossibletofinancethetaxbillentirelythrough adecreaseinconsumptionexpendituresthatleavescash-on-handanddebtcommitmentsunchanged,webelievethis assumptionisjustified(seeCabral&Hoxby(2010)). 12Our review of state statutes suggests that most states’ interest charges and delinquency penalties imply an an annualinterestrateofbetween12%and18%. 13Becomingdelinquentonamortgageentailsatypicalpenaltyofbetween1%and5%ofthemortgagepayment. In thesubprimemarket,itisreasonabletoexpectborrowerstopayaround20%interestoncreditcardbalances. 4

Depending on the financial resources of borrowers and the size of the property tax bill, the liquidity reduction associated with the remittance of property taxes around the property tax due date can be either brief or persistent. When the liquidity reduction is brief, any effect on delinquency and default decisions occurs contemporaneously with the due date when non-escrowed borrowers choose to become delinquent or default on the mortgage to cover other (tax and nontax) payment obligations. When the liquidity reduction is persistent, however, a prolonged state of reduced liquidity after the tax due date makes borrowers’ decisions on delinquency and defaultmoresensitivetoincomeandexpenditureshocks,suchasunemployment,furloughdays,and serious health problems. Tax remittance can produce a persistent liquidity reduction via higher post-due date debt commitments that cause borrowers to have a higher back-end debt-to-income ratio (DTI).14 If subjected to an income shock, a borrower with higher DTI may find it optimal to financehisnon-mortgagedebtcommitments(e.g.,avoidcreditcarddelinquencyordefault)bybecomingdelinquentordefaultingonhismortgage.15 Thus,regardlessofwhetherornottheliquidity reductionisbrieforpersistent,itcanaffectmortgagedelinquencyanddefault. In the empirical analysis we focus on estimating the effect of the timing of the first property tax due date after mortgage origination on the probability of subprime mortgage delinquency and default during the mortgage’s first year. The first property tax due date after mortgage origination offers the best opportunity to identify the effect of an exogenous reduction in liquidity occurring over a finite length of time. During the first year of a mortgage, the first property tax due date cleanly demarcates the months prior to the first bill. During this time, a non-escrowed borrower is notexposedtoaliquidityreductionpriortotheduedatewhereasafterward,heisexposedtoeither a brief or persistent state of reduced liquidity.16 If the liquidity reduction from the first due date ispersistent,subsequentpropertytaxduedatesdonotcleanlydemarcatethetimebeforeandafter a liquidity reduction.17 Subsequent due dates may, however, exacerbate the liquidity reduction associatedwiththefirstduedate. 3 Empirical Strategy 3.1 Identification The objective of the empirical strategy is to estimate the effect of the timing of the first property tax bill on first-year mortgage delinquency and default.18 Differences in the timing of the first property tax due date relative to loan origination and the size of the property tax bill create between-loan variation in the timing (extensive margin) and magnitude (intensive margin) of the 14Theback-endDTIratioisthemortgagepayment(includinganyescrowedinsuranceandtaxes),creditcarddebt, carloans,educationloans,andotherdebtsdividedbyincome. 15Cohen-Cole & Morse (2010) provide evidence that households become delinquent on their mortgage to avoid creditcarddelinquency. 16Theownerofapropertyattheduedateislegallyresponsibleforremittingthepropertytaxpayment.Thus,evenif sellersandbuyersnegotiate,forexample,areductioninclosingcoststo“compensate”thebuyerfortheirfirstproperty taxbill,thebuyer(i.e.,ownerattaxduedate)muststillremitthetaxesandmustpayanydelinquencypenalties. 17Inaddition,sincesomeloanswillnotfaceasecondduedateuntiltheirsecondyearafterorigination,focusingon thefirstduedateallowsustofocusondelinquencyanddefaultinthefirstyearofamortgage. 18Wedefinedelinquencyasoneortwomissedmortgagepaymentsanddefaultasatleast3missedpaymentsora foreclosurestart. 5

post-due date liquidity reduction associated with the property tax bill. We focus on estimating the effect of reduced liquidity along the extensive margin: because homeowners sort into high or low tax jurisdictions based on tastes for public goods, income, and other characteristics that may be correlated with their ability to pay their mortgages, between-loan variation in the timing of due datesisplausiblymoreexogenousthanbetween-loanvariationinthesizeofpropertytaxbills. More explicitly, between-loan variation in the timing of the post-due date liquidity reduction arisesfrombetween-loandifferencesinthemonthoforiginationandbetween-jurisdictionvariation inthemonthofpropertytaxduedates. Propertytaxduedatesvarybetweenstatesandwithinstates. In 33 states, property tax due dates are uniform within the state while the remaining states’ due dates vary within a state because counties or other local governments set their own due dates.19 Table 1, panel A shows that the between-state variation in property tax due dates spans most calendar months as every month except July has at least one state with a due date within it. The most common month for due dates is October and there are fewer states with due dates in the summer. As seen in Table 1, panel B, the origination month of subprime purchase loans varies between loans with a peak in June and a trough in January, similar to the seasonal pattern seen in conformingloans. Together, variation in due dates and origination months generates between-loan variation in loans’ ages at the first property tax due date (“due date age”), as seen in Table 2. All loans face a propertytaxduedatewithinoneyearoforigination. Althoughthemajorityofloansfaceaduedate within the first four months after origination, over 13% of loans face their first property tax due dates nine or more months after origination (panel A). Panel B of Table 2 shows the distribution of due date age, along with some additional borrower characteristics, for each origination month. The average FICO and combined loan-to-value (CLTV) ratio demonstrate that the borrower characteristics identified by the mortgage default literature to be most predictive of delinquency and default are very similar between origination months, suggesting that there is no observable seasonal pattern in the credit quality of borrowers originating mortgages (Mayer et al. (2009)). The within-origination month variation in due date age reported in the last two columns demonstrates that origination month alone does not determine a loan’s due date age. In fact, for all origination months,exceptJune,theduedateageranges,inclusively,fromoneto12.20 Because of the identifying assumption that due date age is as good as random, pre-determined loan and borrower characteristics observed at origination should not be correlated with loans’ due date ages (Holland (1986) and Rubin (1986)). Specifically, loans that are older or younger at the due date should not systematically differ in terms of the differences in borrower characteristics related to delinquency and default decisions such as income, the debt-to-income ratio, and creditworthiness. We test the implications of this identifying assumption in Table 3, which shows loan characteristics for the entire sample by loans’ due date ages. As seen in Panel A, with the exception of the combined loan-to-value (CLTV) ratio at origination, loan and borrower characteristics vary depending on the age of the loan when the property tax bill is due. However, because the maximumduedateagevariesbystate,thecompositionofstateschangesasthenumberofmonths until the property tax due date increases. For example, in Florida, property taxes are due once a yearandaloanmaybe12monthsoldbeforeitfacesitfirstpropertytaxduedatebutinCalifornia, 19Seeappendixtableforalistofpaymentinstallmentsbystate. 20Unliketimeuntilthefirstpropertytaxduedate,timebetweenthefirstandsecondduedatesdoesnotvarymuch within-state. In the 13 states with annual due dates, there is no within-state variation in time between due dates. In stateswithuniformsemi-annualduedatesthetimebetweenduedatestakeson,atmost,severalvalues. 6

