Consumption, Debt and Portfolio Choice: Testing the Effect of Bankruptcy Law

CONSUMPTION, DEBT AND PORTFOLIO CHOICE (cid:3) Testing the Effect of Bankruptcy Law Andreas Lehnert Dean M. Maki Board ofGovernorsofthe Putnam Investments FederalReserve System OnePostOfficeSquare MailStop 93 BostonMA, 02109 WashingtonDC,20551 (617)760-8616 (202)452-3325 Dean Maki@putnaminv.com Andreas.Lehnert@frb.gov This version: February 2002 February 20,2002 (cid:3) The views are expressedin this paper are oursalone and do not necessarily reflectthose of the Board of Governorsor its staff or Putnam Investments. We thank DarrelCohen, Karen Dynan,RonelElul,GaryEngelhardt,DavidLaibson,NickSouleles,RobertTownsendandseminar participants at Syracuse University, the NBER 2000 Summer Institute and the Federal Reserve Board.WereceivedexcellentresearchassistancefromSigurdLund,RichardSauomaandMarcin Stawarcz.Anyremainingerrorsareourownresponsibility.

CONSUMPTION, DEBT AND PORTFOLIO CHOICE Testing the Effect of Bankruptcy Law Abstract Consumerbankruptcylaws,whichvaryacrossstatesandovertime,permitdebtors tokeepassetsbelowastatutoryexemptionwhiledebtsareforgiven. Highexemptions distort household portfolio decisions and tempt households to default on debts; but they also provide a crude form of consumption insurance. We combine information on state-level bankruptcy laws with the Consumer Expenditure Survey from 1984–1999. We find that higher exemptions are associated with (1) Higher bankruptcy rates, (2) Households that are more likely to simultaneously hold low-return liquid assets and owe high-cost unsecured debt, and (3) Slightly betterconsumptioninsurancefor renters and worse consumptioninsurancefor homeowners. Journalof EconomicLiteratureclassificationnumbers: H73,H31,K00, D1. Keywords: Bankruptcylaw,householddebt, portfoliopuzzle, consumption.

1 Introduction IntheUnitedStates,consumer(Chapter7andChapter13)bankruptcyisdesigned to provide debtors a “fresh start.” After a household successfully files a Chapter 7 petition, its unsecured debts are erased, but it must forfeit any assets above an exemption level determined by law. Although the United States constitution specifically grants the Federal government the power to set national bankruptcy law,inpracticetheselawsaremostlysetbytheindividualstates.1 Stateconsumer bankruptcy laws vary most dramatically in generosity–i.e. the exemption levels abovewhichhouseholdsforfeitassets. SomestateChapter7bankruptcycodesare extremely generous, allowing households to keep an unlimited amount of assets after bankruptcy, while others are relatively stingy, allowing households to keep, forexample,only$75in assetsafterbankruptcy. Bankruptcy laws of this type provide households with a crude form of insurance. If householdsface the possibilitythat their non-capitalincomecould dip to zero for an extended period, they would be unwilling to use unsecured debt unless somehow insured. Carroll (1992) uses the PSID to confirm that households do face such a risk, suggesting that in the absence of other forms of insurance, households would be less willing to use unsecured debt in states with less generousbankruptcylaws,althoughlenderswouldbemorewillingtoextendunsecured debtinsuchstates. Further,ahouseholdwithrelativelylargedebtswouldcutcon- 1Article I, Section 8 of the U.S. constitution states “The Congress shall have the power ... [t]o establish a uniform rule of naturalization, and uniform laws on the subject of bankruptcies throughouttheUnitedStates.” ForinformationonconsumerbankruptcylawsoutsidetheUnited States,seeAlexopoulosandDomowitz(1998)andZiegel(1997). 1

sumptionmorein responseto an incomeshockinless generousstates. Atthesametime,though,Chapter7bankruptcylawsdistorthouseholds’portfoliochoices. Householdsthatfileforbankruptcyarebetteroffiftheyhaveassets right up to the exemption level set in law. They carry these assets with them into their post-bankruptcy life, while their debts (largely) vanish. Thus it is in their interesttosimultaneouslyholdlow-returnliquidassetsevenwhiletheyhaveasignificant amountofhigh-interestdebt. Morrison(1999)and Bertaut and Haliassos (2001) have documented the existence of this anomaly using the Survey of Consumer Finances. Bertaut and Haliassos present evidence that the portfolio puzzle may be driven by self control, in which the household may be thought of as dividedbetweentwodecisionmakers,aworkerandashopper. Theworkerchooses nottopay offthehousehold’scredit cards as away to restraintheshopper.2 Finally,allelseequal,householdswillbemorelikelytodeclarebankruptcyin stateswithmoregenerousexemptions. Ofcourse,lendersmayreact toprevailing bankruptcy laws by restricting credit to borrowers living in states with generous laws, so the net effect of bankruptcy law on bankruptcy rates may go in either direction. To test these effects, we collected data on state personal bankruptcy exemptions, and other state-level information, from 1984–1999. We matched these data with household-level responses from the Consumer Expenditure Survey (CE) over the same period. The CE contains detailed information about households’ consumption, along with some information about their geographic location, de- 2SeealsoLaibson,Repetto,andTobacman(1998)andHarrisandLaibson(2001). 2

mographic characteristics, finances, income, occupation, employment and health status. The CE asks in detail about holdings of different asset classes (for example, securities, checking accounts, saving accounts, U.S. savings bonds), and less detailed information about unsecured debt. We use this portfolio information to test whether households in generous bankruptcy law states are more likely to simultaneously hold low-return liquid assets and high-interest unsecured debt. We refer tothispracticeas as “borrowingtosave.” The CE interviews the same household once per quarter for five quarters; the first interview is excluded from the public use microdata, but all other interviews are available. Thus it is possible to construct a short panel for each household, testing how consumption responds to shocks of various types. We use this informationto testtheinsuranceroleofbankruptcylaw. We find: (1) bankruptcy rates are higher in states with higher bankruptcy exemptions;(2)householdsaremorelikelytoborrowtosaveinstateswithgenerous laws;and(3)theconsumptionofrentersisslightlylesssensitivetoincomeshocks in states with generous bankruptcy laws while, by contrast, the consumption of homeowners is slightly more sensitive to income shocks in states with generous bankruptcylaws. The plan of this paper is as follows: in section 2 we briefly review the previous literature and the debate surrounding bankruptcy law, while in section 3 we provide a theoretical framework for our analysis. In section 4 we describe our data sources and the construction of the bankruptcy law database and generosity measures and in section 5 we present our results. Section 6 briefly concludes. 3

An appendix provides the results from alternative specifications of our statistical models and further discussionof our dataset. All tables and figures are located at theend ofthepaper, followingthereferences. 2 The Bankruptcy Debate Much of the current debate over bankruptcy reform is prompted by the extremely high bankruptcy filing rate in the late 1990s.3 Little is known about why the bankruptcyratehasincreased;forexample,usingadatasetofcreditcardrecords, GrossandSouleles(1998)showthattheincreaseindelinquencyratesthroughthe 1990s cannot be explained by observable factors; they posit a decrease in stigma duringthedecade as thecause. 3Thebankruptcyratealsospikedinthefirsthalfof2001.Thismostrecentincreasewaslikely a response to Congressional passage of a comprehensivebankruptcy reform bill. Although the Presidenthassaid hewouldsign thebill, asof thiswriting, a conferencecommitteehadnotyet mettoresolvethesignificantdifferencesbetweentheHouseandSenateversions. 4

Personal Bankruptcy Filings per 100,000 Persons 600 500 400 300 200 100 0 1960 1965 1970 1975 1980 1985 1990 1995 2000 Total U.S. personal bankruptcy filings (Chapter 7 and Chapter 13) per 100,000population, annualrate. Theliteratureonconsumerbankruptcy(and,byextension,consumerdebt)can be roughly divided into two camps. One the one hand, certain authors argue that householdsarenotstrategicintheiruseofbankruptcylaw,whileothersarguethat households value bankruptcy law the same way they would any other financial option, and act strategically given the actions of lenders and the prevailing law. One might characterize the former view as emphasizing the sociological aspects ofconsumption,debtandbankruptcyandthelatterasemphasizingtheireconomic aspects.4 Note that under either view, prevailing bankruptcy law can be critiqued as eithertoo laxortoopunitive. 4Thischaracterizationisabitofanexaggeration,asrecenteconomicresearchhasmovedaway fromtheprevailingmodelofaunitaryrationaldecisionmaker.SeeforexampleHarrisandLaibson (2001). 5

Sullivan, Warren and Westbrook (1989, 1997) advance a theory of consumption and bankruptcy in which a beleaguered middle class, overwhelmed by debt, uses bankruptcy as an insurance mechanism. In this view, most households are not strategic in their use of debt and bankruptcy; instead, they are seen as unfamiliar with the prevailing bankruptcy law in their state. Moreover, in this view, legal culture plays an all-important role in the outcome of a bankruptcy petition. Sullivan, Warren and Westbrook (1997), for example, compare outcomes across district courts in Pennsylvania and find wide disparities, despite having identical statutes. Further, Nelson (1999) finds that legal culture is one of the most importantfactorsaffectingwhetherhouseholdsfileunderChapter7orChapter13ofthe bankruptcycode. Thesestudiescan betaken asevidencethatchangingbankruptcylawswouldhavelittleeffectonhouseholdbehaviorsolongasother,intangible culturalfactors remained constant. Gross (1997) also explicitly rejects the rational actor model, instead arguing that most households sincerely wish to meet their obligations and avoid harming theirlocalcommunities. Fromthispremise,shesuggests,tochooseoneexample, thatbankruptcylawsoughttoaimforequalityofoutcomesratherthanequalityof treatment.5 In the context of bankruptcy law, this could mean that all households areabletosustainroughlysimilarconsumptionlevelsafterbankruptcy,nomatter howdifferenttheirdebtsbeforebankruptcy. 5Notalllegalscholarstakethisapproach;BairdandMorrison(2001)usearealoptionsframeworkto discussreformof corporaterestructuringbankruptcylaw; see also Denning, Ferris, and Lawless(2001)foranempiricalanalysisofcorporatebankruptcy. 6

However, other authors have found that households do indeed respond strategicallytotheprevailingbankruptcylaw. Forexample,usingdatacollectedbythe GAO on a random sample of bankruptcy petitions filed in 1980, Domowitz and Sartain(1999)studythedecisiontofileforbankruptcyand,giventhatahousehold has decided to file for bankruptcy, the choice of Chapter. They find that householdsaremorelikelytofileforbankruptcyafterajoblossoramedicalshock,thus buttressing the view that bankruptcy is an insurance mechanism. However, they alsofindthathouseholdsrespondtoeconomicincentivesbuiltintothebankruptcy system,attemptingto maximizethefinancial benefit to filing. The Panel Survey of Income Dynamics (PSID) follows a large panel of individuals and households over time. In 1996, a special module of the PSID asked whethereachmemberhadeverfiledforbankruptcyand,ifso,inwhichyear. Fay, Hurst,andWhite(1998)useinformationonincome,assets,anddebtinthePSID, matched with information on state bankruptcy laws, to construct measures of the potential economic benefit to declaring bankruptcy for each household.6 The authors find that households are more likely to file for bankruptcy as the financial benefit to doing so increases; and that households are less likely to file for bankruptcy as the stigma to doing so increases. The financial benefit to filing bankruptcydependsinpartonstate-levelbankruptcylaw;thustheauthorsdemonstrate some relationship between state-level bankruptcy law and households’ decisions 6ThereportedrateofbankruptciesinthePSIDiswellbelowthatintheU.S.asawhole;this maybeduetothePSID’ssampleorunder-reportingduetoshameorforgetfulness;however,the PSIDratemaybebelowthenationalratesimplybecausethenationalbankruptcystatisticscontain repeatfilingsofChapter13petitionswhichaPSIDparticipantmightrememberasasingleevent. (Chapter13filersoftenrefilefortechnicalreasons.) 7

tofileforbankruptcy. Elul and Subramanian (1999) use the PSID to determine the extent to which state bankruptcy laws influence households’ location decisions, particularly for households likely to declare bankruptcy. They find statistically significant evidence for forum shopping; that is, households more likely to declare bankruptcy aremorelikelytomovetostateswithmoregenerousbankruptcylaws. Using the 1992 Survey of Consumer Finances, White (1998)finds that a substantialproportion–atleast15%–ofhouseholdsatanygivenpointintimewould benefit financially from declaring bankruptcy. However, by filing for Chapter 7 bankruptcy,householdsgiveup therightto file again forat least six years. White calculates the option value of bankruptcy – a cost to filing – and finds that it can besubstantial.7 Finally,Olney(1999)studiestheeffectofbankruptcylawonhouseholds’consumption decisions during the Great Depression in the U.S. She compares consumption’s sensitivity to income shocks before and after a major bankruptcy law reform in the early 1930s. She finds that households cut consumption more in response to income shocks under the less generous bankruptcy regime. Thus she concludes that punitivebankruptcy laws contributed to the consumption collapse oftheearly 1930s. On balance, it appears that the evidence confirms that households do react strategically to their local bankruptcy laws. This paper is designed to fill three 7Household indebtedness rose through the mid-1990s and then leveled off, suggesting that White’s calculationsremain valid. See also Kennickell, Starr-McCluer, and Surrette (2000)and Durkin(2000). 8

remaininglacunaeintheliterature: 1. Toquantifytherelationshipbetweenstatebankruptcylaws(whicharecomplex and varying along several different dimensions) and the aggregate bankruptcy rate in a given state, allowing us to predict roughly the change inastate’sbankruptcyrateifitchanges itsexemption. 2. To determine the extent to which high bankruptcy exemptions encourage householdsto keep moneyin checking accounts earning low interestrather thanusingitto payofftheirhigh-interestcredit cards. 3. Authors who believe that households do not behave strategically and those who believe that they do agree that bankruptcy law provides an insurance mechanism. Our final goal is to test for the strength of these insurance effects. 3 Model In this section we present a simplemodel of consumptionand portfolio choice in thepresenceof Chapter7 bankruptcy. Weare interested inhow bankruptcyrates, portfolio choice and consumption risk change as the Chapter 7 asset exemption changes. In general, we document that all of these phenomena have, in theory, a non-monotone relationship to the asset exemption. At very low exemptions, few households are willing to borrow, because they face a (small) risk of suffering a low income realization. Without the protection afforded by bankruptcy law, few 9

households are willing to borrow, even at low interest rates. By contrast, at very highexemptions,thezero-profitlendingindustrywillchargeextremeinterestrates and,again, fewhouseholdswillbewillingtoborrow. However,overabroadrangeofintermediatebankruptcyassetexemptions,we showthat: 1. Bankruptcyrates areincreasing inasset exemptions. 2. Borrowing to save–holdinga gross portfolio positiondifferent from the net portfolioposition–isincreasinginasset exemptions. 3. Consumptionislesssensitivetoincomeas asset exemptionsincrease. WerefertothislattereffectasanOlneyeffect,afterOlney’s(1999)paperdemonstrating that tight bankruptcy laws played a part in the consumption collapse of theearly 1930s. Themodelwepresentherefeaturesastronginsuranceroleforunsecureddebt; agents who are unlucky today will borrow to finance consumption in order to smooth consumption across periods. Sullivan (2002) uses the PSID to document that relatively wealthy households in fact do use unsecured debt to smooth consumptionthroughtransitoryspellsofunemployment. 3.1 The Household’s Problem In our model lenders and households will live for exactly two periods. Thus we are ignoring the option value of bankruptcy. In a companion paper to this one, 10

Lehnert and Maki (2000), we explorethedynamicsof borrowingand bankruptcy inastandard permanentincomehypothesismodel.8 Householdsvalueconsumptionstreams f C 0 ; C 1 g as: U (cid:0) C 0 ; C 1 (cid:1) = C (cid:13) (cid:13) 0 + (cid:12) C (cid:13) (cid:13) 1 ; 0 < (cid:12) (cid:20) 1 ; (cid:13) < 0 : (1) Householdsarebornwithaninitialendowment Y 0 ,whichwillvaryacrosshouseholds in the economy, and then receive a second-period endowment of Y 1 , drawn from the known distribution G ( (cid:1) ) . For simplicity, we assume that the initial endowments Y 0 are also distributed as G , and that the distribution of second period endowments is independent of the first period’s endowment. We take the endowment shocks to be distributed lognormally, but with an additional risk of a very lowendowment. Thusendowment Y t , t = 0 ; 1 is: Y t = 8 > < > : y withprobability (cid:18) , and: Y t e " t withprobability 1 (cid:0) (cid:18) (2) " t (cid:24) N ( 0 ; (cid:27) 2 ) : Herewetake y ,theendowmentinthecasethatthehouseholdsuffersthe (cid:18) -shock, to be quite small. We have in mind the small fraction of households that experience near-zero incomes (including transfer income) for more than a fleeting 8Moregenerally,ourmodelisamodified(andfinite-horizon)versionofthemodelsofZeldes (1989a,b) , Deaton (1990), Hubbard, Skinner, and Zeldes (1994), Carroll(1992,1994), Carroll andSamwick(1998),andEngen(1993)tocitejustaportionoftheliteraturethatusesnumerical techniquestostudyconsumptionunderuncertainty. 11

moment.9 In the first period, t = 0 , households know their first-period income draw andchoosetheirconsumptionandportfolio. Householdswillvarytheirdecisions depending on their initial endowment draw. Note that each period’s income distributions are centered about Y t ; we will take Y 1 > Y 0 to reflect growth in the household’spotential labor incomeover time. Most households will want to borrow to smooth their increasing labor income profiles, minimizing consumption variationovertime. Households will have access to two securities in the first period of life: a low return liquid asset, a , which pays a gross return normalized to unity, and unsecured debt, d , which carries an endogenous gross interest rate of r . We will model households as choosing a face value for assets available and debt payable atthebeginningofthesecondperiodoflife,after Y 1 hasbeenobserved. Thusthe household’speriodzero budgetconstraintis: C 0 = Y 0 (cid:0) a 1 + d r 1 : (3) Inthesecondperiodoflife,thehouseholdhasliquidassetsof Y 1 + a 1 andliabilities of d 1 . It has the option of filing for Chapter 7 bankruptcy or repaying its debt. If it files for bankruptcy, all liquid assets above the exogenous exemption level X areseizedtorepaycreditors,evenasitsunsecureddebtsaredischarged. Thusthe 9See Carroll(1992)for an estimate of the extentof this phenomenonusing the PSID, or see figure9(a)forevidencefromtheConsumerExpenditureSurvey. 12

