Social Influence and Bankruptcy - Why Do So Many Leave So Much on the Table

Published on January 2017 | Categories: Documents | Downloads: 23 | Comments: 0 | Views: 189
of 36
Download PDF   Embed   Report

Comments

Content


Social influenceandbankruptcy:
Whydosomanyleavesomuchonthetable?

EthanCohen-Cole
Robert H. SmithSchool of Business
Universityof Maryland- Collegepark
BurcuDuygan-Bump
Federal ReserveBankof Boston
10December 2009
Abstract
Asmuchashalf of theUSpopulationcouldreapmorethan$10,000inimmediatefinancial
gainfromfilingfor bankruptcy. Then, why don’t they file? Among financial decisions, this
ranks as oneof thelargest. Using27millioncredit reports, wefindvis-à-vis thebankruptcy
decision, social spilloversare30-50timeslarger thanfinancial hardship: a1%increaseinlocal
filingratesleadstoa25-40%increaseintheindividual filingprobability. Wealsofindthatthis
influenceis driven by both information diffusion and changes in social stigma. Finally, our
resultsvarystronglyacrossincomeandeducational groups.
JEL Classification Codes: D14, I30, K45, Z13.
Keywords: personal bankruptcy, social interactions, social stigma, informationsharing

Authors: Cohen-Cole: Assistant Professor of Finance, Robert H. SmithSchool of Business. 4420VanMunching
Hall, University of Maryland, CollegePark, MD 20742. email: [email protected]. +1(301) 541-7227.
Duygan-Bump: Federal ReserveBank of Boston, 600 Atlantic AvenueBoston, MA 02210. Email: burcu.duygan-
[email protected]. A prior versioncirculatedunder title"HouseholdBankruptcy Decision: theroleof social stigma
vs informationsharing." Wearegrateful to Nicholas Kraninger, J onathanLarsonandJonathanMorsefor excellent
researchassistance. Wearealsograteful for helpful commentsandsuggestionsfromSumitAgarwal, MaryBurke, J eff
Brown, ChrisCarroll, HansDegryse, StevenDurlauf, Lutz Hendricks, Erik Hurst, Albert Kyle, AndreasLehnert, Vo-
jislavMaksimovic, AnnaPaulson, J ose-Victor Rios-Rull, LemmaSenbet, SophieShive, NickSouleles, Mel Stephens,
J eremy Tobacman, MichelleWhite, as well as seminar participants at theeconomics departments of theUniversity
of Californiaat Davis, University of Cambridge, University of Bonn, Humboldt University, University of Maryland
Smith School of Business, TilburgUniversity, University of Illinois, University of Washington - Seattleeconomics
department andtheUniversityof Washington- EvansSchool of PublicPolicy, theFederal ReserveBanksof Boston,
Cleveland and San Francisco, theNBER Summer Institute, theFinancial Intermediation Research Society, and the
ChicagoFederal Reserve’sForty-FifthAnnual ConferenceonBank StructureandCompetition. Theviewsexpressed
inthispaper arethoseof theauthorsanddonot necessarilyreflect thoseof theFederal ReserveBankof Bostonor the
Federal ReserveSystem.
Evenafter largescalebankrutpcyreformin2005, asmuchashalf of theUSpopulationwould
‘benefit’ financially fromfilingfor bankruptcy. Indeed, theimmediatefinancial benefit of doing
so canbeinthetens of thousands of dollars.
1
Thenwhy don’t they? Despite30years of rising
bankruptcyrates, wefindthat boththesocial stigmaof bankruptcyaswell asapaucityof detailed
informationprevent even higher rates. Whilethesecular increaseinthebankruptcy ratehas led
to a resurgence of research on the determinants of personal bankruptcy, most of these studies
focusedoneconomicfactors: howtodisentanglestrategicbehavior andeconomicshocks, suchas
unemployment andillness, fromtheroleof changesinthecredit market environment itself, from
accesstocredit post-bankruptcy, etc.
Our paper focuses on a different set of drivers—social influence and spillovers.
2
Indeed it
is very likely that social context plays animportant roleinthebankruptcy decision; wefindthe
effectstobeverylarge. Becauseitissolarge, andsomuchbigger thanother driversof bankruptcy,
understanding the composition of these social effects is essential. The underlying idea is that
interactingwithotherswhohavegonebankruptor areintheprocessmayincreasethelikelihoodof
anindividual goingbankruptherself. Onereasonwhysuchaneffectmayexististhestigmaeffect:
being surrounded by many peoplewho havegonethrough bankruptcy decreases theassociated
embarrassment. Inother words, theperceptionthat “everybodydoesit” reducesthepsychological
pressuretofully pay incurreddebtsregardlessof thecircumstances. Similarly, it isalsopossible
that the same causal relation between group and individual outcome is caused by information
sharing. Bankruptcy, after all, is acomplex process which requires somedegreeof specialized
knowledge, andassuchindividualsthathaveaccesstothisinformationmaybemorelikelytofile.
3
Accordingly, people may share information on eligibility, application procedures, bureaucratic
details, etc. with neighbors, friends, or relatives. In general, the role of social phenomena in
individual financial decisionmakingisincreasingly widely accepted.
4
However, theemphasisof
many of thesepapers is to identify thepresenceof aninteractionrather than to label its nature;
thosethat doattributethemtoasinglecause.
5
1
Oneof post-2005criteriafor achapter 7bankruptcyisincomebelowtheareamedian. Anindividual whomeets
this criteriacandischargehalf of his unsecureddebt. The75thpercentileof total credit for individuals who livein
areaswithbelowmedianincomewas$33,000in2006. Thereareof coursemany tradeoffsandother coststofiling,
whichwediscussbelow.
2
A recent paper byLearyandRoberts(2009) oncapital structurespilloversisanoutstandingexampleof arecent
interest insocial spilloversinfinancemodelswithcareful treatment of theidentificationissuesthat arise. Faulkender
andYang(2009) isanother suchexamplewhichtreatscompensationspillovers.
3
Indeed, manycollege-eduated, middleandupper incomeindividualswereunawarethatbankruptcywasanoption
andlikely never consideredit intheir choiceset. Similalrly, someparticularsof bankruptcy aresufficiently nuanced
astomakethedisseminationof informationpotentiallyveryimportant inthedecision. For example, aborrower may
not realizethat acredit cardcanbewrittenoff inachapter 7bankruptcy, but ahomeequitylineof credit canonlybe
writtenoff aspart of alien-strippingactioninachapter 13.
4
KaustiaandKnüpfer (2008) findevidenceinIPOparticipation; KaustiaandKnüpfer (2009), Hong, Kubik, and
Stein (2004) and Brown, Ivkovich, Smith, and Weisbenner (2008) find alink to stock market participation; Shive
(2009) finds interactions in trading activity; Topa, Bayer, and Ross (2009) find alink to workplacechoice. Guiso,
Sapienza, andZingales(2004) findalinkbetweensocial spilloversandfinancial development.
5
Grinblatt, Keloharju, andIkäheimo (2008) is animportant exception. They findinteractions inautomobilepur-
chases, but do not find evidencethat envy plays arole in thesedecisions as theories of conspicuous consumption
wouldsuggest. Another exceptionis Mas andMoretti (2009). They findthat grocery cashiers respondto thespeed
of co-workers. Remarkably, theyalsofindthat beingobservedmatters, anindicationthat observational stigmaisim-
portant. Inadditional evidenceof spillovers, Grinblatt andKeloharju(2001) findthat cultureandlanguagematter for
2
Using acomprehensivedataset of morethan 27 million individual credit reports, weiden-
tify thepresenceof social interactions, measureitsmagnitude, andseparatetherelativeinfluence
of stigmaand information on thebankruptcy decision. Along with recent work by Kaustiaand
Knüpfer (2009), this paper is oneof thefew papers that uselarge-scaledatato understand the
impact of social learning. Inparticular, wetest whether anindividual’sbankruptcydecisionisim-
pactedbytheprior bankruptcydecisionof othersincloseproximityandinvestigatetheempirical
relevanceof thetwomainsocial influencechannels: stigmaandinformationsharing.
Fromanempirical perspective, distinguishingbetweenstigmaandinformationeffectsisquite
difficult. Indeed, botheffectsleadtothesamepositiverelationshipbetweenagroupandindividual
outcome. However, theimplicationsof thetwoarequitedifferent. Inone, diffusionof information
will likelycontinuetooccur, or specialists(lawyers) canprovidethisinformationatacost. Stigma,
however, isafunctionof general perceptionsandmayincreaseor decreasewithtime. Toseparate
thetwo, weturntothework of social psychologists, who havelongstudiedtwoparticular types
of social influence, normative (stigma) and informational. Theformer describes conforming to
normsbasedondoingwhatothersexpect, andthelatter relatestotheuseandexchangeof accurate
information.
Morespecifically, our identificationstrategy exploits thefact that wehavevery detaileddata
onthegeographic locationof individuals. Theprincipal advantageof thisistotakeadvantageof
thesocial psychologistsfindingthat informationisa‘local’ phenomenon, whilestigmaisdefined
morebroadly andisderivedfrommultiplesources. Inpractice, wequantify theimport of others
by distance to the individual in question. Neighbors that are ‘close’ provide both information
andsocial stigma, whilethosefurther away providethebasisfor ageneral averageof population
behavior.
6
Oncestratified, weuseasimpleeconometric techniqueto disentanglethetwo effects
(seeCohen-ColeandZanella, 2008andCohen-ColeandMas, 2009).
Wehavefour primaryresults:
1. Combined social spillovers are30-50 times larger than commonly used measures of eco-
nomicandfinancial hardship: a1%increaseinlocal filingratesleadstoa25-40%increase
intheindividual probabilityof filing.
2. Social influence is driven by both information sharing (diffusion) and changes in social
stigma.
3. Whilebothfactors’ rolehas increasedinrecent years, informationdiffusionis morelikely
responsiblefor thecontinuedincreaseinbankruptcyrates.
4. These effects vary significantly across income and educational groups. We find that so-
cial spillovers, bothinformationandsocial stigma, areas muchas 3times moreimportant
amongst less-educatedandpoorer communities.
To test therobustness of theseresults and our identification strategy, werepeat our analysis
usingasecondidentificationscheme. Becauseour dataallowustoobservewhereindividualslive
stockholdings. Finally, arecent paper by Guiso, Sapienzaand Zingales (2009) uses survey datato unpack various
reasonsfor foreclosure.
6
The useof geographically stratified samples for identification has also been used by Grinblatt, Keloharju and
Ikaheimo(2008), thoughinaslightlydifferent way.
3
at twopointsintime, wecanapplyour analysistothesubset of individualsthat movedduringour
sampletoidentifyandseparatethesesocial effects. Implicithereistheassumptionthatinformation
canbebrought withyouwhenyoumovebut social stigmafromoldneighbors cannot. Wealso
applyavarietyof additional robustnesschecks.
Thepaper proceedsasfollows. Weprovideanoverviewof USbankruptcylawandrelatedlit-
eratureinSectionsI andII, respectively. Themethodologyusedinestimationof social interactions
ispresentedinSectionIII, followedby adiscussionof our datainSectionIV. SectionV presents
our resultsincludingarangeof sensitivityanalyses. SectionVI concludes.
I Personal Bankruptcy in the US: History and Stigma
Prior to theturn of the20thcentury, bankruptcy was alegal conditionrather than an individual
choice. Creditorswouldbeforcedtofilepetitionsprovingthat thedebtor hadcommittedan‘act’
of bankruptcy—typicallysomethingakintofraud(Coleman1974). Theprevailingnotionwasthat
bankruptcywasrootedinfraud(Efrat2006) or inafundamental disregardfor themoralsof society
(McIntyre1989, Channing1921). For example, Efrat(2006) presentsarangeof evidenceshowing
howbankruptcy stigmahashistorically beenparticularly strong. Hefindsquotationsthat refer to
bankruptsasdeservinglower social respect thancriminals(J ones1979). Similarly, AdamSmith,
inhisfamous“Wealthof Nations”, arguesthat bankruptcyisthe‘most humiliatingcalamity’ that
canoccur toanindividual.
Over thepast coupleof hundredyears, legal standards havereflectedsocial efforts to penal-
izeandshamethoseinbankruptcy. Thelaws themselves emphasizethenear criminal natureof
bankruptcy(seeTabb1991for anoverview) andimposedpenaltiesthat wouldnowberegardedas
draconian.
7
Inthe1960s and1970s, bankruptcy policy beganto reflect changes inAmericanperceptions
of bankruptcy stigma. In1978, congresspassedanewbankruptcy law, inpart aimedat reducing
stigma(Efrat 2006). Nonetheless, public views of bankruptcy remainedstrong. And, inspiteof