taxes are due twice a year and a loan can be no older than six months at its first property tax due date. Furthermore, because the composition of states changes across the columns, the distribution of origination year also changes as lending in the subprime market did not decrease as much in 2007 among states with annual due dates (for reasons unrelated to property taxes). Panel B shows the average borrower characteristics after regression adjusting for state, origination year, and the calendar month of the due date. State fixed effects allow for comparisons of average borrower characteristics across the treatment groups holding time-invariant state characteristics fixed. Origination year and due date month fixed effects are also included in the regression adjustment to control for time trends that may be correlated with the composition of states. After regression adjustingtheaveragecharacteristics,weinferthatmostofthedifferencesobservedinPanelAare duetothechangingmixofstatesinthesample. Indeed,thesimilarityintheaveragecharacteristics in Panel B suggests that conditional on state, origination year, and due date month, there is no a priorireasontorejectthevalidityoftheresearchdesign. Weestimatethefollowingequationusingalogitspecification: Dj = α+β Due +β Due +β Due i 4−6 4−6,i 7−9 7−9,i 10−12 10−12,i + γ ·X +ν ·W +(cid:15) (1) i i i where Dj equals one if, at any time during the loan’s first year, borrower i experiences outcome i j and zero otherwise. The four delinquency and default outcomes we are interested in include whether the borrower misses one, two, or three consecutive mortgage payments, leaving him 30, 60, or 90 days delinquent at any point during the first year of the mortgage, as well as whether the lender initiates a foreclosure start. Following conventions in the mortgage default literature, we consider90-daydelinquencyoraforeclosurestartduringaloan’sfirstyearasEPD.Later,wealso examinewhetherthesesameoutcomesoccurduringthefirst2yearsafterorigination. If a loan leaves the sample prior to delinquency or default because the borrower chooses to refinance, then we consider this borrower as not being delinquent or not defaulting (i.e. Dj = 0 i forthisborrower). Becauseweassumethatforanyloaninagivenstate,duedateagesareasgood as random, they are also assumed to be orthogonal to the prepayment incentives borrower i faces, allowing us to estimate (1) in a simple logit framework that need not account for the simultaneity oftheborrower’sprepaymentoption(seeDengetal.(2000)). Inequation1,thethreebinarypropertytaxduedatevariables,Due ,dividethesampleinto t−k,i fourtreatmentgroupsandequaloneifaloanhaspropertytaxesfirstdueatagestthroughk during the first year of the mortgage and zero otherwise. For example, if loan i’s first property tax due date occurs at month 7,8, or 9 since origination, the variable Due = 1 and the other two due 7−9,i datevariablesequal0. Sincetheomittedcategoryrepresentsloanswithduedateagesequalto1,2, or3,theseloanshaveDue = 0. t−k,i The four treatment groups categorize loans according to their due date ages and the maximum potential duration of exposure to a persistent liquidity reduction induced by taxes. The betweenloandifferencesinduedateageproducethebetween-loandifferencesinthedurationofexposure. During the first year after origination, loans that are younger when property taxes are due spend more months after the due date, which is a period when they may be exposed to a persistent liquidityreduction. X is a vector of pre-determined loan and borrower characteristics observed at origination ini cluding the borrower’s FICO score, sales price, a dummy indicating whether the loan was fully 7

documented,anindicatorequaltooneiftheloanisanadjustableratemortgage,theinitialinterest rate,combinedloan-to-valueratio,andfixedeffectsfororiginationyear,thecalendarmonthofthe first property tax due date, and state.21 Note that we do not need to control for loan age because at the end of their first year all loans are one year old. Since all loans face all 12 calendar months, seasonaldifferencesindefaultrateswillnotcreatebetween-treatment-groupdifferencesindefault rates.22 The control variables in X address some reasons that are unrelated to property taxes but i explain why default rates might be higher (or lower) for borrowers in any of the four treatment groups: borrower-specificriskcharacteristicssuchasFICOorCLTV,decliningunderwritingstandardsthatareproxiedforbytheloan’soriginationyear,andstate-specificfactors,suchasmortgage lendinglawsormacroeconomicconditions. W representsavectorofborrowercharacteristicsthatarenotpre-determinedatoriginationbut i maybecorrelatedwithaloan’sageatthepropertytaxduedateandalsoaffectaborrower’sability topaythemortgageandthereforethelikelihoodofdefault. Thesevariablesincludehousingequity at the first property tax due date (measured as the mark-to-market CLTV ratio), first-year house price appreciation, and the size of the property tax burden (measured as the ratio of borrowers’ countymedianpropertytaxtocountymedianincome). Weexploretheextenttowhichloansdiffer along these dimensions in Table 4. Similar to Table 3, loan age is less likely to be correlated with loancharacteristicsuponcontrollingforstate,originationyear,andduedatemonth(PanelB)than inPanelA,wherethecompositionofstateschangesacrossthefourtreatmentgroups. Forexample, within a state, the size of the average annual property tax bill and the number of installments do not vary between groups. The variable (cid:15) is assumed to be a random error term. We show two i setsofestimates: thosethatcontrolforX andW andthosethatcontrolonlyforstate,origination i i year, and the calendar month of first property tax due date. The results are consistent across these twospecifications. In general, identifying the coefficients in equation (1) may be challenging if property taxes do not reduce liquidity and instead, due dates are correlated with loan characteristics, such as down payment amounts or borrowers’ creditworthiness. Our natural experiment and the variables in X i andW allowustocrediblyestimatetheeffectsofthetimingofliquidityreductions. Equation(1), i however, does not control for whether borrower i has an escrow account as this information is not observedinmostpubliclyavailableloan-leveladministrativedata. Henceanimportantidentifying assumptionisthatsubprimeborrowersdonotelecttoopenescrowaccountsbasedontheproperty tax due date, i.e., the fraction of borrowers with escrow accounts is the same across treatment groups. Giventhat veryfew subprime borrowershad escrowaccounts, this assumption is likelyto bemet. However,if,forexample,borrowerswithpropertytaxduedatesthatarefurtherawayfrom the origination date were more likely to set up an escrow with the lender, they may be less likely to default either because they possess better financial management or because they experience no post-duedateliquidityreduction. An alternative methodological approach to estimating the effect of reduced liquidity using equation (1) is an event study along the lines described in Jacobson, LaLonde & Sullivan (1993). 21Forstateswithuniformduedates,thecombinationofstateandthemonthoffirstpropertytaxduedateisaperfect predictoroforiginationmonth. Thus,allthreeindicatorvariablescannotbeincludedinthesameregression. 22There are, however, between-loan differences in loan age for each each calendar month. If the order in which a loan faces each month (i.e., March at age 3 months or at age 10 months) affects first-year default probability, an origination-month dummy will control for any effects. Regressions that include origination month rather than property-tax-due-datemonthfixedeffectsdonotalteranyconclusions. 8

We believe, however, that an event study approach will not identify the causal effect of post-due date liquidity reductions on mortgage delinquency and default because of the difficulty in constructing a counterfactual group of loans that never face property taxes.23 An event study comparing the delinquency rate of loans before and after the property tax due date would incorrectly infer that a higher delinquency rate after the property tax due date owes to the post-due date liquidity reduction. This approach will be misleading because loans observed after the property tax due date are, by construction, older than they were prior to the event and may be more likely to default because they are older. Because all non-escrowed subprime loans are exposed to property tax due dates, leaving no group of counterfactual untreated loans, an event study without a valid counterfactualisnotidentifiedunlessstrongandlikelyinvalidassumptionsaremadeabouttherelationshipbetweenconfoundingfactorsanddelinquency(seeMcCrary(2007)foramoretechnical explanation).24 Insomecasesitmaybepossibletouseacomparisongrouptoidentifyand“differenceout”the confounding effects of age on delinquency. Some obvious comparison groups include subprime loans with escrow accounts, which most loan-level data sets do not identify, or loans that have escrow due to the institutional features of mortgage lending, such as many prime loans. Even if loans’escrowstatuswereobservable,itisnotrandomlyassignedandthusselectionissuesprevent escrowed-loans from offering a valid comparison group. The delinquency and default behavior of subprime and prime loans are so different that the age effects of prime loans that one would “difference out” are not a valid counterfactual and using them as such would lead to misleading inferences. Insum,wethinkaneventstudydoesnotcrediblyestimatethecausaleffectsofapost-due-date liquidityreduction. Althoughwedonotwishtoemphasizethem,intheappendixwepresentevent studyresultsconsistentwithourmainresults. 3.2 Interpretation The marginal effects corresponding to the three βs in equation 1 represent the differences in the first-year delinquency and default probabilities relative to the case where property taxes are due in the first three months after origination. We define a liquidity reduction as brief (and thus not persistent)whenduedatesinpreviousquartershavenotreatmenteffectsinfuturequartersbecause liquidityhasalreadyrecoveredtoitspre-duedatelevel. Between-groupdifferencesinduedateage produce between-group differences in delinquency and default rates because either the liquidity reductionispersistentorthetreatmentstrengthofabriefliquidityreductionvariesbyloanage. As an example, consider the result that the probability of first-year default declines as the due date age increases. This result seems to imply a persistent liquidity reduction, but that is not necessary. In our framework, a declining delinquency or default probability as due date age increasesimpliesthat: β < β < β < 0 (2) 10−12 7−9 4−6 23Inaddition,conceptually,aneventstudywouldcharacterizeanoutcomerelatedtoEPD,suchasathefractionof loansthataredelinquentinaparticularmonth,butnotEPDitselfaroundthepropertytaxduedate. 24For example, Agarwal, Liu & Souleles (2007) and Johnson, Parker & Souleles (2006) are able to exploit the randomtimingoftaxrebatestoestimateeventstudiesoftheconsumptionresponsebecausethereisnotaconfounding variable(e.g. age)correlatedwiththetimingofthetaxrebates. 9