householdfaces twobudgetconstraintsinthesecond(and final) periodoflife: C NoDef 1 = Y 1 + a 1 (cid:0) d 1 ; (4) or: C Default 1 = m i n (cid:8) Y 1 + a 1 ; X (cid:9) : (5) Because we are using a two-period model, the household’ssole motivationto avoid bankruptcy is losing liquid assets above the exemption. If we imagine that householdslivedformultipleperiods,andvaluedtheoptiontodeclarebankruptcy sometimeinthefuture,aswellascontinuedaccesstounsecuredcredit(whichpresumablywouldbecut offfollowingabankruptcyproceeding),householdswould bewillingtotradetheimmediatepecuniarybenefitofbankruptcyinordertokeep their option.10 The option value of bankruptcy would thus depend on the household’sfutureprospects,which, as wehavestressed,isoutsideoftherealm ofthis model. Wecanhoweverproxyforthisfuturevaluebyspecifyinganon-pecuniary stigma penalty, denoted F (for “flogging”), assessed to households that declare 10Evidenceonhouseholds’accesstounsecuredcreditafterbankruptcyismixed.Althoughone hears anecdotes about householdsreceiving large credit lines immediately after filing for bankruptcy, surveyevidenceby Visa, Inc., finds that few filers had access to unsecuredcredit a year afterbankruptcy. 13

bankruptcy.11 Thus thehousehold’ssecond periodvaluefunctionis: W 1 ( Y 1 ; a 1 ; d 1 ) = m a x ( (cid:0) C NoDef 1 (cid:13) (cid:1) (cid:13) ; (cid:0) C Default 1 (cid:13) (cid:1) (cid:13) (cid:0) F ) : (6) Forcertaincombinationsofassets,debtand second-periodincomethehousehold may beinsolvent,that is, unabletorepay itsdebts and achievepositiveconsumption,sothat C NoDef 1 < 0 . Inthatcasewetaketheutilityofnotdeclaringbankruptcyas negativeinfinity,sothat insolventhouseholdsalwaysdeclare bankruptcy. Oneimmediateconsequenceofourtwo-periodformulationisthathouseholds will have a trigger income strategy for declaring bankruptcy. There will be a schedule of trigger incomes, Y ? ( a 1 ; d 1 ; X ; F ) , such that if the household enters theterminalperiodwithassets a 1 ,debts d 1 inasocietywithaChapter7exemption of X and stigma of F , the household will declare bankruptcy if and only if Y ? Y 1 (cid:20) . If the stigma of bankruptcy is zero, households will file for bankruptcy when it affords them the slightest increased consumption. In that case, consumption as 11Althoughwetakethisstigmatobeacombinationoftheforgoneoptionvalueofbankruptcy andthelossofeasyaccesstoconvenientpaymentsystems,therecouldalsobeacertainamountof shameassociatedwithbankruptcy. Bankruptcypetitionslistadebtor’sassets,debtsandongoing liabilities;thesepetitionsarepublic.TheactorBurtReynolds’petition,forexample,listedseveral thousanddollarsowedto the makersofhishairpieces. Further,evenaslate as the18thcentury, defaultersweresubjecttoflogging,transportationorconfinementtoasponginghouse(whichwas asunpleasantasitsounds). 14

afunctionofsecondperiod incomeisgivenby: C 1 = 8 > > > > < > > > > : Y 1 Y + 1 + X a 1 a (cid:0) 1 d 1 Y X Y 1 1 (cid:20) (cid:0) (cid:21) X a X 1 (cid:0) (cid:20) (cid:0) a Y a 1 1 1 (cid:20) + X d 1 : (cid:0) a 1 + d 1 (7) Notice that over a region of width d 1 consumption does not vary with income at all. Oneithersideofthisregion,consumptionmovesoneforonewithincome;for agents with bad income shocks, though, consumption is shifted up by an amount d 1 . This is the sense in which bankruptcy law provides a (crude) form of insurance. In thefirst periodoflife, then,thehouseholdsolvestheproblem: W 0 ( Y 0 ; r ) = a 1 m (cid:21) 0 a ;d x 1 (cid:20) d (cid:26) C (cid:13) (cid:13) 0 + (cid:12) E (cid:8) W 1 ( Y 1 ; a 1 ; d 1 ) (cid:9) (cid:27) : (8) Here consumption in period zero, C 0 , is defined in equation (3) above, and the second-period value function W 1 is defined in equation (6) above. Expectations are taken with respect to hybrid income distribution G defined in equation (2). The credit limit d is taken as exogenous; in practice we set it to the maximum possiblevaluefor Y 1 ,onthetheorythatnolenderwouldmakealoancertaintobe defaulted upon. Denote the household’s optimal choice of portfolio as a ? ( Y 0 ; r ) and d ? ( Y 0 ; r ) .12 12Althoughthe income distribution is lognormaland thus unboundedfromabove, our implementationsuseafinitesupportwithawell-definedupperbound. 15

3.2 The Lender’s Problem We model lenders as competitive entities that make zero profits in expectation. They charge a single rate r on all unsecured loans; this rate does not vary with the borrower’s initial endowment Y 0 or loan demand d 1 . Lenders will also seek to maximize borrowers’ expected utilities; in practice, this means that if lenders havemultipleinterest rates tochoosefrom,theycharge thelowest. In developingthelender’s zero-profit condition. it’suseful toconsidertheexpected profitsassociated withahouseholdwithagivenincomedraw Y 0 inperiod zero: (9) (cid:5) (cid:0) Y 0 (cid:1) = (cid:0) d ? ( Y r 0 ; r ) + d + ? ( E Y (cid:26) 0 ; Y r 1 ) + (cid:2) a P ? 1 r (cid:26) (cid:0) Y C 1 (cid:21) Y ? (cid:0) a ? 1 ; d ? 1 ; X ; F (cid:1) (cid:27) Default 1 (cid:12) (cid:12) Y 1 < Y ? (cid:0) a ? 1 ; d ? 1 ; X ; F (cid:1) (cid:27) : We do not permit lenders to charge credit spreads r that vary with the borrower’s first-period income (although a conceptually simply extension, it generates computational difficulty). Thus the lender may or may not break even on a loan to a borrower of a particular type, Y 0 ; however, across all borrowers of all types, we require the lender to make non-negativeprofits. Thus the lender’s zero-profit conditionis: Z 0 1 (cid:5) ( s ) d G ( s ) (cid:21) 0 : (10) Lenders choose an interest rate r in full knowledge of the borrowers’ problems 16

(including the initial distribution of income in the economy); certain borrowers have a higher probability of declaring bankruptcy than others but, evening out overthewholepopulation,thelenderbreaks even. 3.3 Credit Limits Givenour setup, one might suspect that householdswill pursue a strategy of borrowing as much as possible in the first period and choosing a portfolio in which a 1 = X . Thusthehouseholdisguaranteedpotentiallyunboundedconsumptionin thefirstperiodandpositive(indeed,riskless)consumptioninthesecondperiod.13 In this scenario, our exogenous credit limit d would determine the optimality of the maximal borrowing strategy. This in turn would make our theoretical model much less compelling, because we do not allow the lender to choose d . We now showthat becausewe assumethat therisk aversionparametersatisfies (cid:13) < 0 , the maximalborrowingstrategystandslittlechanceofbeingoptimal. Thusourcredit limitwillnotbebindingin equilibrium. Intuitively, the marginal utility of consumption drops so quickly that, even if consumption is unbounded, utility will still be finite. Pursuing a strategy of heavy borrowing followed by certain default may not be better than a strategy of equating expected marginal utilities across periods in the normal Euler-equation fashion. Thisresultsdependsheavilyonafunctionalformassumptionfortheutilityfunction;withotherutilityfunctionsinwhichmarginalutilitydoesnotdecline 13Thisisequivalenttoastrategyofmaxingoutone’screditcardsandusingtheproceedstobuy ahouseinFlorida(whereallhousingequityisexemptfromseizurebycreditors)beforefilingfor Chapter7bankruptcy. 17

sosharply,themaximal-borrowingstrategymaybemorelikelytobeoptimal. Notice first from the definitions of W 1 , equation (6), and consumption after default,equation(5)that W 1 satisfies: l i d ! m 1 E (cid:26) W 1 (cid:0) Y 1 ; X ; d (cid:1) (cid:27) = X (cid:13) (cid:13) (cid:0) F : This term is clearly finite. Meanwhile, examining the definition of utility in the first period of life, equation (8), we see that the maximal-borrowing strategyyields: l i d ! m 1 ( d (cid:0) (cid:13) X ) (cid:13) = 0 : Thusthemaximal-borrowingstrategygivesan expectedutilityof: X (cid:13) (cid:13) (cid:0) F : Again, this is clearly finite. There is no guarantee, moreover, that it will exceed the interior maximum for some d 1 < 1 . Figure 1 contrasts the interior optimum versus the maximal-borrowing strategy. Notice that even as borrowing goes to infinitywhenassets equaltheexemption,theinteriorsolutionis stilloptimal. 3.4 Numerical Solution We solve our model numerically. Our particular parameter values are shown in thechartbelow. In ourexperimentwevarythebankruptcyexemption X between 18

0 and 2. For exemptions above about 1.45, though, no solutions exist in which lendersmakenon-negativeexpected profits. PreferenceParameters Risk-aversion ......................... (cid:13) (cid:0) 2 : 0 Discountfactor........................ (cid:12) 0 : 9 Bankruptcystigma..................... F 0 : 1 TechnologyParameters Firstperiod meanincome............... Y 0 1 : 0 Second periodmean income............ Y 1 1 : 5 Normalincomeshockvariance.......... (cid:27) 0 : 1 Probabilityoflowshock ............... (cid:18) 0 : 0 1 5 Lowincomerealization ................ y e (cid:0) 5 NOTE. Chart shows the parameter values used in the numerical solution to themodel. We first present some figures establishing the nature of our model, and then we turn to presenting evidence of the various effects that we test in the empirical section. Infigure2weshowtheprobabilitydistributionsoverincomedrawsinthe firstandsecondperiods. Recall thatthedistributioninthefirstperiodisfixedand knownfromthepointofviewallagentsintheeconomy;furtherthesecondperiod probabilitydistributionoverincomedrawsisindependentofthefirstperiod’sand hence the same for all agents in the economy. Notice that mean income grows; because (cid:13) = (cid:0) 2 , this income growth provides a motive for households (even ones with relatively high first-period income draws) to borrow in the first period 19

tosmoothconsumption. Noticealsothedistinctmasspointatalowincomelevel; thisisthecatastrophicallylowincomeidentified byCarroll (1992)and others. Infigure3weshowthelender’sexpectedprofitscheduleasfunctionsofinterestrates r forseveraldifferentvaluesoftheexemptionlevel X . Noticethatwhen X = 0 ,thelendermakesnearlyzeroprofitsatallinterestrates;fewagentswishto borrow. At higher exemptions,by contrast, the lender cannot make positiveprofits at any interest rate. At an intermediate exemption level (marked) the lender has a choice of several interest rates that produce zero profits in expectation. In a competitivemarket, the lender chooses the utility-maximizinginterest rate (the minimuminterestrate). Clearly, both borrower welfare and the interest rates charged by lenders will vary with the exemption. As we have seen, debt with bankruptcy is a crude form ofinsurance;thuswewouldnotexpectthatzeroexemptionstobeParetooptimal. In figure 4 we display the social welfare function as a function of the exemption. The global maximum is marked with a star; note that this optimal exemption is relatively high. In the same way, in figure 5 we plot the interest rate charged by thelenderasafunctionoftheexemption. Theoptimalexemptionisagainmarked with a star. Notice that in both figures, at exemptions just above the optimal, the equilibrium quickly deteriorates. Interest rates jump and expected utilities fall. Someintuitionforthiscanbederivedfromfigure3,whichshowsthatthelender’s profit functionismulti-peaked. As thepeak associatedwithlowinterest ratefalls below zero, the lender switches to a high interest rate strategy. These extreme reactionsaretheresultofthetwo-periodnatureofthemodel;amodelwithricher 20

dynamicswouldfeaturealessextremedeteriorationinsocial welfare. Finally,wearereadytoestablishthecentralfactsofinterest. Thethreepanels infigure6demonstratetherelationshipbetweenbankruptcylawassetexemptions and (a) bankruptcy rates, (b) the percent of agents borrowing to save, and (c) the correlation of second period consumption and income (the Olney effect). All of these relationships are non-monotone in the exemption level; at relatively high levels of the exemption amount, no debt contract exists and the solution collapses to autarky, in which the household must use the low-return liquid asset as a buffer stock. At lower levels of the exemption level, in which solutions do exist, bankruptcy rates, the extent of borrowing to save and the Olney effect all exhibit a“U”-shaped relationshipwiththeexemptionlevel. 4 Data Overview To test the implications of our model, we will use three slightly different datasets. Our first database comprises annual state-level variables from 1984-1999, including income, population, average house prices, unemployment and, most importantly, our measures of each state’s Chapter 7 bankruptcy asset exemption. Weshall referto thisdatabaseas ourbankruptcylawdatabase. Our second two databases are nested subsets of the Consumer Expenditure Survey (known as the CE). The CE interviews a rotating panel of households five times with the interviews spaced three months apart. Responses to the first interview are not part of the public use micro dataset for privacy reasons, but 21

responses in all subsequent interviews (second through fifth) are. Further, for all but about 20% of observations, we know the household’s state of residence. We canthusmatchthesehouseholdswiththeprevailingbankruptcylawthattheyface. FormoredetailontheCE andtheparticularquestionsthatweused,seeappendix A.2. Our second database, which we refer to as the portfolio database, comprises those households in the CE with valid responses to the CE’s questions about portfolios. At the fifth interview, participants are asked about their holdings of financial assets and about their unsecured debts outstanding. We are particularly interestedinholdingsofliquidassets,definedastransactionaccountsplussavings accounts, whichpresumablypay arelativelylowreturn in return fortheirliquidity. Wearealsointerestedinthequantityofunsecureddebtoutstanding,forwhich lenderspresumablychargeaspreadoverthelow-risk,liquidreturnpaidonliquid assets. Ourfinaldatabase,theinsurancedatabase,comprisesthosehouseholdsinthe portfolio database for whom we have valid income and consumptionmeasures at the second and fifth interviews. Thus the insurance database exploits the short panel natureoftheCE. Formoredetailon how weconstructedtheCE databases, seeappendixA.4. Bankruptcy Law Coding bankruptcylawsisa necessarily complexprocedure. For further details on these laws and on the construction of our bankruptcy law database, see appendix A.1. In a nutshell, though, we faced the problem of gen- 22

erating a scalar value of each state’s Chapter 7 bankruptcy asset exemption level for married homeowners, single homeowners, married renters and single renters ineach year. The most crucial distinction in the law is between homeowners and renters. Homeownershaveaccesstoeachstate’shomesteadexemption,theexemptionapplied to equity in a home used as a primary residence. Homestead exemptions vary considerably, from zero in Delaware and Maryland, to explicitly unlimited in Florida, Texas and three or four other states.14 Renters, by contrast, have access onlyto astate’spersonalexemptions,whichare notonlysignificantlylower, on average, than homestead exemptions, but are also often complex and assetspecific. Note that homeowners may claim both the local homestead exemption and thepersonal exemption. States also differ in their treatment of married filers: Some allow doubling of exemptions,sometakenonoticeofafiler’smaritalstatusandothersmakespecial provisionsformarried filers. ExemptionQuartiles Becauseweareprobablymeasuringbankruptcylawwith error, we create exemption quartiles to use as our primary regressors. Moreover, even if we measured bankruptcy law perfectly, there is still the issue of how to classify states with unlimited homestead exemptions. Here, they are simply assignedtothetopquartile. Theprecisedetailsoftheconstructionofthesequartiles, as wellas descriptivestatisticsby quartile,arecontained inappendixA.3. 14Minnesotacappeditshomesteadexemptionin1993,andIowa’shomesteadexemptioncanbe aslargeas$1,000,000. 23

Wecreatedtwoclassesofquartile: theU.S.quartilestreatall51statesequally overtheentire sampleperiod. Thus in constructingthequartileranks for married homeowners (for example), we first deflated the nominal exemptions set by law for married homeowners in all states and all years and then divided the resulting 816state-yearcombinationsintofourgroupsof204state-years each. Bycontrast,CEquartilesarebasedonhouseholdsintheCE.Wematchedeach married homeowner (for example) to the real prevailing bankruptcy exemption available to it. We then produced quartiles from among this sample. Because the CE suppresses the state identifiers of all households from about eight states (for privacy reasons), it is possible that the CE is not a representative sample of the distributionof national bankruptcy law. In appendix A.3 we demonstratethat the distribution of states in the CE quartiles closely approximate the U.S. quartiles. We also show that the quartiles do not generally favor one part of our sample over another; that is, the nominal bankruptcy exemptions changed often enough topreventvariationovertimefrom beingdrivenentirelyby thedeflator. 5 Results 5.1 Testing the Effect of Law on Bankruptcy Rates We begin by showing that our constructed quartile variables explain state-level Chapter 7 bankruptcy rates. A few other papers have examined the link between bankruptcy exemptions and bankruptcy rates, including Mulligan (2001) 24

and Hynes (1998).15 Hynes (1998) (chapter 2) uses panel data from 1980–1998 on states to estimate linear probability and grouped probit models of bankruptcy rates with a variety of measures of the generosity of state-level bankruptcy laws. He finds, generally speaking, that being in the top quartile of states is associated with a higher filing rate; however, his results are of smaller magnitude than ours, and are sensitive to specification, whereas our results are both economically and statistically significant and robust to changes in specification. The main differences in our results appear to be in theeffect of thepersonal (renter) exemptions, wheremeasurementproblems(aswediscussed)aregreatest.16 Inaddition,White (1987)andNelson(2000)examinetheeffectofthe1978BankruptcyReformAct onhouseholdbankruptcyfilings. Wefindthataplausiblebankruptcyreformmeasure(although,weemphasize,nottheonecurrentlybeforeCongress)wouldlower Chapter7 filingsmorethan18 percent. Further, the result provides some assurance that we are actually measuring bankruptcy law fairly well, despite the inherent difficulties in coding the laws. For these empirical results, we cannot use the household-level datasets that we constructed, because they do not contain information on households’ bankruptcy 15Mulligan(2001)uses a cross-sectionof states in 1993andcontrolsforsocio-economicand otherlegalvariables(suchaswagegarnishment)thatdonotchangemuchovertime, butthatdo havepowerfuleffectsonbankruptcyrates. Becausewehaveconstructedapanelofstates,wecan work,inessence,withfirst-differencesbystateandignoresuchfactors. 16It is worth noting that our results are also largely unchanged when we allow for dynamic effects, such as a time trend, serialcorrelationamongthe errorsor a laggeddependentvariable. Bankruptcy filing rates might be auto-regressiveif bankruptcyhad a contagion effect, in which onehousehold’slikelihoodoffilingincreasesif itsneighborfiles. Acompletediscussionofthis subject is beyond the scope of this paper; however, it is the subject of ongoing research using census-tractleveldataandthecross-borderestimatorusedbyPence(2001). 25