evidence of the remaining strong stigma and thealmost complete absence of empirical studies
that measureits fall, therun-up to the2005 bankruptcy reformfound many arguing that rising
bankruptcieswereduetoadeclineinstigma.
8
Currently, theUnitedStateshastwodifferent personal bankruptcyprocedures—Chapter 7and
Chapter 13—and prior to the2005 bankruptcy reform, debtors had great flexibility in choosing
betweenthem. Under bothprocedures, oncethedebtor has filedfor bankruptcy, legal actions to
collect anydebt bycreditorsmust beceased. All unsecureddebt isdischargedinbankruptcywith
some exceptions, such as student loans, debts incurred by fraud, and credit card debt incurred
shortly before filing. On the other hand, secured loans, such as mortgages and car loans, are
not discharged, but bankruptcy generally allows debtors to delay creditors fromforeclosing or
repossessingrelatedassets.
Under bothprocedures, bankruptindividualsmustpayvariousadditional costs; includingcourt
andlawyers’ fees, fees associatedwithgatheringinformationabout thebankruptcy process, and
7
For example, debtors incolonial Americawouldhavetheir hair shaved, bebrandedwithaT for “thief,” andbe
requiredtohaveanear cut off (Pomykala2000).
8
SeeEfrat (2006) for anexhaustivelistingof studiesthat showadeclineinstigmausingindirect methods. Efrat
arguesthat noneof theseestimatetheeffect directly.
4
legal advice. FlynnandBermant (2002) reportthat thesecostsrangedbetween$600for Chapter 7
and$1600for Chapter 13asof 2001. Moreover, debtorswhofileunderChapter7arenotpermitted
tore-fileunder Chapter 7for eightyears, althoughtheymayfileunder Chapter 13asoftenasevery
sixmonths.
As bankruptcy rates rosefivefoldto about 1.5millionper year (seeFigure1), lenders grew
increasinglyaggressiveat lobbying. Incongressional testimonythat predatedthelawbyalmost a
decade, VisaUSA submittedtestimony claimingadeclineinsocial stigmaassociatedwithbank-
ruptcy(seediscussioninEfrat 2006). Thislineof discussionbecameaprinciplemotivatingfactor
behindthenewlegislationthat cameinto effect in2005. Thenameof thenewact reflectedthe
intent to restorethestigmaassociatedwithbankruptcy.
9
TheBankruptcy AbusePreventionand
Consumer ProtectionAct (BAPCPA) took effect inlate2005. Filings reachedabout 2millionin
2005 as debtors rushed to fileunder the old law, and then dropped to 600,000 in 2006, before
beginningtoriseonceagain.
Thenewlegislationmadebankruptcy much less financially attractiveby increasingthetime
andfinancial costsassociatedwithfilingandforcingsomeChapter 7debtorstorepay frompost-
bankruptcyearnings. Thenewlawalsoimposedother requirementsonfilers. Filerscannolonger
choosebetweenthecodes. Instead, onesubmits to ameans test, whereadebtor qualifies to file
under Chapter 7if their averagemonthly family incomeover thepast six months(prior tofiling)
islessthanthemedianmonthlyfamilyincomeintheir state, adjustedfor familysize. Aswell, the
lawabolishedanindividual’sabilitytoproposerepayment plansunder Chapter 13andimposeda
standardizedproceduretodeterminepaymentplans. Finally, thenewlawgreatlyraisedfilingcosts,
mandatesdetailedinformation, andrequiresdebtorstotakeacreditcounselingcourse. Elias(2006)
estimates that thesenewrequirements raiseddebtors’ filingcosts to around$2500for Chapter 7
and$3500for Chapter 13.
Withoutquestion, thelawraisedthefinancial andtemporal costsof filing, and, atleastover the
short run, decreasedthenumber of filings. Itisanopenquestionwhether thelawhasimpactedthe
social stigmaof bankruptcy, thecitedprominent reasonfor itspassage.
II The Bankruptcy Decision: A review of related studies
Followingthedramatic riseinbankruptcies over thelast coupleof decades andthesurrounding
policydiscussions, manyresearchershaveattemptedtostudyhouseholdbankruptcydecisions. As
reviewedingreater detail inWhite(2007), theliteraturetodateonbankruptcycanbegroupedinto
two broad categories, based on theapproaches adopted: i) quantitativemacroeconomic models
that useamodeling/calibration approach to match related stylized facts, such as theincreasein
household debt as well as bankruptcies andii) applied analyses that exploit different sources of
micro datato understandtheempirical factors that drivehouseholds’ bankruptcy decisions. Un-
fortunately, duetolackof data, thenumber of studiesinthissecondgroupisstill quitesmall.
ThequantitativemacroeconomicmodelsincludeLivshits, MacGee, Tertilt(2007a,b) andChat-
terjeeet al. (2005), whichoutlinedynamicequilibriummodelswhereinterest ratesvarywithbor-
rowers’ characteristics. Themodels, for reasonableparameter values, canmatchthelevel of U.S.
9
Sullivan, WarrenandWestbrook (2006) findevidencethat stigmawas unlikely tohavebeentheexplanationfor
theriseinbankruptcyrates.
5
bankruptcyfilingsanddebt-incomeratios.
10
Theexplanationsfor ariseinbankruptcy inthesemodelsmirror thoseinquantitativestudies.
Onepossibilityisanincreaseinidiosyncraticuncertaintyathouseholdlevel duetoincreasedlabor
earnings volatility, anincreaseinthenumber of households without medical insurancecoverage
(seealso Barron, Elliehausen, and Staten, 2000, and Warren and Warren Tyagi, 2003, Sullivan,
Warren, and Westbrook 2006). Similarly, Duygan-Bump and Grant, 2008, exploit institutional
differences in punishment for and legal costs of default across the EU countries. Their results
showthat adverseshocks, suchasunemployment andhealthshocks, areimportant, but theextent
towhichtheymatter dependscruciallyonthepunishment associatedwithdefault.
Another possibilityisachangeinthecredit marketsthat either makesbankruptcymoreattrac-
tiveor expandcredit toabroader set of households, includinghigher-riskones. Thissecondset of
explanationsincludesthestorythatcreditmarketinnovations(suchasthedevelopmentandspread
of credit scoring) or increasedcompetitionfacilitatedtheincreaseincredit grantedtohouseholds
byreducingthetransactioncostsof lending(Athreya2004, DickandLehnert 2009).
Thesefindings areingeneral consistent withthosereportedinthetwo seminal papers inthe
appliedanalysiscategory basedonmicrodata. Fay, Hurst andWhite(2002) estimateamodel of
thehouseholdbankruptcy decisionusingthePSID, andshowthat households aremorelikely to
filefor bankruptcywhentheir financial benefit fromfiling—thevalueof debt dischargedinbank-
ruptcy minusthevalueof nonexempt assets—rises. Theyalsofindthat, evenafter controllingfor
stateandtimefixedeffects, households aremorelikely to filefor bankruptcy if they liveindis-
trictswhichhavehigher aggregatebankruptcyfilingrates. Theyconjecturethat thisresult “could
reflect local differences in thelevel of bankruptcy stigmaor local differences in theadministra-
tionof bankruptcylawthat makethedistrict differ fromthestate, or couldreflect theinfluenceof
informationcascades.”
Gross andSouleles (2002) useadministrativecredit-cardaccount datato analyzecredit card
delinquencyandpersonal bankruptcy. Theyruleout risk compositionof borrowers, andconclude
that householdsdidappear tobemorewillingtodefault inthelate1990’sthaninearlier periods,
all elseequal. Theauthorsacknowledgethat theseresultsdonot directly identify what underlies
theestimateddemandeffect, eventhoughthefindingthat default riseswiththebankruptcy filing
rateinthestateis“suggestive” of adeclineinstigmaor informationcosts.
Inthis paper, wefocus onsocial spillovers inthebankruptcy decisionandtry to disentangle
theroleof stigmafrominformation costs. So far, theliteraturehas used thecoefficient on the
lagged bankruptcy filing rates in the state to capture social influences. This coefficient, while
useful, isacompoundmeasure. Inother words, wedonot knowwhether thesocial effect isdue
to information sharing / social learning (peoplecommunicateand pass along information about
bankruptcyprocedures, for example) or stigmatization(theprevalenceof acertainbehavior makes
its adoption less embarrassing), and so on. Theseparateidentification of thesedifferent social
effects is especially important for policy discussions because different channels will generally
requiredifferent policies, andtheappropriatemeasureswill dependontherelativemagnitudesof
stigmaandinformationeffects. Thegoal of this paper is to shedstronger light ontheempirical
importanceof thesesocial factors.
10
SeealsoAthreya(2002) for ananalysisof thewelfareimplicationsof different bankruptcylawsandLi andSarte
(2006) for ananalysisof consumers’ choiceof Chapter 7versusChapter 13.
6
III Methodology
Understanding Information and Stigma
Bankruptcy intheUnitedStates, as discussedabove, allows for somedischargeof debt, and
evenallowsfor householdstokeepaportionof their homes(if not all) andother assets. Accord-
ingly, even though most researchers study why bankruptcies havebeen rising, themost glaring
questioniswhymorepeopledonotfilefor bankruptcygiventhepotential financial benefits(White
1998). Onepotential explanationcouldbethesocial stigmaassociatedwithbankruptcy, as evi-
dencedby sociologists’ surveysof bankrupts, discussedintheintroduction. Another explanation
isthat anindividual that couldbenefit fromfilingfor bankruptcymaynot besufficientlyawareof
thepossibility or ableto navigatethesystem. To disentangleandidentify therelativeempirical
importanceof theseeffects, wewill exploit thepsychology literaturethat showsthat social influ-
encesarerelatedtothenatureof associationswithinapopulation. Inparticular, individualsdraw
informationandlearnfrompeoplewhoaregeographically closer tothem(e.g. neighbors), while
stigmatizationoccursamongabroader group(family, friends, what “others” ingeneral aredoing,
aswell asneighbors).
11
As mentioned, that financial decision making is intermediated by social factors has largely
beenaccepted. Mechanically, most investigationsof theseeffectsassumethat thepossibility that
othersinfluenceour individual decisionswherethedegreeof influenceisincreasinginthenumber
(or percentage) of others doing acertain action. For example, as bankruptcy rates increase, the
assumptionhasgenerallybeenthat somesocial mechanismleadstoanincreaseintheprobability
of an individual’s own bankruptcy. This fits nicely with experimental evidence. Among many
others, thesocial psychologists LataneandWolf (1981), Latane(1981) andTanfordandPenrod
(1984) foundexperimental evidenceof thisincreasingrelationship.
For identification, wedrawonwhat social psychologistscall the‘growthcurve’ (seeFigure2
for anillustration.) Thefact that theinformational growthcurveismoreconcavethanthestigma
curveallows us to separateand individuals quantify thetwo effects.
12
Thecurvemeasures the
marginal responsiveness of an individual to an additional personal applying the samestimulus.
That is, howmuchmoredoes anindividual respondto thesecondpieceof informationvis-a-vis
thefirst. Weidentifyoff thefact that informationandstigmaoperatedifferently.
That is, weexploit differences in thecurvatureof thegrowth curvewith respect to thetwo
phenomena. If onewishestodeclarebankruptcy, potentially heor shecouldreceiveinformation
thatishelpful bothinmakingthedecisionandinnavigatingthebankruptcyprocess. Thisinforma-
11
Wedirect interestedreaderstoCampbell andFairey (1989), whoprovideauseful overviewanddefinethefirst,
informational social influence, as"influencetoaccept informationobtainedfromanother as evidenceabout reality".
Weusethis as asimpleproxy for thetransfer of practical andrelevant informationabout howonecannavigatethe
bankruptcyprocesswithsuccess. Likewise, inthesamepaper, Campbell andFairey(1989) alsodefinenormativesocial
influence, our secondinfluence, as "influenceto conformto thepositiveexpectationof behavior." Whilenormative
influencecouldwork toreduceor increasetheshameassociatedwithbankruptcy, welabel theresult of thisinfluence
as‘stigma’ toalignour discussionwiththat intheeconomicsliteratureaswell asthepublicdebateonbankruptcy.
12
An example of this in practice is available in Guimond (1997). Heprovides evidence based on students at a
militarycollege. Hefindsthatattitudesandopinionsaboutsubject-specificand/or itemsof relativelysmall importance
wereinfluencedbytherelevant social or educational group: whilefacultycouldinfluenceopinionsabout themilitary
educationinaparticular subject, thefull populationdidnot impact theseopinions. However, onbroader topics, both
local groupsandtheaggregatepopulationhadaninfluence. Thisdichotomymatchesandprovidesfurther support to