The unique circumstances under which conditions in equation 2 hold are infinite but two special cases are worth mentioning. First, consider the case where the liquidity reduction is persistent. In this case, differences in age at first due date differentially expose borrowers to a persistent liquidityreduction. Forexample,duringtheloan’sfirstyearthemaximumdurationofexposureto a persistent liquidity reduction is longer for borrowers facing property taxes in the second quarter than those with a fourth quarter due date. Accordingly, if exposure to reduced liquidity affects delinquency and default, the first-year delinquency and default probabilities of loans with first quarter due dates are higher than those of loans with later due dates. Second, consider the case where the liquidity reduction is brief and not persistent so that the duration of exposure to the liquidityreductionisthesameforallborrowers. Inthiscase,inorderforβ < β < β < 10−12 7−9 4−6 0,itmustbethecasethatthetreatmentstrengthofanidentical(brief)liquidityreductiondeclinesas age-at-due-date increases. Regardless of the specific circumstances under which β < β < 10−12 7−9 β < 0, this result must suggest that property taxes affect mortgage delinquency and default 4−6 througheitherabrieforpersistentliquidityreduction. When the liquidity reduction is persistent, first-year delinquency and default is higher for late due loans because censoring outcomes at one year prevents us from observing the effects of the persistentliquidityreductionforlatedueloans. Iftheliquidityreductionpersistsforafinitenumber ofquarters,theuncensoredeffectsareidenticalacrosstreatmentgroups.25 Iftheliquidityreduction is brief, censoring at one-year will not affect our results because we will have observed the total uncensored effect for all treatment groups. In sum, censoring at one year does not create an effect on delinquency and default where none existed; but censoring can reveal an effect, driven by a persistentliquidityreduction,thatwemightnototherwisesee. The plausibly random assignment of loans into treatment groups implies that if property tax due dates do no affect delinquency and default, loans in the four treatment groups should have equaldelinquencyanddefaultprobabilities. Thatis,thethreeβ coefficientsequalzero. Finding t−k that the β coefficients are not different from zero, however, is a necessary but not a sufficient t−k condition for demonstrating that the post-due date liquidity reduction has no effect on first-year default probabilities. There are two scenarios under which β = β = β = 0. First, this 4−6 7−9 10−12 may happen when property tax due dates and any associated liquidity reductions have no effect on delinquency and default. Second, β = β = β = 0 when the liquidity reduction is 4−6 7−9 10−12 notpersistentandtheeffectofthebriefliquidityreductionishomogenousacrossloansindifferent treatmentgroups. 4 Data and Sample Wecombinedatafrommultiplesourcestoobtaininformationonaloan’spaymentstatusovertime, pre-determined undewriting characteristics, age at due date, and variables that may be correlated with a loan’s age at the property tax due date and also affect default. Loan-level data on payment status are from CoreLogic (formerly known as LoanPerformance) and these data track whether a loan is current, 30/60/90 days delinquent, or in foreclosure.26 These data also contain limited 25Iftheliquidityreductionwereinfinitelypersistent,thedelinquencyanddefaultratesarealwaysdifferentbetween treatmentgroupsbecauseearly-dueloans’wouldhavelongertreatmentdurationthanotherloans. 26The monthly indicator for foreclosure roughly identifies when the foreclosure process starts, not when it ends, whichistypically8to12monthsafterwhentheborrowerstopsmakingpayments(seeCutts&Merrill(2009)). 10

underwriting information, such as the borrower’s credit score (FICO), an indicator for whether the borrower fully documented his income, the combined loan-to-value (CLTV) ratio, and, for about 60% of the observations, the borrower’s stated debt-to-income (DTI) ratio at the time of origination.27 CoreLogic’s database also includes limited information on loan characteristics that may pose risks to the borrower: the initial contract interest rate, an indicator for whether the mortgage has an adjustable rate, and an indicator for whether there is a prepayment penalty. For eachloaninCoreLogic’sdatabase,weknowthemonthandyearinwhichloansareoriginated. Theloan’sageatthepropertytaxduedateisconstructedbycombiningCoreLogic’sdatawith information on property tax due dates, which we obtained from the 2008 U.S. Master Property Tax Guide, internet resources, and phone/email contact with property tax-collecting government officials.28 Appendix Table 1 lists the payment installments by state, which we combine with a loan’s “birthday” to calculate the loan’s age, measured in months, at the time when property taxes arefirstdue. To control for the magnitude of the property tax burden, we use county median property tax amountsrelativetocountymedianincomereportedinthe2005AmericanCommunitySurvey(for calendar year 2004). First-year house price appreciation rates are from CoreLogic’s ZIP code and state house price indexes. Following Foote et al. (2008), we construct a mark-to-market measure of housing equity at the first property tax due date using these house price indexes and the CLTV ratioatorigination. Asalludedtoearlier,akeyfeatureofouridentificationstrategyisthatwedonotneedtoknow whether a loan has an escrow in order to estimate Equation (1). This addresses a limitation of the CoreLogic data where there is no variable indicating the escrow status.29 Because of this data limitation, we analyze all subprime first liens originated for home purchases between 2000 and 2007, including those that may have an escrow.30 We exclude loans originated for refinancing as borrowers with such loans have faced prior property tax due dates and thus the first due date after origination is less likely to provide a clean demarcation of before and after exposure to a liquidity reduction. We also exclude loans packaged into alt-A securities, which typically were originated by investors or borrowers without impaired credit histories who sought out the non-prime market forthemortgageswithexoticfeatures,suchasinterest-onlypaymentsornegativeamortization,or toprovidelittletonodocumentationoftheirincome. Thesesampleexclusionsallowustoidentify agroupofborrowerswhoaremostlikelytobefinanciallystrainedbypropertytaxes. Theloansintheanalysissamplewereoriginatedin40states(includingtheDistrictofColumbia) 27This variable is missing for so many observations because CoreLogic does not require servicers to submit this informationtothedatabase. Becauseitismissing,wedonotincludeitinourmainanalysis. 28In a few states administrative delays sometimes cause actual due dates to differ from due dates in statutes. For example, in Cook County, Illinois, due dates are frequently pushed back. In addition to conversations with local officialsweconsultednewspaperrecordsforreportsofdelaysinduedatesandchangedduedateswhenappropriate. Inthevastmajorityofstatesandcountiesduedateswereneverdelayed. 29An alternative loan-level dataset from LPS Analytics (formerly McDash) has a variable indicating whether the loanhasanescrow. However,thisdatasethaslimitedcoverageofnonprimeloansthatislowrelativetoCoreLogic’s coverage, particularly before 2005. Also, the escrow variable is not well populated for around 70-80% of subprime loans, suggesting that there are serious measurement problems associated with it. Furthermore, it is not possible to infer the existence of an escrow account based on the loan’s monthly payment amount and tracking how the loan’s outstandingbalanceevolvesovertimebecausecertainfieldsintheCoreLogicdataarenotwellpopulatedbythedata provider. 30Becauseofourresearchdesignandtheplausiblyrandomassignmentofloanstotreatmentgroups,theunobserved escrowstatusdoesrequireustoscaleourcoefficientsasintent-to-treatparameters. 11

