The online version of this article can be found at: http ...users.soc.umn.edu/~uggen/Kirk_ASR_09.pdfAMERICAN SOCIOLOGICAL REVIEW, 2009, VOL. 74 (June:484–505) Direct correspondence

http://asr.sagepub.com/American Sociological Review

http://asr.sagepub.com/content/74/3/484The online version of this article can be found at:

DOI: 10.1177/000312240907400308

2009 74: 484American Sociological ReviewDavid S. Kirk

Hurricane KatrinaA Natural Experiment on Residential Change and Recidivism: Lessons from

Published by:

http://www.sagepublications.com

On behalf of:

American Sociological Association

can be found at:American Sociological ReviewAdditional services and information for

http://asr.sagepub.com/cgi/alertsEmail Alerts:

http://asr.sagepub.com/subscriptionsSubscriptions:

http://www.sagepub.com/journalsReprints.navReprints:

http://www.sagepub.com/journalsPermissions.navPermissions:

http://asr.sagepub.com/content/74/3/484.refs.htmlCitations:

at Serials Records, University of Minnesota Libraries on January 10, 2011asr.sagepub.comDownloaded from

http://asr.sagepub.com/

http://asr.sagepub.com/content/74/3/484

http://www.sagepublications.com

http://www.asanet.org

http://asr.sagepub.com/cgi/alerts

http://asr.sagepub.com/subscriptions

http://www.sagepub.com/journalsReprints.nav

http://www.sagepub.com/journalsPermissions.nav

http://asr.sagepub.com/content/74/3/484.refs.html


For the first time in my life, I didn’t have the feel-ing that I had to go to Coxsackie, to Woodburn, andthen to Sing Sing. I had the feeling now that any-thing could happen, anything that I decided to do.It seemed a little bit crazy, but I even had the feel-

ing that if I wanted to become a doctor or some-thing like that, I could go on and do it. This wasthe first time in my life that I’d had that kind of feel-ing, and getting out of Harlem was the first steptoward that freedom.

— Claude Brown (1965:178)

Research reveals that ex-prisoners tend tobe geographically concentrated in a rela-

tively small number of neighborhoods withinmetropolitan areas, often returning to the samecriminogenic neighborhoods where they residedprior to incarceration (La Vigne et al. 2003).Alarmingly, a majority of ex-prisoners will beback in confinement within three years (Langanand Levin 2002). Yet it may be premature toimplicate geographic context as a culprit in thisvicious cycle. Estimating the causal impact ofplace of residence on the likelihood of recidi-vism is complicated by the issue of selectionbias—that is, the possibility that some unmea-sured characteristic of individuals influencesboth where they live and their criminal behav-ior and may account for any relation betweenplace of residence and recidivism. Individualswith a high propensity toward criminal offend-ing, and therefore a high likelihood of recidi-

A Natural Experiment on Residential Change and Recidivism: Lessons from Hurricane Katrina

David S. KirkUniversity of Texas at Austin

Ex-prisoners tend to be geographically concentrated in a relatively small number of

neighborhoods within the most resource deprived sections of metropolitan areas.

Furthermore, many prisoners return “home” to the same criminogenic environment with

the same criminal opportunities and criminal peers that proved so detrimental prior to

incarceration. Yet estimating the causal impact of place of residence on the likelihood of

recidivism is complicated by the issue of selection bias. In this study, I use a natural

experiment as a means of addressing the selection issue and examine whether the

migration of ex-prisoners away from their former place of residence will lead to lower

levels of recidivism. In August 2005, Hurricane Katrina ravaged the Louisiana Gulf

Coast, damaging many of the neighborhoods where ex-prisoners typically reside. The

residential destruction resulting from Hurricane Katrina is an exogenous source of

variation that influences where a parolee will reside upon release from prison. Findings

reveal that moving away from former geographic areas substantially lowers a parolee’s

likelihood of re-incarceration.

AMERICAN SOCIOLOGICAL REVIEW, 2009, VOL. 74 (June:484–505)

Direct correspondence to David Kirk, Departmentof Sociology, University of Texas at Austin, 1University Station A1700, Austin, TX 78712([email protected]). This work has been sup-ported by a grant from the National Poverty Centerat the University of Michigan. I thank John Laub,Andy Papachristos, David Thacher, David Weisburd,Chris Wildeman, the editors, the anonymous review-ers, and participants of the National Poverty Center’s2008 grantee workshop for comments on an earlierdraft of this manuscript. I also thank Bill Sabol andRob Sampson for helpful advice on the conceptualand methodological design of the study, and I amgrateful to Sean Goodison, Maggie Pendzich-Hardy,and Wendy Stickle for research assistance. Finally, Iam indebted to the Louisiana Department of PublicSafety & Corrections for their support, especiallySecretary James LeBlanc and Susan Lindsey. Anyfindings or conclusions expressed are those solely ofthe author.



vism, might self-select into certain geographiccontexts, and the characteristics of these con-texts might have little causal bearing on indi-viduals’ behavior. Moreover, parolees whoreturn “home” following incarceration to thesame neighborhoods where they resided priorto incarceration might be fundamentally dif-ferent on unobserved characteristics relative toparolees who move to different neighborhoodsfollowing incarceration. Without accounting forsuch processes of residential selection, observedcorrelations between recidivism and residen-tial migration likely reflect the effect of unmea-sured individual or family characteristics thatmotivated residential choice. Consequently,empirical estimates of the effects of residentialmigration on recidivism may be biased andinconsistent. If scholars of social context and“neighborhood effects” are to make causalstatements about the impact of place of resi-dence or residential change on a given outcomemeasure, they must address the issue of selec-tion bias by separating out selection effectsfrom true contextual effects.

In this study, I use a natural experiment as ameans of minimizing the potential for selectionbias in a study of recidivism. Hurricane Katrinaand the tragic events that followed have impli-cations for prisoner reentry,1 and, in particular,the selection of residence for those prisonersreleased from incarceration following Katrina.2

In Orleans Parish in Louisiana, 71.5 percent ofhousing units suffered some damage followingHurricane Katrina, with 56 percent of housingunits significantly damaged (U.S. Departmentof Housing and Urban Development 2006). Theextent of housing destruction was similar inadjacent parishes. In both St. Bernard Parish and

Plaquemines Parish, 80 percent of housing unitswere damaged, while 70 percent were damagedin St. Tammany Parish and 53 percent weredamaged in Jefferson Parish.

The tragedy of Katrina presents an opportu-nity—a natural experiment—to assess theeffects of residential change on recidivism. Theresidential destruction resulting from HurricaneKatrina is an exogenous source of variation thatinfluences where a parolee will reside uponrelease from prison. In the absence of completedata on why a given parolee moves to one geo-graphic area versus another, to estimate thecausal effect of residential migration (i.e., thetreatment) it is advantageous to have an exoge-nous source of variation that substantially influ-ences this treatment.

This study aims to answer the followingresearch questions: First, how has the geo-graphic distribution of parolees released fromLouisiana prisons changed following HurricaneKatrina? Second, if there has been a geograph-ic displacement of parolees following the hur-ricane, has this hurricane-induced migrationhad any negative, or even beneficial, effects onthe likelihood of recidivism?

If Claude Brown’s (1965) tale of residentialchange and desistance from crime generalizesto the parolees in Louisiana induced to movebecause of Hurricane Katrina, we may findthat Katrina actually led to positive outcomesfor this particular slice of the population. Thelesson from Katrina may be that residentialchange leads to turning points in the lives ofparolees.

THE GEOGRAPHIC CONTEXT OFPRISONER REENTRY

The Bureau of Justice Statistics reports thatpresently more than 700,000 prisoners arereleased from state and federal facilities eachyear (Sabol and Couture 2008); estimates suggestthat up to half of these releasees will have beenin prison before (Langan and Levin 2002). Bysome estimates, two-thirds of former prisoners inthe United States are rearrested within three yearsof prison release, and half are re-incarcerated(Langan and Levin 2002). These figures shouldnot be separated from the geographic context towhich prisoners return. Ex-prisoners tend to begeographically concentrated within the mostresource deprived sections of metropolitan areas,

RESIDENTIAL CHANGE AND RECIDIVISM—–485

1 “Prisoner reentry” refers to the process of leav-ing prison and returning to the community (NationalResearch Council 2007).

2 Hurricane Katrina initially developed on August23rd, 2005. It made its first landfall on the 25th nearFt. Lauderdale, Florida as a Category 1 hurricane. Thestorm then moved into the Gulf of Mexico, becom-ing a Category 5 hurricane by the 28th. On August29th, Katrina made its second landfall, in southernLouisiana, and then moved inland toward NewOrleans. Levee walls on the Industrial Canal in NewOrleans breached from the storm a couple of hourslater, and additional levee breaches followed soonafter.



often returning to the same neighborhoods wherethey resided prior to incarceration. For instance,research by the Urban Institute on reentry inIllinois reveals that over half of prisoners releasedfrom Illinois prisons return to the city of Chicago,and one-third of those returning to Chicago areconcentrated in just six community areas (LaVigne et al. 2003). These six communities areamong the most economically disadvantaged inthe city. Given that many prisoners return hometo their old neighborhoods, it is important todevelop a theoretical understanding for why thisfact is consequential.

PLACE OF RESIDENCE, RESIDENTIALMIGRATION, AND RECIDIVISM

The alarming rates of recidivism in the UnitedStates demonstrate that there is much continu-ity in criminal behavior over the life course.Sampson and Laub (1993) explain such conti-nuity by the absence of social controls (e.g., sta-ble employment and a healthy marriage). YetSampson and Laub also argue that continuity inbehavior is not inevitable; change is possible,despite the deleterious consequences that themark of a criminal record has for future lifeevents. Crucial to understanding the mecha-nisms underlying changes in criminal behavioris the notion of turning points, defined as “con-sequential shifts that redirect a process” (Abbott1997:101).

Hurricane Katrina may have induced turningpoints in the lives of former prisoners for anumber of different reasons. First, Katrina wasa macro-level shock to the Gulf Coast region.As life-course researchers have shown, histor-ical shocks such as the Great Depression andwars have profound consequences for the lifecourse of affected individuals (Elder 1974). Asecond, but no less important, implication ofHurricane Katrina is residential change.Widespread property destruction in the GulfCoast region had an immense impact on resi-dential relocation.