our separationassumptionof informational influencebeingmore‘local’ thannormative(stigma) factors.
7
tioncouldcomefromlocal, that is, semi-private, sourcessuchasneighbors, friends, family, etc.,
or fromwider, morecommon sources such as television ads, promotional flyers, etc. However,
thetwo sources aredifferent. Most importantly andcentral to our point, informationingeneral
does not increasein influenceas thenumber of its sources increases. In other words, oncewe
knowabout howto filefor bankruptcy, havingthreemorepeopleimpart this informationshould
not changeour understanding. Thus, privateandpublic informationcanbeviewedinfundamen-
tallydifferent ways. Mechanically, weusethenumber of peopleinaneighborhoodthat havegone
bankrupt as being a measure of the probability of receiving non-public and useful information
about bankruptcy. Beyondour neighborhood, increasingly available, by definitionpublic, infor-
mationdoesnot contributedifferentiallytoour abilitytogobankrupt.
13
Thenext section introduces thedetails of this separation methodology froman econometric
perspective.
Basic social effect modeling
Formally, westart bymodelingthebankruptcydecisionof anindividual i, whichwe’ll denote
1
i
. Next, wedenotetherelatively largesocial community anindividual livesinby asuperscript.
Weassumethat thebehaviors of others in this community generatethesocial environment that
contributes to theutility of an individual’s own decision. Wefurther specify two subsets of the
community, a‘local’ group, subscript q, and a‘non-local’ one, subscript o. When referring to
prior workwhichusesthestateasameasureof social group, weusethesubscript :. Weusethese
subscriptstohelpdistinguishthetwokeychannelsof social effectsasweassumethat information
effects arederivedfroma‘close’ social group“q,” whilestigmacancomefromlocal as well as
morediffusesources“o.”
Of course, bankruptcyhasmanypotential causesinadditiontothesocial ones. Tocapturethese
wespecifythebankruptcydecisionproblemintheabsenceof social networkeffectsasfollows.
j: (1
ij
= 1) = Φ(/ + cA
i
+ d1
j
+
ij
) (1)
where1
ij
isanindicator set equal to1 if individual i incommunityq hasdeclaredbankruptcy. To
control for individual differencesincredit quality, onecanincludeavector of individual specific
variables A
i
, suchas ageandindividual credit history. Sinceindividuals arealso impactedas a
groupby theenvironment inwhichthey live, for exampleby changesinemployment conditions,
weincludeavector of variables1
j
that arecommonfor all individualsincommunityq. In1
j
we
alsoincludecommunitylevel demographiccharacteristicsasproxiesfor individual demographics.
For example, weincludeaveragemarriageand divorcerates, educational achievement averages
andincomelevels.
If individuals respond to aggregate behavior in addition to price factors, the estimates of c
in 1will bebiaseddueto correlationwiththeerror term. Thebankruptcy literatureto datehas
augmentedequation1toincludeameasureof averagebankruptcyratesinalarge, non-local area
(stateof residence), :
c
suchthat wecanwrite:
j: (1
ij
= 1) = Φ(/ + cA
ij
+ d1
j
+ J
c
:
c
+
ij
) (2)
13
Any additional variationininformationthat comesfromTV ads, local educationcampaigns, etc. at thecounty,
cityor MSA (or other) level iscapturedbyfixedeffectsat thecorrect level of aggregation.
8
where:
c
=
1
a−1
P
)6=i∈c
1
)c
, and: isthenumber of individualsinthestate. Thus:measuresthe
averagebankruptcyratein: excludingtheindividual i. Notethatthisissimilar tothespecification
usedinFay, Hurst andWhite(2002) andGrossandSouleles(2002).
Composite Social Effects
Our principal twomodifications tothis specificationfollowfromthediscussionabove. First,
usingveryspecificinformationongeographiclocationsof individuals(seedatadescriptioninthe
next section), weareabletoincludecommunity-level information(e.g. income, incomegrowth),
whichhelpsusget closer toindividual level data—animprovement over stateaverages. Second,
wemeasuretheimpact of aggregatebehavior onindividual behavior at twolevelsof aggregation,
lookingat local andnon-local networks.
Mechanically, weaugmentthisspecificationinanumber of waystoallowbothfor interactions
at alevel belowthestate, andto separatethestigmaandinformationeffects. First, wedefinea
vector 1 to captureall community level controls, where1 ≡ (1
j
1
c
)
0
, 1
j
is thesesameset of
community controls but onewherecommunity is defined at somesmall local level, such as a1
mileradius froman individual’s home, and 1
c
captures thesecontrols over alarger community
(exclusiveof thelocal area), such as a1-4mileradius. Wealso allowfor heterogeneous social
interactions amongdifferent local communities. Inorder to do this, wedefine:
c
as avector of
averagebankruptcy rates of other local communities, ¸, withthe1-4mileradius withrespect to
ownlocalityq: :
c
=
1
n
P
¸
{:
¸
}
¸6=j
.
14
A simple choice for estimating equation 1, above, with the addition of our specified social
effects, isafunctionallowingfor local (0-1mile) andnon-local (1-4miles) social coefficients:
j: (1
ijc
= 1) = Φ
³
/
j
+ c
j
A
ij
+ d
j
1 + J
S1
j
:
j
+
e
J
S
c
:
c
+
ijc
´
. (3)
Notethatthisspecificationbringsinadditional notation, whichwebelieveclarifiesourmethod-
ology andassumptions. Morespecifically, notethat :
j
istheaverageof local bankruptcies. Be-
causeweassumethisisassociatedwithbothstigma, o, andinformation, 1; weusethecoefficient
notationJ
S1
j
. Similarly, thecoefficient
e
J
S
c
incorporatesonly stigmatizationeffectsat anon-local
level.
Byconstruction, thetwosetsof coefficientsJ
S1
j
, ascalar, and
e
J
S
c
, a1 × ` vector, capturethe
joint effect of stigma(o) andinformation(1) fromownlocality (q) andof stigmafromother lo-
calities(o). InManski’s(1993) terminology, c
j
expressesindividual effects, d
j
contextual effects,
andJ
S1
j
and
e
J
S
c
endogenous social effects. Wefocus inthis study onthelatter, theendogenous
portion. Itiswell knownthatamodel likethisposesseveral problems. Perhapsthemostdiscussed
inthepeer-effectsliteratureishowtodefinereferencegroups, includingthegeographiclevel. As
wediscuss, wedefinethemas localities within0–1and1–4mileradii. InSectionVI below, we
providesomesensitivityteststosupport thislevel of aggregation.
Threeother econometric problems requiretreatment. Webegin with thereflection problem
(Manski, 1993), whichpotentiallyaffectsanylinearmodel withsocial interactions. Self-consistency
requires that theexpectedparticipationrateof anindividual of locality q beequal to themathe-
matical expectationof theindividual participationindicator inthereferencegroup, that is given
14
Mechanically, wetaketheaveragebankruptcyrateof all censusblocksthat fall intothe1-4mile‘donut’ around
theindividual. Weconductadditional testsonour primaryresultsatdifferentlevelsof aggregation. Theseareavailable
fromtheauthorsonrequest.
9
1
j
:
:
j
= E
¡
1
ijc
|1
c
j
¢
. (4)
Thiscondition, coupledwithequation(3), formsasimultaneousequationsystem. Noticethat
wearetreating:
c
asanother contextual, exogenous, effect. Suppose, asistypicallythecase, that
thegroup-level controls, 1
j
, arethegroup-level mean of theindividual level ones, A
i
. That is,
E(A
i
|1
j
) = 1
j
too. Then, inabsenceof validinstruments, onecannot identify theendogenous
social effects, inour caseJ
S1
j
and
e
J
S
c
, without anexclusionrestriction. Weaddress theproblem
by drawingonthefact that probit modelsarenonlinear inform; thenonlinearity permitsidentifi-
cation.
15
The second problemis the selection problem: individuals in the sample chose to live in a
particular area. If residential choices dependonunobservables that also affect theprobability of
enteringbankruptcy, thengroup-level variables areendogenous, andtheestimatedsocial effects
will beaffectedbyselectionbiases. Howtoget aroundthisselectionprobleminmodelsof social
interactionsbasedonindividual-level dataisacurrent researchtopic—thoughonewithout aclear
solution. A number of methods havebeensuggested, includingastrict characterizationof error
distributions that allows for closed-formidentification of social multipliers (see Zanella 2007).
Inour case, theselectionproblemis thedegreetowhichneighborhoodchoiceis correlatedwith
bankruptcy, anissueminimizedbyfact that it isverydifficult tomoveacrossstatelinesto“shop”
for asset exclusions. Indeed, becausehouseholdexemptionsonly apply toequity accumulatedin
thehousemorethan1000daysprior tothebankruptcy filing, movingtotakeadvantageof filing
differencesisdifficult.
Thethird problemis labeled theconflation problem. As wealready discussed, thedecision
to enter bankruptcy may beinfluencedby themembers of somereferencegroups inavariety of
ways, afactwetakeintoaccountwhendefiningJ
S1
j
: thiscoefficientisthecompositeof stigmaand
informationeffects. Wedrawonprior work toestablishour separationstrategy (Cohen-Coleand
Zanella, 2008) andthecurvaturedifferencesinindividualsresponsestoinformationandstigma.
Consider equation(3) againandlabel astheprimary model:
j: (1
ijc
= 1) = Φ
Ã
/
j
+ c
j
A
i
+ d
j
1
j
+ J
S1
j
:
j
+ J
S
c
P
¸6=j
c
¸
:
¸
+
ijc
!
. (5)
That is, wedefinethestigmaeffect fromother groups as composed of common and group-
specific factors,
e
J
S
c
=
©
J
S
c
c
¸
ª
¸6=j
, wherethespecific factor is thelocal populationshareinthe
0–1 vs. 1–4 mile radii. If proximity generates the feeling of being observed and such feeling
generates stigma, its intensity is plausibly proportional to therelativenumber of individuals ina
15
Therearenowanumber of methods availablefor theidentification of endogenous effects. Brock and Durlauf
(2001) provides theoriginal references inthis literature. Amongrecent work, Bramoulleet al. (2009) identify peer
effectsinnetworksbyutilizingthefact that networkshaveso-calledintransitivelinks(X talkstoY andY toZ, but X
does not talk to Z). This is effectively aninstrument andallows identification. Similarly, Cohen-Cole(2006) finds
that allowinganindividual to beassociatedwithmultiplereferencegroups allows identification(of asingleeffect)
inalinear model. Graham(2008) uses differences invarianceat theaggregatelevel to infer thepresenceof social
networks. Weconstruct asimilar analysisusingalinear model andaBrock-Durlauf styleexclusionrestrictionbased
onidiosyncraticcredit characteristics. Resultsaresimilar tothosehereandareavailableonrequest fromtheauthors.
10
givenouter groupthat canobservesomebody whohasgonebankrupt. Second, wemust definea
parameter c
j
inorder tospecifyafunctional formfor thetotal stigmafunction:
o (:
j
. :
c
) = c
j
:
j
+ (1 −c
j
)
P
¸6=j
c
¸
:
¸
. (6)
Following Cohen-Cole and Zanella (2008), we assume that stigma fromthe local area and
stigmafromsurroundingareas areperfect substitutes, withmarginal rateof substitutionequal to
c
j
. Our basis for choosingthevalueof this parameter comes fromanapproximationof thefre-
quencyof contactwithindividualsinthetworadii. Assuch, our analysisuses0.25tobegin, which
placesa3:1weight onnon-local stigma. Function(6) isusedtoconstruct anewspecification, the
auxiliary model:
j: (1
ijc
= 1) = Φ[/
j
+ c
j
A
i
+ d
j
1
c
j
+ J
1
j
:
j
+
+J
S
jc
Ã
c
j
:
j
+ (1 −c
j
)
P
¸6=j
c
¸
:
¸
!
+
ijc
]. (7)
The total stigma function captures, by construction, all social effects that work within and
acrosslocalities, butexcludessocial effectsthatworkexclusivelywithinalocality. Thisleavesout
informationsharing, whichis capturedby thefunctionJ
1
j
:
c
j
. Inequations (5) and(7) thereare
four distinct endogenoussocial interactionscoefficients: J
S1
j
isthestigmaandinformationeffects
(in thesuperscript) fromone’s own group (in subscript), J
S
c
is thestigmaeffect fromall other
localities, J
1
j
theinformationeffect fromown-group, andJ
S
jc
isthecompoundstigmaeffect from
bothone’sowngroupandother groups.
Conditional on locality, the auxiliary model does not involve new information. Therefore,
thecorrespondingregressionmodels havethesameerrors, whichis also why thecoefficients on
individual andcontextual effectsaredenotedwiththesamesymbol inbothmodels:
j: (1
ijc
= 1) = Φ
Ã
/
j
+ c
j
A
i
+ d
j
1 + J
S1
j
:
j
+ J
S
c
P
¸6=j
c
¸
:
¸
+
ijc
!
, (8)
j: (1
ijc
= 1) = Φ(/
j
+ c
j
A
i
+ d
j
1 + J
1
j
:
j
+ (9)
+J
S
jc
Ã
c
j
:
j
+ (1 −c
j
)
P
¸6=j
c
¸
:
¸
!
+
ijc
).
Inother words, by construction, thesetwomodels areboth“truemodels”. Consequently, we
cancomparethecoefficients of different social effects across themto obtaintheeffect of stigma