identified in Appendix Table 1. We analyze these states because property tax due dates are uniform within either the state or county, which facilitates a merge with the CoreLogic data, where the available geographic identifiers only include state and ZIP code. The 10 excluded states have propertytaxduedatesthatvaryatthelevelofthemunicipalityorschooldistrict,whicharesmaller geographicunitsthantheZIPcode,makingitimpossibletomergewiththeLPdata. With these restrictions, over two million loans remain, which we are able to track monthly for 12months. Inspiteofthesamplerestrictions,therearenearly25millionloan-monthobservations. Duetothecomputationalburdenofthesedata,weconductouranalysisona20%randomsample. Forourmainregressions,thisproducesasampleof480,738loans. 5 Results We now discuss the results of using loan-level variation in the post-due date liquidity reduction to estimate the causal effect of reduced liquidity on subprime mortgage delinquency and default. Each row of Table 5 describes the results of estimating equation 1 and corresponds to one of four different outcome variables. Our regressions compare the first-year delinquency and default outcomes of loans across four treatment groups. For each outcome, Column (1) lists the average delinquency or default rate for loans with due date ages between 1 and 3 months (i.e., “early-due ˆ loans”). Columns (2) through (4) display estimates of the three logit regression coefficients, β , t−k transformedintopercentagepointaveragemarginaleffects. Thethreecolumnsdisplaytheaverage marginal effect, relative to early-due loans, of a loan facing its first due date at ages 4-6 months (i.e., 4-6 month loans), 7-9 months (i.e., 7-9 month loans) and 10-12 months (i.e., late-due loans). Thepercentreportedtotherightofeachmarginaleffectexpressestheaveragemarginaleffectasa percentoftheaveragedelinquencyordefaultrateforearly-dueloans. In Panel A of Table 5, the marginal effects estimated in columns (2) through (4) include only state, origination year, and due date month as controls. The effects shown in Panel B also include the following control variables: sales price, borrower’s FICO score at origination, indicator for full v. no/low documentation, indicator for adjustable rate mortgage, interest rate at origination, combined loan-to-value ratio at origination, mark-to-market combined loan-to-value ratio at the duedate,first-yearhousepriceappreciation,theratioofcountymedianpropertytaxbilltocounty medianincomein2004,andfixedeffectsfororiginationyear,monthofpropertytaxduedate,and state. Again, after the first year, all loans are the same age so the outcomes measure delinquency anddefaultratesforsimilarlyagedloans. Wepresentthesetwosetsofresultstoexplicitlydemonstrate that, consistent with our identifying assumptions, the inclusion of additional controls does notsubstantiallyaltertheresults. InPanelsAandBofTable5,thefirstrowsdescribetheresultsforfirst-year30-daydelinquency rate. Column(1)showsthat30.9%ofearly-dueloansmissonemortgagepaymentatleastoncein thefirstyear. Theresultsincolumns(2)to(4)showthatolderloansatthetimeofthefirstduedate arelesslikelytomissonemortgagepaymentduringthefirstyear. Focusingontheresultswiththe full set of control variables in Panel B, loans with due date ages between 4 to 6 months are 0.004 percentage points less likely to experience a 30-day delinquency in the first year than early-due loans. In column (3), all else equal, loans with due date ages between 7 and 9 months are 0.59 percentage points, or 1.9%, less likely to experience a first-year 30-days delinquency than earlydue loans. Finally, the average marginal effect in column (4) suggests that late-due loans are 0.95 12

percentage points, or 3%, less likely than the early-due loans to experience a 30-day delinquency inthefirstyear. The results for 30-day delinquencies suggest that borrowers facing property taxes earlier are morelikelytodefaultthanthosewithlaterduedates. However,becausemissingonlyonemortgage payment is relatively common and not necessarily indicative of a financial hardship, we examine 60-day delinquencies and EPD in the remainder of Table 5. Focusing on Panel B, we find that for moreseriousdelinquenciesandEPD,late-dueloansare0.59percentagepoints(3.7%)lesslikelyto experienceafirst-year60-daydelinquency,0.34percentagepoints(3.2%)lesslikelytoexperience a 90-day delinquency, and 0.35 percentage points (5%) less likely to experience a foreclosure start. Further, although we do not have enough power to statistically distinguish the size of the ˆ ˆ ˆ coefficients from each other, we find that β < β < β < 0 for 30-day and 60-day 10−12 7−9 4−6 ˆ ˆ ˆ delinquencyandthatβ < β ≈ β < 0for90-daydelinquencyandforeclosurestarts. 10−12 7−9 4−6 Alloftheseresultsleadtothesameconclusion: early-dueloansdisplayhigherfirst-yeardelinquency and EPD rates than late-due loans.31 Our research design and the control variables in X i and W ensure that a plausible interpretation of our results is that liquidity problems contribute to i early payment default. As we argued earlier, between-treatment group differences in average borrowercharacteristics,housingequity,oreconomicconditionsareunlikelytoexplainourresults.32 In addition, we posit that the differences between early-due and late-due loans are attributable to differences in the duration of exposure to a persistent liquidity reduction. We base this interpretation on four reasons. First, in addition to 30-day delinquency, rates of 60-day delinquency and EPD are also higher among early-due loans, which suggests that liquidity reductions were persistent rather than brief. If borrowers were able to quickly restore liquidity following a property tax due date, it would be unlikely to also find higher rates of more serious delinquencies and EPD.Becauseweexpectmortgagedelinquenciesmotivatedbyconsumptionsmoothingtobenonserious, the effects on default are also inconsistent with unconstrained borrowers using mortgage delinquencytosmoothconsumption. Second, we believe it is unlikely for brief liquidity reductions to have effects differing by due date age. While our data do not allow us to directly observe how a borrower’s liquidity changes due to a property tax due date, we are able to control for some plausible sources of heterogeneity in the treatment effect of a brief liquidity reduction such as the amount of equity at the due date orthemagnitudeofthepropertytaxbill.33 Twoothersourcesofheterogeneoustreatmentstrength arising from brief liquidity reductions are surprise and unavoidable low liquidity.34 To rule out the possibility that greater surprise about the property tax bill among early-due borrowers causes higherratesofdelinquency andEPD,wenotethatthe disclosures occurringundertheRealEstate SettlementProceduresAct(RESPA)oughttomakepropertytaxesknownandsalienttoborrowers at closing. Indeed, since early-due borrowers have more recently experienced a home purchase, 31Resultsfromaneventstudyframework,whichareshownintheappendix,areconsistentwiththisinterpretation. Theeventstudyresultsdonotcontrolforloanage.Consistentwithourargumentabove,includingcontrolsforalinear orquadratictrendinloanageproducesestimatesthatdonotaffordusthepowertoidentifythesizeofanyeffect. 32Precautionarysavingsbehaviorcanproduceresultssimilartothoseimpliedbyliquidityconstraints(e.g. Carroll (2001)). 33Notethatduedateageexplicitlydeterminesthemaximumpotentialdurationofexposuretoaliquidityreduction, sowecannotholdduedateageconstantwhilevaryingthedurationofexposure. 34Briefliquidityreductionswithheterogeneouseffectscanalsoexistifthemonthofaduedateaffectsthecontemporaneoustreatmenteffectandmonthoftreatmentdifferamongtreatmentgroups.Ourinclusionofmonth-of-due-date fixedeffectshelpscontrolforanybetween-groupdifferencesinaveragetreatmentmonth. 13