decisions. Instead, we must use the purely state-level database that we constructed. Model andTests We model an individual i living in state s and year t as having a probability (cid:25) i of declaring ( p i = 1 ) or not declaring ( p i = 0 ) Chapter 7 bankruptcy. We take (cid:25) i to beproportionalto thestate’sbankruptcyexemptionquartile, unemployment rate, average real per-capita personal income growth, and house-price growth.17 Unemployment and per-capita income growth we take as measures of transitory income shocks and risk, while house price appreciation we take as a measure of permanent income growth (see Poterba 1991 for evidence that regional house prices are forward-looking). In addition, we include a full set of state and year fixed effects to control for unobserved state-level and annual variation. Sample means of the relevantstate controlsconditional on exemptionstatus are shownin table1. We do not observe the individual’s bankruptcy decision, p i , only the aggregateresultofallindividuals’decisionsinaparticularstate-yearcombination, P s ;t . Moreover, we do not know, in a given state-year, how bankruptcy filers split by housing tenure (i.e. owning vs. renting one’s residence) and marital status. Thus one can envision two separate procedures: testing the effect of each definition of 17Inaddition,wehaveexperimentedwithincludingothervariables,includingstateChapter13 bankruptcyratesandstate populationgrowthratesamongmany,manyothervariables,andnonlineartransformationsandlaggedvaluesofourexplanatoryvariables;thesehadsomeexplanatory powerbutdidnotaffectthecoefficientsofinterest. 26

theU.S. quartilesseparatelyorjointly. One potential problem with any joint test of the effect of bankruptcylaw provisionsfordifferenttypesoffilersisthatstates’quartileranksdonotdiffersignificantly by marital status. In other words, states that are generous towards single filersarelikelytobegeneroustomarriedfilersaswell. Thismakesidentifyingthe effectofmarriageprovisionsinthelawdifficult. Table2showsthedistributionof states across married homeowner quartiles conditional on their single homeowner quartile rank and their single renter rank. As the table clearly shows, there is little variation between the single and married homeowner quartiles, but significant variation between the quartiles for married homeowners and single renters. Indeed,thereappearsatfirstglancetobenorelationshipbetweenthetwoquartile rankings.18 Further,studiesofbankruptcyfilersshowthatapluralityofChapter7 filersaremarriedhomeowners,followedbysinglerenters. Inourjointstudy,then, wewillusethequartiledummiesformarriedhomeownersand singlerenters. Empirical Specification and Results Given that our data are a series of individual decisions aggregated within states, we use a grouped version of a limited dependent estimator. We use a logistic modelspecificationforease ofcomputation: l o g (cid:18) N s P ;t s (cid:0) ;t P s ;t (cid:19) = b Top Æ Top s ;t + b 2nd Æ 2nd s ;t + b 3rd Æ 3rd s ;t + X s ;t (cid:1) B + (cid:15) s ;t : (11) 18Aneasywaytochecktherelationshipistoexaminetheprincipaldiagonalsofeachmatrix;in thefirstcase,neverlessthan80%ofobservationssharethesamerank,inthesecond,nevermore than50%do. 27

Here N s ;t is the total population in state s and year t ; the variables f Æ i s ;t g are indicator variables set to unity if state s in year t is in bankruptcy exemption quartile i ; and X s ;t is a vector of other explanatory variables, which contains a full set of state and year dummy variables. The weights associated with each observation, ! s ;t ,are: ! s ;t = N s ;t P s ;t (cid:0) 1 (cid:0) P s ;t (cid:1) : Notethat ! s ;t is theinverseofthelarge-samplevarianceof (cid:15) s ;t . Coefficient estimates and robust standard errors for the four separate regressions are in table 3. Table 4 displays the same results for the joint regression, whichincludesquartilerankdummiesformarriedhomeownersandsinglerenters. Notice that bankruptcy laws do have a powerful effect on bankruptcy rates. Further,thelargestcoefficientisassociatedwiththetopquartileformarriedhomeowners, the largest group of Chapter 7 filers. Finally, the coefficient estimates generallybecomesmalleras onemovesdownexemptionquartiles. Comparing the results from the separate regressions and the joint regression, (table3versus4)weseethattheestimatedcoefficientsdonotsubstantiallychange fromonespecification totheother. PolicyExperiment With the coefficient estimates from the joint study, table 4, we can determine the effectofthefollowingpolicyexperiment: Cappinghomesteadexemptionssothat 28

all states currently in the top three quartiles are forced into the bottomquartile of homeownerexemptions,butleaving thepersonalexemptionsuntouched. Proposals of roughly this form have been floated; indeed, it is precisely the treatment of thehomesteadexemptionthatdividestheSenateandHouseversionsofthecurrent bankruptcy law reform bill. We emphasize, however, that our policy experiment is at best stylized, and is mainly intended to demonstrate the relative importance ofbankruptcylaw inourempiricalresults. Withthelogisticspecification,thepredictednumberofbankruptciesinastate inthetopexemptionquartile(formarried homeowners)is: P N b s s ;t ;t = e x p (cid:0) b Top mh 1 + e x p (cid:0) b (cid:1) e x p (cid:0) A s ;t (cid:1) Top mh (cid:1) e x p (cid:0) A s ;t (cid:1) ; where A s ;t is the net effect of all other variables and their estimated coefficients and b Top isthe coefficient on thetop quartileindicatorvariablefor married homemh owners. Imagine forcing this state into the bottom quartile of bankruptcy exemptions for married homeowners only while leaving the state’s relative rank among singlerenters untouched. Thenewpredictednumberofbankruptciesis(with b Top mh setto zero): P N e s s ;t ;t = 1 e + x p e x (cid:0) A p (cid:0) s ;t A s (cid:1) ;t (cid:1) : 29

Therelationshipof P b to P e isthus: P P e b s s ;t ;t = e x p (cid:2) (cid:0) (cid:0) b Top mh (cid:1) (cid:3) (cid:2) 1 + e x p (cid:0) b Top mh 1 + e x p (cid:1) (cid:0) e A x s p ;t (cid:0) (cid:1) A s ;t (cid:1) : Because Chapter 7 bankruptcy rates are small (no state’s bankruptcy rate ever exceedsone-halfofonepercent ofitspopulation),wecan ignorethesecondterm inthisrelation. Similarargumentscanbeusedfortheotherexemptionquartiles,replacing b Top mh with b 2nd for the second rank of states and so on. Thus, in period mh t , the predicted numberofbankruptciesfollowingsuchapolicyexperimentwouldbe: (12) P e t = = s = s X 5 1 1 ( Æ Top ; mh s ;t e x p (cid:0) (cid:0) b Top mh (cid:1) P s ;t + Æ 2nd ; mh s ;t e x p (cid:0) (cid:0) b 2nd mh + Æ (cid:1) P s ;t 3rd ; mh s ;t e x p (cid:0) (cid:0) b 3rd mh (cid:1) P s ;t ) We conduct the experiment in 1999 (the last year in our database). To conduct the experiment we need to know the aggregate number of Chapter 7 filings in 1999 by quartile, the estimated coefficients b Top mh ; b 2nd mh ; b 3rd, and their exponential mh transformations. Given these facts we can compute the filings in each quartile undertheexperimentalconditions: MarriedHomeowner ExemptionQuartile Top 2nd 3rd Bottom Chapter7 filings .............. 188,665 328,378 144,483 224,045 Coefficient: b i ............... 0.4540 0.2856 0.0803 0 mh Effect: e x p (cid:0) (cid:0) b i mh (cid:1) ............ 0.6351 0.7516 0.9228 1 Post-reform Chapter7 filings... 119,821 246,808 133,329 224,045 30

In 1999, there were 885,571 total Chapter 7 bankruptcy filings. Under the proposedreform, wepredictthattherewouldinsteadhavebeenonly724,003filings, adecrease of161,584filingsormorethan18 percent. However,ifinsteadoftighteningtheirbankruptcylaws,stateswereinsteadto loosen them, so that all states moved to the top quartile of married homeowner exemptions, the effects would be relatively larger. In that case, we predict that therewouldhavebeenapproximately250,000moreChapter7bankruptcyfilings, an increase of 29 percent. The effect of loosening exemptions is larger in part because of the pattern of bankruptcy filings across quartiles. The bottom quartile of states actually had the second-highest number of bankruptcy filings; these are thestates thatwouldbemostaffected byan increase inbankruptcyexemptions. 5.2 Testing the Portfolio Choice Effects of Bankruptcy Law As shown in section 3, bankruptcy law encourages households to simultaneously hold low-return liquid assets and high-interest unsecured debt. In this section we present evidence that households living in states with higher bankruptcy exemptionsare morelikelytoengagein thisbehavior. One of our primary sources for bankruptcy law is the attorney’s handbook by Williamson(publishedannually).19 Thesebooksexplicitlyrecommendthatbankruptcy lawyers advise their clients to convert as many of their assets as possible into exemptforms before filing for bankruptcy,a practice known as “negativeestateplanning.” Thus,onecantakeourresultshereasevidencethatnegativeestate 19See,forexample,Williamson(1999). 31

planningis morecommoninstates withmoregenerous bankruptcylaws. Definitions and Empirical Specification We first determinewhethera givenhouseholdis, in fact, “borrowingto save.” As discussed in section A.2, the CE asks about the balance on checking and savings accounts, as well as for unsecured debt. For each household, we generate an indicatorvariable,BORRSAVE, set to unityif(1)both liquidassets and unsecured debtexceedathresholdleveland,(2)liquidassetsexceed3%ofgrossincome. We vary thethreshold in thousand-dollarincrements, from $2,000 to $5,000. Sample meansfrom thedataset are presentedin table5. Our general strategy is to divide households by housing tenure (owners and renters) and estimate probit regressions of the borrowing to save indicator variable on bankruptcy law variables and a full set of controls. In addition to bankruptcy law, we included indicator variables for whether the household head was marriedinthefifthinterview,hadahighschooldiploma,acollegedegree,orwas a minority; we also included a full set of indicators for the nine different family types recorded by the CE, indicators for the month of the fifth interview (to pick up seasonal effects) and indicators for the year of the fifth interview. We also included the level and the log of real total family income before taxes, the number offamilymembers,thenumberofearners, theaverageageofthehouseholdhead and spouse and age squared. For brevity, coefficient estimates for these control variablesarenot presented. As bankruptcy law variables we used both the CE and the U.S. quartiles de- 32

scribed in section A.3. In addition, some specifications include a full set of state fixed effects; identification in these specifications comes only from states that switchexemptionquartilesoverthetimeperiod. Results In table 6 we display selected results for regressions using the $2,000 threshold for renters and the $5,000 threshold for homeowners. (Complete results for all thresholds, as well as alternate specifications, are in appendix B below.) The tableshowsresultsfromeightregressions: specificationsestimatedseparatelyfor homeownersandrenters,bothwithandwithoutafullsetofstatefixedeffectsand usingeithertheU.S. orCE quartiles. Wefindampleevidencethathouseholdsinthetopquartileofbankruptcygenerosity are more likely to borrow to save than households in the bottom quartile of bankruptcy generosity. We estimate that homeowners living in states in the top quartile of bankruptcy exemptions are between 1% and 4.5% more likely to borrow to save than homeowners living in states in the bottom quartile (the excluded category). For the full sample, the incidence of borrowing to save among homeowners at the $5,000 threshold is 7.5%. Renters living in states in the top quartileofbankruptcyexemptionsareatmost1.7%morelikelytoborrowtosave than renters living in states in the bottom quartile; at the higher thresholds the effect vanishes, suggesting that renters, who are poorer on average than homeowners, are doing their borrowing to save at a lower level. The sample incidence ofborrowingto saveamongrenters at the$2,000thresholdis aboutninepercent. 33

Robustness Tests Wehavealreadypresentedresultsfromawidevarietyofempiricalspecifications; inappendixBwepresentresultsfromallthresholds. InspectingtablesB.1through B.4, we see that for homeowners our results generally appear at all thresholds, whileforrenterstheyvanishatthehigherthresholds. Thisisnotsurprisingasthe number of renters who borrow to save in our dataset also grows extremely small at highincomes. However, so far we have not considered the possibility that bankruptcy law only affects one side of households’ balance sheets. For example, households in high-exemption states might hold more debt than those in low-exemption states. One could imagine the story going in the other direction as well; households in generousstatesmighthaverestrictedaccesstounsecuredcreditandsoholdlarger precautionary balances. Ifthe othersideof thebalancesheet is subject toenough measurementerror, morepeoplemayappeartoborrowtosaveingenerousstates. We tackle this possibility directly by testing the effect of bankruptcy law on eachsideofthebalancesheet,assetsandliabilities. Werepeatalloftheexercises fromourprimarystudyabove,exceptthatnowwereplacethedependentvariable BORRSAVE with strictly an asset or a debt version. That is, instead of testing whetherbothassetsanddebtexceedathreshold,wetestonlywhethereitherdoes. Tosavespace, weuseonlytheCE quartiles. Theresultsusingeitherpurelyadebtoranassetmeasureasadependentvariable are shown in table B.5 below. These regressions are precisely the same as the previous borrowing to save regressions, except that now the dependent vari- 34

able is set to unity ifeither assets or debts exceed theindicated threshold. Notice that none of the bankruptcy law variables are significantly different from zero. As a statistical matter, bankruptcy law appears to have no effect on households’ equilibrium debt choices. However, notice that for homeowners, the coefficients tend to be positive while for renters they tend to be negative. This accords with the findings of Gropp, Scholz, and White (1997), who documented that generous bankruptcylawstend torestrict credit tothepoor(mostlyrenters). 5.3 Testing the Insurance Role of Bankruptcy Law The theory developed in section 3 also predicts the possibility of Olney effects, the increased sensitivity of consumption to income shocks in tight bankruptcy states. InstateswithlowChapter7exemptions,lendersaremorewillingtoextend credit, which some householdsuse to bring forward consumption. As a result, ex ante they are better off, but at the cost of servicing a large ex post debt burden. Almostmechanically, such a result requires householdsto cut consumptionmore in response to shocks. Note that our theory only raises the possibility that these effects wouldappearin equilibrium;itis byno meanscertain. Definitions and Measurement Issues We can use the CE’s short-panel nature to test for the presence of these Olney effects; we know the household’s reported gross total family income as it enters the survey and at its final interview. Moreover, we know whether the household becomes too sick to work or becomes unemployed over the course of the CE. In 35

order to use the CE’s short panel nature, we must delete many observations used intheanalysisofportfoliosintheprevioussection;fulldetailsontheconstruction oftherestricted datasetare inappendix A.4. As we noted in section A.2, the CE also asks about the household’s debt outstanding at the beginning of the second interview. Thus we can split households on the basis of their outstanding debt on the eve of entering the CE. One complication with this strategy is that we do not know whether a household has zero debt because it is thrifty or is credit constrained. Unlike the Survey of Consumer Finances, the CE does not ask whether a participant has been turned down for credit, or is discouraged from borrowing. Thus we will treat the approximately 37% of the sample that report having no debt at the second interview differently thanthosethat report havingsomedebt. We define consumption as real expenditures on non-durables, excluding expenditures on educational services, health services, charitable contributions, and any housing-related expenditures, including rent, equivalent rent and imputed rents. Themajorcomponentsofthisconsumptionmeasurearefood(bothathome and out), clothing, footwear, alcohol and tobacco. This is the same consumption measureusedbyParker(1999)andDynanandMaki(2001);butisslightlybroaderthan themeasureused bySouleles (1999). Before we can analyze the effect of debt on consumption, we have to determine which households have an unusually high debt burden at the second interview. We choose two broad approaches, setting an absolute dollar threshold and setting a threshold debt to income ratio. For the absolute thresholds, we choose 36

$3,000 and $6,000; as shown in figure 7, these are about the 75th and 87th percentiles of real debt holdings. For the debt to income ratios we labeled those households with debt-to-income ratios above the median for all households with positivedebtoutstandingashighdebt. Wechoosetwomeasuresofincome: actual reported second interview income, and potential income at the second interview. Weformedpotentialincomebyregressinglogincomeonavarietyofexplanatory variables with a restricted data set that excluded those households who were too sick to work or involuntarily unemployed; we estimated the model’s coefficients separately for homeowners and renters. The relationship between actual and potential income is displayed in figure C.1. The full specification and parameter estimatesare reported inappendixC. Thus we constructed four separate indicator variables of high debt. Table 7 displays sample means for a variety of variables in each realization of the debt indicator. One arresting observation is that households with zero debts have the fastest income and consumption growth; at the same time, they have the lowest level of income, homeownership rate and are least likely to be married. These zero-debt households, however, are more likely to be minorities and less likely to have high school or college degrees. Thus we conclude that zero-debt households are in general relatively poor and high-risk. However, we shall also argue that the group of zero-debt households conceals heterogeneity, with some of the householdsbeingrelativelywealthy. Note also that households with high debt in the sense of a high absolutelevel of debt are different from households with high debt in the sense of a high ratio 37

of debt to income. These differences are robust to changes in the threshold level of debt, and to whether actual or potential income is used. Households with high levels of debt appear have higher permanent incomes than households with high ratios of debt to income. Under the level criterion, high debt households are more likely to own their homes, have college degrees and to be married; under the ratio criterion these differences are exactly reversed. In addition, while high debt households under the level criterion have much higher incomes than their low-debt counterparts, high debt households under the ratio criterion have only slightlyhigherincomes. Thisfactreassuresusthattheratiocriterionisnotdriven byhouseholdswithextraordinarilylowincomes. We can address the nature of zero-debt households, and the differences between householdssatisfyingour varioushigh-debt criteria, by plottingtheempirical distribution of income (figure 8) for each group. Three striking facts emerge from these figures. (1) The group of zero debt households does appear to contain a wide variety of household types. The income distribution of zero debt households is more or less flat among incomes from $18,000 to $75,000. (2) Using a levelcriterionproducestwosetsofhouseholdswithdissimilarincomes;high-debt households clearly have higher incomes. (3) By contrast, using a ratio criterion produces two sets of households with similarincomes. In the same vein, the particular choice of threshold, or income (for the denominator), does not seem to affect thedistributionofhouseholds. Consideringthesefacts,weusearatiomeasuretoclassifyhouseholdsas high debt or low debt. To avoid misclassifying households with temporarily low in- 38