In a study of the life course of crime throughage 70, Laub and Sampson (2003:149) contendthat, “offenders desist [from crime] in responseto structurally induced turning points that serveas the catalyst for sustaining long-term behav-ioral change.” Based on interviews with 70-year-old desisters, they find that residentialchange is often a fundamental turning point

leading to desistance from crime.3 Similarly, inan analysis of data from the Cambridge Studyof Delinquent Development, Osborn (1980)finds that delinquents who subsequently movedaway from London were significantly less like-ly to be reconvicted of a crime than were delin-quents who stayed in London. Laub andSampson, as well as Osborn, find that a changeof residence allows individuals to separate frompast situations and criminal peers, thus elimi-nating some of the factors contributing to thepersistence of criminal behavior.

Claude Brown’s (1965) memoir, Manchildin the Promised Land, illustrates the benefits ofseparating from criminogenic environments.During his troubled youth, Brown committedcountless crimes, used and sold a variety ofdrugs, was expelled from school, asserted hisdominance over both the Wiltwyck School forBoys and the Warwick State Training School,and was even shot during a burglary. Yet afterthis antisocial childhood, Brown was able todesist from crime while many of his peers ulti-mately went to prison, became heroin addicts,or died (or some combination thereof). Onesure reason why Brown avoided such a fate isbecause he did not take to heroin as did manyof his peers (Brown did not enjoy his briefexperimentation with the drug). Yet anothercompelling reason why he avoided a life ofcrime is change of residence. As he enteredadulthood, Brown made a conscious decision tomove away from Harlem and the peers, rou-tines, and temptations of a familiar environ-ment:

One thing began to scare me more than anythingelse about jail. This was the fact that if I went tojail and got that sheet on me, any time I decidedthat I didn’t want to go the crime way, that I want-ed to do something that was straight, I’d have a lotof trouble doing it behind being in jail. I didn’t wantthat sheet on me, and I knew if I kept hangingaround Harlem I was going to get busted. (P. 177)

By moving to Greenwich Village, Brown bothphysically and mentally separated himself frommany of the criminogenic influences in his life.

486—–AMERICAN SOCIOLOGICAL REVIEW

3 Desistance refers to the causal process that sup-ports the termination of criminal offending (Laub andSampson 2003).



There are a number of potential mechanismsby which residential change may lead to a turn-ing point in the life course of crime, includinga change in association with criminal peers anda change in one’s routine activities. Associationwith criminal peers may influence an individ-ual’s criminality, in part, through a contagionprocess whereby individuals learn the motiva-tions and techniques for crime. Association withcriminal peers may also be consequential tocrime because such associations provide accessto criminal opportunities. With respect to theformer causal mechanism, contagion modelsof crime generally posit that the likelihood ofcriminal and antisocial behavior increases withexposure to others who engage in similar behav-ior. As Warr (2002:3) convincingly demon-strates, “Criminal conduct is predominantlysocial behavior. Most offenders are imbeddedin a network of friends who also break the law,and the single strongest predictor of criminalbehavior known to criminologists is the numberof delinquent friends an individual has.”

If crime is in fact a behavior learned andfacilitated through group interaction, then itfollows that removing individuals from theircriminogenic social networks should reducetheir likelihood of engaging in criminal behav-ior. Research evidence supports such a con-tention. For instance, Warr (1993) finds thatexposure to delinquent peers declines with age,and this explains the decline in crime with age.In a later work, Warr (1998) dissects the reasonswhy marriage is such a powerful predictor ofdesistance from crime, arguing that it is becausemarriage often disrupts peer relationships. Hefinds that marriage reduces the likelihood thatan individual will associate with delinquentpeers, and it alters the amount of time individ-uals spend with delinquent peers.

In addition to a learning process, peers mayinfluence an individual’s criminal conductthrough several other mechanisms, includingthe creation of criminal opportunities (Lauband Sampson 2003; Warr 2002). Associationwith delinquent peers does not necessarily meanthat an individual is learning the motivations andtechniques for committing crime; rather, it mayjust mean that the person is exposed to moreopportunities for committing crime than wouldbe available in the absence of criminal peers. Forinstance, an individual may use a criminal asso-ciate to fence stolen goods or to purchase drugs.

Separating individuals from criminal networksvia residential change may thus reduce theircriminal behavior by eliminating opportunitiesfor engaging in crime.

Besides altering peer associations and anycorresponding criminal opportunities, residen-tial change may lead to desistance by disrupt-ing routine activities and daily temptations thatare conducive to criminal conduct. For example,the cue-reactivity paradigm in addiction researchsuggests that drug addicts are vulnerable torelapse in the presence of familiar environ-mental stimuli (Carter and Tiffany 1999; Wikler1948). Through a process of classical condi-tioning, drug addicts come to associate certainstimuli with the use of a drug. These stimuli cantrigger physiological reactions, including anintense craving for drugs. Addicts are thus morelikely to relapse in environments associatedwith prior drug use. Given that over half ofprisoners have some form of drug dependence,and that reduced consumption of drugs is onekey factor leading to desistance from crime(Mumola and Karberg 2006; National ResearchCouncil 2007), residential change may lower thelikelihood of recidivism by separating drugaddicts from familiar contexts associated withpast drug use.

The life-course literature suggests that todecrease the probability of recidivism, it wouldbe beneficial to separate parolees from theircriminal past and their peers. Yet, sorting out thecausal consequences of residential migrationand peer effects requires a research design thatexplicitly addresses the issue of selection bias.While I am unaware of any studies that express-ly account for selection bias in investigations ofthe association between residential migrationand recidivism, several studies of residentialmobility have found that selection bias can bequite consequential for estimates of neighbor-hood effects. As but one example, with datafrom the Moving to Opportunity (MTO) hous-ing mobility experiment, Ludwig and Kling(2007) examine the effects of neighborhoodconditions on the likelihood of arrest for a vio-lent crime. In their analysis, Ludwig and Klingcompare nonexperimental results estimatedthrough ordinary least squares (OLS) withexperimental results based on instrumental vari-able (IV) estimates. For male youths, they findsubstantially lower treatment effects of exposureto neighborhood violence on arrest through IV




estimates than for OLS estimates. Specifically,the size of the treatment effect declined by near-ly 40 percent after accounting for the differen-tial sorting of youth into neighborhoods.Accounting for selection is thus vital to uncov-ering unbiased estimates of the effects of mov-ing on recidivism.

In this study, I address the first part of a two-part causal story. I seek to uncover whether sep-arating individuals from their formercriminogenic environments reduces their like-lihood of recidivism. I do this through a researchdesign that expressly accounts for selectionbias. If residential migration does reduce thelikelihood of recidivism, the second part of thecausal story is to investigate why. Two potentialanswers worthy of future exploration arechanges to routine activities and peer associa-tions.

DATA AND RESEARCH DESIGN

This study exploits the benefits of a naturalexperiment in order to characterize the geo-graphic distribution of prisoner reentry inLouisiana and assess the repercussions of geo-graphic displacement from Hurricane Katrinaon the probability of re-incarceration. The ana-lytic sample is drawn from prisoners releasedfrom Louisiana correctional facilities. Becausemy interest is in residential displacement due toHurricane Katrina, I restrict analyses to thoseprisoners who resided in affected metropolitanareas prior to incarceration. Accordingly, theanalytic sample includes only ex-prisoners whowere committed to prison from Orleans Parishand the four parishes adjacent to Orleans(Jefferson, Plaquemines, St. Bernard, and St.Tammany). These areas sustained substantialhousing damage in the days following HurricaneKatrina. Therefore, the residential options ofprisoners released post-Katrina are significant-ly different than their options if they had beenreleased prior to the hurricane, resulting in somemeasure of geographic displacement. Finally, Ialso restrict the sample to exclude sex offend-ers. Given the nature of their offense, sex offend-ers face a number of constraints on theirresidency choices upon release from prison.

For the purposes of descriptive analyses, Iconstruct three cohorts of prison releasees, twoof which were released from prison prior toHurricane Katrina and one released afterward.

The first cohort is made up of releases from aLouisiana prison to parole supervision anytimebetween September 2001 and February 2002(hereafter called the 2001 to 2002 cohort).4 Thesecond cohort consists of releases betweenSeptember 2003 and February 2004 (the 2003to 2004 cohort), while the third cohort consistsof releases onto parole supervision betweenSeptember 2005 and February 2006 (the post-Katrina cohort). This study uses two pre-Katrinacohorts to more fully establish whetherHurricane Katrina altered prior trends in paroleemigration. Sample sizes equal 1,538, 1,731,and 1,370 for the 2001 to 2002, 2003 to 2004,and post-Katrina cohorts, respectively.

I use three varieties of data in this study: (1)individual-level data on parolees from theLouisiana Department of Public Safety &Corrections (DPS&C) and the Division ofProbation and Parole (DPP), (2) zip code andparish-level characteristics from the U.S.Department of Housing and UrbanDevelopment, the Louisiana Department ofLabor, and ESRI,5 and (3) Louisiana criminal


4 I use the term “parole supervision” to refer to allrelease mechanisms from incarceration that resultin a term of supervision for the returning prisoner.Releases to supervision in Louisiana are generallymade through parole board action or through thediminution of the sentence via good time credit.“Parole supervision” encompasses both types ofreleases. Roughly 90 percent of prisoners releasedeach year from Louisiana prisons are released ontoparole supervision (in contrast to unconditionalreleases, which do not require post-incarcerationsupervision). To distinguish my analytic sample ofparoled releases from unconditional releases, I use theterm “parolee.”

5 ESRI is a developer of geographic informationsystems and associated software applications, includ-ing ArcGIS. ESRI also compiles and distributes anassortment of geographically referenced data, includ-ing yearly demographic estimates of the U.S. popu-lation at county and zip code levels. I use theseestimates to produce measures of racial segregationand household income. To derive post-Katrina esti-mates, ESRI (2006a, 2006b) augmented their stan-dard methodology used to compute intercensaldemographic estimates by incorporating data from theFederal Emergency Management Agency (FEMA) onproperty damage and applications for assistance,data from the United States Postal Service NationalChange of Address file, and data from the American



justice system data from the Supreme Court ofLouisiana, DPS&C, DPP, and the UniformCrime Reports.