theyoughttobelesssurprisedthanlate-dueborrowers,whichworksagainstusfindinganeffect.35 Unavoidable low liquidity arises when early-due borrowers are unable to prepare for fully anticipated property taxes because they are treated so soon after origination. We find unavoidable low liquidity unlikely because we are unaware of any evidence on whether substantive differences in liquidity exist between, for example, borrowers two months after origination compared with 10 months after origination. Furthermore, if property taxes are not a surprise and all borrowers know whentheirtaxesaredue,allborrowersshouldbeequallycapableofaccumulatingadequateliquidity at the due date. For example, early-due borrowers can save more in advance of origination or theycannegotiatewiththesellertolowerclosingcoststohelpfinancethepropertytaxpayment. Third, we extend the period of time during which we observe delinquency and default outcomes to two years after origination. The additional year exposes loans to additional periods of possible default risk and, for many loans (see Table 3), an impending interest rate reset at the end of the second year. As Tables 3 and 4 show, these default risks and impending interest rate resets do not differ across the treatment groups and thus do not directly pose a threat to our estimation strategy. Observing loans for two years reveals whether small or temporary effects of property taxes are overwhelmed by other more important shocks, such as those stemming from unemployment. However, as shown in Table 6, we find that early-due loans are more likely to become delinquent and default than late-due loans after two years, which is consistent with a persistent liquidityreduction. Fourth, the results in Table 7 show that once a loan becomes delinquent or defaults, late-due borrowersaremorelikelyto“cure”thesedelinquenciesbymakingupmissedpaymentsandavoid default (in the first year). The regressions in columns (1) through (4) include only loans that have become delinquent or have reached default status. Column (4) shows that conditional on a borrower missing one payment, borrowers with late-due loans are 3.8 percentage points (12%) more likely to cure that delinquency than are early-due borrowers. The results are similar for the other three outcomes, consistent with the interpretation that prolonged exposure to a persistent liquidity reduction makes early-due borrowers less likely to cure first-year delinquencies. For example,supposeaborrowerexperiencesanincomeshockataloanageof6monthsandbecomes delinquent. At a loan age of 6 months, an early-due borrower has already faced property taxes whereas a late-due borrower has yet to face them. If it is more difficult for borrowers with less liquidity (i.e., post-due date) to cure a delinquency, then early-due borrowers will be less likely to cure this delinquency. If the liquidity reduction were brief rather than persistent, however, we expectearly-dueandlate-dueborrowerstohaveequalliquidityataloanageof6monthsandthus theyshouldbeequallylikelytocurethisdelinquency,whichisnotwhatwefind. To summarize, we interpret our results as driven primarily by the between-group differences inthedurationofexposuretoapersistentliquidityreductionratherthanabriefliquidityreduction with heterogeneous treatment effects by loans’ due date age. That is, early-due loans become delinquent and default more than late-due loans in their first-year because they are exposed to a 35To check our assumption that property taxes are not a surprise we estimated our regressions on a sample of refinanceloans. Borrowerswhorefinanceare,bydefinition,notfirst-timehomeownersandarethuslesslikelytobe surprisedbypropertytaxes. Ifsurprisealoneexplainsthehighdefaultrateamongearly-dueborrowerswithsubprime purchaseloans,wewouldexpecttofindnodifferenceindefaultratesforearly-dueandlate-duerefinanceloans,where the borrowers are not surprised. Instead, we find results similar to our purchase loan results. Early-due refinance borrowersdefaultatahigherratethanlate-duerefinanceborrowers,consistentwithsurprisenotplayingamajorrole inourresults. 14

low-liquiditystateforanadditionalsevento11monthsduringtheirfirstyear,notbecausetheyare treated“early.” Ourresultscontrolforduedateequity,thesurprisestoryseemsimplausible,andwehavelittlea priorireasontobelievethatdifferencesinunavoidablelowliquidityexplainourfindings. However, regardless of the precise liquidity mechanism by which property taxes increase delinquency and default,ourresultsindicatethatliquidityingeneralhasacausaleffectonmortgagedefault. 6 Conclusion We use property tax due dates to observe borrowers’ liquidity reductions at the loan-level. We exploit exogenous variation in the timing of these loan-level liquidity reductions to estimate the causal effect of liquidity reductions on mortgage delinquency and EPD for the average borrower. Prior work has been unable to estimate the causal effects of illiquidity on delinquency and default becauseavailablemeasuresofbetween-loandifferencesinliquidityareendogenous. Ourregressionresultsdemonstratethatloansfacingapropertytaxduedatewithinonetothree months after origination have at least 3% percent higher first-year delinquency and default rates than loans that face property tax due date 10 to 12 months after origination. Since we control for differences in property tax bills and borrower characteristics, we argue the most plausible mechanism for these results is the additional 7 to 11 months of exposure to a persistent post-due date liquidityreductionamongearly-dueloans. Thatis,allelseequal,moremonthsofreducedliquidity increasestheprobabilityoffirst-yeardelinquencyanddefaultasborrowersaremoresusceptibleto incomeandexpenditureshocks,suchasunemployment,furloughdays,andmedicalexpenses. Onewaytointerpretthesizeofthiseffectistocomparetheincreaseindelinquencyanddefault probabilitiesassociatedwithadditionalexposuretoreducedliquiditywithpreviousestimatesofthe effectofnegativeequity. Early-dueloansareexposedtoreducedliquidityfor3quarterslongerthan late-dueloansandare0.59percentagepointsmorelikelytobecome60-daysdelinquentduringthe first year. In contrast, the estimates in Elul et al. (2010) suggest that an increase in CLTV from 90 to 120, i.e., moving from positive to negative equity, is associated with a 1.9 percentage point increase in the probability of loans becoming at least 60-days delinquent during a year.36 Thus, the effect of three additional quarters of exposure to the post-due-date liquidity reduction is about one-thirdaslargeastheeffectofatransitiontonegativeequity. Tointerprettheresultswithrespecttothemagnitudeoftheliquidityreductionassociatedwith property taxes,consider a subprimehousehold that faceslower liquidity afterthe property taxdue date because they use their credit card to pay a property tax bill comprising 3% of their annual income. The back-end-debt-to-income-ratio (DTI) contains the minimum required monthly paymentoncreditcardbalancesinitsnumerator. Iftheminimumpaymentis2%ofthebalance,since borrowing to pay the tax bill increases the balance by 36% of monthly income, the household’s DTI increases by 0.0072. Suppose there are two identical households with the above characteristics, one that pays property taxes at 2 months, the other at age 11 months. Assuming that each household had the average pre-due date DTI of 0.40, the household that pays its property tax bill 36For this approximation, we calculate the difference in their Table 1 quarterly default estimates (d = 1.343− 0.872=0.471)andtheimpliedleveldifferenceintheannualhazardrate1−(1−d/100)4 =0.0119. Althoughthese effectsareestimatedonadifferentsampleofloans,asamplethatincludesmanyprimeloans,theestimatesusedata fromapproximatelythesameperiodasourestimates. 15