come,weusetheratioofdebt topotentialincomeas ourmeasure. Empirical Test We wish to test whether high debt households’ consumption in tight bankruptcy statesismoresensitivetochangesinincomethanhighdebthouseholds’consumptioningenerousbankruptcystates. Asapreliminarystep,wedisplaytheempirical distributionsofincomeand consumptionchanges in figure9. Theincomeshocks have the unusual feature that they are either quite small (tightly centered around zero) or quite large, with income increasing or decreasing by more than a factor of five in about 2% of cases (marked with the grey dots in the figure). These largechangesreflecthouseholdsthatexperiencedorrecoveredfromadevastating income shock, e.g. unemployment or illness, or misreported their income at one interview. Notice that the distribution of consumption assumes a more conventionalshape,not showingtheextremepointsthatincomeshocksdo. As an initial trial of the hypothesis, we can simply calculate the average log difference in consumption for every combination of debt status and bankruptcy exemptionquartile(table9). In our analysis we want to measure the treatment effect of tight bankruptcy law on high-debt households. Clearly, households’ access to credit and desire to borrowwillinturndependonthelocalbankruptcylaw. Thustheroadtodebtwill bedifferentindifferentstates;however,thereactionofconsumptiontoanincome shock ought to allow us to measure the effect of bankruptcy law. Because consumption’s reaction to income shocks may well depend on state-specific effects, 39

wealsoincludea fullsetofstate-levelindicators. Wethusestimatetheparametersforthefollowingregressionequationforeach ofthethreedebt groupsthatwehavedefined (zero debt,lowdebtand highdebt): (13) (cid:1) l o g (cid:0) c i (cid:1) = a 0 + Æ Top i (cid:20) b Top + (cid:13) Top + Æ (cid:1) l o g (cid:0) y i (cid:1) (cid:21) 2nd i (cid:20) b 2nd + (cid:13) 2nd (cid:1) l o g (cid:0) y i (cid:1) (cid:21) + Æ 3rd i (cid:20) b 3rd + (cid:13) 3rd + Æ (cid:1) l o g (cid:0) y i (cid:1) (cid:21) Bottom i (cid:20) b Bottom + (cid:13) Bottom (cid:1) l o g (cid:0) y i (cid:1) (cid:21) + X i (cid:1) B + (cid:15) i : Wearemostinterestedinhowtheconsumptionofahighdebthouseholdreactsto incomeshocks relativeto the consumptionreaction of a low debt household. Because we will be cutting the dataset into fairly fine partitions, we provide sample sizesforall oftherelevantbinsin table10. The estimated parameters using our ratio to potential income criterion are shown in table 11. We find, broadly speaking, that the consumption of homeowners is more sensitive to income in high exemption states. For renters, by contrast,wedofindsomeevidenceOlneyeffects,withmoderate-debtrenterscuttingconsumptionmoreintightbankruptcystatesthaninlaxstates. Forhigh-debt renters this pattern breaks down, but still holds on average between the top two and bottomtwoquartiles. First, consider only high debt homeowners: their log consumptiongrowth responds at a rate of 0.0426 to their log income growth in generous bankruptcy states; at the sametime, consumptiongrowthresponds only at a rate of 0.0037 in 40

tight bankruptcy states, which is not statistically different from zero. Thus consumption growth is actually less sensitive to income growth in tight bankruptcy statesthaninloosebankruptcystates. Thepatternisthesameforlowdebthomeowners,sothatthedifferencebetweenhighandlowdebthouseholds’responseto incomegrowthdoesnotvary across statesby bankruptcygenerosity. Forrenterstheresultsaremorecomplex;perversely,lowdebthouseholdsfeel theeffectsofincomeshocksmorekeenlyintightbankruptcystates. Forhighdebt households, the average consumption response to income growth is lower in the top two quartiles than in the bottom two quartiles, but the high sensitivity in the top quartile is still evidence against an Olney effect. Relatively low debt renters thus appear to be the single group that exhibits a clear form of the Olney effect, although an argument can be made that high-debt renters also appear to suffer froma formofOlneyeffect. 6 Conclusion In this paper we analyzed the relationships among bankruptcy law, bankruptcy rates, portfoliochoiceand consumption. Weused state-leveland household-level data to test the effect of bankruptcy law variations. In particular we tested three propositionsof interest to policymakers: (1) whether loosebankruptcylaws lead to increased bankruptcy rates, (2) what effect bankruptcy laws have on household portfolios and (3) whether bankruptcy law provides an appreciable form of insuranceto households. 41

We find that generous bankruptcy laws indeed lead to increased state-level bankruptcy rates and that they also discourage households from using their lowreturn liquid assets to pay off their high interest unsecured debt bills. We find some evidence that for renters, generous bankruptcy laws do appear to protect household consumption from income shocks. For homeowners, though, if anything the opposite seems to hold true, with generous bankruptcy laws actually makingconsumptionmoresensitiveto income. Our results bear on the debate surroundingbankruptcy law in the U.S. In particular, they shed light on the differential effects of bankruptcy law on homeowners and renters. If states severely restricted their homestead exemptions, our resultssuggestthatChapter7filingswoulddecreaseby18%,theincidenceofborrowing to save would fall dramatically, and homeowners’ consumption would be lesssensitivetoincomeshocks. Atthesametime,ifstatesliberalizedtheirpersonal exemptions(the exemptionsthat affect renters directly), our evidence suggests that bankruptcy filings and borrowing to save would increase only slightly,while renters’consumptionwouldbelesssensitiveto incomeshocks. Further, our paper is one of the few that attempts to discern an effect of unsecured debt on consumption. Although much work obviously remains to be done inthisarea,ourresultscanbetakenasevidencethathighlevelsofdebtdonot,by themselves,threaten households’consumption. 42

A Data Issues A.1 State Bankruptcy Laws Bankruptcylawsarerepletewitharchaicorquirkyexemptions,andthusdifficulttocode. Asan example, wewerefaced withcoding state bankruptcy statutes that permitted filers toexempt100% ofthevalueofafamilypet,twomules, andabuckboard. Inthissection weexplainourcoding procedure andpresentsomesummarystatistics. Chapter 7 asset exemptions fall into two categories: Homestead and non-homestead exemptions. Homesteadexemptionsareappliedtotheequityinahomeusedasaprimary residence. These exemptions are especially generous in the farm states of the Midwest (and often contain explicitly higher limits for homesteads used as farms), while Atlantic coast states havenohomestead exemptions. Thestatutes creating homestead exemptions are explicit about the dollar amount protected from creditors. However, Florida, Kansas, Oklahoma, South Dakota, and Texashave unlimited homestead exemptions, as did Minnesota until 1993. Potentially, a debtor could convert millions of dollars in otherwise non-exempt assets to cash and purchase a house with the proceeds, leaving nothing for creditors. (In certain high-profile bankruptcies, this has indeed happened.) We coded these states ashaving ahomestead exemption level equal tothe largest single homestead exemption(withanexplicitlimit)amongallotherstatesinthatyear. Thismaximumstate was always Iowa, which, under certain circumstances, allows a homestead exemption of $1,000,000.20 Non-homestead(orpersonal)exemptionsaremurkier. Inadditiontoaproliferationof different classes ofassets, state lawsfrequently allow households toexempt100% ofthe value of a specific type of asset. For example, many states exempt 100% of the value of clothing for personal use. In principle, adebtor could shift assets into such categories to eludecreditors, although inpractice bankruptcy courtsmightviewsuchmassiveshiftsas abuse of the system. Wecoded these categories as being equal to the maximum allowed exemption (again, with an explicit limit) in that category among all states in each year, reasoning that this amount represented the largest acceptable amount that a bankruptcy courtwouldpermit. Some (but not all) states allow debtors to use the federal bankruptcy exemptions. In those states, if the federal exemptions were more generous, wecoded the states’ exemptionsasequaltothefederalinthatyear. Ifthestate’sownexemptionsweremoregenerous, weignoredthispossibility. Finally,somestatesallowmarrieddebtorstodoubletheirexemptionsiftheyfilejointly, while other states make explicit provisions for married filers, usually additional exemptions. Ineachstate,wecalculatedtheeffectiveexemptionsforsingleandformarried 20Thisissolargethatinourempiricalwork,weclassifiedIowaashavinganunlimitedhomesteadexemption. 44

households. Thustheexemptionsavailable tohouseholds underChapter7ofthebankruptcy code dependonthehousehold’shousingtenure(ownerorrenter)andmarriagestatus. Toreflect this,ineachstateandineachyear,weproducedfourseparatemeasuresoftheexemption level, one each for married homeowners, single homeowners, married renters and single renters. Note that homeowners have available to them all of the non-homestead exemptions,sotheseareincluded intheirexemptionmeasures. A.2 The Consumer Expenditure Survey The Consumer Expenditure Survey (CE) is a rotating panel of about 5,000 households, each of whom is interviewed five times, at three month intervals. Each calender month, about a third of the sample is interviewed. The responses to the first interview are not partofthepublic usemicrodata set(forprivacyreasons), butallinterviews thereafter are available. Duringtheinterviews,CEparticipantsareaskedaboutmonthlyperiodicexpenditures (e.g. housing, apparel, transportation, health care, insurance, and entertainment). In addition, CE participants are provided with diaries and asked to record higher-frequency expenditures (e.g. food, beverages, personalcareproductsandtobacco). The CE also contains financial information about participants. At the fifth interview only,participants areaskedabouttheirholdingsofavarietyoffinancialassets:21 (1)savings accounts, (2) checking, brokerage and other transactions accounts, (3) U.S. savings bonds, and (4) stocks, bonds, mutual funds and “other such securities.” We are particularly interested in the first two categories, which are the most liquid and lowest-return assetclasses.22 Therecorded levelsforallassetclassesaretop-coded, withthetopcodes changingovertime;before1996allseriesweretop-codedat$100,000(innominalterms). Because weform variables central to our analysis from the CE’sfinancial data, weeliminate all top-coded observations. To keep our sample consistent over time, we drop all observations with nominal values of $100,000 or greater in any year. (This had no discernibleeffectsonourpointestimates.) At the fifth interview, participants are also asked retrospective questions about their asset holdings (for each asset class) as of the first of the month one year ago. In other words, the question seeks the asset balances on the morning of the household’s first day in the CE sample. However, these questions are subject to the well-known problems of poorrecallandunderreporting ofsensitivebehaviors orevents. Forexample,participants maybeloathtoadmittoinvestment losses.23 21Thequestionasksaboutholdings“asofthefirstofthismonth.” 22DynanandMaki(2001)usetheCE’sinformationonsecuritiesholdingstostudyhowstockholders’wealthreactstostockmarketfluctuations. 23Subjectsareseparatelyaskedwhetherthebalancehasincreased,decreasedorstayedthesame; 45

Participantsareaskedabouttheirdebtsatthesecondandfifthinterviews;specifically, theyareaskedaboutthetotalamountowedtocreditorsasofthefirstdayoftheinterview month. Thus, thefirstinformation about debts isavailable three months after thefirstinformation aboutassets (derived fromtheretrospective fifthinterview questions). TheCE classifies creditors into eight categories: (1) revolving credit cards (including store, gas and general-purpose cards); (2) store installment credit accounts; (3) banks and savings and loan companies; (4) credit unions; (5) finance companies; (6) insurance companies; (7) medical practitioners not covered by insurance; and (8) other credit sources. Loans secured byhousing assets are covered elsewhere intheCE.Thesedebts cantherefore be thought of as largely unsecured. As with asset variables, before 1996, debts were topcoded at $100,000; wealso excluded any observation with top-coded debts or debts that wouldhavebeentop-coded underthemorerestrictivepre-1996 level. We do not include information on housing assets (beyond tenure status) in our analysis. Although the CE provides consistent information about whether aparticipant owns his or her primary residence, it lacks detailed information about house value and financing for much of the sample. Starting with the 1988 waves of the CE, the BLS included self-reported information on house price, length of ownership and purchase price for homeowners. Andstartingwiththe1992wavesoftheCE,theBLSincludedself-reported information on mortgage payments and principle outstanding. Even after 1992, though, thisinformation isnotavailable forallhomeowners. To be included in the sample, observations had to satisfy the following criteria: (1) haveavalidstateidentifier,(2)beacompleteincomereporter,(3)haveonlyoneconsumer unitlivingintheresidence, (4)haveavalidfifthinterview, (5)havepositivegrossfamily income, (6) have avalid age variable, and (7) have asset and debt information that isnot missingandnottop-coded. Formoredetailontheselection ofobservations fromtheCE, seesectionA.4. A.3 Constructing Exemption Quartile Ranks The CE provides limited information on the household’s geographic location, including the state of residence. Thus we can potentially match the household with the prevailing statebankruptcy law. Because of the BLS’sprivacy requirements when constructing the CE, our ability to matchislimited. Inparticular,theBLSstructurestheCEsothatthecombinedgeographic information(state,urban/ruralstatusandcitysize)cannotbeusedbyanalyststouniquely determine a geographic area of less than 100,000 persons. Thus the BLS suppresses the state code for all households living in AR, IA, ME, MS, NM, and SD. The BLS further suppresses the state codes of some households living in CA, FL, GA, IL, KS, MI, MO, theresponsestothesequestionsmaybemorerobusttotheseproblems. 46

MN,NY,NC,OK,OR,TN,VA,andWA.Finally,forasmallnumberofobservations, the BLSreplacesthehousehold’s truestateidentifierwithanotherstate’sinordertopreserve information about urban/rural status or population size; the BLSincluded information to identifysuchrecords. Wediscardallobservationswhosestateidentifierswereunavailable ortamperedwith,about21%ofthetotalsample. We divided households into four quartiles, based on the generosity of the prevailing bankruptcy law for households of that type (owner or renter, married or single), referred to as the CE quartiles. Any particular household’s quartile rank is unaffected by the state’s treatment ofhouseholds ofother types. Forexample, amarried homeowner could be classified in the top quartile while next door a single renter could be classified in the bottomquartile. Wealsoclassifiedstates(ineachyear)intoquartilesseparatelyforeachpossiblecombinationoftenureandmaritalstatus(referredtoastheU.S.quartiles.) TheU.S.quartiles are not population-weighted, and each state-year combination is weighted equally. For boththeU.S.andCEquartiles, werefer tothetopquartile astheoneassociated withthe most generous bankruptcy laws (the highest asset exemptions) and the bottom quartile as the one associated with the least generous bankruptcy laws (the lowest asset exemptions). The CE and U.S.quartiles will differ because the CE (as we have seen) excludes eightstatesfromitsgeographically identifiedsampleandbecausetheCEquartiles divide households evenly. The U.S. quartiles, by contrast, use all 51 states and are not population weighted. Table A.3.1 gives means of state-level characteristics conditional on U.S. bankruptcy exemption quartile. Characteristics such as unemployment rates, population, and per-capita income do not appear to vary systematically with bankruptcy exemption quartile; surprisingly, though, neither do bankruptcy filing rates. The analysis in section 5.1, though, demonstrates that there is a causal link from bankruptcy law to Chapter 7 filingrates. For our study to be a fair test of the effect of prevailing bankruptcy laws, the CE quartiles should do a good job of approximating the U.S. quartiles. We compare these quartiles intwoways;(1)bycomparing theirquartile rankcutoffsand(2)bydistributing households from the CE survey into their appropriate U.S. quartile. First, figure A.3.1 displays the empirical cumulative distribution functions of exemptions (by tenure and marital status) for both the U.S. and the CE quartiles; table A.3.2 lists the associated dollarvalues. Second,tableA.3.3showshowCEhouseholds wouldbedistributed across U.S.quartiles, conditional ontheirCEquartiles. ExaminingfigureA.3.1andtableA.3.2,weseethatthedifferencesinquartilecutoffs are small (less than 5% of the U.S. exemption level); and that the CE quartile cutoffs (markedineachpanelonthefigure)alsodoagoodjobofsortingtheU.S.exemptions. Examining table A.3.3 we see that the bulk of CE households are located along the principal diagonals; in other words, they would have the same quartile rank whether the U.S.orCEquartileswereused. 47

We construct our quartiles treating each year-state (for the U.S.) or month-state (for the CE) combination as a separate observation. However, we are also deflating the exemptions faced by households, while the nominal exemptions (set by law) are changing relatively infrequently. Thusitcould bethat weareclassing allstates atthebeginning of oursample(forexample)asthemostgenerousandallstatesattheendofoursampleasthe least generous. In table A.3.4 we display the distribution of married homeowner households from the CE across quartiles and years (the distributions for other tenure-marital statuscombinations arequitesimilar). TableA.3.5displaysthedistribution ofU.S.states inoursampleacrossyearsandquartiles(usingtheexemptionsformarriedhomeowners). 48

FIGURE A.3.1: EmpiricalcumulativedistributionsofCE andU.S. quartiles Cumulative Empirical Distribution 100 U.S. CE 75 50 Married Homeowners: N(CE)=8663 25 0 31 59 115 tnecreP Cumulative Empirical Distribution 100 U.S. CE 75 50 Single Homeowners: N(CE)=3072 25 0 15 30 70 Bankruptcy Asset Exemption (real $000s) (a) Marriedhomeowners tnecreP Bankruptcy Asset Exemption (real $000s) (b)Singlehomeowners Cumulative Empirical Distribution 100 Married Renters: N(CE)=2348 75 50 25 CE U.S. 0 12 20 30 tnecreP Cumulative Empirical Distribution 100 Single Renters: N(CE)=3489 75 50 25 CE U.S. 0 6 10 15 Bankruptcy Asset Exemption (real $000s) (c)Married renters tnecreP Bankruptcy Asset Exemption (real $000s) (d)Singlerenters NOTE. Figure gives the empirical cumulative distributions of bankruptcy asset exemption laws for households in the CE and for states. CE quartile breaks are marked. 49

TABLE A.3.1: Mean forU.S. States, 1984–1999 BankruptcyExemptionQuartile(U.S.) (Married Homeowners) Bot. 3rd 2nd Top Variable Chapter 7FilingRatea........................... ( 0 0 : : 2 1 2 1 ) ( 0 0 : : 1 1 8 1 ) ( 0 0 : : 2 1 2 1 ) ( 0 0 : : 2 1 3 2 ) Chapter 13FilingRate a .......................... ( 0 0 : : 1 1 2 3 ) ( 0 0 : : 0 0 8 9 ) ( 0 0 : : 0 0 6 5 ) ( 0 0 : : 0 0 6 6 ) Population(millions)............................ ( 4 3 : : 5 2 6 7 ) ( 5 4 : : 5 9 1 2 ) ( 5 7 : : 3 6 5 5 ) ( 4 5 : : 4 1 8 0 ) Per-capita incomeb .............................. 2 ( 3 3 : : 5 9 8 4 ) 2 ( 4 4 : : 5 4 3 0 ) 2 ( 5 4 : : 1 9 8 1 ) 2 ( 4 4 : : 5 7 8 3 ) UnemploymentRate a ............................ ( 5 1 : : 6 9 4 5 ) ( 6 1 : : 1 7 3 5 ) ( 6 1 : : 1 9 4 8 ) ( 5 1 : : 3 5 3 6 ) NOTE. Tablegivesmeansofselectedstate-levelvariablesfromadatasetofall51statesovertheyears1984–1999; standarddeviationsare inparentheses. aPercent. bThousandsofreal1996dollars. 50