fromgroupq only. Our estimator for thestigmaeffect fromgroupq only, J
S
j
, isthefollowing:
J
S
j
≡ J
S
jc
−J
S
c
. (10)
11
IV Data
A Credit Bureau Data
Becausebankruptcy is principally adecisionto expungeexistingdebt, to gleaninsight into this
particular financial decision, weturntoacomprehensivedatabaseof debt records. Individual debt
records for most types of debt arerecorded by threelargecredit bureaus. Our principal datais
avery largesampleof this dataset providedunder contract by Transunion. Thedataaredrawn
fromgeographicallystratifiedrandomsamplesof individualsandincludeinformationonvariables
commonly availableinapersonal credit report. Inparticular, thefileincludes individual dateof
birth, avariety of account and credit quality information such as thenumber of open accounts,
defaulted accounts, current and past delinquencies, sizeof missed payments, credit lines, credit
balances, etc. The information spans all credit lines, frommortgages, bank cards, installment
loanstodepartment storeaccounts. Transunionalsoprovidesasummary measureof default risk
(aninternal creditscore). Asiscustomary, accountfileshavebeenpurgedof names, social security
numbers, andaddressestoensureindividual confidentiality. However, theydoprovidegeo-coding
informationthat allows us to matchthesepersonal credit history files withinformationfromthe
USCensus, andtoinfer social networks.
The data were drawn fromfour time periods in 18 month intervals—J une 2003, December
2004, J une2006, andDecember 2007. Thefirst twoportionsof thedataprovideabalanced, short
panel of 285,780individuals, whilethesecondtwo compriseavery largerepeatedcross-section
withabout27millionindividuals, aswell asasmaller shortpanel of about 2.2millionindividuals.
Thevery largesizeof therepeatedcrosssectionisespecially important for our analysisof social
interactions, because it allows us to be more confident that the sample average of community-
level effects arevery closeapproximations of thetruepopulation means. Twenty seven million
individualsamount toanapproximate1in9drawof all individualswithacredit history.
Oneof thebenefitsof thecredit databaseusedhereisthat it includesameasureof credit risk.
For eachindividual, Transunionincludes aproprietary credit score. Credit scores ingeneral are
inverseordinal rankingsof risk. That is, anindividual withacredit scoreof 200isviewedtohave
higher riskof default thananindividual of score201. However, whilemost credit scoringsystems
inusearebasedonalogarithmicscale, thedifferenceinriskbetween200and201mayor maynot
beequal tothechangefrom201to202. AsinGrossandSouleles(2002), weusethisproprietary
credit scoreas acontrol for changes intherisk compositionof borrowers, together withaccount
informationoncredit lines, balances, andutilizationrates.
The data set also includes information on individual public bankruptcy filings. Transunion
keepsthebankruptcyonfilefor atleast7yearsafter thefiling, soour dataencompassbankruptcies
as early as 1996. We use all historical bankruptcies in our analysis. Given the availability of
geo-codinginformationfor theindividuals, weareableto computelocal bankruptcy rates. The
bankruptcyvariableusedisanindicator of whether anindividual hasfiledbankruptcyinthepast7
years. Thishastheadvantageof capturinglingeringstigmaandinformationeffectsof individuals
that filedover thepast fewyears.
Our administrativedataprovidesanadvantageoverpublicmeasuresof bankruptcy, particularly
whenonewantstounderstandtheroleof social networks. Usingour owncredit bureaudata, we
12
are able to construct bankruptcy rates at a very low level of aggregation, which allow precise
interpretationsof local or networkeffectsthanthestate-level average. Weuseconstantgeographic
radii of 1 mileand 4 miles as measures of relevant referencerates for social information. The
Transuniondatafieldsusedfor thisstudydonotdistinguishbetweentypesof bankruptcy(Chapter
7vs. Chapter 13), assuch, our measureisatotal personal bankruptcyrate.
B Census Data and Other Information
Asalreadymentioned, weuseanindividual’sgeo-codedcensusblockaddressfromtheTransunion
data, andlink awidevariety of informationonlocationcharacteristics. Inparticular, becausewe
donot haveindividual-level dataonvariablessuchasincomeandeducation, weusethefollowing
variables to control for local economic and demographic conditions. For demographic controls
(education, race, andmarital status), weusedatafromtheUS2000Censusnational summaryfiles
andmergeinformationat theneighborhoodlevel (definedas a1mileradius) averages. Weuse
dataonmedianhouseholdincomesandpovertyratesfromtheUS2000Censusandthe2005and
2006AmericanCommunitySurveysat thecountylevel. WealsomatchinformationfromtheCur-
rent PopulationSurvey andLocal AreaUnemployment Statisticsof theBLS onhealthinsurance
coverage(at thestatelevel) and unemployment rates (at thecounty level), respectively, for the
correspondingyears. Thekey advantagehereisthat weareabletolink informationat agranular
level that controlsfor thewideheterogeneityineconomicshocksfacedintheUSeconomy.
Whenall this informationhas beenmerged, of theoriginal sampleof observations, acertain
number of individualsget droppedduetomissingdata, for exampleoncredit scores. Oncethese
andother similar missingobservationsareremoved, wehaveabout150,000observationsavailable
for 2003and2004, andabout12millionfor 2006and2007.
16
TableI providesdetaileddescription
of all thevariablesweuseinour analysesaswell astheir respectivesources, andTableII presents
somesummarystatistics.
V Results
Inthissection, wepresent detailsonour four results.
A Result 1: Social Spillovers are Very Large
Our initial resultsisthat combinedsocial spilloversarebotheconomically largeand30-50times
larger thancommonly consideredrisk factors suchas unemployment rates andcredit utilization.
Wefind(seeTableIII) that a1%changeinlocal bankruptcyratesleadstoanincreasedprobability
of bankruptcyat theindividual level of 0.25-0.4%. Incomparison, a100percent increaseincredit
16
Missinginformationoncredit fileinformationcomes fromgaps intheoriginal data. Missinginformationfrom
thedemographic files isduetodiscrepancies betweenthegeo-codesfromthecredit bureauandthecensus. Whena
geo-codefromthecredit bureaulay morethanamilefromtheclosest censusblock groupcentroidfromthecensus,
thedatapoint is excluded. Onecanalsomatchtheseremainingpointsby associatingtheindividual withtheclosest
centroidandruntheriskof connectingtheindividual withanincorrectneighborhood. Nonetheless, thekeycoefficients
onaregressionusingthismethodologyaresubstantivelyunchangedfromthebaselinesbelow.
13
utilizationleadstoachangeintheprobabilityof abankruptcyfilingthat is1/1000themagnitude
of thesocial effect. Similarlysmall magnitudesareobservablefor unemployment rates.
Toshowthis, weestimatetheequationsusedinelsewhereintheliterature. Thespecification
is:
j: (1
iIc
) = Φ(/ + cA
i
+ d1
j
+ J
c
:
c
+
iI
) (11)
whereA
i
areindividual-specific credit characteristics taken fromour credit file. Theseinclude
dateof birthof theaccount holder, andamount of outstandingdebt, total credit lineandutiliza-
tion rates for revolving credit, mortgageline, as well as an aggregatemeasureof credit quality
(theinternal credit score). Thesevariables correspond to therisk-controls used in theGross &
Souleles(2002) model, andcapturedifferencesinriskcompositionsof borrowers. Wealsoinclude
community-level controls to proxy for local economic conditions anddemographic composition
of theneighborhoodandthecounty, labeled1
j
. This vector includes controls for neighborhood
race, education, andmarital statuscomposition, together withmedianhouseholdincomeandun-
employment rateinthecountyof residence, averageincomegrowthintheneighborhoodbetween
2000and2005, thepercentageof peoplewithouthealthinsuranceinthestateof residence, andthe
percentageof peopleonpublicassistanceintheneighborhood. Finally, weincludethebankruptcy
ratefor thestateof residence, computed usingour own sampleaverages fromthecredit bureau
data.
TableIII presents theresults fromthis exerciseineachof our four datedobservations (J une
2003, December 2004, June2006, December 2007). Ineachof thefour timeperiods, almost all
of thecredit risk controlsaresignificant albeit small.
17
For example, theTransunionscoreissig-
nificant andisinlinewithexpectations: peoplewithhigher credit scoresarelesslikelytofilefor
bankruptcy. Individualswithhigher limits(revolve_cred) arelesslikely todefault, andincreased
utilization, particularlyintheextremes(credit_utilsq), leadstoincreasedbankruptcyprobabilities.
Theagevariables arealso inlinewithexpectations, whereprobability of default increases with
agebut thenflattensout. Interestingly, communitieswithhigher proportionsof black populations
arelesslikelytodefault, whichwebelieveisconsistent withevidencefoundinforthcomingwork
(Cohen-Cole, 2010) that access to credit is differentiated by location, implying that only rela-
tivelyhigher qualityborrowersinminorityareashaveaccesstocredit. Theeffect of incomeisas
expected: bankruptcyratesarelower inneighborhoodswithhighmedianincome. Similar toprevi-
ousfindings, wealsoshowthattheneighborhoodswithhighpovertyandunemploymentratesalso
seemto havehigher proportionof individuals that becomebankrupt. A key thingto noteinthis
tablehowdemographic andeconomic factors seemto dominateinmagnitudetheeffects of risk
controls, suchasoutstandingdebt balances. Theseresultsalsoshowthat social context andaggre-
gatebehavior indeedplay asignificant roleinindividuals’ bankruptcy decisions: thecoefficients
17
An additional disadvantageto our datais that it includes contemporaneous measures of credit risk rather than
trailingones. Sincethebankruptcyeventoccurswell after individuals’ credithasdeterioratedduetomissedpayments,
increasedutilizationandother factors, asimultaneousmeasureof bankruptcytodayandcredit qualitytoday may not
provideanaccuratereflectionof theroleof risk. Totest theimportanceof thistimingproblem, werepeat our analysis
bylookingattheriskfactorsof individualsinour 2006sampleonthebankruptcyinformationfrom2007. Thisallows
us to account for thefact that theinformationinthe2007datamay beafter individuals havealready changedtheir
behavior. Theresultsfromthisexercise(availablefromtheauthorsuponrequest) showlittlechangeinour estimates
andconfirmthat thefindingsbasedontheprimarysamplearerobust.
14
of theaveragebankruptcyrateinthestateareall highlysignificant andpositive.
18
B Result 2: Both Stigma and Information are Important
Inadditiontothejointeffectof spilloversseenabove, bothinformationandstigmaeffectsindivid-
ually swamprisk controlsandother local measures. UsingthemethodologyfromSectionIII, we
findthata1%changeinlocal bankruptcyratesleadstoanincreasedprobabilityof bankruptcydue
tostigmaof about 3-11%, andduetoinformationof about 5-8%(seeTableIV).
19
Thenumbers reportedarethemarginal effects basedoncoefficients estimatedusingaprobit
model. This regression includes all the independent variables fromthe baseline specifications,
together withaconstant term, but wereport only themarginal effects relatedto thevariables of
interest—information(J
1
j
) andstigma(J
S
jc
).
Theseresultsshowthatthesocial effectsof bothstigmaandinformationarestatisticallysignifi-
cantandhighlyrelevant. Intheearlyportionsof thedata, theeffectof informationappearsslightly
larger thanstigma, withtherelationshipreversedin2006and2007, whichsuggests alarger in-
creaseintheroleof stigmainthis timeframe. Thesenumbers alsoshowthat bothsocial stigma
andinformationcostshaveindeeddecreasedonanational basis: in2006and2007, themagnitudes
of boththestigmaandtheinformationeffectsweresomewhatlarger thanthoseestimatedusingthe
2003and2004samples. Inother words, inthelastfewyearscommunityperceptionshavebecome
increasinglyimportant inhouseholdbankruptcydecisions.
C Result 3: Information Explains the Trend
Our thirdresult isthat informationisthemorelikely explanationfor risingbankruptcy rates, and
lack of informationfor thecontinuedfailureof individuals to file. Themini-trends identifiedin
thelast sectionimply that bankruptcy might indeedbelosingits stigma, as has beenspeculated.
Figure3 illustrates how thechanges in stigmacorrelate with thetrend in bankruptcy rates and
showsthat theestimatedstigmacoefficientsactuallymoveintheoppositedirectiontobankruptcy