Given that macro-level social and economicconditions in Louisiana changed drasticallyimmediately following Hurricane Katrina, onemust control for such temporal changes to iso-late the effect of residential change on recidi-vism. The statistical models include controlsfor segregation, average household income, theunemployment rate, average weekly wages, andfair market rents. Massey and Denton (1993)argue that racial segregation has been instru-mental in creating and perpetuating “under-class” geographic areas characterized by povertyand social isolation. To the extent that underclassareas are associated with high rates of crime,whether from a lack of informal social controlor a lack of police protection, temporal changesin segregation may influence temporal changesin the likelihood of recidivism (Massey andDenton 1993; Sampson and Wilson 1995). Lowincome levels, as well as poverty, are related tocrime by a number of mechanisms. Specifically,average income is negatively related to socialdisorganization (Shaw and McKay 1942) andeconomic strain (Merton 1938), which are pos-itively predictive of crime. Unemployment mayhave contrasting effects on recidivism. Highunemployment rates may lead to an increase inrecidivism as individuals turn to crime in theabsence of legitimate employment opportunities(Cantor and Land 1985; Grogger 1998).However, while criminal motivation mayincrease with the unemployment rate, criminalopportunity may decline. Homes are more like-ly to be occupied during the day, and potentialvictims are less likely to be away from home ortheir neighborhoods, when unemployment ishigh (Cantor and Land 1985). Declining wagesand rising rents, by increasing deprivation, mayalso be related to recidivism (Blau and Blau1982; Merton 1938).

Hurricane Katrina also had many implica-tions for temporal changes in Louisiana’s crim-

inal justice system (for detailed discussions,see Garrett and Tetlow 2006; Roman, Irazola,and Osborne 2007). Given Katrina’s impact onthe criminal justice system, particularly inOrleans and adjacent parishes, it is vital toaccount for temporal variation in the operationof the justice system to draw causal inferencesabout the effect of residential change on recidi-vism. The analyses thus include control variablesrelated to parole practices, court operations,and the probability of arrest given the com-mission of a crime.

INDIVIDUAL-LEVEL VARIABLES

RE-INCARCERATION. Re-incarceration refers towhether a parolee returned to a Louisiana prisonfor a new criminal conviction or a parole vio-lation within one year of prison release. This isthe primary measure of recidivism used in thestudy. In addition to estimating the effect ofmigration on the combined measure of re-incar-ceration, I also perform separate analyses fornew criminal convictions and parole revoca-tions.

RE-INCARCERATED OR DETAINED. I rely onLouisiana data in this study, but a proportion ofthe release cohorts from Louisiana may becommitted to prison in another state or in thefederal system. However, since the three cohortsinclude only those prisoners released to parolesupervision, and parolees are required to remainin Louisiana and have periodic visits with theirparole officer, the commitment of parolees toother state correctional systems may be limit-ed. Information on non-Louisiana incarcera-tions, as well as incarcerations in a local facility(as opposed to state facilities), is captured in theDPP case management system in a field forsupervision level. A parolee’s supervision levelis marked as “detained” if that parolee is con-fined in another jurisdiction.6 In the interest of


Red Cross on housing unit damage. Estimates for2006 reflect data current as of February 2006. Dataused in this study are published in annual editions ofthe Community Sourcebook of County Demographicsand the Community Sourcebook of ZIP CodeDemographics.

6 Information on other jurisdictional detention isobtained if the DPP directly transfers a given paroleeto the other jurisdiction, or through a process oftracking absconders. In the event that a paroleeabsconds and misses a periodic visit with a paroleofficer, the officer may decide to issue a warrant forthe parolee through the FBI’s National CrimeInformation Center (NCIC) and through local lawenforcement. To track absconders, DPP parole offi-



determining whether my inferences are sensi-tive to the operationalization of recidivism, I re-estimate my statistical model with a secondmeasure of recidivism. This second, more inclu-sive, measure combines the re-incarcerationmeasure described previously with informationon other detentions and indicates whether agiven parolee was re-incarcerated in a Louisianaprison within one year or was detained in anoth-er jurisdiction.

DIFFERENT PARISH FROM CONVICTION. This isa binary treatment variable indicating whetherparolees moved to a different parish followingincarceration relative to where they were orig-inally convicted. This variable equals zero ifparolees returned to the same parish where theywere convicted, and one if they migrated to a dif-ferent parish.

DISTANCE OF MIGRATION. This is a continu-ous treatment variable indicating how farreleased parolees reside relative to where theywere originally convicted. This variable equalszero if parolees return to the same parish wherethey were convicted. Otherwise, it is computedas the Euclidean distance (in miles) between thecentroid of the parish where a parolee was orig-inally convicted and the centroid of the zip codewhere the parolee resided upon release fromprison.

POST-KATRINA RELEASE. This is a binary vari-able indicating whether the parolee was releasedfrom prison following Hurricane Katrina. Thisvariable is used as an instrument in analyses.

FIRST RELEASE. This is a binary variable indi-cating whether parolees were released fromtheir first term of incarceration (equals one) orfrom their second or greater term (zero).Parolees as a whole are a heterogeneous groupwith widely varying rates of recidivism(National Research Council 2007). Parolees

released from their first term of imprisonmenthave, on average, substantially lower rates ofrecidivism than do parolees with multiple priorincarcerations (Rosenfeld, Wallman, andFornango 2005). Controlling for prior incar-cerations is thus vital in a study of recidivism.

The study employs five additional individual-level measures as correlates of recidivism: race,gender, age at time of release, marital status, andtime served. Black parolees make up 74.2 per-cent of the sample, with Whites making up 25.7percent. Other races make up .1 percent of thesample. I use a binary indicator (Black) in analy-ses to compare Black versus White–other racecategories. Male is a binary indicator of gender.Males make up 87 percent of the sample.Married is a binary variable indicating the mar-ital status of parolees at the time of their release(not married equals zero). Ten percent of thesample was married at the time of release. Timeserved refers to the amount of time a paroleeserved in prison (in years or fraction thereof)until release.7 Controlling for time served iscrucial to account for any differences betweencohorts in the average severity of prior offend-ing. If the likelihood of prisoners being releasedfrom custody once they were eligible declinedin the post-Katrina period, relative to the pre-Katrina period, accounting for time served willcapture differences across release cohorts inprior offending severity.

CONTEXTUAL-LEVEL VARIABLES8

DISSIMILARITY. Dissimilarity (D) is a measureof the evenness of population distribution


cers also search the National Law EnforcementTelecommunications System (NLETS) and localwarrant systems. If absconders are detained in a localor non-Louisiana facility, information in the DPPdata system is updated accordingly.

7 Time served is highly associated with the offenseof conviction (e.g., prisoners in Louisiana convictedof violent offenses serve more time relative to otheroffenses). In the interest of minimizing collinearity,I use time served as a control in the analyses, but Ido not use indicators of offense of conviction.

8 Research shows that the level of aggregation(e.g., census tract, zip code, county) used in studiesof contextual effects influences inferences derivedfrom the association between a given contextual char-acteristic and a dependent variable (Hipp 2007). Inthis study, I use zip code and parish levels of aggre-gation. I do so partially because measures such asunemployment and wages are reflective of metro-politan-wide economic conditions, and also because



(Duncan and Duncan 1955). In the case here, itis a measure of segregation between Blacks andWhites, reflecting their relative distributionsacross zip codes within each parish. D can rangein value from 0, indicating complete integration,to 100, indicating complete segregation. Forthe 2001 to 2002 cohort, I use an estimate of Dfor a parolee’s parish of residence calculatedfrom the 2002 ESRI data, while for the 2003 to2004 and post-Katrina (i.e., 2005 to 2006)cohorts I use estimates of D from the 2004 and2006 data, respectively.

AVERAGE HOUSEHOLD INCOME. This is a meas-ure of the average household income (in 2000adjusted dollars) in the zip code in which aparolee resides. For the 2001 to 2002 cohort, Iuse the 2002 ESRI household income estimates,and I use 2004 and 2006 estimates for the 2003to 2004 and post-Katrina cohorts, respectively.

UNEMPLOYMENT RATE. This is a measure ofthe parish unemployment rate in parolees’parishof residence in the quarter during which theywere released from prison. Data are drawn fromthe Louisiana Department of Labor.

AVERAGE WEEKLY WAGE. This is a measure ofthe average weekly wage (in 2000 adjusted dol-lars) in the parish where parolees reside for thequarter during which they were released fromprison. Data are drawn from the LouisianaDepartment of Labor.

FAIR MARKET RENT. This is a measure of theaverage fair market rent in the parish where aparolee resides, taken from data compiled by theU.S. Department of Housing and UrbanDevelopment. For the 2001 to 2002 cohort, I usethe fair market rent average in 2002, while forthe 2003 to 2004 and post-Katrina cohorts, I userent averages from 2004 and 2006, respective-ly. All figures are adjusted to 2000 dollars.

CRIMINAL JUSTICE SYSTEM VARIABLES

AVERAGE PAROLE CONTACTS. This measure isdrawn from the DPS&C Quarterly Statistical

Performance Report. Greater scrutiny influ-ences the likelihood that a parolee will getcaught for violating conditions of parole (Turnerand Petersilia 1992), yet increased scrutiny mayalso have a deterrent effect on parolees thatlowers the likelihood of recidivism. It is thusimportant to control for temporal variations inaverage parole contacts (i.e., the total contactsparole officers have across their entire case-loads), as well as between parole district vari-ation, to account for these contrasting influencesof scrutiny on recidivism. For analyses, I use theaverage number of contacts made across paroleofficers in a parole district during the quarter inwhich parolees were released from prison.Contacts include those that occurred within theparole office, as well as field contacts to aparolee’s residence or workplace.

JUDGE CASELOADS. This measure derivesfrom the Supreme Court of Louisiana’s annualreport. There is a vast body of research exam-ining the relation between judge caseloads andcriminal case dispositions (e.g., Heumann1975). One key issue in this regard is whetherlarge caseloads, and the need to process largenumbers of defendants, pressure courts towardthe use of plea bargaining and result in lenientsentencing practices (Eisenstein, Flemming,and Nardulli 1988). Recent research (see, e.g.,Johnson 2005; Ulmer and Johnson 2004) revealsthat judge caseloads are inversely related to thelikelihood of incarceration and the severity ofcriminal sentences. I thus include a control forjudge caseloads given that such caseloads like-ly influence whether a convicted offender issentenced to a term of imprisonment or someother sanction, such as probation. For the 2001to 2002 cohort, I use the number of cases perjudge in 2002 in a parolee’s parish of residence;for the 2003 to 2004 and post-Katrina cohorts,I use caseload figures from 2004 and 2006,respectively.