TABLE A.3.2: QuartileRank Cutoffs Quartile Bottom 3rd 2nd Top min max min max min max min max Marriedhomeowners (ex. unlim. exemptions) U.S. 13.0 31.1 31.4 61.1 61.4 130.7 131.6 503.4 CE 11.9 31.2 31.3 59.2 59.3 114.6 114.7 471.0 Singlehomeowners (ex. unlim. exemptions) U.S. 8.1 16.3 16.4 31.2 31.3 70.7 71.2 251.7 CE 7.5 15.4 15.4 29.3 29.4 68.8 69.9 235.5 Marriedrenters (allstates) U.S. 2.3 13.2 13.2 20.2 20.2 30.4 30.5 99.0 CE 2.2 12.6 12.7 20.0 20.1 31.1 31.2 92.3 Singlerenters (allstates) U.S. 1.2 6.7 6.7 10.4 10.4 15.2 15.3 49.5 CE 1.1 6.1 6.1 9.7 9.7 14.6 14.7 46.2 NOTE. Table gives the range of real bankruptcy exemptions in each quartile (for the indicated debtor type) in two different data sets. The first comprises all 51 U.S. states weighted equally, the second comprises the CE sample. Exemptions are in thousands of real 1996 dollars, deflated by the non-durables consumption deflator. 51

TABLE A.3.3: DistributionofhouseholdsacrossCE andU.S. quartiles ExemptionQuartiles Married Single Homeowners U.S. U.S. CE Bot. 3rd 2nd Top Bot. 3rd 2nd Top Bot. 3,520 565 0 0 1,577 126 0 0 3rd 0 3,737 377 0 12 1,653 63 0 2nd 0 0 4,053 0 0 6 1,363 309 Top 0 0 350 3,706 0 0 0 1,700 Renters U.S. U.S. CE Bot. 3rd 2nd Top Bot. 2nd 2nd Top Bot. 1,405 25 0 0 2,395 0 0 0 3rd 0 1,263 168 0 77 1,168 25 0 2nd 0 0 995 458 0 76 1,896 423 Top 0 0 0 1,395 0 0 0 2,368 NOTE. TablecomparestheCEandU.S.bankruptcyexemptionquartilesbyshowingthedistributionofCEhouseholdsamongU.S.quartileranks. TheCEquartiles are formed from the states in the CE; the U.S. quartiles use all 51 states. See text forfurtherinformation. 52

TABLE A.3.4: DistributionofMarried HomeownersinCE ExemptionQuartiles Bot. 3rd 2nd Top Total Year 1984 30 127 76 43 276 1985 138 438 357 180 1,113 1986 223 434 311 255 1,223 1987 304 383 316 312 1,315 1988 274 319 243 260 1,096 1989 290 320 254 262 1,126 1990 307 343 190 326 1,166 1991 282 318 262 287 1,149 1992 379 236 202 274 1,091 1993 424 143 246 313 1,126 1994 257 169 362 312 1,100 1995 257 280 214 250 1,001 1996 272 228 170 356 1,026 1997 301 215 288 287 1,091 1998 276 118 427 260 1,081 1999 71 43 135 79 328 Total 4,085 4,114 4,053 4,056 16,308 NOTE. Table gives the distribution of observations (for married homeowners) from the CE across years and quartile exemption ranks. Note that there is no particularbias towardsonequartileoranotherovertime. 53

TABLE A.3.5: DistributionofU.S. States ExemptionQuartiles Bot. 3rd 2nd Top Total Year 1984 10 19 12 10 51 1985 11 17 13 10 51 1986 11 17 11 12 51 1987 11 18 10 12 51 1988 13 15 10 13 51 1989 14 15 11 11 51 1990 15 14 10 12 51 1991 16 12 10 13 51 1992 15 13 10 13 51 1993 13 12 13 13 51 1994 13 6 20 12 51 1995 13 7 18 13 51 1996 13 8 17 13 51 1997 14 12 10 15 51 1998 12 8 15 16 51 1999 10 11 14 16 51 NOTE. Table gives the distribution of U.S. states across exemption quartiles (ranked bytheexemptionsformarried homeowners)and overtime. 54

A.4 Constructing the CE Microdata Sample Ourinitialsamplegivesusover100,000observationstobeginwith. However,werequire that observations satisfy several criteria in order to be included in the final data sets. We havetwosetsofcriteria: abasesetandanadditionalseriesofextrarestrictions. Thebase set of criteria comprise the bare minimum required for an observation to be included in ourborrowing tosaveanalysis insection 5.2above. Thatis,theserestrictions aremerely enough to guarantee that we can form the dependent and control variables. In addition, we impose an extra set of restrictions to implement the consumption analysis in section 5.3above. Thustobeincluded intheborrowingtosaveanalysis, recordshadtohave: 1. Avalidstateidentifier. 2. Exactlyoneconsumerunit(family)intheresidence. 3. Avalidfifthinterview. 4. Anon-missing valueforhousehold head’sage. 5. Positiveandnon-missing incomeatthefifthinterview. 6. Non-missingvaluesforcheckingandsavings accounts. 7. Non-topcoded checking, savings accountanddebtvalues. In addition, to be included in the consumption risk analysis, in section 5.3 above, observationshadtosatisfyafurthersetofcriteria. Inparticular, recordshadtohave: 1. Avalidsecondinterview. 2. Alogincome difference between 1/3and3;aswellaspositive incomeatboth the secondandthefifthinterview. 3. Non-topcoded incomeinboththesecond andfifthinterviews. 4. Agebetween21and70. 5. Completedebtinformation atthesecond interview. 6. Second interview non-durable consumption could not be below $1,000 (in real 1996 dollars) or greater than $20,000; in addition, second interview consumption hadtoexist. 7. Fifthinterview consumption hadtomeetthesamecriteria. FromexaminingtableA.4.1below,weseethatthemostcommonreasonsforeliminatinga recordfromourdatasetsaremissinginterviews,amissingstateidentifierormissingasset, debtorincomeinformation. Wedideliminatesomerecordsbecausetheirresponsesmade us suspect misreporting; this is absolutely standard in the CE literature, see for example Parker(1999)orDynanandMaki(2001). 55

TABLE A.4.1: ConstructionofDataset from theCE NetLoss Remaining Condition Owners Renters Total Initialsamplea 51,538 63,855 115,393 Initialset ofrestrictions Invalidstate 25,758 39,937 50,573 90,510 Duplicaterecord 1,165 38,787 50,558 89,345 MultipleCU 6,904 38,411 44,030 82,441 No 5thinterview 22,872 38,411 21,158 59,569 Missingage 237 38,260 21,072 59,332 Bad 5th interviewincome 8,826 32,392 18,114 50,506 Bad 5th interviewassets 10,566 24,517 15,423 39,940 Topcoded assets 1,547 23,141 15,252 38,393 Topcoded debts 39 23,117 15,237 38,354 Additionalsetof restrictions No 2ndinterview 10,322 18,878 9,154 28,032 Consumptionviolationb 191 18,774 9,067 27,841 Incomeviolationc 1,741 17,665 8,435 26,100 Incometopcoded 2,755 15,099 8,246 23,345 Age > 70or < 21d 2,994 13,084 7,267 20,351 Bad 2nd interviewchecking 2,152 11,677 6,522 18,199 Bad 2nd interviewsaving 2,403 10,019 5,777 15,796 Bad 2nd interviewcons.e 78 10,002 5,716 15,718 Bad 5th interviewcons.e 53 9,973 5,692 15,665 aEveryonewho is notreportedas a homeowner,includingthose with missing or incomplete tenurerecordsandthoseneitherowningnorrentingtheirresidences,arehereclassedas“renters.” bLogconsumptiondifferenceexceeded3orwaslessthan1/3. cSecondorfifthinterviewincomenegative,zeroormissing. dAgedefinedastheaverageofhouseholdheadandspouse(ifpresent). eConsumptionlessthan$1,000orgreaterthan$20,000ormissing. 56

B Further Portfolio Results We use four different thresholds for the borrowing to save variable; we display the estimated coefficients and cluster-adjusted standard errors for each threshold level in table B.1($2,000threshold), tableB.2($3,000threshold), tableB.3($4,000threshold) andtableB.4($5,000threshold). Eachtableshowsresultsfromeightregressions: specifications estimatedseparately forhomeownersandrenters,bothwithandwithoutafullsetofstate fixed effects and using either the U.S. or the CE quartiles. In some states, no household everborrowstosave;wedropped allobservations fromsuchstates. 57

TABLE B.1: ResultsofBorrowingto SaveModel: $2,000Threshold ProbabilityDerivatives @ F = @ x HomeownersOnly RentersOnly Quartile Quartile CE U.S. CE U.S. Top ( 0 0 : : 0 0 2 1 7 2 3 4 ) ( 0 0 : : 0 0 1 2 2 3 8 3 ) ( 0 0 : : 0 0 2 1 6 3 4 0 ) ( 0 0 : : 0 0 0 2 7 4 5 4 ) ( 0 0 : : 0 0 0 0 5 4 4 6 ) ( 0 0 : : 0 0 0 1 2 3 0 7 ) ( 0 0 : : 0 0 0 0 7 4 8 1 ) ( 0 0 : : 0 0 1 1 7 4 3 7 ) 2nd ( 0 0 : : 0 0 3 1 7 0 3 7 ) ( 0 0 : : 0 0 1 2 0 0 8 8 ) ( 0 0 : : 0 0 3 1 4 1 4 2 ) ( 0 0 : : 0 0 1 2 1 1 2 2 ) ( 0 0 : : 0 0 0 0 0 4 0 5 ) ( 0 0 : : 0 0 0 1 5 1 1 8 ) (cid:0) ( 0 0 : : 0 0 0 0 4 4 9 9 ) ( 0 0 : : 0 0 0 1 8 2 4 8 ) 3rd X 2 ( ( 0 0 : : 0 0 1 1 0 1 6 1 ) (cid:0) ( 0 0 : : 0 0 0 1 2 3 9 3 ) ( 0 0 : : 0 0 0 1 3 2 0 2 ) (cid:0) ( 0 0 : : 0 0 1 1 8 5 4 1 ) (cid:0) ( 0 0 : : 0 0 0 0 3 4 2 4 ) (cid:0) ( 0 0 : : 0 0 0 0 3 9 2 7 ) (cid:0) ( 0 0 : : 0 0 0 0 0 4 4 6 ) ( 0 0 : : 0 0 0 1 4 1 5 0 ) State ) 112.43 115.42 52.68 54.89 ( 0 : 0 0 0 0 ) ( 0 : 0 0 0 0 ) ( 0 : 0 8 6 4 ) ( 0 : 0 3 7 4 ) States 43 43 43 43 43 41 43 41 Successes 4,788 4,788 4,788 4,788 1,361 1,361 1,361 1,361 Observations 23,117 23,117 23,117 23,117 15,237 15,231 15,237 15,231 NOTE. TablegivesprobabilityderivativesfromaprobitregressionofBORRSAVE(withacutoffof$3,000)against explanatory variables, including bankruptcy exemption quartiles. CE quartiles are formed from the CE sample while U.S. quartiles are formed from the U.S. states over the sample period (see text for details). Some state dummy variables predict failure perfectly; all such states are dropped for regressions using state fixed effects. For regressionswithoutstatefixedeffects,standarderrorsarecorrectedforclustering;robust,cluster-adjusted,standard errors arein parentheses. 58

TABLE B.2: ResultsofBorrowingto SaveModel: $3,000Threshold ProbabilityDerivatives @ F = @ x HomeownersOnly RentersOnly Quartile Quartile CE U.S. CE U.S. Top ( 0 0 : : 0 0 1 1 6 1 7 7 ) ( 0 0 : : 0 0 2 2 0 0 5 1 ) ( 0 0 : : 0 0 1 1 7 2 0 2 ) ( 0 0 : : 0 0 2 2 2 1 9 0 ) ( 0 0 : : 0 0 0 0 6 4 7 2 ) ( 0 0 : : 0 0 0 1 7 0 9 3 ) ( 0 0 : : 0 0 0 0 7 4 4 0 ) ( 0 0 : : 0 0 1 1 7 1 6 2 ) 2nd ( 0 0 : : 0 0 2 0 6 8 1 7 ) ( 0 0 : : 0 0 1 1 7 7 6 9 ) ( 0 0 : : 0 0 2 0 5 9 1 5 ) ( 0 0 : : 0 0 2 1 6 8 8 2 ) (cid:0) ( 0 0 : : 0 0 0 0 0 3 2 5 ) ( 0 0 : : 0 0 0 0 9 8 3 9 ) (cid:0) ( 0 0 : : 0 0 0 0 3 3 5 6 ) ( 0 0 : : 0 0 1 0 0 9 3 6 ) 3rd X 2 ( ( 0 0 : : 0 0 0 0 9 8 9 1 ) ( 0 0 : : 0 0 0 1 4 1 6 3 ) ( 0 0 : : 0 0 0 0 6 8 6 5 ) ( 0 0 : : 0 0 0 1 3 2 2 9 ) (cid:0) ( 0 0 : : 0 0 0 0 0 4 2 2 ) ( 0 0 : : 0 0 0 0 5 7 1 5 ) ( 0 0 : : 0 0 0 0 0 4 8 5 ) ( 0 0 : : 0 0 0 0 8 8 4 4 ) State ) 96.88 98.80 57.95 54.89 ( 0 : 0 0 0 0 ) ( 0 : 0 0 0 0 ) ( 0 : 0 2 0 1 ) ( 0 : 0 3 7 4 ) States 43 43 43 43 43 39 43 39 Successes 3,344 3,344 3,344 3,344 860 860 860 860 Observations 23,117 23,117 23,117 23,117 15,237 15,196 15,237 15,196 NOTE. TablegivesprobabilityderivativesfromaprobitregressionofBORRSAVE(withacutoffof$3,000)against explanatory variables, including bankruptcy exemption quartiles. CE quartiles are formed from the CE sample while U.S. quartiles are formed from the U.S. states over the sample period (see text for details). Some state dummy variables predict failure perfectly; all such states are dropped for regressions using state fixed effects. For regressionswithoutstatefixedeffects,standarderrorsarecorrectedforclustering;robust,cluster-adjusted,standard errors arein parentheses. 59

TABLE B.3: ResultsofBorrowingto SaveModel: $4,000Threshold ProbabilityDerivatives @ F = @ x HomeownersOnly RentersOnly Quartile Quartile CE U.S. CE U.S. Top ( 0 0 : : 0 0 1 1 2 0 3 6 ) ( 0 0 : : 0 0 4 1 2 8 8 9 ) ( 0 0 : : 0 0 1 1 2 0 3 8 ) ( 0 0 : : 0 0 3 1 4 8 9 9 ) ( 0 0 : : 0 0 0 0 2 3 3 2 ) (cid:0) ( 0 0 : : 0 0 0 0 0 6 0 8 ) ( 0 0 : : 0 0 0 0 2 3 9 1 ) ( 0 0 : : 0 0 0 0 6 7 1 7 ) 2nd ( 0 0 : : 0 0 1 0 3 6 6 4 ) ( 0 0 : : 0 0 3 1 5 6 1 5 ) ( 0 0 : : 0 0 1 0 1 6 3 8 ) ( 0 0 : : 0 0 3 1 0 6 5 0 ) (cid:0) ( 0 0 : : 0 0 0 0 1 2 8 8 ) ( 0 0 : : 0 0 0 0 0 5 8 8 ) (cid:0) ( 0 0 : : 0 0 0 0 4 2 0 8 ) ( 0 0 : : 0 0 0 0 2 6 4 4 ) 3rd X 2 ( ( 0 0 : : 0 0 0 0 3 6 6 0 ) ( 0 0 : : 0 0 1 1 7 0 5 2 ) ( 0 0 : : 0 0 0 0 1 6 8 4 ) ( 0 0 : : 0 0 1 1 7 1 0 3 ) (cid:0) ( 0 0 : : 0 0 0 0 1 3 9 4 ) ( 0 0 : : 0 0 0 0 0 5 0 2 ) (cid:0) ( 0 0 : : 0 0 0 0 0 3 7 5 ) ( 0 0 : : 0 0 0 0 4 6 2 2 ) State ) 107.49 106.33 46.74 44.24 P r > X 2 0 : 0 0 0 0 0 : 0 0 0 0 0 : 1 3 0 9 0 : 1 9 2 4 States 43 43 43 43 43 38 43 38 Successes 2,372 2,372 2,372 2,372 578 578 578 578 Observations 23,117 23,117 23,117 23,117 15,237 15,169 15,237 15,169 NOTE. TablegivesprobabilityderivativesfromaprobitregressionofBORRSAVE(withacutoffof$4,000)against explanatory variables, including bankruptcy exemption quartiles. CE quartiles are formed from the CE sample while U.S. quartiles are formed from the U.S. states over the sample period (see text for details). Some state dummy variables predict failure perfectly; all such states are dropped for regressions using state fixed effects. For regressionswithoutstatefixedeffects,standarderrorsarecorrectedforclustering;robust,cluster-adjusted,standard errors arein parentheses. 60