trends. Inother words, eventhoughstigmaisvery important andhavedecreasedingeneral over
thelast 5years, thedecreasesinit donot matchtheperiodsof risingbankruptcy rates. However,
informationdisseminationhasuniformlyfollowedthebankruptcychangesduringthetimeperiod
under evaluation. Inthenext section, weprovideadditional evidencetoillustratethesedifferences
bylookingmorecloselyat howbankruptcychangedinsub-samplesof thepopulation.
18
Itisworthnotingthatour baselineresultsshowsimilar directional social effectsasFayet al. (2002) andGrossand
Souleles (2002). However, wefindlarger impacts. Weattributethisfindingtodifferences indataandspecification.
Principally, wenotedagreat deal of sensitivity inthemagnitudeof thecoefficient inthis specification, particularly
withrespect totheinclusionof nonlinear credit scoreterms. Inclusionof thesquaredor cubedcredit scoreleadsto
adrop in themagnitudeof thesocial coefficient. Sincecredit scores areordinal scales, non-linear terms areakin
to rescalingof thevariable. This sensitivity is muchlower inour detailedspecifications below. At lower levels of
aggregation, our coefficientsmatchtherest of theliterature.
19
Becausetheestimatesintheabovesectionweredoneat thestatelevel andtheseat alocal level, theeffectswill
not equivalent inmagnitude.
15
D Result 4: Social Spillovers Vary by Group
It hasbeenwell establishedthat financial decisionsdiffer byincomeandeducation.
20
Indeed, we
canseeinTableV that 14%of bankruptcies occur amongst thepoorest andleast well education
individuals. Accordingly, it isnot surprisingthat social intermediationof thesesamefinancial de-
cisionsmayalsodiffer bysocioeconomicgroup.
21
Wefindthat social factors, bothsocial learning
or social stigma, haveahigher impact onindividual decisions amongst less-educatedandpoorer
communities. In2007, thecoefficientsfor stigmainthepoorest, leasteducatedcell (0.17) arethree
timeslarger thaninitscomplement (0.06). Weseeasimilar patterninthecaseof theinformation
coefficient (seeTableVI).
Thesocial context inwhichindividuals livemay beimportant tounderstandingthenatureof
thesocial interactions guidingtheir decisionmaking. As anexample, onemight imaginethat an
individual facing an adverse shock, such as unemployment, may speak to his or her neighbors
for advicemoreoftenif heknowsthat they arealsoexperiencinghardship. Thisisimportant for
theunderstandingof social interactionsasit impliesthat theestimatesof social effectsmaydiffer
basedonmacroeconomic circumstances. Noticethat thereareacoupleof ways that individuals
mayreacttoaneconomicshock. First, their individual actionssuchasadeclarationof bankruptcy
maychange. Second, anindividual’seconomicdecisionsmaybeinfluencedbythecollectivedeci-
sionsof hisor her social group. Thisisthebasisfor nowcommonestimatesof social interactions
and aretheresults shown in theprior section. Finally, their social behavior itself may change,
whichinturnmayimpact howoftenor intenselytheyrelatetoothers, whichcanthenimpact their
economicdecisionsover andabovethetwoforcesabove. Thatis, thestrengthof thesocial interac-
tionscoefficient (theManski endogenouseffect) maychangeover timeasafunctionof economic
conditions(theManski contextual effects), or varyinthecrosssectioninwaysthat correlatewith
contextual factors. Broadly, thisisanargument that thestrengthof social interactionsmaynot be
universal, andthatunderstandinghowtheseinteractionsdiffer acrossthepopulationmaybeuseful
inunderstandingtheeconomicphenomenainquestion.
Welook at thispossibilitybyparsingour dataalongtwodimensions, incomeandeducational
levels. That is, we subdivide the individuals in our data set into five quintiles of income and
education, creatingatotal of 25groups. Thenwere-estimatetheprincipal modelsabovefor each
subset andreport thesocial spillover coefficientsfor stigmaandinformationinTableVI, panelsA
andB. Panel A includesinformationfrom2006andpanel B for 2007.
22
Wecanalsoseechanges inthesecoefficients inTableVII, withstigmainpanel A andinfor-
mationinpanel B. Thefirst point tonoticeisthat theincreasesinthestigmacoefficient (adecline
in social stigmaassociated with bankruptcy) occur through many of thecells, except theupper
left corner. Inother words, thelargest declinesinsocial stigmaseemtohaveoccurredamongthe
more-educatedandricher individuals, whilethevery poorest showtheoppositeeffect. Informa-
20
Somerecent evidenceisavailableinBertrandandMorse(2009).
21
Guiso, Sapienza, andZingales(2004) findthat social spillovers, describedassocial capital, variesbyregionand
socio-economicconditionsinItaly.
22
Unfortunately, wearenotabletorepeatour analysisof thetemporal changesusingthe2003and2004samplesdue
tolimitednumber of observationsinthoseyears. Despitehavingmorethan250,000observations, the2003and2004
dataarenot sufficientlydensetoallowfor apreciseestimationof theseeffects. Inother words, theeducation-income
“cells” arevery sparsely populated, especially becauseweareinterestedinbankruptcy—analready low-probability
event.
16
tionpatternsshowauniformity acrosssocioeconomic groupsreflectinganincreasingimportance
of informationsharing.
Thesepatternsareilluminatinginthecontextof therecentcreditcrisisinthattheysuggestboth
anincreaseinthevalueof financial education, particularlyfor at-risk segmentsof thepopulation,
andapatternof stigmatizationchanges. Theyimplydeclinesinstigmanot amongst thepoorest or
least well educatedindividuals, but insteadamongst themoreeducatedinsociety. Thesefindings
arealsoconsistent withthediscussioninZywicki (2005). Hearguesthat stigmaplaysalargerole
formiddle-classborrowersbecausetheyhaverelativelythestrongestincentivestofilestrategically.
Ontheother hand, low-incomeborrowershavelower incentivestofilestrategically andaremore
likelytobeconstrainedbyinformationandtransactioncosts.
E Using Movers
Werepeatour analysisfocusingonmoversalonetoidentifythetwosocial effectsbyexploitingthe
availabilityof locationinformation. Bycomparingtheresidential locationin2006with2007, we
candeterminewhichindividualshavemovedinthe18monthperiodbetweenour samples. Using
moverstoidentify hasalonghistory ineconomicsandfinance. Perhapsmost notably, anumber
of papers, includingCarroll, Rhee, andRhee(1994) andFernandez (2007), haveusedimmigrant
behavior to identify theroleof culturein various decision making processes relying on thefact
that immigrantscometonewcountrieswitharangeof information, beliefs, andexpectationsthat
impact their behavior. Overlayingthesepriorswiththeir newenvironment allowsonetoseparate
local phenomenafromthecultural antecedents. Similarly, Guiso, SapienzaandZingales (2004)
exploit thebehavior of moversseparatelywhenanalyzingtheeffectsof social capital onfinancial
development.
Accordingly, to disentanglethetwo social effects of stigmaand informationsharing, were-
peat our analysisusingasubset sampleof movers. Theunderlyingassumptioninthisstrategy is
that whatever informationpeoplehavelearnedregardingthemethodsandproceduresof filingfor
bankruptcy stayswiththem. However, therelevant context inasocial sensefor thesocial stigma
channel is theonefromthenewenvironment. In other words, if information about bankruptcy
cannotbeunlearned, butsocial stigmavariesbylocation, thenwecanre-estimatethesocial effects
usingtheprior areaof residenceasour measureof informational influenceandthecurrent areaas
our measureof stigma.
In each case, welook at 0-4 mileradii and estimateour baselinespecification with asocial
effect onapproximately 100,000‘movers’ inthedata. As showninTableVIII, theresults from
thisexerciseareverysimilar tothestigmacoefficient from2007inTableIV. Theinformationco-
efficientonthemoversisquiteabitlarger. Weinterpretthisdifferenceasafunctionof thefactthat
theindividual mayinfact haveaccesstomorethanasinglesourceof information. Nonetheless, it
reinforcesour primaryfindingsandaddsfurther support toour earlier identificationandlabelling
strategy.
F Additional Robustness
We look in Table IX at varying definitions of stigma. We began by defining a 0-1 and a 1-4
mileradii. As discussedabove, thewider-level effects arethosewhichwill besharedacross the
17
population and, froma social psychology perspective, have no additional impact on individual
decisions aboveandbeyondagivensaturationpoint. Thequestionremains as to theappropriate
areaof referencefromwhichto consider social effects? If indeedstigmais formedbothlocally
and non-locally, what is theappropriateradius? To illustratetherobustness of our findings, we
look at an alternativeconstruction of theradii. Referring to Figure2, this is not aquestion of
identification, rather aquestionof theprecisionof theestimates basedthedividinglinebetween
local andnon-local.
TableIX showsthestigmaandinformationcoefficientsfor all four yearsof our data(2003-07).
Ineachrow, weshowtheresultsfor ayear of data. Eachsub-rowshowsfirst thecoefficientsfrom
thebaselinemodel and second theresults for thestigmacoefficient that is now formed at four
levels insteadat two. Stigmahereis permittedto beacombinationof 0-1, 1-4, county andstate
level averages. Inother words, wedecomposethestigmacomponentsandshowfour social impact
coefficientsinrow2. Thefirstisstill interpretableasaninformationeffect, andisverycomparable
tothebaselinemodel, whiletheother threerepresentvariousstigmaeffects. Thestigmacoefficient
growswiththeareaof relevanceaswouldbeexpectedbythefactthatstate-level interactionswere
larger inmagnitudethanlocal level ones. It alsohighlightsthefact that normativeinfluencesare
definedonabroadersphereof influence; informationeffectsbydefinitiondieoutatlargedistances.
Welook inTableX at thetimingconventionof our information. Recall that inour data, we
flagas bankrupt any individual withabankruptcy declarationintheprior 7years. Notethat the
bankruptciesweobserveinthissamplemay haveoccurredprior toour observationof individual
level credit datain2006. However, inadditiontoprovidinginformationonhowsocial responses
may accumulateover alongperiodof time, weareconfident that our results arerepresentative.
Noteinparticular thevery lowcorrelationbetweentheindividual level credit characteristics and
thebankruptcychoice. After bankruptcy, thesecharacteristicsdeteriorateuniformly, soonewould
expect to seeahigh correlation. Nonetheless, wesubdivideour datato ensurethat wearenot
capturing credit effects post-bankruptcy for someindividuals, were-assess in TableX.
23
In this
table, welook only at individualsthat went bankrupt between2006and2007. Todoso, wedraw
onthefact that 2.2millionof the27millionindividuals inour 2006-2007dataset appear twice.
As a result, we can evaluate our same model above using only those individuals with a recent
bankruptcy, usingonly credit informationavailableprior tothetimeof bankruptcy. Asshouldbe
apparent, thecoefficientsarenearlyidentical acrossthetwomethods.
VI Conclusions
Many individual financial decisionsseempuzzlingtoeconomist asthey repeatedly imply money
left onthetable. Inthis paper weinvestigatetheroleof social factors inhouseholdbankruptcy
decisions as many individuals chosenot tofileevenif they may significantly benefit fromdoing
so. Put differently, in spite of a rapidly rising bankruptcy rate, we emphasize the absence of
additional filingstohighlight thefact that social spilloverssuchasstigmacanprevent individuals
fromtakingadvantageof avaluablebenefit. Inparticular, weanalyzetheempirical relevanceof
thetwoprimarychannelsof social influence—stigmaandsocial learning, andfindthat bothhave
23
Giventheverylowcorrelationsbetweencredit variablesandthebankruptcy decision, it isunlikely that our data
arebiasedinthisfashion.
18
aquantitativelylargeimpact onthebankruptcydecision.
Weencouragecontinued attempts to understand thesourceand natureof social effects at a
level deeper than what has been done in this literature to date. Since the effects appear to be
non-stableover timeandtheir strengthconditional onsocial context, webelieveanunderstanding
of thefeedback betweentheseeffects is essential, especially for understandingthedistributional
implicationsof policychanges.
19
References
[1] Agarwal, S. Chomsisengphet, S. andC. Liu, 2009. “Consumer BankruptcyandDefault: The
Roleof Individual Social Capital”. Federal ReserveBankof Chicago, Mimeo.