UCR ARRESTS PER CRIME. I use this measure,computed from the FBI’s Uniform CrimeReports, to control for the temporal and geo-graphic variation in the likelihood of gettingarrested for a Part I offense (i.e., murder, rape,robbery, aggravated assault, burglary, larceny,motor vehicle theft, or arson) given the com-mission of such a crime. For the 2001 to 2002


monthly, quarterly, and yearly time series data are notavailable at a lower level of aggregation.



cohort, I construct this measure using the ratioof Part I arrests per reported Part I crimes in2001 for a parolee’s parish of residence; for the2003 to 2004 and post-Katrina cohorts, I use the2003 and 2005 ratios, respectively.

ANALYTIC STRATEGY

Analyses follow two paths to coincide with theobjectives of the study. First, I expect that pris-oners released soon after Katrina faced severe-ly constrained opportunities to reside in NewOrleans, resulting in far fewer returns to theNew Orleans metropolitan area than in the pre-Katrina period and therefore fewer returns toformer parishes. This expectation leads toHypothesis 1:

Hypothesis 1: Because of hurricane-relateddestruction, significantly fewer paroleesmoved to New Orleans upon release fromprison during the time period immediate-ly following Hurricane Katrina, relative tothe pre-Katrina period, and significantlyfewer returned to the parish they inhabitedprior to incarceration.

To examine this hypothesis, I geocode andmap the post-release addresses for the threeseparate cohorts of parolees, and I provide illus-trations of how the geographic distribution ofprisoner reentry changed post-Katrina.

With regard to my second research question,concerning the impact of residential change onrecidivism, I expect that separating paroleesfrom their former residential environments willbe beneficial with respect to desisting fromcrime. Induced migration due to HurricaneKatrina allows for a separation between paroleesand their criminal past, thus reducing the like-lihood of re-incarceration. These propositionsyield Hypothesis 2:

Hypothesis 2: The likelihood of re-incarcerationis lower when parolees reside in a geo-graphic area different from where theyresided prior to incarceration.

To test Hypothesis 2, I compare the likelihoodof re-incarceration within one year of releasefrom prison for parolees who resided in thesame parish upon release as where they wereoriginally convicted versus parolees who movedto a different parish. Before describing the spe-

cific statistical analyses that will be performed,it is first relevant to highlight the benefits ofusing a natural experiment to examine the influ-ence of residential migration on recidivism.There are numerous potential threats to internalvalidity, but in studies of residential migration,the threat of selection bias is of particular con-cern, where selection refers to the process ofassigning individuals to conditions (i.e., treat-ment versus control groups). One approach toremedy the threat of selection bias is simply tomeasure more characteristics of individuals andmore fully specify statistical models with thosemeasures; in essence, the strategy is to subsetthe data such that two or more individuals areidentical except with respect to the treatmentcondition. The assumption with such anapproach is that selection into treatment andcontrol groups is ignorable (i.e., assignment tocontrol and treatment groups is random) afterconditioning on the observed characteristics ofparolees. The threat of omitted variables is stillprobable, however, even with extensive use ofstatistical controls.

An alternative approach to alleviating thethreat of selection bias is the use of instrumen-tal variables. With an IV approach, a variable (orvariables) that is unrelated to the outcome vari-able is used as a predictor (i.e., instrument) ofthe key explanatory variable (i.e., the treat-ment), and the outcome variable is thenregressed on the predicted treatment measure.Conceptually, this approach removes the spuri-ous correlation between the explanatory variableand unobserved characteristics, in this caseunobservable characteristics of parolees. An IVremedies the issue of omitted variables by usingonly that portion of the variability in the treat-ment variable that is uncorrelated with omittedvariables to estimate the causal relation betweenthe treatment and outcome. The key criticism ofthis approach is that the assumption about thelack of relation between the instrument and theoutcome variable may be problematic. However,the use of an instrument derived from a naturalexperiment obviates this issue. We can havemore confidence that the instrument and out-come variable are unrelated if the instrumentderives from a random force of nature like a hur-ricane (for a discussion, see Angrist and Krueger2001). This assumption is known as the exclu-sion restriction—that is, cov(Zi,ui) = 0, where Zi




is the instrument and ui is the model error thatincludes the omitted variables—but it is notdirectly verifiable (Angrist, Imbens, and Rubin1996). The validity of inferences from an IVanalysis depends on the appropriateness of thisassumption. I address potential violations ofthe exclusion restriction in the Discussion andAppendix.

Using Hurricane Katrina as an exogenoussource of variation that influences where aparolee resides, I combine a natural experi-ment with an instrumental variable approach toprovide a consistent estimate of the effect ofmoving to a different parish (i.e., the treat-ment) on re-incarceration (Angrist et al. 1996).9

Angrist and Krueger (2001:77) note that,“instrumental variables provide an estimate fora specific group—namely, people whose behav-ior can be manipulated by the instrument.” Inthe present context, Hurricane Katrina affect-ed the residential behavior of those paroleeswho could not or would not have moved awayfrom the New Orleans metropolitan area fol-lowing incarceration, but it did not affectparolees who would have moved even in theabsence of the hurricane.10 In short, using anIV approach allows me to compute an estimateof the effect of migrating to a different parishfor parolees who otherwise would have movedback to the same parish, had Hurricane Katrinanot occurred (this is the Local AverageTreatment Effect, or LATE). I do not providean estimate of the effect of moving to a differ-ent parish on re-incarceration for parolees whowould have moved regardless of the hurricane.

I specify in Equations 1 and 2 the two-stageestimation process with the IV technique. Thefirst equation models the key explanatory vari-able Si (different parish from conviction) as afunction of an instrumental variable Zi (post-Katrina release) and a vector of control vari-

ables Xi used to account for any observed dif-ferences between the treatment and controlgroups. I assume that where parolees residedepends, in part, on whether they were releasedfrom prison before or after Hurricane Katrina:11

Si = Zi�1 + Xi� + �i (1)

The second-stage of the two-stage estima-tion process models the dependent variable Yi(re-incarceration) as a function of the predict-ed Si from Equation 1 and a vector of controlvariables Xi:12

Yi* = � S^i + Xi� + ui (2)

where

Yi = 0 if Yi* < 0, Yi = 1 if Yi* ≥ 0,and ui ~ N(0,1).

The coefficient � is the key parameter ofinterest and will be used to determine whethermigrating to a new parish upon release influ-ences a parolee’s likelihood of recidivism.


09 Consistency means that the IV estimate con-verges to the population parameter as the samplesize increases.

10 As an example, Angrist (1990) uses Vietnam-eradraft lotteries as an instrument to assess the effect ofmilitary service on future earnings. Some individu-als joined the military voluntarily, and some joinedbecause of the draft. Angrist’s IV estimates of theeffects of military service on future earnings areapplicable for draftees but not volunteers.

11 I estimate Equations 1 and 2 in Stata using theivprobit function. Via the cluster estimation com-mand with the ivprobit function, I adjust model stan-dard errors to account for the clustering (i.e.,interdependence) of parolees within parishes. Notethat the ivprobit function treats the endogenous treat-ment variable Si in Equation 1 as continuous, and alinear regression model is used to predict treatmentassignment. As Angrist and Krueger (2001:80) reveal,using a linear regression model to produce the first-stage estimates generates consistent second-stageestimates even when the endogenous treatment vari-able is binary (as is the case here). Stata do-files forall models estimated in analyses are available fromthe author on request.

12 With a true randomized experiment, it isassumed that treatment and control groups are bal-anced (i.e., equivalent) with respect to observable andunobservable characteristics, aside from random vari-ation. Even though I use an exogenous source ofvariation—a hurricane—to predict treatment assign-ment, in the absence of random assignment it is notnecessarily the case that treatment and control groupsare identical. It is thus beneficial to include statisti-cal controls (Xi) for key parolee characteristics, aswell as for contextual and criminal justice covariates,to account for any observed differences across groups.The benefit of statistical controls is a gain in statis-tical efficiency (i.e., a reduction in sampling vari-ance).



RESULTS

HYPOTHESIS 1

Figure 1 provides a snapshot of the geograph-ic redistribution of parolees post-Katrina. Thisfigure shows which parish parolees resided inimmediately after their exit from prison. I usedata from two pre-Katrina cohorts to establishthe extent to which Hurricane Katrina prompt-ed a shift in the geographic distribution ofparolees. The release figures from 2001 to 2002and 2003 to 2004 reveal that roughly 50 percentof prisoners convicted in the five-parish area(i.e., Orleans, Jefferson, Plaquemines, St.Bernard, and St. Tammany) subsequentlyreturned to New Orleans. Post-Katrina, thisnumber drops to 20 percent. In the post-Katrinaperiod, proportionally more parolees opted toreside in Baton Rouge (2 percent before Katrinaversus 11 percent after), and many otherparolees dispersed throughout the state. Thisfigure demonstrates that the proportion of newlyreleased prisoners (who were originally com-mitted from the New Orleans metropolitan area)who returned to the five-parish area upon releasedeclined drastically immediately followingKatrina. This finding provides support forHypothesis 1. Moreover, recall that one bene-fit of using an instrument derived from a natu-

ral experiment is to assure that the instrumentis correlated with the treatment condition (i.e.,migration) but otherwise unrelated to the out-come variable (i.e., re-incarceration). Figure 1suggests that the time period of release (pre- ver-sus post-Katrina) is highly correlated with thetreatment condition (see the Appendix for adetailed assessment of the assumptions of theIV framework).

Table 1 contrasts the place of residence acrossthe three cohorts through a cross-tabulation ofthe proportion of members of each cohort whomoved to the same parish following incarcera-tion as where they were originally convicted, rel-ative to the proportion who migrated to adifferent parish. Results reveal that prior toHurricane Katrina, roughly three-quarters ofparolees returned to their parish of convictionupon release from prison. Post-Katrina, this dis-tribution changed significantly (Chi-Square =286.65, p < .001), with 49.9 percent of paroleesreturning to the same parish and 50.1 percentmigrating to a different parish. This findingprovides additional support for Hypothesis 1.