TABLE B.4: ResultsofBorrowingto SaveModel: $5,000Threshold ProbabilityDerivatives @ F = @ x Homeowners Only RentersOnly Quartile Quartile CE U.S. CE U.S. Top ( 0 0 : : 0 0 0 0 9 7 2 8 ) ( 0 0 : : 0 0 4 1 4 7 8 6 ) ( 0 0 : : 0 0 0 0 7 7 3 9 ) ( 0 0 : : 0 0 2 1 7 6 3 0 ) (cid:0) ( 0 0 : : 0 0 0 0 0 2 2 7 ) (cid:0) ( 0 0 : : 0 0 0 0 2 4 4 7 ) ( 0 0 : : 0 0 0 0 0 2 3 5 ) ( 0 0 : : 0 0 0 0 3 5 1 8 ) 2nd ( 0 0 : : 0 0 0 0 8 4 0 4 ) ( 0 0 : : 0 0 2 1 9 4 2 4 ) ( 0 0 : : 0 0 0 0 6 4 3 3 ) ( 0 0 : : 0 0 2 1 2 3 2 3 ) (cid:0) ( 0 0 : : 0 0 0 0 0 2 3 1 ) (cid:0) ( 0 0 : : 0 0 0 0 1 4 6 3 ) (cid:0) ( 0 0 : : 0 0 0 0 4 1 4 9 ) (cid:0) ( 0 0 : : 0 0 0 0 0 4 1 6 ) 3rd X 2 ( ( 0 0 : : 0 0 0 0 5 4 2 7 ) ( 0 0 : : 0 0 2 0 0 9 7 2 ) ( 0 0 : : 0 0 0 0 1 4 6 6 ) ( 0 0 : : 0 0 1 0 1 9 5 3 ) (cid:0) ( 0 0 : : 0 0 0 0 1 2 0 6 ) ( 0 0 : : 0 0 0 0 0 4 4 2 ) (cid:0) ( 0 0 : : 0 0 0 0 0 2 3 6 ) ( 0 0 : : 0 0 0 0 3 4 4 9 ) State ) 90.96 87.06 45.82 42.11 P r > X 2 0 : 0 0 0 0 0 : 0 0 0 0 0 : 1 0 4 3 0 : 1 9 0 2 States 43 42 43 42 43 36 43 36 Successes 1,735 1,735 1,735 1,735 401 401 401 401 Observations 23,117 23,092 23,117 23,092 15,237 15,051 15,237 15,051 NOTE. TablegivesprobabilityderivativesfromaprobitregressionofBORRSAVE(withacutoffof$5,000)against explanatory variables, including bankruptcy exemption quartiles. CE quartiles are formed from the CE sample while U.S. quartiles are formed from the U.S. states over the sample period (see text for details). Some state dummy variables predict failure perfectly; all such states are dropped for regressions using state fixed effects. For regressionswithoutstatefixedeffects,standarderrorsarecorrectedforclustering;robust,cluster-adjusted,standard errors arein parentheses. 61

TABLE B.5: Probitson PureAssetorDebt Measures Threshold Quartile $2,0000 $3,0000 $4,0000 $5,0000 PureDebtMeasure StateFixedEffectsIncluded HomeownersOnly TopQuartile........... ( 0 0 :0 :0 2 7 3 9 7 5 ) ( 0 0 :0 :0 3 8 8 2 6 5 ) ( 0 0 :0 :0 1 8 7 5 9 8 ) ( 0 0 :0 :0 4 8 4 9 1 8 ) SecondQuartile ....... ( 0 0 :0 :0 1 7 0 0 6 7 ) ( 0 0 :0 :0 0 7 7 3 3 5 ) ( 0 0 :0 :0 1 7 1 6 0 7 ) (cid:0) ( 0 0 :0 :0 1 8 0 0 6 4 ) ThirdQuartile......... ( 0 0 :0 :0 1 4 0 6 6 2 ) ( 0 0 :0 :0 0 4 7 8 3 1 ) ( 0 0 :0 :0 1 5 1 0 0 2 ) (cid:0) ( 0 0 :0 :0 1 5 0 2 6 8 ) RentersOnly TopQuartile........... (cid:0) ( 0 0 :0 :1 8 0 5 1 0 7 ) (cid:0) ( 0 0 :0 :1 2 0 2 5 4 8 ) (cid:0) ( 0 0 :0 :1 8 1 9 2 6 3 ) (cid:0) ( 0 0 :1 :1 2 1 9 7 7 3 ) SecondQuartile ....... (cid:0) ( 0 0 :0 :0 7 8 2 5 1 6 ) (cid:0) ( 0 0 :0 :0 1 8 7 8 5 5 ) (cid:0) ( 0 0 :1 :0 0 9 2 4 3 1 ) (cid:0) ( 0 0 :1 :0 2 9 6 7 1 9 ) ThirdQuartile......... (cid:0) ( 0 0 :0 :0 7 7 2 6 1 6 ) (cid:0) ( 0 0 :0 :0 1 7 7 9 5 1 ) (cid:0) ( 0 0 :1 :0 0 8 2 4 3 3 ) (cid:0) ( 0 0 :1 :0 2 8 6 7 1 6 ) StateFixedEffectsIncluded HomeownersOnly TopQuartile........... ( 0 0 :0 :0 2 7 3 9 7 5 ) ( 0 0 :0 :0 3 8 8 2 6 5 ) ( 0 0 :0 :0 1 8 7 5 9 8 ) ( 0 0 :0 :0 4 8 4 9 1 8 ) SecondQuartile ....... ( 0 0 :0 :0 1 7 0 0 6 7 ) ( 0 0 :0 :0 0 7 7 3 3 5 ) ( 0 0 :0 :0 1 7 1 6 0 7 ) (cid:0) ( 0 0 :0 :0 1 8 0 0 6 4 ) continuedonnextpage 62

continuedfrompreviouspage Threshold Quartile $2,0000 $3,0000 $4,0000 $5,0000 ThirdQuartile......... ( 0 0 :0 :0 1 4 0 6 6 2 ) ( 0 0 :0 :0 0 4 7 8 3 1 ) ( 0 0 :0 :0 1 5 1 0 0 2 ) (cid:0) ( 0 0 :0 :0 1 5 0 2 6 8 ) RentersOnly TopQuartile........... (cid:0) ( 0 0 :0 :1 8 0 5 1 0 7 ) (cid:0) ( 0 0 :0 :1 2 0 2 5 4 8 ) (cid:0) ( 0 0 :0 :1 8 1 9 2 6 3 ) (cid:0) ( 0 0 :1 :1 2 1 9 7 7 3 ) SecondQuartile ....... (cid:0) ( 0 0 :0 :0 7 8 2 5 1 6 ) (cid:0) ( 0 0 :0 :0 1 8 7 8 5 5 ) (cid:0) ( 0 0 :1 :0 0 9 2 4 3 1 ) (cid:0) ( 0 0 :1 :0 2 9 6 7 1 9 ) ThirdQuartile......... (cid:0) ( 0 0 :0 :0 7 7 2 6 1 6 ) (cid:0) ( 0 0 :0 :0 1 7 7 9 5 1 ) (cid:0) ( 0 0 :1 :0 0 8 2 4 3 3 ) (cid:0) ( 0 0 :1 :0 2 8 6 7 1 6 ) AssetsOnly StateFixedEffectsIncluded HomeownersOnly TopQuartile........... (cid:0) ( 0 0 :0 :0 2 8 9 2 6 3 ) ( 0 0 :0 :0 2 8 0 0 3 2 ) ( 0 0 :0 :0 1 7 3 9 9 7 ) (cid:0) ( 0 0 :0 :0 5 7 5 9 7 8 ) SecondQuartile ....... (cid:0) ( 0 0 :0 :0 4 7 8 2 4 1 ) (cid:0) ( 0 0 :0 :0 0 7 5 0 6 6 ) (cid:0) ( 0 0 :0 :0 1 7 4 0 2 6 ) (cid:0) ( 0 0 :0 :0 4 7 8 0 0 9 ) ThirdQuartile......... (cid:0) ( 0 0 :0 :0 4 4 8 6 4 9 ) (cid:0) ( 0 0 :0 :0 0 4 5 6 6 1 ) (cid:0) ( 0 0 :0 :0 1 4 4 6 2 3 ) (cid:0) ( 0 0 :0 :0 4 4 8 6 0 8 ) RentersOnly TopQuartile........... (cid:0) ( 0 0 :0 :1 2 0 1 8 9 5 ) (cid:0) ( 0 0 :0 :1 2 1 0 4 8 1 ) ( 0 0 :0 :1 1 2 2 0 0 4 ) (cid:0) ( 0 0 :0 :1 0 2 6 3 8 7 ) SecondQuartile ....... ( 0 0 :0 :0 4 9 4 1 0 2 ) ( 0 0 :0 :0 1 9 6 6 5 7 ) ( 0 0 :0 :1 8 0 0 2 1 4 ) ( 0 0 :0 :1 4 0 9 4 3 8 ) ThirdQuartile......... ( 0 0 :0 :0 4 8 4 1 0 3 ) ( 0 0 :0 :0 1 8 6 7 5 2 ) ( 0 0 :0 :0 8 9 0 2 1 9 ) ( 0 0 :0 :0 4 9 9 5 3 4 ) StateFixedEffectsNotIncluded continuedonnextpage 63

continuedfrompreviouspage Threshold Quartile $2,0000 $3,0000 $4,0000 $5,0000 HomeownersOnly TopQuartile........... (cid:0) ( 0 0 :0 :0 2 8 9 2 6 3 ) ( 0 0 :0 :0 2 8 0 0 3 2 ) ( 0 0 :0 :0 1 7 3 9 9 7 ) (cid:0) ( 0 0 :0 :0 5 7 5 9 7 8 ) SecondQuartile ....... (cid:0) ( 0 0 :0 :0 4 7 8 2 4 1 ) (cid:0) ( 0 0 :0 :0 0 7 5 0 6 6 ) (cid:0) ( 0 0 :0 :0 1 7 4 0 2 6 ) (cid:0) ( 0 0 :0 :0 4 7 8 0 0 9 ) ThirdQuartile......... (cid:0) ( 0 0 :0 :0 4 4 8 6 4 9 ) (cid:0) ( 0 0 :0 :0 0 4 5 6 6 1 ) (cid:0) ( 0 0 :0 :0 1 4 4 6 2 3 ) (cid:0) ( 0 0 :0 :0 4 4 8 6 0 8 ) RentersOnly TopQuartile........... (cid:0) ( 0 0 :0 :1 2 0 1 8 9 5 ) (cid:0) ( 0 0 :0 :1 2 1 0 4 8 1 ) ( 0 0 :0 :1 1 2 2 0 0 4 ) (cid:0) ( 0 0 :0 :1 0 2 6 3 8 7 ) SecondQuartile ....... ( 0 0 :0 :0 4 9 4 1 0 2 ) ( 0 0 :0 :0 1 9 6 6 5 7 ) ( 0 0 :0 :1 8 0 0 2 1 4 ) ( 0 0 :0 :1 4 0 9 4 3 8 ) ThirdQuartile......... ( 0 0 :0 :0 4 8 4 1 0 3 ) ( 0 0 :0 :0 1 8 6 7 5 2 ) ( 0 0 :0 :0 8 9 0 2 1 9 ) ( 0 0 :0 :0 4 9 9 5 3 4 ) NOTE. Table gives coefficient estimates from probit regressions in which the dependent variable is an indicator variable set to unity if the household’s unsecured debt or asset holdings exceed the indicated thresholds. The regressors include the household’s bankruptcy exemption (shown) and the same set of controls used in the borrowing to save regressions (suppressed). 64

C Potential Income Regression To construct our estimate of the potential income that a household could earn, we first restrictedourCEsamplefromsection5.3tothosehouseholds whowerenotunemployed or too sick to work. Among the remaining households, weregressed log total household incomeasreportedatthesecondinterviewondemographiccharacteristicsandasetofcohortdummies. (Theexcludedcohortcomprisesthosehouseholdsborninorbefore1929.) We also included some household characteristics, such as marital status and the number of earners. We estimated the coefficients separately for renters and homeowners. The coefficient estimates and robust standard errors are shown in table C.6. Most coefficient estimates are roughly the same for owners and renters, with the notable exception of the cohort dummies, which are well below the levels for homeowners. Presumably, being a renterdespitehavingbeenborninthe1950sisassociated withlowerincome. Figure C.1 compares the distributions (empirical pdfs) of actual and predicted incomes. Recallthatwepurposely excludedallhouseholds sufferingsevereincomeshocks (unemployment orillness) from our sample, thus potential income is skewed to the right relativetoactualincome. TABLE C.6: PotentialIncomeRegressions Variable Owners Renters Constant .................................... ( 8 0 :9 :1 0 6 2 8 1 8 ) ( 8 0 :2 :2 1 0 5 9 5 0 ) Age......................................... ( 0 0 :0 :0 3 0 5 9 3 4 ) ( 0 0 :0 :0 4 0 7 8 5 7 ) Age 2 ........................................ (cid:0) ( 0 0 :0 :0 0 0 0 0 4 1 ) (cid:0) ( 0 0 :0 :0 0 0 0 0 5 1 ) Black....................................... (cid:0) ( 0 0 :1 :0 6 2 2 8 4 3 ) (cid:0) ( 0 0 :1 :0 8 3 8 1 1 5 ) College ..................................... ( 0 0 :6 :0 1 5 2 3 2 1 ) ( 0 0 :6 :0 5 4 6 2 2 3 ) HighSchool................................. ( 0 0 :3 :0 4 4 6 9 3 1 ) ( 0 0 :3 :0 2 3 7 3 7 9 ) Married..................................... 0 :2 8 1 1 0 :2 1 6 2 continuedonnextpage 65

continuedfrompreviouspage Variable Owners Renters ( 0 :0 1 7 0 ) ( 0 :0 2 3 0 ) NumberofEarners........................... ( 0 0 :2 :0 0 1 5 1 8 4 ) ( 0 0 :3 :0 5 1 1 5 8 4 ) CohortDummyVariables Cohort2(1930 (cid:20) YearBorn (cid:20) 1939)........... ( 0 0 :0 :0 5 3 5 6 2 8 ) (cid:0) ( 0 0 :1 :0 1 6 2 6 6 6 ) Cohort3(1940 (cid:20) YearBorn (cid:20) 1949)........... ( 0 0 :0 :0 5 4 4 8 7 8 ) (cid:0) ( 0 0 :0 :0 3 9 8 0 1 1 ) Cohort4(1950 (cid:20) YearBorn (cid:20) 1959)........... ( 0 0 :1 :0 3 6 0 2 2 9 ) ( 0 0 :0 :0 1 9 5 8 0 7 ) Cohort5(1960 (cid:20) YearBorn (cid:20) 1969)........... ( 0 0 :1 :0 2 7 6 4 4 9 ) ( 0 0 :0 :1 3 1 4 5 4 8 ) Cohort6(YearBorn (cid:21) 1970).................. (cid:0) ( 0 0 :0 :0 1 9 5 0 0 0 ) (cid:0) ( 0 0 :0 :1 4 3 6 1 1 0 ) R 2 .......................................... 0 :2 6 5 2 0 :3 0 6 2 Observations................................. 9,000 5,235 NOTE. Table gives regression results of log of second interview income on explanatory variables;robust,cluster-adjusted standarderrorsareinparentheses. Regressionexcluded allhouseholds involuntarily unemployed ortoosicktowork. 66

FIGURE C.1: Actualand estimatedpotentialincome 8 Potential 7 6 5 4 3 2 1 Actual 0 1 10 50 100 Real Gross Total Family Income ($000s) tnecreP Empirical Distributions 8 7 6 Potential 5 4 3 2 1 Actual 0 1 10 50 100 Real Gross Total Family Income ($000s) (a)HomeOwners tnecreP Empirical Distributions (b) Renters 8 7 6 Potential 5 4 3 2 1 Actual 0 1 10 50 100 Real Gross Total Family Income ($000s) tnecreP Empirical Distributions (c) All NOTE. Figures give the probability distributions (pdfs) of actual and estimated potential total gross household familyincome. Theregressionsused toproducepotentialincomeareshowninappendixC. 67

References Alexopoulos, M. and I. Domowitz (1998). Personal liabilities and bankruptcy reform: Aninternationalperspective.InternationalFinance1(1),127–159. Baird, D. G. and E. R. Morrison (2001, June). Bankruptcy decision making. University of Chicago, John M. Olin Law & Economics Working Paper Series #126;forthcomingJournalof Law, EconomicsandOrganization. Bertaut, C. C. and M. Haliassos (2001). Debt revolvers for self-control. Manuscript,Federal ReserveBoard, Washington,D.C. Carroll, C. D. (1992). Buffer stock saving: Some macroeconomic evidence. BrookingsPaperson EconomicActivity1992(2),61–156. Carroll, C. D. (1994, February). How does future income affect current consumption? QuarterlyJournalof Economics109(1),111–47. Carroll, C. D. and A. A. Samwick (1998). How important is precautionary saving? Reviewof EconomicsandStatistics80(3),410–19. Deaton, A. (1990). Saving in developing countries: Theory and review. In Proceedings of the World Bank Annual Conference on Develo pment Economics1989.TheWorld Bank. Denning,K. C., S. P. Ferris, and R. M. Lawless(2001,February). Serial bankruptcy: Planinfeasibilityorjustbadluck? AppliedEconomicsLetters8(2), 105–109. Domowitz, I. and R. L. Sartain (1999). Determinants of the consumer bankruptcydecision.Journalof Finance54(1),403–20. Durkin, T. A. (2000). Credit cards: Use and consumer attitudes, 1970–2000. FederalReserveBulletin86(1), 623–634. Dynan, K. and D. M. Maki (2001). Does stock market wealth matter for consumption? FinanceandEconomicsDiscussionSeries2001-23,FederalReserveBoard. Elul,R.andN.Subramanian(1999).Forum-shoppingandpersonalbankruptcy. Brown UniversityWorkingPaper #99-1. Engen, E. (1993). Consumptionand saving in alife-cyclemodel withstochastic earnings and uncertain lifespan. Manuscript, Federal Reserve Board, WashingtonD.C. 68

Fay, S., E. Hurst, and M. White (1998, January). The bankruptcy decision: Doesstigmamatter? DepartmentofEconomicsWorkingPaper98-01,UniversityofMichigan. Gropp,R.,J.K.Scholz,andM.J.White(1997).Personalbankruptcyandcredit supplyand demand.QuarterlyJournalofEconomics112(1), 217–51. Gross, D. B. and N. S. Souleles (1998). Explaining the increase in bankruptcy and delinquency: Stigma versus risk-composition. The Wharton School, UniversityofPennsylvania. Gross, K. (1997). Failure and Forgiveness: Rebalancing the Bankruptcy System. NewHaven, CT: YaleUniversityPress. Harris, C. and D. Laibson (2001, July). Dynamic choices of hyperbolic consumers.Econometrica69(4), 935–57. Hubbard, R. G., J. Skinner, and S. P. Zeldes (1994, May). Expanding the lifecycle model: Precautionary saving and public policy. American Economic Review84(2),174–79. Hynes,R.(1998).ThreeEssaysonExemptionsandConsumerBankruptcy.Ph. D. thesis,UniversityofPennsylvania. Kennickell,A.B.,M.Starr-McCluer,andB.J.Surrette(2000).Recentchanges in U.S. family finances: Results from the 1998 Survey of Consumer Finances. FederalReserve Bulletin86(1), 1–29. Laibson,D.I.,A.Repetto,andJ.Tobacman(1998).Self-controlandsavingfor retirement. BrookingsPapers onEconomicActivity1998(1),91–172. Lehnert, A. and D. M. Maki (2000). The great American debtor. Manuscript, Federal ReserveBoard ofGovernors,Washington,DC. Morrison, A. K. M. (1999). Two Essays on Consumer Credit. Ph. D. thesis, UniversityofChicago. Mulligan,J.V.(2001,May).Theimpactofstateexemptionsonpersonalbankruptcyfilings.MA thesis,VirginiaPolytechnicalInstitute. Nelson, J. P. (1999). Consumer bankruptcy and chapter choice: State panel evidence. ContemporaryEconomicPolicy17(4),552–566. Nelson,J.P.(2000).Consumerbankruptciesand thebankruptcyreformact: A time-seriesinterventionanalysis,1960-1997.JournalofFinancialServices Research 17(4),181–200. 69