[2] Alesina, A., and E. Glaeser. 2004. “Fighting Poverty in the US and Europe,” New York:
OxfordUniversityPress.
[3] Athreya, K., 2002, “WelfareImplicationsof theBankruptcyReformActof 1999,”Journal of
MonetaryEconomics, 49: 1567–95.
[4] Athreya, K., 2004, “ShameAsIt Ever Was: StigmaandPersonal Bankruptcy,” Federal Re-
serveBankof RichmondEconomicQuarterly, 90(2): 1–19.
[5] Barron, J . M., G. Elliehausen, and M. E. Staten, 2000, “Monitoring theHousehold Sector
withAggregateCredit BureauData,” BusinessEconomics.
[6] Bertrand, M andA. Morse, “Financial Literacy, CognitiveBiases andPayday Borrowing,”
Universityof Chicago, Mimeo.
[7] Bramoullé, Y., H. Djebbari, andB. Fortin, 2009, “Identificationof PeerEffectsthroughSocial
Networks,” J ournal of Econometrics, forthcoming.
[8] Brock, W. andS. Durlauf, 2001, “DiscreteChoiceWithSocial Interactions,” Reviewof Eco-
nomicStudies68(2): 235–60.
[9] Brown, J., Ivkovich, Z., Smith, P., Weisbenner, S. 2008. “NeighborsMatter: Causal Commu-
nityEffectsandStockMarket Participation,” J ournal of Finance, LXII: 1509–1531.
[10] Calvo-Armengol, A., E. Patacchini, andY. Zenou, 2009, “Peer EffectsandSocial Networks
inEducation,” Reviewof EconomicStudies, forthcoming.
[11] Campbell, J . D., &P. J . Fairey, 1989, “Informational andnormativeroutestoconformity: the
effect of factionsizeasafunctionof normextremity andattentiontothestimulus.” J ournal
of PersonalityandSocial Psychology, 57: 457–468.
[12] Carroll, C.D., B. Rhee, andC. Rhee, 1994, “AreThereCultural Effects onSaving? Some
Cross-Sectional Evidence,” QuarterlyJ ournal of Economics, 109(3): 685–699.
[13] Channing, Edward, 1921. A Historyof theUnitedStates. Macmillan.
[14] Chatterjee, S., D. Corbae, M. Nakajima, andJ .-V. Rios-Rull, 2007, “A QuantitativeTheory
of UnsecuredConsumer Credit withRiskof Default,” Econometrica75(6): 1525-89.
[15] Cohen-Cole, E., 2006, “MultipleGroupsIdentificationintheLinear-in-MeansModel,” Eco-
nomicsLetters92(2), 753–58.
[16] Cohen-Cole, E., 2010, “Credit CardRedlining,” Reviewof EconomicsandStatistics, forth-
coming.
[17] Cohen-Cole, E. and G. Zanella, 2008, “Unpacking Social Interactions,” Economic Inquiry
46(1), 19–24.
[18] Cohen-Cole, E. andA. Mas, 2009, “Social NormsandProductivity,”Universityof Maryland,
mimeo.
[19] Coleman, P., 1974, DebtorsandCreditorsinAmerica. BeardBooks.
20
[20] Dick, A., andA. Lehnert. (2009), "Product Market CompetitionandPersonal Bankruptcy,"
Journal of Finance
[21] Dick, A., A. Lehnert, andG. Topa(2008), “Social SpilloversinPersonal Bankruptcy,”Federal
ReserveBankof NewYork, mimeo
[22] Duflo, EstherandEmmanuel Saez, 2002, “ParticipationandInvestmentDecisionsinaRetire-
mentPlan: TheInfluenceof Colleagues’ Choices”, Journal of PublicEconomics85, 121-148.
[23] Durlauf, S. (2004) "NeighborhoodEffects"in: J . V. Henderson&J. F. Thisse(ed.), Handbook
of Regional andUrbanEconomics, 1(4), chapter 50, 2173-2242.
[24] Duygan-Bump, B. andC. Grant, 2009, “HouseholdDebt Repayment Behaviour: what role
doinstitutionsplay?” EconomicPolicy.
[25] Efrat, R., 2006, “The Evolution of Bankruptcy Stigma,” Theoretical Inquiries in Law
7(2):365–393.
[26] Elias, S., 2006, The New Bankruptcy: Will it Work for You? NoloPress.
[27] Faulkender, M and J. Yang, 2009, "Inside the Black Box: The Role and Composition of
CompensationPeer Groups" J ournal of Financial Economics, forthcoming
[28] Fay, S., E. Hurst, and M. White, 2002, “TheHousehold Bankruptcy Decision,” American
EconomicReview, 92(3): 706–18.
[29] Fernandez, R., 2007, “CultureandEconomics,” inNewPalgraveDictionary of Economics,
2ndedition.
[30] Flynn, E., andG. Bermant, 2002, “Bankruptcy by theNumbers: A Taleof Two Chapters:
Financial Data,” AmericanBankruptcyInstituteJ ournal, 21(8): 20and38.
[31] Graham, B., 2008, "Identifying Social Interactions through Conditional Variance Restric-
tions" Econometrica76(3): 643–660.
[32] Grinblatt, MarkandMatti Keloharju, 2001, “HowDistance, LanguageandCultureInfluence
StockholdingsandTrades”, J ournal of Finance56, 1053–1073.
[33] Grinblatt, Mark, Matti Keloharju, and Seppo Ikäheimo, 2008, “Social Influenceand Con-
sumption: EvidencefromtheAutomobilePurchases of Neighbors”, Reviewof Economics
andStatistics90, 735–753.
[34] Gross, D. andN. Souleles, 2002, “AnEmpirical Analysisof Personal BankruptcyandDelin-
quency,” TheReviewof Financial Studies, 15(1):319–47.
[35] Guimond, S., 1997, “AttitudeChangeDuring College: Normativeor Informational Social
Influence?” Social Psychologyof Education, 2(3-4):237–261.
[36] Guiso, L., P. Sapienza, andL. Zingales, 2004, “TheRoleof Social Capital inFinancial De-
velopment,” AmericanEconomicReview, 94(3): 526–556.
[37] Guiso, L., P. Sapienza, and L. Zingales, 2009, "Moral and Social Constraints to Strategic
Default onMortgages" ChicagoBoothWorkingPaper.
[38] Hong, Harrison, J effreyD. Kubik, andJ eremyC. Stein, 2004, “Social InteractionandStock-
Market Participation”, J ournal of Finance59, 137-163.
[39] Huberman, Gur, 2001, “Familiarity Breeds Investment”, Review of Financial Studies 14,
659–680.
21
[40] J ones, W. J .,1979, “TheFoundations of EnglishBankruptcy: Statutes andCommissions in
theEarlyModernPeriod,” Transactionsof theAmericanPhilosophical Society, NewSeries,
69(3): 1-63.
[41] Kaustia, MarkkuandSami Torstila, 2008, “Political PreferencesandStockMarketParticipa-
tion”, WorkingPaper, Helsinki School of Economics.
[42] Kaustia, MarkkuandSamuli Knüpfer, 2008, “DoInvestorsOverweightPersonal Experience?
EvidencefromIPOSubscriptions”, Journal of Finance63, 2679-2702.
[43] Kaustia, MarkkuandSamuli Knüpfer, 2009, "Learningfromtheoutcomesof others: Stock
market experiencesof local peersandnewinvestors’ market entry," WorkingPaper, Helsinki
School of Economics.
[44] Latane, B., 1981, “Thepsychology of social impact,” AmericanPsychologist, 36, pp. 343–
356.
[45] Latane, B., andS. Wolf, 1981, “Thesocial impact of majoritiesandminorities”, Psychologi-
cal Review, 88, pp. 438–453.
[46] Leary, M and M Roberts, "Strategic Interaction in Corporate Capital Structure," Working
Paper, Cornell UniversityandUniversityof Pennsylvania
[47] Li, W., andP.-D. Sarte, 2006, “U.S. consumer bankruptcychoice: TheImportanceof General
EquilibriumEffects,” J ournal of MonetaryEconomics, 53(3): 613–31.
[48] Livshits, I., J . MacGee, andM. Tertilt, 2007a, “Consumer Bankruptcy: A FreshStart,”Amer-
icanEconomicReview, 97: 402–18.
[49] Livshits, I., J . MacGee, andM. Tertilt, 2007b, “AccountingFor TheRiseInConsumer Bank-
ruptcies,” NBER WorkingPaper, no: 13363.
[50] Manski, C., 1993, “Identifications of Endogenous Social Effects: TheRefectionProblem,”
Reviewof EconomicStudies, 60: 531–42.
[51] MacIntyre, L. 1989, “A Sociological PerspectiveonBankruptcy,” IndianaLawJ ournal 65:
123–136.
[52] Pomykala, J . 2000, “BankruptcyLaws: theNeedfor Reform,”inLegal Environmentof Busi-
ness, K. Stanberry, ed.: 178–180.
[53] Shive, Sophie, 2009, “AnEpidemic Model of Investor Behavior”, J ournal of Financial and
QuantitativeAnalysis, Forthcoming.
[54] Sullivan, T. A., E. Warren, andJ . L. Westbrook, 2006, “LessStigmaor MoreFinancial Dis-
tress: AnEmpirical Analysisof theExtraordinaryIncreaseinBankruptcyFilings,” Stanford
LawReview59(2): 213-256.
[55] Tabb, C., 1991. “TheHistorical Evolution of theBankruptcy Discharge,” American Bank-
ruptcyLawJ ournal 65: 325–330.
[56] Tanford, S, andS. Penrod, 1984, “Social influencemodel: A formal integrationof research
onmajorityandminorityinfluenceprocesses,” Psychological Bulletin, 95, pp. 189-225.
[57] Thorne, D. andL. Anderson, 2006, “ManagingtheStigmaof Personal Bankruptcy,” Socio-
logical Focus39(2):77–97.
22
[58] Topa, Giorgio, Patrick Bayer, andStephenL. Ross, 2009, “Placeof Work andPlaceof Res-
idence: Informal HiringNetworksandLabor Market”, Journal of Political Economy, Forth-
coming
[59] Wall Street J ournal, December 19, 2007, “Now, Even Borrowers With Good Credit Pose
Risks,” byGeorgeAnders, p. A2.
[60] White, M., 1998, “WhyDon’t MoreHouseholdsFilefor Bankruptcy?” J ournal of Law, Eco-
nomics, andOrganization, 14(2): 205–31.
[61] White, M., 2007, “BankruptcyReformandCreditCards,”Journal of EconomicPerspectives,
21(4): 175–199.
[62] Warren, E., andA. WarrenTyagi, 2003, "TheTwoIncomeTrap: WhyMiddle-ClassMothers
& Fathers AreGoing Broke(With Surprising Solutions That Will ChangeOur Children’s
Futures)", BasicBooks, NewYork.
[63] Zanella, G., 2007, “DiscreteChoicewithSocial InteractionsandEndogenousMemberships,”
J ournal of theEuropeanEconomicAssociation, 5(1): 122–153.
[64] Zitek, E. andM. Hebl, 2007, “TheRoleof Social NormClarityintheInfluencedExpression
of Prejudiceover Time,” J ournal of Experimental Social Psychology, 43: 867–867.
[65] Zywicki, ToddJ ., 2005, “AnEconomicAnalysisof theConsumer BankruptcyCrisis,”North-
westernUniversityLawReview, Vol. 99, No. 4, pp. 1463–1541.
23
TABLE I: VARIABLE DEFINITIONS
VARIABLES DEFINITION SOURCE
age2 age of individual squared authors' calculation based on credit bureau data
avgbkrpt_state average number of bankruptcies filed in the state authors' calculation based on credit bureau data
BRP_ind indicator of public record bankruptcies authors' calculation based on credit bureau data
mortgage_limit mortgage high credit/credit limit authors' calculation based on credit bureau data
credit_util credit utilization, in thousands of dollars authors' calculation based on credit bureau data
credit_utilsq credit utilization, in thousands of dollars, squared authors' calculation based on credit bureau data
age age of individual credit bureau data
revolve_cred total revolving high credit/credit limit, in thousands of dollars credit bureau data
c.score internal credit score credit bureau data
gt_eq_HS_01
percentage of residents in a one mile radius who have achieved high
school equivalency or greater authors' calculation based on data fromU.S. Census 2000
married_01 percentage of residents in a one mile radius who are married authors' calculation based on data fromU.S. Census 2000
divorced_01 percentage of residents in a one mile radius who are divorced authors' calculation based on data fromU.S. Census 2000
perc_black_01 percentage of residents in a one mile radius who are black authors' calculation based on data fromU.S. Census 2000
perc_hispanic_01 percentage of residents in a one mile radius who are Hispanic authors' calculation based on data fromU.S. Census 2000
public_assistance_01 percentage residents who receive public assistance in a one mile authors' calculation based on data fromU.S. Census 2000
incgrowth_inflation average income growth authors' calculation based on data fromACS 2000 & 2005
median household income median household income in county of residence U.S. Census 2000, 2005-2006 American Community Survey
poverty_rate percentage of people below poverty level in county of residence U.S. Census 2000, 2005-2006 American Community Survey
unemployment percentage of unemployed residents in county of residence Bureau of Labor Statistics: Local Area Unemployment Statistics
uninsured percentage of residents in the state who are uninsured U.S. Census Bureau: Current Population Survey
24
TABLE II: SUMMARY STATISTICS
VARIABLES MEAN SD MEAN SD MEAN SD MEAN SD
BRP_ind 0.054 0.226 0.057 0.232 0.054 0.227 0.049 0.215
mortgage_limit ($ thousands) 56.104 121.326 69.965 140.755 71.648 161.225 82.598 181.627
revolve_cred ($ thousands) 35.310 49.141 40.544 59.715 24.741 29.857 25.539 30.465
credit_util ($ thousands) 6.852 14.087 7.968 17.286 7.405 18.536 8.203 20.624
credit_utilsq ($ thousands) 245.40 2,639.08 362.29 4,030.66 398.42 7,989.51 492.65 10,670.38
c. score 648.080 140.447 650.194 139.487 697.180 142.987 696.443 145.356
age 48.798 17.133 49.661 17.032 37.379 11.221 37.405 11.314
age2 2,674.74 1,843.14 2,756.26 1,852.51 1,523.08 898.49 1,527.15 900.33
perc_blac~01 0.094 0.169 0.096 0.172 0.103 0.176 0.099 0.172
perc_hisp~01 0.108 0.167 0.110 0.169 0.124 0.181 0.123 0.181
gt_eq_HS_01 0.828 0.117 0.821 0.119 0.827 0.121 0.829 0.120
married_01 0.577 0.108 0.572 0.106
divorced_01 0.096 0.034 0.097 0.034
public_as~01 0.030 0.032 0.031 0.032 0.030 0.032 0.031 0.032
incgrowth_inflation 1.004 2.940 0.995 2.917 0.996 2.931 0.959 2.898
median_HH_inc 45,016 10,803 44,827 10,820 50,090 12,309 52,516 12,614
unemployment 5.788 1.433 5.993 1.496 5.038 1.323 4.599 1.283
poverty_rate 11.676 5.131 11.708 5.144 12.481 4.893 12.487 4.642
