mopdf.com

Smith, 2006a; Schuetz, Been, and Ellen, 2008; Harding, Rosenblatt, and Yao, 2009;Haughwout, Mayer, and Tracy, 2009; Lin, Rosenblatt, and Yao, 2009; Rogers andWinter, 2009; Hartley, 2010; Wassmer, 2010; Campbell, Giglio and Pathak; 2011;Gerardi, Rosenblatt, Willen, and Yao, 2012). The papers vary in their methods butseveral use statistical techniques to demonstrate that foreclosures actually lead toreductions in with such reductions.2 However, while these papers persuasivelydemonstrate causality, few explore the mechanism through which foreclosures reduceproperty values. One exception is Gerardi et al (2012), who find that estimated priceeffects are highly sensitive to the reported condition of the foreclosed property,suggesting that reduced home maintenance and investment by owners of properties inforeclosure play a large role in driving price effects.II. MechanismsWe model the relationship between foreclosures and crime by focusing on thedecision-making process of potential offenders, borrowing from Becker’s theory ofcriminal behavior (Becker, 1968) and the framework of routine activity theory commonlyused in criminology research (Cohen and Felson, 1979). The assumptions of routineactivity theory, which are that criminal acts require potential offenders, suitable targets,and the absence of “capable guardians” who prevent crime, provide a set of theoreticalpathways through which foreclosures may affect crime by changing the expected payoffsand costs of committing crime.2Note that several other recent papers have used difference-in-difference regression methods to identify theexternal effects of property characteristics on surrounding property values (Schwartz, Ellen, Voicu, andSchill, 2006; Linden and Rockoff, 2008).6

The structure of a routine activities model most closely resembles search modelsof economic decision making (e.g. job search models or consumer search models).Specifically, we assume that the population of potential offenders moves acrossblockfaces, each potential offender encountering opportunities to commit crime.3 Eachopportunity to commit crime presents a new optimization problem. Potential offenders inthis model are rational agents, and commit crime if the payoff from the crime minus thecost of committing the crime, exceeds the payoff from not committing the crime. Thecost of not committing the crime is normalized to zero, and most of the cost ofcommitting a crime is the perceived chance of being caught.4Foreclosures potentially change the benefits and costs of committing a crime on ablockface by affecting both the availability of suitable targets for criminal activity and theperceived presence (or absence) of “capable guardians” against crime. The foreclosureprocess varies across states, but lenders typically initiate a foreclosure (by issuing anotice of foreclosure or lis pendens) after a borrower defaults on his or her mortgage, orstops making payments for 90 days. Borrowers may default on their mortgages whenthey cannot afford mortgage payments and they cannot sell their home for a price thatwill cover the mortgage amount. Alternatively, some borrowers with negative equitymay default because they recognize that they owe more than the home is worth, and theirties to the neighborhood are weak.53Potential offenders are assumed to move around the city as a part of their routine activities, and so thecosts of searching for a criminal opportunity (relative to not committing a crime) are zero.4This is a common result in the economic literature on crime deterrence; see Grogger, 1991.5Recent research suggests that such strategic default behavior is relatively rare. See Bhutta, Dokko, andShan, 2011.7

An owner who receives a foreclosure notice from a lender may cut back onmaintenance of her building or grounds either because she needs to save money to payback arrears or because she expects she will lose the property and no longer has anincentive to keep up its value. The visible deterioration of the property that follows maysignal to potential offenders that local residents are less invested in the block and lesslikely to intervene in or report crime, which decreases the perceived chances of beingcaught (Harcourt and Ludwig, 2006).Foreclosures may also reduce the number of ‘capable guardians’ or monitors ofpublic spaces in the neighborhood, as households with homes in the foreclosure processwithdraw from the neighborhood, either due to stress or because they simply care lessabout the neighborhood once their financial investment in it is imperiled. Similarly, theelevated residential turnover that may result as owners sell or lose their homes, or asowner-occupied properties are converted to rental units may make it more difficult forlocal residents to recognize outsiders and to maintain the effective social controls (suchas neighborhood watch associations) that help to cut off opportunities for crime(Sampson, Raudenbush, and Earls, 1997; Taylor, 1997).The effects of foreclosures are likely to be magnified for properties that linger inthe foreclosure process, as the problems resulting from reduced maintenance will mount.Further, properties that go through lengthy foreclosures are more likely to become vacant.Vacant properties clearly provide easy targets for vandalism and trespassing, and mayalso provide a safe haven for prostitution and drug-related crimes that can potentially lead8

to more serious, violent crimes (Spelman, 1993).6 Vacant properties may also facilitatecertain types of property crime, such as theft of wires and appliances. That said, thepayoff from stealing from vacant buildings may be lower than that from stealing fromoccupied buildings, because vacant buildings include fewer valuable and marketableitems (e.g. jewelry, laptop computers, and other electronics). As a result, a larger numberof foreclosed properties on a block might actually reduce the number of targets orperceived payoff to property crimes and decrease their number.III. Analysis: Differences-in-Differences ModelGeographic Unit of AnalysisOur primary unit of analysis is the blockface, a street segment that is bounded bythe two closest cross-streets, and which incorporates buildings on both sides of the street(see Figure 1). We believe blockfaces are preferable to the more commonly-used Censusblocks (encompassing all buildings on the interior of a square city block bordered by fourstreet segments) because foreclosures are more likely to affect behavior and crime justacross the street than around the corner (or two corners). In a study of crime on streetsegments in Seattle, Groff, Weisburd, and Yang (2010) found that crime patterns variedwidely between street segments, reinforcing the importance of using small-scalegeographies in research on crime. We employed New York City street shapefiles and GISanalysis to create blockfaces, which are not captured in standard mappingshapefiles. These geographic units are very small – in New York City, there are over6Indeed, vacant properties may attract more lucrative or dangerous forms of criminal activity (for example,a vacant building can house a drug lab, while a street corner can only provide a “retail” site), which mightincrease the seriousness of crime as well as increasing the amount of crime.9