HYPOTHESIS 2

Results thus far reveal that there has been ageographic redistribution of prisoner reentry


Figure 1. Parish of Release for Prisoners Committed from the Five-Parish Area



post-Katrina. Whether the residential migrationunderlying this redistribution is causally relat-ed to the likelihood of recidivism is an empir-ical question, which I will address through anIV analysis. Before moving onto the IV results,it is necessary to address the topic of treatmentnoncompliance, so as to clarify precisely whichtype of treatment effect is estimated in thisstudy. Recall that in the present case the treat-ment Si is represented by a given parolee resid-ing in a different parish upon release relativeto where he was convicted prior to incarcera-tion. Perfect treatment compliance would rep-resent the situation where all parolees releasedpost-Katrina (Z = 1) moved to a different parish(S = 1), and all parolees released pre-Katrina(Z = 0) moved to the same parish as where theywere originally convicted (S = 0). As revealedin Table 1, this situation certainly does nothold; 49.9 percent of parolees released post-Katrina (Z = 1) moved to the same parish (S =0), and 24.7 percent of parolees from the twopre-Katrina cohorts (Z = 0) moved to a differ-ent parish (S = 1). While we can assume thatthe assignment to treatment is ignorable (i.e.,conditional on the instrument Z and the othercovariates from Equation 1, assignment to con-trol and treatment groups is random), a con-sequence of noncompliance is that the receiptof treatment is nonignorable (Angrist et al.1996). If this is the case, simply computing thedifference between the pre- and post-Katrinacohorts on recidivism will not provide an unbi-ased or consistent estimate of the averagecausal effect of migrating to a different parishon recidivism. Yet, by using IV methods, I cancompute a consistent estimate of the effect ofmigrating to a different parish on re-incarcer-ation for parolees who would not have movedhad it not been for Hurricane Katrina (i.e., an

estimate of LATE).13 Results in the ensuinganalyses represent the LATE estimate.

Table 2 presents the IV probit results of re-incarceration (a detailed assessment of theassumptions of the IV framework is presentedin the Appendix). Focusing on the first columnof results (Re-incarceration), males are morelikely than females to be re-incarcerated, andmarried parolees are less likely to be re-incar-cerated. Age and time served in prison are neg-atively related to re-incarceration. As expected,first releases from prison are significantly andsubstantially less likely to recidivate than arerepeat offenders (this relation will be exploredin greater detail in Table 3). With respect to thecontextual-level and criminal justice systemcovariates, only fair market rent is significant-ly associated with re-incarceration. In this case,higher rents correspond to a lower likelihood ofre-incarceration, which may indicate that areaswith higher rent have relatively more informalsocial control and surveillance, and thereforeless crime.

Turning to the main finding of the study,results show that individuals who migrated to


Table 1. Post-incarceration Place of Residence

Post-incarceration Residence

Cohort Same Parish Different Parish

2001 to 2002 (N = 1,538) 75.5% 24.5%2003 to 2004 (N = 1,731) 75.2% 24.8%Post-Katrina (N = 1,370) 49.9% 50.1%

Note: “Same” means the proportion of parolees in a given cohort who moved to the same parish following incar-ceration as where they were originally convicted. “Different” refers to the converse. Chi-Square = 286.65, p <.001.

13 Through the IV method, I adjust the recidivismdifference between the pre- and post-Katrina cohortsfor the fact that not all post-Katrina parolees movedto a different parish and not all pre-Katrina paroleesreturned to the same parish. This adjustment yieldsthe Local Average Treatment Effect, which is com-puted as follows (Angrist et al. 1996):

{E[Yi | Zi =1] – {E[Yi | Zi = 0]} / {E[Si | Zi =1] – {E[Si | Zi = 0]} = � ,

where Yi is the outcome measure of re-incarcera-tion, Zi is the instrument (post-Katrina release), andSi is the key explanatory variable (different parishfrom conviction).



a different parish were significantly and sub-stantially less likely to be re-incarcerated with-in one year of release from prison. With respectto the marginal effect, the probability of re-incarceration is .15 lower for parolees who didnot move back to the parish where they wereoriginally convicted, relative to parolees who didreturn (net of individual, contextual, and crim-inal justice system correlates). The 95 percentconfidence interval of the marginal effect rangesfrom –.082 to –.221. The predicted probabilityof re-incarceration for male parolees whoreturned to the same parish as where they wereconvicted is .265. By contrast, the predictedprobability for males released to a differentparish is .110. As noted, these inferences pertainto the local average treatment effect for com-pliers—that is, the effect of migrating to a dif-ferent parish for parolees who otherwise wouldhave moved back to the same parish hadHurricane Katrina not occurred.

Re-incarceration typically results from one oftwo paths: a new criminal conviction or a parolerevocation due to violations of the conditions of

parole (e.g., a failed drug test). To determine ifthe significant treatment effect of migrationidentified in the first column of results is robustto these two different forms of recidivism, I re-estimated the model using parole revocationand new conviction as dependent variables. Thesecond and third columns in Table 2 reveal thatindividuals who migrated to a different parishwere significantly less likely to be re-incarcer-ated, whether through parole revocation or anew criminal conviction.

Given that offenders with multiple prior incar-cerations have drastically higher rates of recidi-vism than do first releases (Rosenfeld et al.2005), it may be the case that repeat offendersare relatively immune to the apparent benefit ofresidential migration. The risk factors that pro-pel repeat offenders toward crime, and the dis-advantages associated with the mark of acriminal record, may render the influence of achange of residence essentially meaningless.Table 3 explores this possibility by estimatingEquations 1 and 2 separately by group.


Table 2. Instrumental Variable Probit Estimates of Re-incarceration

Re-incarceration Parole Revocation New Conviction

Coef. Robust SE Coef. Robust SE Coef. Robust SE

Different Parish from Conviction –.597 (.156)*** –.586 (.151)*** –.481 (.226)*Individual-Level—Black –.004 (.084) –.009 (.090) .024 (.066)—Male .163 (.054)** .132 (.063)* .316 (.109)**—Married –.205 (.056)*** –.236 (.051)*** –.019 (.098)—Age at release –.008 (.002)*** –.008 (.002)*** –.009 (.003)**—Time served –.034 (.013)** –.031 (.015)* –.045 (.007)***—First release –.304 (.076)*** –.320 (.085)*** –.127 (.046)**Context and Criminal Justice System—Unemployment rate .067 (.072) .037 (.085) .198 (.129)—Average weekly wage –.003 (.005) –.004 (.005) .004 (.006)—Average household income .003 (.002) .003 (.002) .007 (.005)—Dissimilarity .021 (.040) .021 (.044) –.025 (.073)—Fair market rent –.008 (.004)* –.007 (.004)* –.011 (.006)*—Average parole contacts –.005 (.012) –.009 (.011) .017 (.022)—Judge caseloads .000 (.001) .000 (.001) –.004 (.002)—UCR arrests per crime (parish) –.244 (.245) –.407 (.508) .017 (.108)Intercept .382 (.402) .436 (.469) –1.260 (.663)*N 4,639 4,514 3,817

Note: Re-incarceration includes both parole revocations and new convictions. The instrument Zi is a binary indi-cator of the release period (pre- versus post-hurricane). The coefficient and standard error for average householdincome are multiplied by 1,000. Coefficients and standard errors for all other context and criminal justice systemmeasures, except UCR arrests per crime, are multiplied by 10. Significance tests are calculated from robuststandard errors. The analytic sample for the Parole Revocation model excludes new conviction recidivists. Thesample for the New Conviction model excludes parole revocation recidivists.* p ≤ .05; ** p ≤.01; *** p ≤ .001 (one-tailed test).



Results reveal that all parolees, whether firstreleases or those with multiple prior incarcera-tions, are significantly less likely to be re-incar-cerated if they migrate to a different parish.While the likelihood of re-incarceration with-in one year differs for these two groups, with aprobability of .174 for first releases and a prob-ability of .270 for repeat offenders, the marginaleffect of migration is equivalent for the groups(.15).

SENSITIVITY ANALYSES

To assess whether findings are robust to dif-ferent operationalizations of the recidivismmeasure, the treatment variable, the sample,and the instrumental variables, I perform a seriesof sensitivity analyses, with results presented inTable 4.14 First, I re-estimate Equations 1 and2 with an alternative measure of recidivism,

which combines indicators of re-incarcerationand detainers. Recall that detainers refer toincarcerations in another jurisdiction (i.e., fed-eral, another state, or local). Results in the firstcolumn of Table 4 reveal that there is some vari-ation in the correlates of this second measure ofrecidivism relative to the re-incarceration modelin Table 2, but I still find that parolees whomigrate to a different parish relative to wherethey were originally convicted are significant-ly less likely to recidivate (re-incarcerated ordetained) within one year of prison release. Theprobability of re-incarceration or detention is .08lower for parolees who did not move back to theparish where they were originally convicted,relative to parolees who did return. The declinein the treatment effect of migration relative tothe .15 effect from the re-incarceration modelin Table 2 is likely due to the detained measurecontaining relatively more minor forms of deten-tion (e.g., short-term stays in jail following anarrest). Residential migration thus appears to bemore consequential for desistance from seri-ous forms of criminal offending. ClaudeBrown’s (1965) memoir is instructive in thisregard. After moving to Greenwich Village,


Table 3. IV Probit Estimates of Re-incarceration, First Release versus Repeat Offenders

First Release Repeat Offenders

Coef. Robust SE Coef. Robust SE

Different Parish from Conviction –.684 (.195)*** –.499 (.202)**Individual-level—Black .044 (.084) –.141 (.110)—Male .208 (.069)** .075 (.072)—Married –.191 (.063)** –.230 (.093)**—Age at release –.008 (.004)* –.009 (.003)**—Time served –.035 (.014)** –.042 (.047)Context and Criminal Justice System—Unemployment rate –.013 (.088) .204 (.185)—Average weekly wage .001 (.005) –.007 (.004)*—Average household income .002 (.002) .006 (.003)*—Dissimilarity –.003 (.045) .056 (.053)—Fair market rent –.009 (.005)* –.007 (.008)—Average parole contacts –.008 (.007) .000 (.025)—Judge caseloads .000 (.001) .000 (.001)—UCR arrests per crime (parish) –.482 (.636) –.136 (.108)Intercept .134 (.376) .410 (.693)N 3,275 1,364

Note: The instrument Zi is a binary indicator of the release period (pre- versus post-hurricane). The coefficientand standard error for average household income are multiplied by 1,000. Coefficients and standard errors for allother context and criminal justice system measures, except UCR arrests per crime, are multiplied by 10.Significance tests are calculated from robust standard errors.* p ≤ .05; ** p ≤ .01; *** p ≤ .001 (one-tailed test).