uninsured 15.020 4.091 15.355 3.879 15.729 4.188 15.619 4.486
avgbkrpt_state 0.048 0.012 0.053 0.013 0.054 0.012 0.049 0.011
Number of observations 145,567 145,567 152,441 152,441 16,801,971 16,801,971 17,051,621 17,051,621
Notes: Based on authors' calculations using credit bureau data, Census and other information as described in thedata section, and TableI.
2003 2004 2006 2007
25
TABLE III: BASELINE SPECIFICATION
2003 2004 2006 2007
mortgage_limit ($ thousands) 0.00000426** 0.00000599*** -0.00000327*** -0.00000762***
(0.000002) (0.0000022) (0.00000023) (0.00000020)
revolve_cred ($ thousands) -0.000572*** -0.000497*** -0.000467*** -0.000499***
(0.000014) (0.000014) (0.0000014) (0.0000012)
credit_util ($ thousands) 0.0000508 -0.00000479 -0.0000416*** 0.000278***
(0.000038) (0.000039) (0.0000038) (0.0000028)
credit_utilsq ($ thousands) 0.000000864*** 0.000000933*** 0.000000374*** 0.000000108***
(0.00000013) (0.000000088) (0.0000000032) (0.0000000025)
c.score -0.000117*** -0.000150*** -0.000138*** -0.0000967***
(0.0000042) (0.0000042) (0.00000040) (0.00000033)
age 0.00274*** 0.00318*** 0.00833*** 0.00766***
(0.00011) (0.00011) (0.000026) (0.000024)
age2 -0.0000243*** -0.0000281*** -0.0000928*** -0.0000858***
(0.000001) (0.0000011) (0.00000031) (0.00000029)
perc_black_01 -0.00875*** -0.0101*** -0.0107*** -0.00738***
(0.0014) (0.0016) (0.00017) (0.00016)
perc_hispanic_01 -0.000654 -0.00132 0.00108*** 0.000534**
(0.0019) (0.0022) (0.00024) (0.00023)
gt_eq_HS_01 0.0139*** 0.0135*** 0.00350*** 0.00236***
(0.0028) (0.0032) (0.00037) (0.00034)
married_01 0.00333 0.00149
(0.0024) (0.0028)
divorced_01 0.0389*** 0.0359***
(0.00088) (0.00081)
public_assistance_01 0.0236*** 0.0361*** 0.0442*** 0.0376***
(0.0086) (0.01) (0.0012) (0.0011)
incgrowth_inflation 0.000148** 0.000159* 0.0000749*** 0.0000537***
(0.000075) (0.000091) (0.00001) (0.00001)
median_HH_inc 0.0000000184 -0.0000000476 -0.0000000657*** -0.000000104***
(0.000000034) (0.000000041) (0.000000004) (0.000000004)
unemployment 0.0000237 0.0000246 0.00000124 0.000138***
(0.00017) (0.00019) (0.00002) (0.00002)
poverty_rate -0.000214*** -0.000348*** -0.000367*** -0.000397***
(0.000078) (0.000091) (0.00001) (0.00001)
uninsured -0.000326*** -0.000453*** -0.000248*** -0.000182***
(0.000063) (0.000079) (0.0000079) (0.0000066)
avgbkrpt_state 0.345*** 0.404*** 0.289*** 0.260***
(0.019) (0.021) (0.0024) (0.0024)
Number of observations 145,567 152,441 12,300,000 12,400,000
Notes: Thedependent variableisanindicator for existenceof abankruptcyfilinginthe7yearsprior tothedateof thecredit report.
Thereportedcoefficients arethemarginal effects at themeanestimatedusingaprobit model. SeeTableI for adetaileddescription
of eachof thevariables. A constant termwasalsoincludedbut isnot reportedhere. Standarderrorsarereportedinparentheses, and
we adopt the usual convention: *** p<0.01, ** p<0.05, * p<0.1.
26
TABLE IV: TOTAL STIGMA AND INFORMATION
2003 2004 2006 2007
Stigma 0.0275** 0.0384** 0.118*** 0.106***
(0.0141) (0.0157) (0.0018) (0.0016)
Information 0.0532*** 0.0638*** 0.0948*** 0.0746***
(0.00612) (0.00709) (0.0014) (0.0013)
Number of Observations: 131,430 135,046 12,300,000 12,300,000
Notes: Thedependent variableis an indicator for existenceof abankruptcy filingin the7years prior to the
date of the credit report. The reported coefficients are the marginal effects at the mean estimated using a
probit model. Thisregressionincludestheindependent variablesfromthespecificationsinTableIII, together
with a constant term, but are not reported here for brevity. We report the marginal effects related to the
variables of interest – information and stigma. Theseresults arebased on equation 7 in thetext, wherewe
assume α=0.75, which denotes the marginal rate of substitution between stigma fromlocal and non-local
groups, and puts 3:1 weight on non-local stigma. The stigma variable shown in this table refers to 'total
stigma' as defined in the paper. Local and non-local stigma estimates are available fromthe authors upon
request. Standard errors are reported in parentheses, and we adopt the usual convention: *** p<0.01, **
p<0.05, * p<0.1.
27
TABLE V: DISTRIBUTION OF BANKRUPTCIES BY EDUCATION AND INCOME QUINTILES
1 2 3 4 5 1 2 3 4 5
Education Education
1 13.74 5.33 0.91 0.21 0.03 1 14.73 5.20 1.13 0.20 0.07
2 4.80 10.93 6.31 1.36 0.30 2 5.23 10.31 6.56 1.45 0.25
3 1.24 5.65 8.67 4.96 0.90 3 1.25 5.78 8.70 5.04 0.89
4 0.47 1.75 5.70 8.17 3.62 4 0.40 1.92 5.85 7.38 3.52
5 0.30 0.41 1.28 5.08 7.88 5 0.35 0.38 1.16 4.57 7.70
1 2 3 4 5 1 2 3 4 5
Education Education
1 14.62 4.41 0.75 0.16 0.05 1 14.62 4.39 0.76 0.17 0.06
2 3.78 9.04 5.44 1.40 0.35 2 3.80 9.03 5.39 1.42 0.36
3 0.93 4.83 7.75 5.14 1.35 3 0.96 4.85 7.70 5.12 1.36
4 0.37 1.38 4.94 8.26 5.05 4 0.35 1.39 5.00 8.23 5.03
5 0.30 0.34 1.12 5.04 13.21 5 0.27 0.33 1.15 5.06 13.20
Income Quintile Income Quintile
2003 2004
% of Total
Bankruptcies:
% of Total
Bankruptcies:
Notes: Thevaluesreportedarethepercentageof all bankruptciesinour samplefor eachof theyears2003, 2004, 2006and2007attributabletoeachincome/educationgroup. Thevaluesareaggregatedacross
twodimensions, lowest tohighest incomequintiles (basedonaggregatehouseholdincomeinazerotoonemileradius) andlowest tohighest educationquintiles (basedonpercentageof residents withhigh
school equivalency or greater in azero to onemileradius).
Income Quintile Income Quintile
2006 2007
% of Total
Bankruptcies:
% of Total
Bankruptcies:
28
TABLE VI: STIGMA AND INFORMATION ACROSS EDUCATION AND INCOME QUINTILES
Stigma: Stigma:
1 2 3 4 5 1 2 3 4 5
Education Education
1 0.31*** 0.20*** 0.18*** 0.08** 0.01 0.17*** 0.09*** 0.12*** 0.13*** 0.05
2 0.11*** 0.25*** 0.2*** 0.16*** 0.14*** 0.15*** 0.21*** 0.17*** 0.12*** 0.04*
3 0.13*** 0.14*** 0.15*** 0.12*** 0.11*** 0.15*** 0.17*** 0.15*** 0.13*** 0.07***
4 0.11*** 0.17*** 0.18*** 0.11*** 0.06*** 0.1*** 0.19*** 0.19*** 0.13*** 0.07***
5 0.05*** 0.13*** 0.14*** 0.13*** 0.06*** 0.05*** 0.17*** 0.17*** 0.14*** 0.06***
Information: Information:
1 2 3 4 5 1 2 3 4 5
Education Education
1 0.14*** 0.12*** 0.08*** 0.02 0.00 0.14*** 0.12*** 0.02* -0.02 0.02
2 0.22*** 0.13*** 0.07*** 0.05*** 0.00 0.15*** 0.1*** 0.07*** 0.04*** 0.02*
3 0.14*** 0.13*** 0.12*** 0.08*** 0.04*** 0.11*** 0.09*** 0.09*** 0.05*** 0.04***
4 0.01 0.09*** 0.07*** 0.09*** 0.07*** 0.04** 0.07*** 0.05*** 0.05*** 0.05***
5 0.04*** 0.05*** 0.07*** 0.08*** 0.07*** 0.04** 0.03 0.02* 0.04*** 0.05***
Notes: Thedependent variableineachregressionis anindicator for existenceof abankruptcyfilinginthe7years prior to thedateof thecredit report. Thereportedcoefficients arethemarginal
effects at themeanestimated using aprobit model. Eachcell pair (onestigmaandoneinformationinagivenyear) represents asingleregressionof theformseeninTableIV. Eachregressions
includes all the independent variables fromthe baseline specifications, together with a constant term. As in Table IV, we report only the marginal effects related to the variables of interest –
informationandstigma. Theseresults arebasedontheauxiliarymodel, whereweassumeα=0.75, whichdenotes themarginal rateof substitutionbetweenstigmafromlocal andnon-local groups,
andputs 3:1weight onthenon-local stigma. Thetableisconstructedto showvaluesacross two dimensions, lowest to highest incomequintiles(basedonaggregatehouseholdincomeinazero to
onemileradius) andlowest to highest educationquintiles(basedonpercentageof residentswithhighschool equivalencyor greater in azero to onemileradius). Thestigmavariableshowninthis
tablerefersto 'total stigma' asdefinedabove. Local andnon-local stigmaestimatesareavailablefromtheauthorsuponrequest. Standarderrorsarereportedinparentheses, andweadopt theusual
convention: *** p<0.01, ** p<0.05, * p<0.1.
2006 2007
Income Quintile Income Quintile
Income Quintile Income Quintile
29
Stigma:
1 2 3 4 5
Education
1 (0.146) (0.107) (0.060) 0.046 0.036
2 0.047 (0.033) (0.031) (0.042) (0.104)
3 0.026 0.029 (0.008) 0.013 (0.035)
4 (0.008) 0.021 0.015 0.017 0.013
5 0.003 0.038 0.036 0.006 (0.004)
Information:
1 2 3 4 5
Education
1 0.009 (0.006) (0.053) (0.037) 0.023
2 (0.072) (0.029) (0.005) (0.007) 0.026
3 (0.031) (0.042) (0.031) (0.037) 0.009
4 0.021 (0.023) (0.024) (0.033) (0.019)
5 (0.003) (0.023) (0.048) (0.033) (0.014)
Notes: Thevaluesreportedarethedifferenceinsetsof informationandstigmacoefficientsfrom2006to2007.
The values are aggregated across two dimensions, lowest to highest income quintiles (based on aggregate
householdincomeinazeroto onemileradius) andlowest to highest educationquintiles(basedonpercentage
of residents with high school equivalency or greater in a zero to one mile radius).
TABLE VII: CHANGES IN STIGMA AND INFORMATION
Change in Information:
2006 - 2007
Income Quintile
Change in Stigma:
2006 - 2007
Income Quintile
30
Baseline Movers
Stigma 0.155*** 0.167***
(0.0250) (0.015)
Information 0.0952*** 0.263***
(0.0190) (0.017)
Number of Observations 108,700 109,023
TABLE VIII: MOVERS
Notes: Thenumbersreportedarethemarginal effectsbasedoncoefficientsestimatedusing
a probit model. This regression includes all the independent variables fromthe baseline
specifications, together with a constant term, however only the stigma and information
coefficients arereported here. Thedataset used for theseregressions containindividuals
who have a primary residence four miles or more fromtheir 2006 residence. The first
columnshowstheresultsfromtheauxiliarymodel whentherestricteddataset isused. The
second column uses 2006 controls, defines the information group as those located 0-4
milesfromanindividuals' 2006residence, anddefinesthestigmagroupasthoselocated0-
4 miles froman indviduals' 2007 residence. Local and non-local stigma estimates are
availablefromtheauthors upon request. Standard errors arereported in parentheses, and
we adopt the usual convention: *** p<0.01, ** p<0.05, * p<0.1
31
Information Stigma Stigma1 Stigma2 Stigma3
Baseline 0.0532*** 0.0275*
Stigma (Multiple) 0.0549*** 0.00477 0.210*** 0.293***
Information Stigma Stigma1 Stigma2 Stigma3
Baseline 0.0638*** 0.0384**
Stigma (Multiple) 0.0664*** 0.0107 0.262*** 0.376***
Information Stigma Stigma1 Stigma2 Stigma3
Baseline 0.0948*** 0.118***
Stigma (Multiple) 0.116*** 0.0634*** 0.0901*** 0.110***
Information Stigma Stigma1 Stigma2 Stigma3
Baseline 0.0746*** 0.106***
Stigma (Multiple) 0.0938*** 0.0622*** 0.0711*** 0.0979***
Notes: The values reported are the marginal effects based on coefficients estimated using a probit
model. This regression includes all the independent variables fromthe baseline specifications,
together with a constant term, however only the stigma and information coefficients are reported
here. The definition of stigma is particular to each column heading. 'Stigma' refers to 'total
stigma' as defined above. 'Stigma (4 Level)' is similar to stigma as defined in the auxiliary model
with the sole exception that the non-local group is an equal weight average of 1-4 mile radius,
county, and state level bankruptcy averages. 'Stigma1' refers to the 1-4 mile bankruptcy average.
'Stigma2' refers to the county bankruptcy average. 'Stigma3' refers to the state bankruptcy
average. Local and non-local stigma estimates are available fromthe authors upon request.
Standard errors are reported in parentheses, and we adopt the usual convention: *** p<0.01, **
p<0.05, * p<0.1
TABLE IX: ALTERNATIVE STIGMA DEFINITIONS
2004
2007
2003
2006
32
2007 controls 2006 controls
Stigma 0.106*** 0.122***
(0.0016) (0.0059)
Information 0.0746*** 0.0878***
(0.0013) (0.0046)
Number of Observations: 12,300,000 1,093,448
Notes: Thenumbers reportedarethemarginal effects basedoncoefficients estimatedusing
a probit model. This regression includes all the independent variables fromthe baseline
specifications, together with a constant term, however only the stigma and information
coefficients are reported here. For comparison, the first column shows the 2007 results
fromTable IV. In the second column, 2007 bankruptcy is regressed on 2006 controls.
Local andnon-local stigmaestimates areavailablefromtheauthors uponrequest. Standard
errors are reported in parentheses, and we adopt the usual convention: *** p<0.01, **
p<0.05, * p<0.1
TABLE X: 2007 BANKRUPTCY, 2006 CONTROLS
33
FIGURE 1: QUARTERLY NONBUSINESS BANKRUPTCY FILINGS (IN THOUSANDS)