at age 2 has an average monthly first-year DTI of 0.4066, while the household that pays at age 11 months has an average monthly first-year DTI of 0.4012.37 This corresponds to a 1.3% higher average first-year DTI for the early-due household, which in turn is associated with a 3.7% increase intheprobabilityofafirst-year60-daydelinquency,oranelasticityofapproximately2.9.38 ThisdiscussionsuggestsanimportantroleforilliquidityinEPDandperhapsmortgagedefault more generally. These tax-induced liquidity reductions are much smaller in magnitude than the liquidityreductionswecannotobserveattheloanlevel,suchasthoseproducedbyunemployment, healthissues,ordivorce. Observingtheseshocksattheloan-levelmightproduceevenlargerloanlevel estimates of the effect of reduced liquidity on delinquency and default. On the other hand, borrowers with prime loans may be less sensitive to liquidity reductions since they are generally lesssensitivethansubprimeborrowerstoincomeandexpenditureshocks. Finally,givensurveyevidencethathouseholdsprefersmoothpaymentsoffinancialobligations to lump sum payments, it appears puzzling that escrow was so uncommon in the subprime mortgage market. Unlike the prime mortgage market, where Fannie, Freddie, and the FHA have long had strict escrow account guidelines, until the Federal Reserve revised the HOEPA rules in July 2008,thesubprimemortgagemarketwasdevoidofanybroad-reachingescrowaccountguidelines. The lack of escrow accounts may have been peculiar to the dramatic rise in housing prices during 2000 to 2006. Once prices began to decline, some lenders, such as Washington Mutual (now JPMorgan Chase), began requiring escrow accounts on all new subprime loans. The HOEPA rule revisions, phased in during 2010, require escrow accounts for property taxes and homeowner’s insuranceforallfirst-lien“higher-pricedmortgageloans.”39 37Thefirsthousehold hasaDTIof0.4072for 11monthsand0.40for 1month; theotherhas0.40for10 months and0.4072for2months. Thisassumesfixedmortgagepaymentsandmonthlyincome. 38Theaverageback-endDTIratioinoursampleisapproximately40%whencomputedamongtheobservationsthat providenon-missinginformation. Footeetal.(2008)andAmronin&Paulson(2009)providesimilarestimatesofDTI ratiosamongsubprimeborrowers. 39PressRelease,BoardofGovernorsoftheFederalReserveSystem,July14,2008. 16

References Agarwal, Sumit, Liu, Chunlin & Souleles, Nicholas S. (2007). ‘The Reaction of Consumer SpendingandDebttoTaxRebates-EvidencefromConsumerCreditData’,JournalofPoliticalEconomy115(6),986–1019. Amronin, Gene & Paulson, Anna. (2009). ‘Comparing Patterns of Default Among Prime and SubprimeMortgages’,EconomicPerspectives.FederalReserveBankofChicagopp.18–37. Anderson, Nathan B. & Dokko, Jane K. (2009). ‘Mortgage Delinquency and Property Taxes’, StateTaxNotes52(1),49–57. Bajari, Patrick, Chu, Sean & Park, Minjung. (2008). ‘An Empirical Model of Subprime MortgageDefaultFrom2000to2007’,NBERWorkingPaper#14625. Cabral, Marika & Hoxby, Caroline. (2010), The Hated Property Tax: Salience, Tax Rates, and TaxRevolts. unpublishedmanuscript. Carroll, Christopher D. (2001). ‘A Theory of the Consumption Function, with and without LiquidityConstraints’,JournalofEconomicPerspectives15(3),23–45. Cohen-Cole,Ethan&Morse,Jonathan.(2010),YourHouseorYourCreditCard,WhichWould YouChoose? unpublishedmanuscript. Cutts, Amy Crews & Merrill, William A. (2009). ‘Interventions in Mortgage Default: Policies andPracticestoPreventHomeLossandLowerCosts’,FreddieMacWorkingPaper#08-01. Deng, Yongheng, Quigley, John M. & van Order, Robert. (2000). ‘Mortgage Terminations, Heterogeneity,andtheExerciseofMortgageOptions’,Econometrica68(2),275–307. Elul, Ronel, Souleles, Nicholas S., Chomsisengphet, Souphala & Glennon, Dennis. (2010). ‘What“Triggers”MortgageDefault?’,AmericanEconomicReview100(2),490–494. Experian-Oliver Wyman. (2009). ‘Understanding Strategic Default in Mortgages Part I’, Experian-OliverWymanMarketIntelligenceReport2009TopicalReportSeries. Fay, Scott, Hurst, Erik & White, Michelle J. (2002). ‘The Household Bankruptcy Decision’, AmericanEconomicReview92(3),708–718. Foote, Christopher, Gerardi, Kristopher & Willen, Paul. (2008). ‘Negative Equity and Foreclosure: TheoryandEvidence’,JournalofUrbanEconomics64(2),234–245. Ghent, Andra C. & Kudlyak, Marianna. (2009). ‘Recourse and Residential Mortgage Default: Theory and Evidence from U.S. States’, The Federal Reserve Bank of Richmond Working Paper. Holland, Paul W. (1986). ‘Statistics and Causal Inference’, Journal of the American Statistical Association81(396),945–960. 17

Jacobson, Lous S., LaLonde, Robert J. & Sullivan, Daniel G. (1993). ‘Earnings Losses of DisplacedWorkers’,AmericanEconomicReview83(4),685–709. Johnson, David S., Parker, Jonathan A. & Souleles, Nicholas S. (2006). ‘Household ExpenditureandtheIncomeTaxRebatesof2001’,AmericanEconomicReview96(5),1589–1610. Keys, Benjamin. (2010). ‘The Credit Market Consequences of Job Displacement’, Finance and EconomicsDiscussionSeries,FederalReserveBoard. Mayer, Chris, Pence, Karen & Sherlund, Shane. (2009). ‘The Rise in Mortgage Defaults’, JournalofEconomicPerspectives23,27–50. McCrary, Justin. (2007). ‘The Effects of Court-Ordered Hiring Quotas on the Composition and QualityofPolice’,AmericanEconomicReview97(1),318–353. Rubin, Donald B. (1986). ‘Statistics and Causal Inference: Comment: Which Ifs Have Causal Answers’,JournaloftheAmericanStatisticalAssociation81(396),961–962. Vandell, Kerry. (1995). ‘How Ruthless is Mortgage Default? A Review and Synthesis of the Evidence’,JournalofHousingResearch6(2),245–264. 18

Table1: PropertyTaxDueDatesandLoanOriginationMonths PanelA.DistributionofPropertyTaxDueDatesin2007 Month #ofStatesw/DueDateinMonth January 4 February 7 March 4 April 6 May 7 June 2 July 0 August 2 September 4 October 9 November 7 December 6 Source: Authorstabulations. Notes. Tableonlyincludesstates32withuniformduedatesandWashingtonDC,fora totalof33. Columntotaldoesnotaddupto33becausestateswithsemi-annualinstallmentsarecountedmorethan once. Analysissamplealsoincludes7stateswithnon-uniformduedates,whereinsteadduedatesvarybycounty. PanelB.DistributionofLoanOriginationMonths,2000-2007 TypeofLoan Subprime Conforming MonthofOrigination(%) Purchase Refi Purchase Refi January 6.5 7.6 5.7 6.8 February 6.7 7.4 6.2 7.2 March 8.9 8.6 8.3 9.1 April 8.5 8.3 8.5 9.2 May 8.8 8.6 9.3 8.5 June 9.6 8.6 10.0 8.5 July 8.7 8.2 9.5 8.7 August 9.2 8.7 9.8 8.7 September 8.8 8.2 8.6 7.9 October 8.4 8.5 8.5 8.6 November 7.8 8.5 7.9 8.3 December 8.1 8.9 7.8 8.3 Source. Fornon-primeloans,CoreLogic. Notes. Forconformingloans,LPSAppliedAnalytics. Tableentriesshow, forfourtypesofloans,thepercentageofloansoriginatedineachmonthduringtheperiod2000-2007,inclusive. 19