Olney, M. L. (1999). Avoiding default: The role of credit in the consumption collapseofthe1930s.QuarterlyJournalof Economics114(1),319–35. Parker,J.A.(1999,September).Thereactionofhouseholdconsumptiontopredictablechangesinsocialsecuritytaxes.AmericanEconomicReview89(4), 959–73. Pence, K.(2001).Foreclosingonopportunity? Statelawsandmortgagecredit. Manuscript,Federal ReserveBoard, WashingtonDC. Poterba, J. M. (1991). House price dynamics: The role of tax policy and demography.BrookingsPaperson EconomicActivity0(2), 143–83. Souleles, N. (1999, September). The response of household consumption to incometax refunds.AmericanEconomicReview89(4),947–58. Sullivan, J. X. (2002). Borrowing during unemployment: Unsecured debt as a safety net. Department of Economics, Northwestern University, Evanston IL. Sullivan, T. A., E. Warren, and J. L. Westbrook (1989). As We Forgive Our Debtors: Bankruptcy and Consumer Credit in America. New York, NY: Oxford UniversityPress. Sullivan, T. A., E. Warren, and J. L. Westbrook (1997). Consumer bankruptcy in the United States: A survey of alleged abuse and of local legal culture. JournalofConsumerPolicy20(2), 223–68. Visa U.S.A. Inc. (1998, September). 1997 Annual Bankruptcy Survey. Visa consumerbankruptcyreports, VisaU.S.A. Inc. White, M. (1998, March). Why don’t more households file for bankruptcy. DepartmentofEconomicsWorkingPaper98-03, UniversityofMichigan. White,M.J.(1987).Personalbankruptcyunderthe1978bankruptcycode: An economicanalysis.IndianaLaw Journal63(1), 1–57. Williamson, J. H. (1999). The Attorney’s Handbook on Consumer Bankruptcy and Chapter13. Lakewood,Colorado: ArgylePublishingCompany. Zeldes,S.P. (1989a,April).Consumptionandliquidityconstraints: Anempirical investigation.JournalofPoliticalEconomy97(2),305–46. Zeldes, S. P. (1989b, May). Optimal consumption with stochastic income: Deviations from certainty equivalence. Quarterly Journal of Economics104(2), 275–98. 70

Ziegel, J. S. (1997). Canadian perspectives on the challenges of consumer bankruptcies.Journalof ConsumerPolicy20(2),199–221. 71

72

TABLE 1: State-levelSampleMeans U.S. QuartileRank Bot. 3rd 2nd Top Variable MarriedHomeowners Exemptiona................ 24,580 40,648 91,779 212,703 (4,471) (7,802) (20,222) (90,140) Chapter 7Rate............. ( 0 0 : : 2 1 2 1 ) ( 0 0 : : 1 1 8 1 ) ( 0 0 : : 2 1 2 1 ) ( 0 0 : : 2 1 3 2 ) Incomegrowth............. ( 2 2 : : 4 1 6 5 ) ( 2 2 : : 9 4 1 2 ) ( 2 2 : : 3 3 4 9 ) ( 2 3 : : 4 5 3 8 ) Unemploymentrate......... ( 5 1 : : 6 9 5 5 ) ( 6 1 : : 1 7 3 5 ) ( 6 1 : : 1 9 5 8 ) ( 5 1 : : 3 5 3 6 ) Housepricegrowth......... ( 1 3 : : 3 9 7 8 ) ( 2 5 : : 4 5 1 0 ) ( 1 5 : : 7 9 9 7 ) ( 0 4 : : 5 5 4 0 ) SingleRenters Exemption................. 5,201 8,676 12,928 23,696 (1,258) (1,089) (1,583) (9,113) Chapter 7Rate............. ( 0 0 : : 2 1 3 0 ) ( 0 0 : : 1 1 7 1 ) ( 0 0 : : 2 1 2 2 ) ( 0 0 : : 2 1 3 2 ) Incomegrowth............. ( 2 2 : : 5 5 4 3 ) ( 2 3 : : 6 5 8 6 ) ( 2 2 : : 4 1 5 6 ) ( 2 2 : : 4 3 7 0 ) Unemploymentrate......... ( 5 1 : : 4 7 0 3 ) ( 6 2 : : 0 0 9 0 ) ( 6 1 : : 0 7 9 0 ) ( 5 1 : : 6 8 6 6 ) Housepricegrowth......... ( 1 3 : : 6 5 9 7 ) ( 1 6 : : 7 4 1 5 ) ( 1 4 : : 8 9 4 6 ) ( 0 4 : : 8 8 8 9 ) NOTE. Table gives sample means and standard deviations from a sample of 51 states over the 16 years 1984-1999, conditionalon the state’s exemptionquartile. Notethatstatesswitchquartiles,sothatanyonestate’sdatamaybepresentinthe resultsformanydifferentquartiles. aForstateswithfiniteexemptionsonly. 73

TABLE 2: Relation ofU.S. quartiles SingleHomeowners SingleRenters Bottom 3rd 2nd Top Bottom 3rd 2nd Top MarriedHomeowners MarriedHomeowners Bottom 189 15 0 0 Bottom 105 60 38 1 3rd 15 169 20 0 3rd 40 85 51 28 2nd 0 12 170 21 2nd 13 38 60 92 Top 0 8 13 184 Top 46 25 52 82 NOTE. Table gives cross-tabulations of U.S. quartiles. Note that the quartiles for single and married homeowners are quite similar, while those for married homeownersand singlerenters (thetwo largestgroups ofdebtors)arevery different. 74

TABLE 3: Separate EffectsofBankruptcyLaw onChapter 7 BankruptcyRates Exemptions Homeowners Renters Married Single Married Single Variable Top quartile......... ( 0 0 : : 4 0 9 5 8 5 8 7 ) ( 0 0 : : 1 0 5 5 0 6 6 1 ) ( 0 0 : : 1 0 4 5 7 0 2 2 ) ( 0 0 : : 1 0 5 4 7 7 5 3 ) 2nd quartile......... ( 0 0 : : 3 0 2 4 5 6 7 6 ) ( 0 0 : : 2 0 9 4 6 1 8 3 ) ( 0 0 : : 1 0 1 4 7 9 9 9 ) ( 0 0 : : 1 0 3 4 2 5 1 9 ) 3rd quartile ......... ( 0 0 : : 0 0 9 2 4 8 1 5 ) ( 0 0 : : 0 0 9 2 0 6 4 6 ) ( 0 0 : : 0 0 2 4 5 1 5 4 ) ( 0 0 : : 0 0 2 4 0 0 9 8 ) Incomegrowtha ..... (cid:0) ( 0 0 : : 0 0 0 0 4 4 8 6 ) (cid:0) ( 0 0 : : 0 0 0 0 1 3 6 9 ) (cid:0) ( 0 0 : : 0 0 0 0 6 4 1 6 ) (cid:0) ( 0 0 : : 0 0 0 0 5 4 4 5 ) UnemploymentRate. ( 0 0 : : 0 0 7 1 6 0 6 8 ) ( 0 0 : : 0 0 8 0 0 9 5 9 ) ( 0 0 : : 0 0 7 1 8 1 1 1 ) ( 0 0 : : 0 0 8 1 0 1 1 0 ) House-pricegrowth.. F ( (cid:0) ( 0 0 : : 0 0 0 0 6 2 9 7 ) (cid:0) ( 0 0 : : 0 0 0 0 8 2 5 3 ) (cid:0) ( 0 0 : : 0 0 0 0 6 3 5 0 ) (cid:0) ( 0 0 : : 0 0 0 0 6 3 3 0 ) State ) b 134.54 137.87 122.97 123.84 F ( Year ) c 141.96 186.89 140.71 142.79 NOTE. Tablegivescoefficientestimatesfromestimatingequation(11)usingdata from51 statesfrom 1984–1999;robuststandarderrors are inparentheses. aStateaveragerealper-capitaincomegrowth. b F (cid:0) teststatisticofhypothesisthatallstatedummiesarejointlyequaltozero. c F (cid:0) teststatisticofhypothesisthatallyeardummiesarejointlyequaltozero. 75

TABLE 4: JointEffect ofHomeownerandRenter Exemptions Coefficient Robust Estimate Std. Err. Variable MarriedHomeowner Quartiles Top..................................... 0 : 4 5 4 0 0 : 0 7 0 5 2nd..................................... 0 : 2 8 5 6 0 : 0 5 9 9 3rd...................................... 0 : 0 8 0 3 0 : 0 2 8 8 SingleRenter Quartiles Top..................................... 0 : 0 8 4 6 0 : 0 4 6 2 2nd..................................... 0 : 0 6 6 5 0 : 0 4 5 1 3rd...................................... 0 : 0 3 1 0 0 : 0 3 7 6 Real per-capitaincomegrowth............. (cid:0) 0 : 0 0 5 3 0 : 0 0 4 3 Unemploymentrate ...................... 0 : 0 7 8 5 0 : 0 1 0 6 Real house-pricegrowth .................. F ( (cid:0) 0 : 0 0 6 4 0 : 0 0 2 8 State ) a................................ 130.50 F ( Year ) b................................ 146.46 NOTE. TablegivescoefficientestimatesandrobuststandarderrorsfromaweightedleastsquaresregressionofChapter7bankruptcyratesontheU.S.quartiledummies for both married homeowners and single renters (the two largest groups of bankruptcyfilers). a F (cid:0) teststatisticofhypothesisthatallstatedummiesarejointlyequaltozero. b F (cid:0) teststatisticofhypothesisthatallstatedummiesarejointlyequaltozero. 76

TABLE 5: Samplemeans fromCE portfoliodatabase Tenure All Rent Own Variable Age ................................. 45 40 49 (16) (13) (15) Familysize........................... 2.74 2.48 2.92 (1.57) (1.60) (1.52) —Real1996— Unsecured debt....................... 3,081 2,408 3,525 (6,720) (6,031) (7,103) Liquidassets ......................... 8,446 3,727 11,556 (15,897) (10,176) (18,073) Grosstotalfamilyincome(annual) ..... 45,369 29,695 55,700 (36,391) (23,791) (39,407) —Percent— Married.............................. 57 37 71 Black................................ 13 20 9 Highschoolgraduate.................. 54 55 54 Collegegraduate...................... 24 19 28 Borrowingtosave(variousthresholds) $2,000............................... 16.0 8.9 20.7 $3,000............................... 11.0 5.6 14.5 $4,000............................... 7.7 3.8 10.3 $5,000............................... 5.6 2.6 7.5 Observations......................... 38,354 15,237 23,117 NOTE. Table gives sample means from the subset of the CE used in estimating the propensity of borrowing to save in section 5.2 of the text. See appendix A.4 fortheconstructionofthisdataset. 77

TABLE 6: Selected resultsfrom theborrowingtosavemodel ProbabilityDerivatives @ F = @ x Renters: $2,000Threshold Homeowners: $5,000Threshold Quartile Quartile CE U.S. CE U.S. Top.......... ( 0 0 : : 0 0 0 0 5 4 4 6 ) ( 0 0 : : 0 0 0 1 2 3 0 7 ) ( 0 0 : : 0 0 0 0 7 4 8 1 ) ( 0 0 : : 0 0 1 1 7 4 3 7 ) ( 0 0 : : 0 0 0 0 9 7 2 8 ) ( 0 0 : : 0 0 4 1 4 7 8 6 ) ( 0 0 : : 0 0 0 0 7 7 3 9 ) ( 0 0 : : 0 0 2 1 7 6 3 0 ) 2nd.......... ( 0 0 : : 0 0 0 0 0 4 0 5 ) ( 0 0 : : 0 0 0 1 5 1 1 8 ) (cid:0) ( 0 0 : : 0 0 0 0 4 4 9 9 ) ( 0 0 : : 0 0 0 1 8 2 4 8 ) ( 0 0 : : 0 0 0 0 8 4 0 4 ) ( 0 0 : : 0 0 2 1 9 4 2 4 ) ( 0 0 : : 0 0 0 0 6 4 3 3 ) ( 0 0 : : 0 0 2 1 2 3 2 3 ) 3rd.......... X 2 ( (cid:0) ( 0 0 : : 0 0 0 0 3 4 2 4 ) (cid:0) ( 0 0 : : 0 0 0 0 3 9 2 7 ) (cid:0) ( 0 0 : : 0 0 0 0 0 4 4 6 ) ( 0 0 : : 0 0 0 1 4 1 5 0 ) ( 0 0 : : 0 0 0 0 5 4 2 7 ) ( 0 0 : : 0 0 2 0 0 9 7 2 ) ( 0 0 : : 0 0 0 0 1 4 6 6 ) ( 0 0 : : 0 0 1 0 1 9 5 3 ) State ) .... 52.68 54.89 90.96 87.06 P r > X 2 0 : 0 8 6 4 0 : 0 3 7 4 0 : 0 0 0 0 0 : 0 0 0 0 States........ 43 41 43 41 43 42 43 42 Successes.... 1,361 1,361 1,361 1,361 1,735 1,735 1,735 1,735 Observations . 15,237 15,231 15,237 15,231 23,117 23,117 23,117 23,117 NOTE. Table gives probability derivatives from a probit regression of BORRSAVE(with a cutoff of $2,000 for renters and $4,000 for owners)against explanatory variables, including bankruptcy exemptionquartiles. CE quartilesareformedfromtheCEsamplewhileU.S.quartilesareformedfromtheU.S.statesoverthesampleperiod(see text for details). Some state dummy variables predict failure perfectly; all such states are dropped for regressions using state fixed effects. For regressions without state fixed effects, standard errors are corrected for clustering; robust,cluster-adjusted,standard errors arein parentheses. 78

TABLE 7: Samplemeansby debtstatusfortheinsurancemodel DebtMeasure Threshold IncomeRatio $3,000 $6,000 Potential Actual ZeroDebt Low High Low High Low High Low High Incomea ( 3 2 1 3 :6 :2 9 9 ) ( 4 2 1 3 :7 :0 2 9 ) ( 5 2 0 2 :0 :6 5 1 ) ( 4 2 3 3 :2 :0 8 5 ) ( 5 2 1 2 :8 :7 6 0 ) ( 4 2 4 3 :5 :5 6 0 ) ( 4 2 6 2 :0 :9 3 8 ) ( 4 2 4 3 :6 :5 0 1 ) ( 4 2 5 3 :9 :0 3 0 ) a Debt ( 1 0 :0 :8 9 4 ) ( 1 9 0 :5 :1 2 1 ) ( 1 1 :9 :6 0 2 ) ( 1 1 3 2 :8 :0 6 3 ) ( 1 0 :0 :8 0 9 ) ( 8 9 :4 :7 2 5 ) ( 0 :9 :8 4 2 ) ( 8 9 :1 :6 8 3 ) a Assets ( 1 8 6 :0 :4 5 4 ) ( 1 8 4 :3 :8 4 6 ) ( 1 6 1 :0 :2 0 6 ) ( 1 7 4 :7 :0 6 7 ) ( 1 5 1 :9 :2 7 4 ) ( 1 9 5 :0 :4 7 3 ) ( 1 5 0 :6 :9 0 2 ) ( 1 9 5 :2 :5 0 5 ) ( 1 5 0 :6 :9 2 6 ) Age (cid:1) lo g ( y ) ( 4 1 ( 5 4 0 0 :6 :0 :0 :6 8 5 7 7 ) ) ( 4 1 ( 3 2 0 0 :6 :8 :0 :5 1 2 4 7 ) ) ( 4 1 ( 1 1 0 0 :5 :0 :0 :5 4 8 4 3 ) ) ( 4 1 ( 3 2 0 0 :0 :4 :0 :5 4 9 4 5 ) ) ( 4 1 ( 1 0 0 0 :7 :9 :0 :5 1 2 5 5 ) ) ( 4 1 ( 3 2 0 0 :4 :3 :0 :5 0 9 4 6 ) ) ( 4 1 ( 2 1 0 0 :0 :8 :0 :5 5 6 4 4 ) ) ( 4 1 ( 3 2 0 0 :4 :4 :0 :5 7 4 4 6 ) ) ( 4 1 ( 2 1 0 0 :0 :8 :0 :5 4 3 4 4 ) ) —Percent— Ownhome 5 2 :0 1 6 8 :5 0 7 3 :6 8 6 9 :4 1 7 5 :0 1 7 4 :2 2 6 7 :2 2 7 4 :3 7 6 7 :3 7 Married 4 9 :8 4 6 0 :8 4 6 9 :4 4 6 2 :6 0 7 0 :8 5 6 7 :0 9 6 1 :9 8 6 7 :1 2 6 2 :1 6 Black 1 8 :5 4 1 1 :1 9 9 :0 0 1 0 :9 0 8 :1 4 8 :7 4 1 1 :7 6 8 :6 4 1 1 :7 3 H,S. 5 1 :2 9 5 8 :7 4 6 0 :8 7 5 9 :8 2 5 9 :0 8 5 8 :6 8 6 0 :6 2 5 8 :4 3 6 0 :7 8 College 1 7 :4 7 2 3 :7 8 2 7 :6 6 2 3 :8 5 3 0 :6 3 2 7 :3 4 2 3 :5 5 2 7 :5 6 2 3 :4 9 Topquart. 2 3 :7 7 2 4 :6 0 2 6 :9 4 2 4 :8 7 2 8 :0 1 2 4 :4 2 2 6 :7 9 2 4 :3 9 2 6 :7 2 Bot. quart. N 2 4 :4 7 2 5 :7 6 2 5 :3 1 2 5 :8 9 2 4 :5 1 2 5 :3 4 2 5 :7 9 2 5 :3 1 2 5 :8 0 5,937 5,552 4,176 7,443 2,285 4,865 4,863 4,662 5,066 NOTE. Tablegivessamplemeansandstandarddeviationsforzero-debt, low-debtandhigh-debthouseholds,using three different definitions of high debt. See text for details. Assets are liquid assets at fifth interview and debts are unsecured debtsat secondinterview. aReal1996,$000s 79