Source: American Bankruptcy Institute.



0
100
200
300
400
500
600
700
1
9
9
4

Q
1
1
9
9
5

Q
1
1
9
9
6

Q
1
1
9
9
7

Q
1
1
9
9
8

Q
1
1
9
9
9

Q
1
2
0
0
0

Q
1
2
0
0
1

Q
1
2
0
0
2

Q
1
2
0
0
3

Q
1
2
0
0
4

Q
1
2
0
0
5

Q
1
2
0
0
6

Q
1
2
0
0
7

Q
1
2
0
0
8

Q
1
F
i
l
i
n
g
s

(
t
h
o
u
s
a
n
d
s
)
34
FIGURE 2: SOCIAL INFLUENCE

Normative
Social
Influence
Probability
Of
Action
Number
of People
Influencing
Informational and Normative
Social
Influence
Local Non-local


Notes: Authors' illustration.



35
FIGURE 3: STIGMA AND BANKRUPTCY



Source: American Bankruptcy Institute & Author’s calculations using Credit Bureau and Census 2000 data. The left
hand side axis is the inverse of the estimated sigma coefficient, such that an increase in the coefficient corresponds to a
decline in stigma – and vice versa. The left hand side axis represents the number of bankruptcy filings and corresponds
to the continuous time series variable represented in the figure.

0.82
0.84
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
2
0
0
0

Q
4
2
0
0
1

Q
2
2
0
0
1

Q
4
2
0
0
2

Q
2
2
0
0
2

Q
4
2
0
0
3

Q
2
2
0
0
3

Q
4
2
0
0
4

Q
2
2
0
0
4

Q
4
2
0
0
5

Q
2
2
0
0
5

Q
4
2
0
0
6

Q
2
2
0
0
6

Q
4
2
0
0
7

Q
2
I
n
v
e
r
s
e

o
f

t
h
e

E
s
t
i
m
a
t
e
d

S
t
i
g
m
a

c
o
e

0
100
200
300
400
500
600
700
N
o
.

o
f

B
a
n
k
r
u
p
t
c
y

F
i
l
i
n
g
s

(
0
0
0
s
)
36

Sponsor Documents

Or use your account on DocShare.tips

Hide

Forgot your password?

Or register your new account on DocShare.tips

Hide

Lost your password? Please enter your email address. You will receive a link to create a new password.

Back to log-in

Close