Table2: LoanAge,OriginationMonths,andLoan/BorrowerCharacteristics,2000-2007 PanelA.AgeDistributionat1st PropertyTaxDueDate TypeofLoan: SubprimePurchase SubprimeRefi LoanAge DistributionofAgeat1st DueDate(%): 1 14.4 15.0 2 14.4 15.1 3 14.0 14.6 4 10.3 10.2 5 9.7 9.6 6 8.8 8.6 7 7.2 6.9 8 5.9 5.8 9 3.5 5.5 10 3.6 3.0 11 3.2 3.0 12 3.0 2.8 Source. CoreLogic and authors’ compilation of tax due dates. Notes. Loan age is in months. Table entries are the percentageofsampleloansthatfacetheirfirstpropertytaxduedateateachage. PanelB.Loan/BorrowerCharacteristicsbyOriginationMonth SubprimePurchaseLoans LoanAgeat1st TaxDueDate: OriginationMonth FICO CLTV %Missa1st yearpayment mean 25th p-tile 75th p-tile January 637 91.0 0.319 6.5 2 9 February 636 91.3 0.308 8 7 11 March 636 91.1 0.296 7.4 6 10 April 638 91.0 0.292 6.7 6 9 May 639 91.5 0.290 5.9 5 8 June 640 91.7 0.290 4.9 4 7 July 639 91.6 0.294 4 3 6 August 641 91.8 0.300 3.2 2 5 September 641 91.7 0.294 3.2 1 5 October 639 91.8 0.316 3.7 3 4 November 638 91.4 0.310 3.5 2 3 December 638 90.9 0.307 4.9 1 4 Sources. CoreLogicandauthors’compilationoftaxduedates. Notes. Statisticscomputedfrom20%randomsample of subprime purchase loans originated between 2000 and 2007 in 40 states. See text for additional details. Unless otherwise noted, table entries are means conditional on loans’ origination month. Percentiles (p-tile) refer to the percentilesoftheconditionaldistributionofloanageatfirstpropertytaxduedate. Foralloriginationmonthsexcept June,therangeofloanageatfirstduedateis[1,12]. 20

Table 3: Average Origination Characteristics of Subprime Purchase Loans by Number of Months Until1st PropertyTaxDueDate #MonthsUntil1st PropertyTaxDueDate: PanelA.FullSample 1-3 4-6 7-9 10+ SalePrice 163,220 142,127 142,048 100,354 FICO 641 637 636 626 CLTV 91 92 91 92 %w/FullDocumentation .576 .624 .615 .673 %w/PPPenalty .746 .791 .796 .803 InitialInterestRate 7.9 8.0 8.0 8.4 %ARM .862 .850 .841 .803 %Missa1st yearpayment .305 .306 .312 .338 #Statesobserve 40 40 31 25 PanelB.FullSample,Adjusted SalePrice 148,718 148,781 142,361 141,675 FICO 638 638 637 636 CLTV 92 92 91 91 %w/FullDocumentation .600 .6091 .615 .615 %w/PPPenalty .775 .774 .775 .773 InitialInterestRate 8.0 8.0 8.0 8.1 %ARM .849 .849 .847 .849 %Missa1st yearpayment .313 .309 .306 .305 SampleSize 203,242 137,480 91,517 48,499 Source. CoreLogicandauthors’compilationoftaxduedates. Notes. Statisticscomputedfrom20%randomsampleofsubprimepurchaseloansoriginatedbetween2000and2007 in 40 states. All characteristics are averages at origination except for % Miss a 1st year payment, which equals the share of loans that experience at least one 30-day delinquency during their first year after origination. See text for additionaldetails. 21

Table4: EvidenceonMagnitudeofPropertyTaxPaymentbyNumberofMonthsUntil1st Property TaxDueDate #MonthsUntil1st PropertyTaxDueDate: PanelA.FullSample 1-3 4-6 7-9 10+ (2004)MedianAnnualPropertyTax as%ofMedianIncome .045 .040 .039 .037 (2004)MedianPropertyTaxBill as%ofMedianIncome .024 .026 .030 .036 #Installments 2.0 1.7 1.4 1.0 Back-endDTI 41.0 40.7 40.7 39.9 Mark-to-MarketCLTV 90.0 88.9 87.3 88.0 PanelB.FullSample,Adjusted (2004)MedianAnnualPropertyTax as%ofMedianIncome .027 .027 .027 .027 (2004)MedianPropertyTaxBill as%ofMedianIncome .041 .041 .041 .041 #Installments 1.7 1.7 1.7 1.7 Back-endDTI 40.8 40.7 40.7 40.7 Mark-to-MarketCLTV 90.0 89.0 87.5 87.3 SampleSize 203,242 137,480 91,517 48,499 Source. CoreLogic and authors’ compilation of tax due dates. Data on county median property taxes and county medianincomein2004arefromthe2005AmericanCommunitySurvey. Notes. Statisticscomputedfrom20%randomsampleofsubprimepurchaseloansoriginatedbetween2000and2007 in 40 states. See text for additional details. The # of installments is the number of times per year property tax payments are due. Property taxes and income are measured at the county level. The median property tax bill is the medianannualpropertytaxdividedbythenumberofinstallments. Back-endDTIisthemortgagepayment(including escrowedinsuranceandtaxes), creditcarddebt, carloans, educationloans, andotherdebtsdividedbyincome. The back-end DTI variable is missing for 60% of observations because it was either not recorded by the lender or not reportedbytheservicer. Insomecases,thisvariableisbasedonstatedincomeratherthanverifiedincome. 22

Table5: 1st-yearDelinquencyandDefaultRatesbyTimingof1st PropertyTaxDueDate (1) (2) (3) (4) #MonthsUntil1st DueDate: PanelA.LimitedControls 1-3 4-6 7-9 10-12 Outcome Mean RegressionAdjustedDifferencew/r/tMean: 30-day .3087 -.0055*** -1.8% -.0096*** -3.1% -.0116*** -3.8% (.001) (.0017) (.0020) (.0025) 60-day .1589 -.0028** -1.7% -.0069*** -4.3% -.0072*** -4.5% (.0008) (.0013) (.0016) (.0020) 90-day .1060 -.0021* -1.9% -.0034*** -3.2% -.0045*** -4.2% (.0007) (.0011) (.0013) (.0017) FCStart .0702 -.0013 -1.9% -.0027** -3.8% -.0038*** -5.4% (.0006) (.0009) (.0011) (.0014) PanelB.FullControls Outcome Mean RegressionAdjustedDifferencew/r/tMean: 30-day .3087 -.004*** -1.3% -.0069*** -2.2% -.0101*** -3.3% (.001) (.0018) (.0021) (.0027) 60-day .1589 -.0026* -1.6% -.0052*** -3.3% -.0059*** -3.7% (.0008) (.0014) (.0016) (.0021) 90-day .1060 -.0026** -2.5% -.0022 -2.1% -.0034* -3.2% (.0007) (.0012) (.0014) (.0018) FCStart .0702 -.0019* -2.7% -.0017 -2.4% -.0035** -5.0% (.0006) (.0010) (.0011) (.0015) N=480,738. Source. CoreLogic. Note. Sampleincludessubprimeloansoriginatedforpurchasesbetween2000and 2007. Estimates obtained from logit regression and calculating average marginal effects. Standard errors obtained using delta method. Column (1) lists the average first-year default rate, for each outcome, among loans with age at due date between 1-3 months. Columns (2)-(4) list the percentage point marginal effects from logit estimation. The number to the right of each percentage point marginal effect is the marginal effect as a percentage of the mean in column (1). Limited set of control variables include fixed effects for state, origination year, and calendar month of property tax due date. Full set of control variables include these covariates as well as sales price, borrower’s FICOscore, fullv. no/lowdocumentationdummy, indicatorforadjustableratemortgage, initialinterestrate, initial combined loan-to-value ratio, mark-to-market combined loan-to-value ratio at due date, the ratio of county median propertytaxbilltocountymedianincomein2004,andfirst-yearhousepriceappreciationrate. *** indicates result statistically significant from 0 at the 1% significance level.** indicates result statistically significant from 0 at the 5% significance level.* indicates result statistically significant from 0 at the 10% significance level. 23