TABLE 8: (cid:1) l o g ( c ) by debtand bankruptcyexemptionusingthresholdmeasures ofdebt. Owners Renters $3,000Threshold DebtMeasure AssetExemptionQuartile Asset ExemptionQuartile Bottom 3rd 2nd Top Bottom 3rd 2nd Top Debt Zero ( 0 0 : : 0 2 1 9 4 4 0 8 ) (cid:0) ( 0 0 : : 0 3 1 0 2 1 8 0 ) (cid:0) ( 0 0 : : 0 2 0 8 9 4 3 5 ) (cid:0) ( 0 0 : : 0 3 1 0 1 4 1 9 ) ( 0 0 : : 0 3 0 1 5 5 5 3 ) ( 0 0 : : 0 3 1 4 4 0 9 4 ) ( 0 0 : : 0 3 0 0 7 8 6 6 ) ( 0 0 : : 0 3 2 2 4 7 4 8 ) Low (cid:0) ( 0 0 : : 0 2 0 6 4 1 9 1 ) (cid:0) ( 0 0 : : 0 2 0 7 9 4 9 5 ) (cid:0) ( 0 0 : : 0 2 2 7 0 4 7 6 ) (cid:0) ( 0 0 : : 0 2 0 8 4 5 7 3 ) (cid:0) ( 0 0 : : 0 2 1 8 3 8 6 8 ) (cid:0) ( 0 0 : : 0 3 0 0 5 2 5 6 ) (cid:0) ( 0 0 : : 0 2 3 9 2 1 4 5 ) (cid:0) ( 0 0 : : 0 2 0 9 5 6 7 5 ) High (cid:0) ( 0 0 : : 0 2 0 6 9 2 8 0 ) (cid:0) ( 0 0 : : 0 2 0 7 9 2 7 4 ) (cid:0) ( 0 0 : : 0 2 0 6 8 3 0 9 ) (cid:0) ( 0 0 : : 0 2 2 6 2 6 2 1 ) ( 0 0 : : 0 2 0 8 4 4 0 9 ) (cid:0) ( 0 0 : : 0 2 1 6 3 9 1 7 ) ( 0 0 : : 0 2 0 7 6 9 3 7 ) ( 0 0 : : 0 3 1 1 3 2 6 4 ) $6,000Threshold DebtMeasure AssetExemptionQuartile Asset ExemptionQuartile Bottom 3rd 2nd Top Bottom 3rd 2nd Top Debt Zero ( 0 0 : : 0 2 1 9 4 4 0 8 ) (cid:0) ( 0 0 : : 0 3 1 0 2 1 8 0 ) (cid:0) ( 0 0 : : 0 2 0 8 9 4 3 5 ) (cid:0) ( 0 0 : : 0 3 1 0 1 4 1 9 ) ( 0 0 : : 0 3 0 1 5 5 5 3 ) ( 0 0 : : 0 3 1 4 4 0 9 4 ) ( 0 0 : : 0 3 0 0 7 8 6 6 ) ( 0 0 : : 0 3 2 2 4 7 4 8 ) Low (cid:0) ( 0 0 : : 0 2 0 6 7 1 6 9 ) (cid:0) ( 0 0 : : 0 2 1 7 5 1 2 0 ) (cid:0) ( 0 0 : : 0 2 1 7 6 1 8 8 ) (cid:0) ( 0 0 : : 0 2 0 8 7 4 6 2 ) (cid:0) ( 0 0 : : 0 2 0 8 5 9 2 9 ) (cid:0) ( 0 0 : : 0 3 0 0 7 1 7 0 ) (cid:0) ( 0 0 : : 0 2 3 9 2 0 6 0 ) ( 0 0 : : 0 2 0 9 4 8 8 0 ) High (cid:0) ( 0 0 : : 0 2 0 6 5 0 4 2 ) ( 0 0 : : 0 2 0 8 8 1 2 5 ) (cid:0) ( 0 0 : : 0 2 0 6 8 3 9 5 ) (cid:0) ( 0 0 : : 0 2 2 5 7 4 2 2 ) (cid:0) ( 0 0 : : 0 2 1 7 4 7 1 0 ) (cid:0) ( 0 0 : : 0 2 1 4 1 2 0 3 ) ( 0 0 : : 0 2 4 6 8 7 5 3 ) (cid:0) ( 0 0 : : 0 3 0 2 4 0 4 4 ) NOTE. Tablegivesmeansand standard deviations(inparentheses) forthelog differencein non-durableconsumption from the second to the fifth interviews of the CE depending on (1) the household’s debt status and (2) the quartilebankruptcyexemptionrank ofthestate-yearinwhich thehouseholdresides. 80

TABLE 9: (cid:1) l o g ( c ) by debtand bankruptcyexemption. Asset ExemptionQuartile Bottom 3rd 2nd Top Homeowners Debt Status Zero d = 0 ( 0 0 : : 0 2 1 9 4 4 0 8 ) (cid:0) ( 0 0 : : 0 3 1 0 2 1 8 0 ) (cid:0) ( 0 0 : : 0 2 0 8 9 4 3 5 ) (cid:0) ( 0 0 : : 0 3 1 0 1 4 1 9 ) Low d = y b (cid:20) 6 : 4 % (cid:0) ( 0 0 : : 0 2 0 5 2 9 1 6 ) (cid:0) ( 0 0 : : 0 2 0 7 8 0 4 5 ) (cid:0) ( 0 0 : : 0 2 2 6 0 9 3 1 ) (cid:0) ( 0 0 : : 0 2 0 8 3 9 4 0 ) High d = y b > 6 : 4 % (cid:0) ( 0 0 : : 0 2 1 6 2 3 5 4 ) (cid:0) ( 0 0 : : 0 2 1 7 1 7 6 5 ) (cid:0) ( 0 0 : : 0 2 0 7 8 0 9 2 ) (cid:0) ( 0 0 : : 0 2 2 6 2 3 5 1 ) Renters Debt Status Zero d = 0 ( 0 0 : : 0 3 0 1 5 5 5 3 ) ( 0 0 : : 0 3 1 4 4 0 9 4 ) ( 0 0 : : 0 3 0 0 7 8 6 6 ) ( 0 0 : : 0 3 2 2 4 7 4 8 ) Low d = y b (cid:20) 6 : 4 % (cid:0) ( 0 0 : : 0 2 0 8 3 7 5 4 ) (cid:0) ( 0 0 : : 0 3 0 0 6 6 2 4 ) (cid:0) ( 0 0 : : 0 2 4 9 0 5 9 1 ) ( 0 0 : : 0 2 0 9 1 9 9 5 ) High d = y b > 6 : 4 % (cid:0) ( 0 0 : : 0 2 0 8 9 7 6 5 ) (cid:0) ( 0 0 : : 0 2 1 7 0 6 1 8 ) ( 0 0 : : 0 2 0 8 1 0 2 1 ) ( 0 0 : : 0 3 0 0 3 6 0 3 ) NOTE. Tablegivesmeansandstandarddeviations(inparentheses)forthelogdifference innon-durableconsumptionfrom thesecond tothefifth interviewsofthe CE depending on (1) the household’s debt status and (2) the quartile bankruptcy exemptionrank ofthestate-yearinwhich thehouseholdresides. 81

TABLE 10: Distributionofobservationsby debtand bankruptcylawstatus. Owners Renters Quartile Quartile Bot. 3rd 2nd Top Bot. 3rd 2nd Top Debt Zero 778 783 782 745 675 785 723 666 Low 911 949 875 876 322 307 313 312 High 835 748 840 846 419 345 373 457 NOTE. Tablegivesnumberofobservationsineachcombinationofdebtstatusand bankruptcy exemptionquartile. Here we use the ratio of debt to potential income as ourdebt statusmeasure. 82

TABLE 11: TestforOlneyEffects Owners Renters Zero Low High Zero Low High (cid:13) Top (cid:13) ( 0 0 : : 0 0 3 2 5 2 8 2 ) ( 0 0 : : 0 0 5 1 0 8 7 3 ) ( 0 0 : : 0 0 4 1 2 8 6 8 ) (cid:0) ( 0 0 : : 0 0 0 2 8 0 7 7 ) ( 0 0 : : 0 0 1 2 0 5 1 5 ) ( 0 0 : : 0 0 4 2 1 2 7 3 ) 2nd (cid:13) ( 0 0 : : 0 0 6 1 4 6 6 3 ) ( 0 0 : : 0 0 4 1 6 9 3 5 ) ( 0 0 : : 0 0 1 2 8 7 1 8 ) (cid:0) ( 0 0 : : 0 0 0 1 6 7 0 1 ) ( 0 0 : : 0 0 2 3 7 6 9 4 ) ( 0 0 : : 0 0 0 3 2 2 5 2 ) 3rd (cid:13) ( 0 0 : : 0 0 2 1 3 3 4 5 ) ( 0 0 : : 0 0 3 1 4 7 9 0 ) ( 0 0 : : 0 0 2 2 1 0 8 7 ) ( 0 0 : : 0 0 7 1 6 9 8 9 ) ( 0 0 : : 0 0 3 2 2 8 6 9 ) ( 0 0 : : 0 0 3 3 0 2 9 1 ) Bottom b ( 0 0 : : 0 0 5 1 9 9 8 8 ) ( 0 0 : : 0 0 1 1 2 6 7 0 ) ( 0 0 : : 0 0 0 2 3 0 7 9 ) ( 0 0 : : 0 0 0 2 2 2 8 0 ) ( 0 0 : : 0 0 7 2 0 3 3 6 ) ( 0 0 : : 0 0 4 2 6 1 8 7 ) Top b (cid:0) ( 0 0 : : 1 1 4 1 8 0 9 3 ) (cid:0) ( 0 0 : : 1 1 5 0 2 2 2 1 ) (cid:0) ( 0 0 : : 0 1 7 1 5 9 8 2 ) (cid:0) ( 0 0 : : 0 1 4 0 6 9 2 8 ) ( 0 0 : : 3 1 8 4 5 6 1 2 ) (cid:0) ( 0 0 : : 0 1 9 2 9 3 3 0 ) 2nd b (cid:0) ( 0 0 : : 1 1 5 0 9 5 4 2 ) (cid:0) ( 0 0 : : 1 0 5 9 0 8 1 8 ) (cid:0) ( 0 0 : : 0 1 8 1 2 5 2 9 ) (cid:0) ( 0 0 : : 0 1 7 0 4 6 3 4 ) ( 0 0 : : 3 1 6 4 9 2 5 5 ) (cid:0) ( 0 0 : : 0 1 9 2 6 3 6 5 ) 3rd b (cid:0) ( 0 0 : : 1 0 0 9 2 9 3 6 ) (cid:0) ( 0 0 : : 1 0 1 9 7 4 7 7 ) (cid:0) ( 0 0 : : 1 1 2 1 4 3 7 2 ) (cid:0) ( 0 0 : : 0 1 7 0 2 8 8 8 ) ( 0 0 : : 4 1 7 4 2 5 5 9 ) (cid:0) ( 0 0 : : 1 1 1 2 8 9 9 3 ) Bottom F ( (cid:0) ( 0 0 : : 0 1 6 0 6 4 0 0 ) (cid:0) ( 0 0 : : 1 0 1 9 1 7 8 5 ) (cid:0) ( 0 0 : : 1 1 2 1 9 5 1 2 ) (cid:0) ( 0 0 : : 1 1 0 1 0 8 4 0 ) ( 0 0 : : 5 1 2 5 6 5 2 4 ) (cid:0) ( 0 0 : : 1 1 0 4 1 0 9 1 ) State ) 2 : 0 5 1 : 2 3 2 : 5 0 3 : 7 6 3 : 0 7 2 : 1 2 Prob > F ( 0 : 0 0 0 1 ) 0 : 1 4 7 1 ( 0 : 0 0 0 0 ) ( 0 : 0 0 0 0 ) ( 0 : 0 0 0 0 ) ( 0 : 0 0 0 1 ) Observations 3,088 3,611 3,269 2,849 1,254 1,594 NOTE. Tablegivesselected regression coefficients and robust standard errors from estimatingequation(13) under the indicated subsets of the data. The dependent variable in all cases is log consumptiongrowth; the (cid:13) parameters givethecoefficienton logincomegrowthinteracted withbankruptcygenerosityquartile,the b parametersgivethe intercepts(theconstantissuppressed). Herelowdebtisdefinedashavingpositivedebt,butbelowthemediandebttopotentialincomeratiointhesample. Those households with debt to potential income ratios above the median among those households with positive debtare classedas highdebt. Onlythosehouseholdswithexactlyzero unsecured debtare classedas zero debt. 83

FIGURE 1: Contrastingan interioroptimalchoiceofassets anddebt withamaximal-borrowingstrategy. −1 −2 0 1 2 4 8 12 16 20 Debt Choices: d ytilitU detcepxE y=0.99 a*=0.55: EW *=−0.70316 0 b E{W (a*,d,y)} 1 i U (a*,d,y) 0 i E{U + b W} 0 1 (a)Interioroptimum a ; d −1 −2 0 1 2 4 8 12 16 20 Debt Choices: d ytilitU detcepxE y=0.99 a=EXEMPT=0.75: EW *=−0.71227 0 b E{W (a*,d,y)} 1 i U (a*,d,y) 0 i E{U + b W} 0 1 (b)Maximalborrowingstrategy: a = X ; d ! 1 NOTE. Figures give examples of the utility in the first (the green lines) and second (the blue lines) periods of life as a function of the quantity of borrowing in the first period of life; the present discounted value is given by the magentalines. Panel (a) displays utilitiesas a function ofdebt choices when assets are set to theirinterioroptimal level; panel (b) when assets are set to the exemption level. In neither case does the maximal borrowing strategy dominatetheinterioroptimum;further, thepeak ofthebottompanel isbelowthepeak ofthetoppanel. 84

FIGURE 2: Incomeshocks 8 6 4 2 0 y0 y1 Y, Y 0 1 tnecreP FIGURE 3: Lenderprofits Probability Distribution of Period 0,1 Income First Period 0.5 Second Period 0 −0.5 −1 −1.5 −2 −2.5 0 100 200 300 400 500 600 700 800 Credit Spread } {p E Expected Lender Profits X=0 X=2 X=X* NOTE. Figures plot the distribution of period 0 and period 1 incomes (note the masspointsatalowlevel)andlenders’profitschedulesasfunctionsoftheirinterest rates (in basis points) under different bankruptcy law asset exemptions. Note thatatthehighestexemptionlevel( X = 1 )itisimpossibleforthelendertomake positiveprofits;also,sometimesmultipleequilibriuminterestrates resultsinzero profits. 85

FIGURE 4: Social welfare 0 −20 −40 −60 −80 −100 −120 −140 −160 −180 0 0.5 1 1.5 2 Bankruptcy Asset Exemption }))X(*r;y(W{E 0 0 FIGURE 5: Lenderinterest rates g=−2 b =0.9 STIGMA=0.1 s =0.1 s =0.1 0 1 20 15 10 5 0 0 0.5 1 1.5 2 Bankruptcy Asset Exemption )stniop sisab( daerpS tiderC g=−2 b =0.9 STIGMA=0.1 s =0.1 s =0.1 0 1 NOTE. Figures gives the expected present discounted value of all agents in the economyandequilibriuminterestratesonunsecureddebtunderdifferentlevelsof assetexemptions. Thebankruptcyexemptionassociatedwiththeglobalmaximum socialwelfare ismarked withastar. 86

FIGURE 6: Modelresults 10 8 6 4 2 0 0 0.5 1 1.5 2 Bankruptcy Asset Exemption yctpurknaB gniralceD tnecreP Bankruptcy Rate 7 6 5 4 3 2 1 0 0 0.5 1 1.5 2 Bankruptcy Asset Exemption (a)Bankruptcyrates and exemptions noitpmusnoC enogroF Cost of Borrowing to Save (b) Borrowingtosave Correlation of Consumption and Income 1 0.9 0.8 0.7 0.6 0.5 0 0.5 1 1.5 2 Asset Exemption X )Y,C( r 1 1 (c) Olneyeffect NOTE. The three panels of the figures show the theoretical foundation that we investigate empirically. Panel (a) showstheaggregatebankruptcyrateasafunctionoftheassetexemption,panel(b)showsthequantityofborrowing to save in the economy, measured in terms of forgone consumption and panel (c) shows the correlation of second period consumptionand income, theso-called “Olneyeffect.” In all cases, theexemptionlevelassociated with the globalsocialwelfare optimumis markedwitha star. 87

FIGURE 7: Cumulativedistributionsofdebt anddebt toincomeratios 100 75 50 25 0 0 3 6 25 100 Unsecured Debt (real, $000s) tnecreP evitalumuC Empirical Cumulative Distribution (a)Unsecured debt, d 100 75 50 25 0 0 0.01 0.055 1 2 4 Ratio . tnecreP evitalumuC Empirical Cumulative Distribution (b) Debt toactual income d = y . 100 75 50 25 0 0 0.01 0.064 1 2 4 Ratio tnecreP evitalumuC Empirical Cumulative Distribution (c)Debt topotentialincome d = y b . NOTE. Figures give the empirical cumulative distribution of unsecured debt (panel a) in the CE, the ratio of unsecured debt to actual income (panel b), and the ratio of unsecured debt to potential income (panel c). About 37%oftheCE samplereport owingzero unsecured debt(indicatedby thelargegreydots). 88

FIGURE 8: DistributionofIncomeAcrossDebt Measures 8 Debt Exceeds $3,000 Threshold 7 6 5 4 3 2 1 0 5 10 20 100 Income (real, $000s) tnecreP Empirical Distribution 9 Debt Exceeds $6,000 Threshold 8 Zero Debt Low Debt 7 High Debt 6 5 4 3 2 1 0 5 10 20 100 Income (real, $000s) (a)$3,000Threshold tnecreP Empirical Distribution Zero Debt Low Debt High Debt (b) $6,000Threshold 7 6 Ratio to Potential Income 5 4 3 2 1 0 5 10 20 100 Income (real, $000s) tnecreP Empirical Distribution 7 Ratio to Actual Income 6 Zero Debt 5 Low Debt High Debt 4 3 2 1 0 5 10 20 100 Income (real, $000s) (c)Debt topotentialincomeratio tnecreP Empirical Distribution Zero Debt Low Debt High Debt (d)Debt toactual incomeratio NOTE. Figure givesthe distributionof real incomes among thosewith zero debt, lowdebt and high debt forfour different definitionsofhigh debt. Notice(1) That the income distribution of zero debt holders is relatively flat; and (2) Separating based on a threshold criterion produces two relatively dissimilar income distributions, while separating on a debt to income ratio criterion two more-similar incomedistributions. 89

FIGURE 9: Consumptionand IncomeShock Distributions Empirical Distributions 6 4 Consumption 2 Income 0 −2 −1 0 1 2 D log(c) and D log(y) tnecreP Empirical Cumulative Distributions 100 Consumption 75 50 Income 25 0 −2 −1 0 1 2 D log(c) and D log(y) (a) tnecreP evitalumuC (b) NOTE. Figuresgivetheempiricalpdf(panela)andcdf(panelb)ofthelogdifferencesinconsumptionandincome. Notethatthedistributionofincomedifferences features very fat tails (indicated by the grey dots) but also less central variation. Thusincomeshocksare eithersmallorquitelarge. 90