14 For the four different sensitivity analyses, onlyone model specification is varied at a time (e.g., I donot test alternate measures of recidivism and treat-ment at the same time).




Tabl

e 4.

Sen

siti

vity

An

alys

es: I

nst

rum

enta

l V

aria

ble

Pro

bit

Est

imat

es o

f R

e-in

carc

erat

ion

(1)

Alt

ern

ate

Ou

tcom

e(2

) A

lter

nat

e Tr

eatm

ent

(3)

Alt

ern

ate

Sam

ple

(4)

Alt

ern

ate

IVR

e-in

carc

erat

ion

/Det

ain

edR

e-in

carc

erat

ion

(D

ista

nce

)R

e-in

carc

erat

ion

(O

rlea

ns)

Re-

inca

rcer

atio

n (

5 IV

s)

Coe

f.R

obu

st S

EC

oef.

Rob

ust

SE

Coe

f.R

obu

st S

EC

oef.

Rob

ust

SE

Tre

atm

ent

—D

iffe

rent

par

ish

from

con

vict

ion

–.21

7(.

073)

**–.

388

(.20

7)*

–.53

4(.

188)

**—

Dis

tanc

e of

mig

rati

on (

mil

es)

–.00

4(.

001)

***

Indi

vidu

al-L

evel

—B

lack

.047

(.08

4).0

21(.

058)

–.35

8(.

092)

***

–.00

1(.

088)

—M

ale

.240

(.04

0)**

*.1

72(.

050)

***

.203

(.07

4)**

.165

(.05

8)**

—M

arri

ed–.

112

(.04

9)*

–.24

2(.

065)

***

–.34

4(.

128)

**–.

209

(.05

2)**

*—

Age

at r

elea

se–.

013

(.00

2)**

*–.

008

(.00

3)**

*–.

006

(.00

3)*

–.00

8(.

002)

***

—T

ime

serv

ed–.

007

(.00

6)–.

034

(.01

3)**

–.04

3(.

012)

***

–.03

5(.

012)

**—

Firs

t rel

ease

–.22

5(.

047)

***

–.32

0(.

071)

***

–.21

0(.

053)

***

–.30

3(.

073)

***

Con

text

and

Cri

min

al J

usti

ce S

yste

m—

Une

mpl

oym

ent r

ate

–.16

7(.

059)

**.0

01(.

050)

.220

(.17

9).0

41(.

083)

—A

vera

ge w

eekl

y w

age

.005

(.00

5)–.

007

(.00

4)–.

017

(.00

8)*

–.00

3(.

005)

—A

vera

ge h

ouse

hold

inco

me

–.00

1(.

002)

.002

(.00

2).0

11(.

006)

*.0

03(.

002)

*—

Dis

sim

ilar

ity

.003

(.02

9).0

28(.

030)

.062

(.08

3).0

18(.

039)

—Fa

ir m

arke

t ren

t–.

001

(.00

2)–.

006

(.00

3)*

–.00

1(.

006)

–.00

7(.

004)

—A

vera

ge p

arol

e co

ntac

ts.0

11(.

013)

–.00

1(.

011)

–.00

2(.

023)

–.00

4(.

010)

—Ju

dge

case

load

s.0

00(.

001)

.000

(.00

1)–.

004

(.00

2)**

.000

(.00

1)—

UC

R a

rres

ts p

er c

rim

e (p

aris

h)–.

169

(.10

4)–.

260

(.31

0).0

14(.

117)

–.25

4(.

263)

Inte

rcep

t–.

122

(.46

5).3

57(.

392)

.654

(.65

8).3

15(.

314)

N4,

639

4,63

92,

207

4,63

9

Not

e: I

n M

odel

4, “

5 IV

s” d

enot

es th

at th

e m

odel

was

est

imat

ed w

ith

five

inst

rum

ents

mea

suri

ng th

e in

tera

ctio

n be

twee

n ti

me

peri

od o

f re

leas

e (p

re-

vers

us p

ost-

Kat

rina

) an

dpa

rish

of

conv

icti

on (

Orl

eans

, Jef

fers

on, P

laqu

emin

es, S

t. B

erna

rd, o

r S

t. Ta

mm

any)

. The

coe

ffic

ient

and

sta

ndar

d er

ror

for

aver

age

hous

ehol

d in

com

e ar

e m

ulti

plie

d by

1,0

00.

Coe

ffic

ient

s an

d st

anda

rd e

rror

s fo

r al

l oth

er c

onte

xt a

nd c

rim

inal

just

ice

syst

em m

easu

res,

exc

ept U

CR

arr

ests

per

cri

me,

are

mul

tipl

ied

by 1

0. S

igni

fica

nce

test

s ar

e ca

lcul

ated

from

rob

ust s

tand

ard

erro

rs. F

or M

odel

3, t

he s

ampl

e is

res

tric

ted

to p

arol

ees

orig

inal

ly c

onvi

cted

in O

rlea

ns P

aris

h.*

p ≤

.05;

**

p≤

.01;

***

p≤

.001

(on

e-ta

iled

test

).



Brown avoided the serious forms of criminali-ty that had led to his detention in juvenile reforminstitutions. However, he did not abstain fromcriminal behavior altogether (e.g., he frequent-ly used marijuana and occasionally got intofights). Residential migration may thus lead toa de-escalation to less serious forms of offend-ing, rather than a total cessation from criminalbehavior.

To assess the sensitivity of results to the typeof treatment variable used, I re-estimatedEquations 1 and 2 with an alternative treatmentmeasure. For analyses performed up to thispoint, I used a binary treatment measure indi-cating whether parolees moved to a differentparish relative to where they were originallyconvicted. As an alternate treatment, I use ameasure of distance indicating how far paroleesmoved relative to the parish where they wereoriginally convicted (i.e., distance of migra-tion). In this sense, the measure of distance isanalogous to a treatment dosage, and the modelreveals whether the level of dosage affects thelikelihood of re-incarceration. Results in thesecond set of columns in Table 4 once againdemonstrate the significant effect of migrationon re-incarceration. In this case, the marginaltreatment effect of migration is equivalent to a.01 reduction in the likelihood of re-incarcera-tion for every 10 miles parolees migrate awayfrom their prior parish.

Recall that the sample used in this study ismade up of prisoners originally convicted inmetropolitan areas affected by HurricaneKatrina (i.e., Orleans, Jefferson, Plaquemines,St. Bernard, and St. Tammany parishes). As athird sensitivity analysis, I restrict the sample toindividuals who were convicted in OrleansParish and re-estimate Equations 1 and 2 withthis restricted sample. Given that Orleans Parishtypically has a higher crime rate than the otherparishes (Federal Bureau of Investigation 2007),residential migration away from criminogenicareas may be particularly advantageous for indi-viduals convicted in Orleans Parish, while lessadvantageous for individuals convicted in lowercrime areas. The third model in Table 4 revealsthat there is a significant negative effect ofmigration on re-incarceration for parolees whowere originally convicted in Orleans Parish,although the treatment effect is smaller rela-tive to the full sample (the marginal effect oftreatment is .15 for the full sample and .11 for

the Orleans Parish sample). For comparisonpurposes, I also estimated the model excludingthe Orleans convictions but including individ-uals convicted in the four other parishes, and Ifound a marginal treatment effect of .19 (resultsavailable from the author on request). In sum,I find a negative effect of migration regardlessof the pre-incarceration location of the parolee,although the size of the treatment effect variesby parish. One potential reason for this diver-gence may be related to distance. If the likeli-hood of recidivism declines with distance fromthe original parish, then the lower marginaleffect of treatment for parolees convicted inOrleans Parish may be due to the fact thatmovers did not migrate very far when they leftthe parish (at least relative to movers from otherparishes). Such a scenario is a worthy topic forfuture investigation.

As a final sensitivity analysis, I re-estimatedEquation 1 using an alternative specification ofinstrumental variables. As described earlier,property damage from Hurricane Katrina var-ied by parish; St. Bernard and Plaqueminesparishes suffered the most damage, whileJefferson Parish had the least amount of dam-age. As depicted in Figure 1, the likelihood ofa parolee moving to a given parish post-incar-ceration was influenced by the extent of hous-ing destruction in the parish. Given that both thetime period of release from prison (pre- versuspost-Katrina) and the geographic variation inproperty damage affect whether parolees willreturn to the same parish where they were orig-inally convicted, I use an interaction betweentime period of release and parish of convictionas an alternative specification of instrumentalvariables.15 With this specification, I assume that


15 In essence, I have five instruments. The firstinstrument (post-Katrina release and Orleans con-viction) equals 1 for all parolees who were releasedfollowing Katrina and who were originally convict-ed in Orleans Parish; it equals 0 otherwise. I constructsimilar instruments for parolees originally convict-ed in Jefferson, Plaquemines, St. Bernard, and St.Tammany parishes. I entered all five instrumentsinto Equation 1 simultaneously. One key benefit ofestimating a model with more instruments thanendogenous regressors (i.e., treatments) is that I cantest the IV assumption that the dependent variable andinstruments are unrelated (see the Appendix fordetails).



whether parolees migrate to a different parishor not depends on whether they were releasedfrom prison before or after Hurricane Katrinaand where they were originally convicted.

Findings presented in the fourth model inTable 4 are consistent with inferences derivedin Table 2. Parolees who migrated to a differ-ent parish were significantly less likely to be re-incarcerated, such that the marginal effect ofre-incarceration is .14 lower for parolees whodid not move back to the parish where theywere originally convicted, relative to paroleeswho did return.

In summary, findings suggest that movingaway from former places of residence servessome benefit for parolees. A change of resi-dence allows individuals to separate from theirpeers and the temptations that contributed totheir criminality in the first place. This findingholds for both first releases and repeat offend-ers, and it holds under alternative specifica-tions of the dependent variable, treatmentvariable, sample, and instruments.

DISCUSSION

Selection bias is very much an issue withresearch on both residential migration andneighborhood effects. The innovation of thisstudy has been to use a natural experiment toinvestigate a theoretical question whose answerhas largely eluded researchers because of selec-tion—just how consequential is a change ofresidence to behavioral outcomes such as crime?In the absence of perfect information on why agiven individual moves to one place of resi-dence versus another, it is difficult to answer thisquestion because omitted variables may intro-duce bias into estimates of the effects of migra-tion. While tragic, the residential destruction andmigration resulting from Hurricane Katrinapresent a unique opportunity for understandinghow place and migration affect outcomes suchas crime and recidivism.