Table6: 2nd-yearDelinquencyandDefaultRatesbyTimingof1st PropertyTaxDueDate (1) (2) (3) (4) #MonthsUntil1st DueDate: 1-3 4-6 7-9 10-12 Outcome Mean RegressionAdjustedDifferencew/r/tMean: 30-day .4430 -.0034* -0.8% -.0087*** -2.0% -.0131*** -3.0% (.0011) (.0018) (.0022) (.0029) 60-day .2880 -.0031* -1.1% -.0086*** -2.9% -.0097*** -3.4% (.0010) (.0017) (.0019) (.0025) 90-day .2270 -.0010 -0.4% -.0074*** -3.3% -.0068*** -3.0% (.0009) (.0015) (.0018) (.0024) FCStart .1741 .0001 0.0% -.0052** -3.0% -.0059** -3.4% (.0008) (.0014) (.0017) (.0022) N=480,738. Source. CoreLogic. Note. Sampleincludessubprimeloansoriginatedforpurchasesbetween2000and 2007. Estimates obtained from logit regression and calculating average marginal effects. Standard errors obtained usingdeltamethod.Column(1)liststheaveragefirst-yeardefaultrate,foreachoutcome,amongloanswithageatdue datebetween1-3months.Columns(2)-(4)listthepercentagepointmarginaleffectsfromlogitestimation.Thenumber totherightofeachpercentagepointmarginaleffectisthemarginaleffectasapercentageofthemeanincolumn(1). Control variables include sales price, borrower’s FICO score, full v. no/low documentation dummy, indicator for adjustableratemortgage,initialinterestrate,initialcombinedloan-to-valueratio,mark-to-marketcombinedloan-tovalueratioatduedate,theratioofcountymedianpropertytaxbilltocountymedianincomein2004,two-yearhouse priceappreciationrate,andfixedeffectsforstate,originationyear,andcalendarmonthofpropertytaxduedate. ***indicatesresultstatisticallysignificantfrom0atthe1%significancelevel. **indicatesresultstatisticallysignificantfrom0atthe5%significancelevel. *indicatesresultstatisticallysignificantfrom0atthe10%significancelevel. 24

Table 7: Probability of Making Up Missed Mortgage Payments (“Curing”) During 1st Year of MortgageAmongBorrowerswithSubprimePurchaseLoans (1) (2) (3) (4) #MonthsUntil1st DueDate: 1-3 4-6 7-9 10-12 Outcome Mean RegressionAdjustedDifferencew/r/tMean: N 30-day .3090 .0022 .0186*** .0380*** 158,042 (.0019) (.0032) (.0038) (.0048) 60-day .1960 .0052 .0192*** .0306*** 76,540 (.0023) (.0040) (.0045) (.0056) 90-day .1198 .0009 .0116*** .0162*** 50,792 (.0023) (.0040) (.0045) (.0056) FCStart .1515 .0044 .0030 .0142* 32,999 (.0031) (.0054) (.0062) (.0077) Source. CoreLogic. Note. Sampleincludessubprimeloansoriginatedforpurchasesbetween2000and2007thathaveexperiencedaparticular outcome. The sample size (N) varies between the four regressions because, for example, fewer loans have experienced 90-day delinquency than 30-day delinquency. Estimates obtained from logit regression and calculating average marginal effects. Standard errors obtained using delta method. Control variables include sales price, borrower’sFICOscore,fullv. no/lowdocumentationdummy,indicatorforadjustableratemortgage,initialinterestrate, initialcombinedloan-to-valueratio,mark-to-marketcombinedloan-to-valueratioatduedate,theratioofcountymedianpropertytaxbilltocountymedianincomein2004,first-yearpercentageappreciationinhousingvalue,andfixed effectsfororiginationyear,monthofpropertytaxduedate,andstate. ***indicatesresultstatisticallysignificantfrom0atthe1%significancelevel. **indicatesresultstatisticallysignificantfrom0atthe5%significancelevel. *indicatesresultstatisticallysignificantfrom0atthe10%significancelevel. 25

TableA.1: PropertyTaxDueDatesintheUnitedStates State IncludedinAnalysis? #Installments UniformwithinState? √ AL Annual Yes √ AK MultipleVariations VariesbyBorough √ AZ Semi-annual Yes √ AR Annual Yes √ CA Semi-annual Yes √ CO Semi-annual Yes √ CT Semi-annual/Quarterly VariesbyTaxDistrict √ DE Annual Yes √ DC Semi-annual Yes √ FL Annual Yes √ GA Annual/Semi-annual VariesbyCounty √ ID Semi-annual Yes √ IL Semi-annual VariesbyCounty √ IN MultipleVariations VariesbyCounty √ IA Semi-annual Yes √ KS Semi-annual Yes √ KY Annual Yes √ LA Annual Yes √ MD Semi-annual Yes √ MN Semi-annual Yes √ MS Annual Yes √ MO Annual Yes √ MT Semi-annual Yes √ NE Semi-annual VariesbyCounty √ NV Quarterly Yes √ NJ Quarterly Yes √ NM Semi-annual Yes √ NC Annual Yes √ ND Semi-annual Yes √ OH Semi-annual VariesbyCounty √ OK Semi-annual Yes √ OR Tri-annual Yes √ SC Annual Yes √ SD Semi-annual Yes √ TN Annual Yes √ TX Annual Yes √ UT Annual Yes √ WA Semi-annual Yes √ WV Semi-annual Yes √ WY Semi-annual Yes Sources. 2008 U.S. Master Property Tax Guide, state websites, county websites, email correspondence, and telephone conversations. The ten states not in the table, and not in the analysis, have duedatesthatcanvarywithincounty. 26

Figure A.1: Event Study Results: Probability of Delinquency Relative to the 1st Property Tax Due Date 0=t t/r/w ffiD 60. 20. 20.− 40. 0 40.− 30−day Delinquency −5 0 5 Months Before/After Property Taxes are 1st Due 0=t t/r/w ffiD 20. 20.− 40. 0 40.− 60−day Delinquency −5 0 5 Months Before/After Property Taxes are 1st Due 0=t t/r/w ffiD 20. 20.− 40. 0 40.− 90−day Delinquency −5 0 5 Months Before/After Property Taxes are 1st Due 0=t t/r/w ffiD 30. 10. 10.− 20. 0 20.− Foreclosure Start −5 0 5 Months Before/After Property Taxes are 1st Due Source. CoreLogic. Note. Wetransformourdatatopaneldatawithloan-monthobservationsanddroploansoncethey become delinquent or default as is common in standard hazard analysis (thus a different set of loans is dropped in each period for each outcome we examine). Using these data, we use ordinary leastsquarestoestimatethefollowingmodel(Jacobsonetal.(1993)): 6 (cid:88) Y = β ∗1(RelTime = j) +γX +δW +(cid:15) , it j i i i it j=−6 whereY equalsoneifloaniisdelinquentordefaultsinperiodt,X isavectorofpre-determined it i loan characteristics, and W is a vector of borrower characteristics that are not pre-determined at i origination but may be correlated with a loan’s age at the property tax due date. The regression does not control for loan age. The indicator function 1(RelTime = j) equals one in period j for i loani,wherej denotestheperiodbeforeorafterpropertytaxesarefirstdue. Theerrorterm,(cid:15) ,is it assumedtobeuncorrelatedacrossloansandovertime. Thedummiesβ representthedelinquency j rate,relativetotherateatthe1st propertytaxduedate,j periods(i.e. months)beforeandafterthe firstpropertytaxduedate. The figures are consistent with our main set of results because the slope of the delinquency and defaultfunctionbecomessteeperafterthefirstpropertytaxduedate,suggestingthatthepost-duedateliquidityreductionquickensthepaceofmortgagedelinquencyanddefault. 27