This study addresses whether separating indi-viduals from their former residential environ-ments reduces their likelihood of recidivism. Myfindings support this assertion. In particular,the marginal effect of migrating to a differentparish upon release from prison is a .15 declinein the probability of re-incarceration. This find-ing holds for both first releases from incarcer-

ation and parolees with multiple prior incar-cerations.

We must consider whether such findings areinternally valid. Recall my discussion of theexclusion restriction, which states that the instru-ment and the outcome variable are unrelatedexcept indirectly through the effect of the instru-ment on the causal treatment. The validity of theIV analysis rests on the assumption that anyeffect of the time period of release (pre- versuspost-Katrina) on the likelihood of re-incarcer-ation must be captured through the effect oftime period on residential migration. Tests pre-sented in the Appendix provide support for theassumption that the instruments and the re-incarceration measure are not significantly cor-related. While I cannot completely rule out thepossibility of violating the exclusion restric-tion, my findings suggest that even with such aviolation, the potential bias would not whollyeliminate the sizable treatment effect observedin this study.

The next step in this causal story is to furtherinvestigate why there is such a powerful treat-ment effect from changing place of residence.For instance, the causal mechanism underlyingthe effect of migration may be related to socialties. Parolees who migrate to a different parishmay be more likely to sever ties with criminalpeers than would parolees who return to thesame parish. They would therefore have lessopportunity and provocation for engaging incrime. While I reason that changes to peer asso-ciations and routine activities explain whymigration is consequential to recidivism, Iattempted to account for several alternativeexplanations for the reduced likelihood of recidi-vism among parolees who migrated. Forinstance, judges may be less likely to convict orsentence an offender to prison and parole offi-cers may be less likely to revoke the parole ofparole violators in destination parishes, rela-tive to the original parishes. By controlling forcourt, parole, and police practices in my statis-tical models, I attempted to determine if migra-tion has a causal effect on recidivism afterconsideration of the variation in criminal justicepractices across space and time.

Similarly, to account for the possibility thatparolees who migrated were more likely toabscond and leave Louisiana, I examined theeffect of migration not only on re-incarcera-tions in Louisiana (i.e., Table 2), but also on




detentions in local, federal, or other state facil-ities (i.e., the first column in Table 4). Still, thepotential for unmeasured variation in the oper-ation of the criminal justice system across spaceand time presents a possible alternative expla-nation for the effect of migration on recidivism(in contrast to the hypothesized mechanisms ofchanges in peer association and routine activi-ties). Exploration of mechanisms underlyingthe causal effect of migration is an importantavenue for future research.

Yet, even before the full causal story bearsout, the findings established thus far with respectto residential migration and recidivism havesignificant policy implications. If separatingex-prisoners from their former residential envi-ronments actually benefits those prisoners with-out sacrificing public safety, then a logical nextstep is to consider how to disperse the popula-tion of ex-prisoners on a large scale. Researchby the Urban Institute suggests that substantialproportions of returning prisoners would wel-come the opportunity to move away from theirformer criminogenic environments (Visher andFarrell 2005). For instance, in a sample ofrecently released males from the IllinoisDepartment of Corrections, Visher and Farrell(2005) find that 45 percent of returning pris-oners explicitly expressed a desire to move toa different neighborhood than where they livedprior to prison. Over half of this group wishedto live in a different neighborhood to avoiddrugs and the other temptations and troubles offamiliar settings.

Yet many ex-prisoners still end up movingback to their former counties and neighbor-hoods despite an expressed interest to avoidsuch places. One prime reason is because oflegal barriers. In most states, prisoners releasedto some form of parole supervision are legallyrequired to return to their county of last resi-dence (National Research Council 2007).Criminal justice policies in most states aredesigned to return prisoners to the same famil-iar surroundings where they got into troublewith the law in the first place. In Louisiana, thereare no such geographic restrictions. The find-ings from Louisiana presented in this study sug-gest that allowing ex-prisoners to move todifferent parishes or counties will reduce recidi-vism. It may thus be fruitful for states to recon-sider the residency restrictions imposed onreturning prisoners.

As noted, however, the sizable treatmenteffect of moving found in this study is onlyapplicable to parolees who moved to a differentparish post-Katrina who otherwise would havemoved back to where they lived before. To fullyexplore the policy implications of this finding,we must understand why prisoners return totheir former parishes and neighborhoods evenin a state like Louisiana where there are nolegal barriers preventing parolees from movingto a different parish. We know from the UrbanInstitute research that another key reason whyreleased prisoners move back to their formerneighborhoods is to be close to family (Visherand Farrell 2005). To capitalize on the apparentbenefits of residential migration, it may thus benecessary to both remove legal barriers pre-venting ex-prisoners from moving and provideex-prisoners and their families with opportuni-ties and incentives to move away from old neigh-borhoods. While forcing ex-prisoners to moveaway from their old neighborhoods is neitherrealistic nor ethical, providing opportunities forex-prisoners to move away from old neighbor-hoods is a policy prescription that may be worthpursuing.

Before enacting policy initiatives to promoteresidential change among the parole popula-tion, more research must be done to both vali-date the findings demonstrated in this studyand uncover the mechanism underlying the treat-ment effect. With respect to the former, repli-cating the current study may prove challenginggiven that the research design is based on a nat-ural experiment. One possibility for futureresearch is to contrast recidivism rates of statesthat require prisoners to return to their countyof last residence with states that do not. A sec-ond approach would be to identify states thatchanged the residency restrictions imposed onreleased prisoners and then compare recidivismrates before and after the changes were made.

Two particular extensions to the current studywould add another layer of nuance to under-standing the relation between migration andrecidivism. First, results presented in Table 4reveal a negative relationship between distanceof migration and recidivism, yet it remainsunclear just how far parolees must move to sep-arate themselves from past situations and crim-inal associates. For Claude Brown (1965), theroughly seven miles between Harlem andGreenwich Village was sufficient to separate




him from the daily criminal temptations of hisformer social environment. The key question iswhether there is a distance threshold after whichparolees have a substantial decline in the like-lihood of recidivism. Moreover, does a parolee’slikelihood of desisting from crime increase witheach increment after this threshold, or does itremain approximately the same after passing thethreshold? Further research is needed to answersuch questions. Second, as Abbott (1997:89)notes, “what makes a turning point a turningpoint rather than a minor ripple is the passageof sufficient time ‘on a new course’ such that itbecomes clear that direction has indeed beenchanged.” To more definitively conclude thatHurricane Katrina and the resulting migrationproduced turning points in the life course ofcrime for many former prisoners, it is thus nec-essary to follow up with the same cohorts ofindividuals in future analyses to determine if theobserved differences in recidivism acrosscohorts remain after a sufficient passage oftime (e.g., three to five years).

In addition to these various opportunitiesto further investigate the relation between res-idential migration and recidivism, other wor-thy research topics include the repercussionsof macro-structural changes due to Katrinaand the psychological strain associated withsuch a tragedy. With respect to macro-struc-tural change, Katrina not only influenced res-idential decisions at the individual-level butalso the macro-level distribution of paroleesand criminals more generally. De-concentrat-ing the spatial clustering of criminals, just likede-concentrating poverty, may attenuate thecultural disorganization that characterizes somany urban neighborhoods, as well as reversethe deleterious consequences of social isolation(Sampson and Wilson 1995). With respect topsychological strain, one plausible hypothesisis that strain associated with the hurricanewould increase the likelihood of recidivism.The fact that I found a sizable treatment effectdespite the likely psychological strain fromKatrina means that the effect of migration mayin fact be underestimated. Still, the psycho-logical impact of Hurricane Katrina may bearupon recidivism, and it is a worthy topic ofinvestigation.

While there are plenty of opportunities forfuture research on the effects of Katrina, as wellas the relation between migration, place of res-

idence, and recidivism, findings thus far suggestthat successful prisoner reentry and reintegra-tion depend on providing opportunities for pris-oners to separate from their criminogenic past.Such a finding provides initial support for thelife-course propositions outlined earlier. If crim-inal peers and familiar criminogenic contextscausally influence criminal behavior, separatingindividuals from their peers and familiar con-texts should lead to a reduced likelihood ofcriminal behavior. That is precisely what I find.In a practical sense, to lessen the likelihood ofrecidivism and to foster the path to desistance,it is beneficial to separate ex-prisoners fromtheir criminal past.

David S. Kirk is an Assistant Professor in theDepartment of Sociology and a Faculty ResearchAssociate of the Population Research Center at theUniversity of Texas at Austin. His current researchexplores the influence of social context and neigh-borhood change on criminal behavior. One ongoingproject examines the structural and cultural predic-tors of neighborhood violence. Kirk’s recent researchhas appeared in Demography and Criminology.

APPENDIX: ASSUMPTIONS OF THEIV FRAMEWORK

We must consider potential violations of theIV framework before concluding that the treat-ment effect of residential migration describedin Tables 2, 3, and 4 is in fact valid. As men-tioned previously, one key assumption is that theinstruments and the outcome variable are unre-lated, except through the treatment condition(Angrist and Krueger 2001). This is known asthe exclusion restriction. While the exclusionrestriction is not directly verifiable, it is possi-ble to indirectly assess this assumption in overi-dentif ied models where there are moreinstruments than endogenous regressors.Specifically, for the results presented in Table2, models are just identified—that is, the modelhas the same number of instruments as endoge-nous variables (one). Yet with five instrumentsfor the fourth model estimated in Table 4, themodel is overidentified. It is thus possible toindirectly test the exclusion restriction—thatis, the instruments are uncorrelated with themodel error—by correlating the residuals fromthe estimation of Equation 2 with the excludedinstruments. I use Hansen’s (1982) J statistic(see also Baum, Schaffer, and Stillman 2003) to




test the joint null hypothesis that the instru-ments are in fact valid instruments (i.e., they areuncorrelated with the error term).16 The J sta-tistic for the fourth model estimated in Table 4equals 5.996 (p = .200). I thus fail to reject thenull hypothesis, thereby providing necessarysupport for the exclusion restriction.

A second key assumption of the IV frame-work is that the covariance between the treat-ment and instrument differs from zero: cov(Si,Zi ) ≠ 0. If the instrument (i.e., post-Katrinarelease) does not affect residential migration,then use of Katrina as an instrument is inap-propriate. Results from Figure 1 support theassumption that the treatment and instrument co-vary, in that the spatial distribution of the parolepopulation changed following HurricaneKatrina. However, even if the treatment andinstrument are significantly associated, prob-lems may arise in the second-stage estimation(Equation 2) if they are only weakly related(i.e., a nonzero yet small correlation). This isknown as the weak instruments problem (Staigerand Stock 1997). A primary consequence isthat weak instruments produce inconsistent IVestimators (Bound, Jaeger, and Baker 1995).Additionally, a weak instrument increases stan-dard errors of the IV estimates and thereforeaffects hypothesis testing.

To assess the potential for a weak instru-ment, it is useful to examine the explanatorypower of the instrument in the first-stage (i.e.,Equation 1) of the two-stage approach. Forthe re-incarceration model from Table 2, toassess the instrument’s explanatory power, I usean F-test of the significance of the instrument.Results reveal that the instrument, post-Katrinarelease, is significantly correlated with thetreatment variable (F = 60.25; df = 1, 46; p <.001). Yet, a statistically significant F statisticmay still represent bias in the IV estimator, sothe F statistic must be compared against a suit-able threshold. Staiger and Stock (1997) sug-gest that an F statistic below 10 is indicativeof a weak instrument. In the case here, theinstrument post-Katrina release is statistical-ly significant and far exceeds Staiger andStock’s threshold. Such findings alleviate con-

cerns over violating assumptions of the IVapproach.17

REFERENCES

Abbott, Andrew. 1997. “On the Concept of TurningPoint.” Comparative Social Research 16:85–105.

Angrist, Joshua D. 1990. “Lifetime Earnings and theVietnam Era Draft Lottery: Evidence from SocialSecurity Administration Records.” AmericanEconomic Review 80:450–77.

Angrist, Joshua D., Guido W. Imbens, and Donald B.Rubin. 1996. “Identification of Causal EffectsUsing Instrumental Variables.” Journal of theAmerican Statistical Association 91:444–55.

Angrist, Joshua D. and Alan B. Krueger. 2001.“Instrumental Variables and the Search ofIdentif ication: From Supply and Demand toNatural Experiments.” Journal of EconomicPerspectives 15:69–85.

Baum, Christopher F., Mark E. Schaffer, and StevenStillman. 2003. “Instrumental Variables and GMM:Estimation and Testing.” Stata Journal 3:1–31.

Blau, Judith R. and Peter M. Blau. 1982. “The Costof Inequality: Metropolitan Structure and ViolentCrime.” American Sociological Review 47:114–29.

Bound, John, David A. Jaeger, and Regina M. Baker.1995. “Problems with Instrumental VariablesEstimation When the Correlation Between theInstruments and the Endogenous ExplanatoryVariable Is Weak.” Journal of the AmericanStatistical Association 90:443–50.

Brown, Claude. 1965. Manchild in the PromisedLand. New York: MacMillan.

Cantor, David and Kenneth C. Land. 1985.“Unemployment and Crime Rates in the Post-World War II United States: A Theoretical andEmpirical Analysis.” American SociologicalReview 50:317–32.

Carter, Brian L. and Stephen T. Tiffany. 1999. “Meta-Analysis of Cue-Reactivity in Addiction Research.”Addiction 94:327–40.

Duncan, Otis D. and Beverly Duncan. 1955. “AMethodological Analysis of Segregation Indexes.”American Sociological Review 20:210–17.

Eisenstein, James, Roy Flemming, and Peter Nardulli.


16 Tests of IV assumptions are available in Statathrough the ivreg2 function.

17 Another common strategy for assessing weakinstruments is to examine the partial R2 of the first-stage equation, which isolates the squared partialcorrelation between the instrument and the endoge-nous regressor (Bound et al. 1995). The magnitudeof the bias of IV estimators increases as the partialR2 approaches zero. First-stage results from the re-incarceration model in Table 2 reveal a partial R2 of.173.



1988. The Contours of Justice: Communities andTheir Courts. Boston, MA: Little, Brown.

Elder, Glen H., Jr. 1974. Children of the GreatDepression: Social Change in Life Experience.Chicago, IL: University of Chicago Press.

ESRI. 2006a. ESRI Demographic UpdateMethodology: 2006/2011. Redlands, CA: ESRI.Retrieved October 24, 2007 (http://www.esri.com/library/whitepapers/pdfs/ demographic-update-methodology.pdf).

———. 2006b. ESRI Gulf Coast UpdatesMethodology: 2006/2011. Redlands, CA: ESRI.Retrieved October 24, 2007 (http://www.esri.com/library/whitepapers/pdfs/gulf-coast-methodology.pdf).

Federal Bureau of Investigation. 2007. Uniform CrimeReporting Program Data [United States]: County-Level Detailed Arrest and Offense Data, 2005[Computer File]. Ann Arbor, MI: Inter-universityConsortium for Political and Social Research.

Garrett, Brandon L. and Tania Tetlow. 2006.“Criminal Justice Collapse: The Constitution afterHurricane Katrina.” Duke Law Journal 56:127–78.

Grogger, Jeff. 1998. “Market Wages and YouthCrime.” Journal of Labor Economics 16:756–91.

Hansen, Lars P. 1982. “Large Sample Properties ofGeneralized Method of Moments Estimators.”Econometrica 50:1029–54.

Heumann, Milton. 1975. “A Note on Plea Bargainingand Caseload Pressure.” Law & Society Review9:515–28.

Hipp, John R. 2007. “Block, Tract, and Levels ofAggregation: Neighborhood Structure and Crimeand Disorder as a Case in Point.” AmericanSociological Review 72:659–80.

Johnson, Brian D. 2005. “Contextual Disparities inGuidelines Departures: Courtroom SocialContexts, Guidelines Compliance, and ExtralegalDisparities in Criminal Sentencing.” Criminology43:761–96.

Langan, Patrick A. and David J. Levin. 2002.Recidivism of Prisoners Released in 1994.Washington, DC: Bureau of Justice Statistics.

Laub, John H. and Robert J. Sampson. 2003. SharedBeginnings, Divergent Lives: Delinquent Boys toAge 70. Cambridge, MA: Harvard UniversityPress.

La Vigne, Nancy G., Cynthia A. Mamalian, JeremyTravis, and Christy Visher. 2003. A Portrait ofPrisoner Reentry in Illinois. Washington, DC:Urban Institute.

Ludwig, Jens and Jeffrey R. Kling. 2007. “Is CrimeContagious?” Journal of Law & Economics50:491–518.

Massey, Douglas S. and Nancy A. Denton 1993.American Apartheid: Segregation and the Makingof the Underclass. Cambridge, MA: HarvardUniversity Press.

Merton, Robert K. 1938. “Social Structure and

Anomie.” American Sociological Review3:672–82.

Mumola, Christopher J. and Jennifer C. Karberg.2006. Drug Use and Dependence, State andFederal Prisoners, 2004. Bureau of JusticeStatistics Special Report. Washington, DC: U.S.Department of Justice.

National Research Council. 2007. Parole, Desistancefrom Crime, and Community Integration.Committee on Community Supervision andDesistance from Crime. Committee on Law andJustice, Division of Behavioral and Social Sciencesand Education. Washington, DC: The NationalAcademies Press.

Osborn, S. G. 1980. “Moving Home, Leaving Londonand Delinquent Trends.” British Journal ofCriminology 20:54–61.

Roman, Caterina Gouvis, Seri Irazola, and Jenny W.L. Osborne. 2007. After Katrina: Washed Away?Justice in New Orleans. Washington, DC: UrbanInstitute.

Rosenfeld, Richard, Joel Wallman, and RobertFornango. 2005. “The Contribution of Ex-Prisoners to Crime Rates.” Pp. 80–104 in PrisonerReentry and Crime in America, edited by J. Travisand C. Visher. New York: Cambridge UniversityPress.

Sabol, William J. and Heather Couture. 2008. PrisonInmates at Midyear 2007. Bureau of JusticeStatistics Bulletin. Washington, DC: U.S.Department of Justice.

Sampson, Robert J. and John H. Laub. 1993. Crimein the Making: Pathways and Turning Pointsthrough Life. Cambridge, MA: Harvard UniversityPress.

Sampson, Robert J. and William J. Wilson. 1995.“Toward a Theory of Race, Crime, and UrbanInequality.” Pp. 37–54 in Crime and Inequality,edited by J. Hagan and R. D. Peterson. Stanford,CA: Stanford University Press.

Shaw, Clifford R. and Henry D. McKay. 1942.Juvenile Delinquency and Urban Areas. Chicago,IL: University of Chicago Press.

Staiger, Douglas and James H. Stock. 1997.“Instrumental Variables Regression with WeakInstruments.” Econometrica 65:557–86.

Turner, Susan and Joan Petersilia. 1992. “Focusingon High-Risk Parolees: An Experiment to ReduceCommitments to the Texas Department ofCorrections.” Journal of Research in Crime andDelinquency 29:34–61.

Ulmer, Jeffery and Brian D. Johnson. 2004.“Sentencing in Context: A Multilevel Analysis.”Criminology 42:137–75.

U.S. Department of Housing and UrbanDevelopment. 2006. Current Housing UnitDamage Estimates: Hurricanes Katrina, Rita, andWilma. Washington, DC: U.S. Department ofHousing and Urban Development. Retrieved




December 8, 2006 (http://www.huduser.org/publications/destech/GulfCoast_HsngDmgEst.html).

Visher, Christy and Jill Farrell. 2005. ChicagoCommunities and Prisoner Reentry. Washington,DC: Urban Institute.

Warr, Mark. 1993. “Age, Peers, and Delinquency.”Criminology 31:17–40.

———. 1998. “Life-Course Transitions andDesistance from Crime.” Criminology 36:183–216.

———. 2002. Companions in Crime: The SocialAspects of Criminal Conduct. Cambridge, UK:Cambridge University Press.

Wikler, Abraham. 1948. “Recent Progress inResearch on the Neurophysiologic Basis ofMorphine Addiction.” American Journal ofPsychiatry 105:329–38.




Documents

The online version of this article can be found at: http ...users.soc.umn.edu/~uggen/Kirk_ASR_09.pdfAMERICAN SOCIOLOGICAL REVIEW, 2009, VOL. 74 (June:484–505) Direct correspondence