96,000 blockfaces. We also map foreclosures and crimes to police precincts and Censustracts.Baseline Model and Identification StrategyEmpirically identifying the causal effect of foreclosures on crime is a challenge,as elevated crime on a block might reduce demand and prices and potentially triggerdefaults and foreclosures as some borrowers find themselves in negative equity.7 Butsince borrowers might need time to adjust their perceptions of blockface-level houseprices and then to make the decision to stop making mortgage payments, and lendersgenerally must wait for the loan to enter formal default status (90 days past due) beforeissuing a foreclosure notice, this reverse-causal mechanism would likely take some timeto unfold – so crime would not immediately lead to foreclosures. Perhaps moreworrisome is the possibility that very localized distress (economic, social, or otherwise)might lead to both elevated crime and foreclosure activity.We take several steps to address concerns about causality. First, we lag ourforeclosure measures, which means that any confounding, localized distress would haveto have led to elevated crime in the current quarter but elevated foreclosures in the priorsix quarters. Second, we include blockface fixed effects to take into account pre-existing,time invariant, blockface-specific contributions to the payoffs and costs of committingcrime, such as geographic features, proximity to commercial areas and transit, and thedistribution of building and occupancy types.8 Further, we control for the characteristicsof the neighborhood that change over time (such as household structure, age, income, and78White (1986) shows that a failure to lower assessed values can accelerate the pace of abandonment.We also estimate models without blockface fixed effects that include these measures directly.10

other local socioeconomic drivers of mobility) by also including neighborhood-specifictime trends (either police precinct-by-quarter, or Census tract-by-quarter, depending onthe specification) as fixed effects. Through including these neighborhood-by-quarterfixed effects, we capture most of the localized social and economic trends that might leadto both elevated foreclosures and crime. As discussed below, Census tracts are quitesmall in New York City given the city’s high population density. The average Censustract in the city covers just 0.14 square miles and includes only 14 blocks and 29blockfaces.As for trends that are specific to the blockface, we include some attributes ofblockfaces that change over time, which may reflect population and economic shifts,including changes in the total number of units, building demolitions and newconstruction, and new liquor licenses. Finally, in some models, we also include futureforeclosure starts to control for unobserved trends on a blockface that might be correlatedwith both crime and foreclosures. If unobserved trends in blockface-level economicdistress or social cohesion are causing both foreclosures and crimes, then futureforeclosure activity also should be correlated with current crime (through the effects ofthe unobserved, persistent trends).Intuitively, then, our baseline empirical strategy in answering this question is tocompare changes in crime levels on blockfaces experiencing an increase in foreclosureactivity to changes in crime levels on nearby blockfaces that are not experiencing anincrease in foreclosures, but are located within the same small neighborhood (defined as apolice precinct or Census tract). Specifically, we estimate the following model:11

where ybnt is the level of criminal activity on blockface b in neighborhood n and quarter t.We focus primarily on simple counts of crimes per quarter, rather than rates, in partbecause we do not have quarterly population estimates for blockfaces, but we control forthe number of units on a blockface in a given quarter to capture changes in density.Note that our use of crime counts instead of crimes per capita will bias ourestimated coefficients towards zero if foreclosures lead to reductions in the population,since our estimates will understate the true impact of foreclosures on victimization riskby not accounting for the reduction in population. Similarly, to the extent that a smallerpopulation means fewer people available to report crime, any reporting bias will also bein the direction of not finding an effect. It is possible that policing strategies havechanged in response to the foreclosure crisis, and that areas with high foreclosure activityare more heavily monitored by the police. However, our dataset of crime reports arethose initiated by complaints made by crime victims and observers, and excludes to theextent possible crime reports that were initiated solely by police officers.9 Thus, our dataare less affected by any monitoring bias.On the right hand side of the equation, Xbnt-1 is a measure of foreclosure activityin the previous quarter on blockface b; Zbnt represents our set of time-varying blockfacecharacteristics (including total units) to control for other observable changes in theblockface over time that might affect the payoffs and costs to committing crime; Tnt is avector of fixed effects indicating the quarter for each neighborhood n, which controls forcrime, policing, and other trends in the neighborhood; Bb are blockface fixed effects,9False or prank calls are not filed as official crime complaints, and are not included in our data. It ispossible that some of the drug crime reports were initiated by police officers but they are likely a minority.12

which control for time-invariant differences between blockfaces with more and lessforeclosure activity; and εbnt is the random error term.As noted, we proxy for neighborhoods with both police precincts and Censustracts. While the Census tract-by-quarter variables allow us to control for trends in asmaller geographic area, there are also good arguments for also using police precincts.First, policing is managed in New York at the precinct level, and thus controlling forprecinct time trends will capture any differences in trends in policing practices or crimereporting across precincts. In addition, Census tracts are quite small in New York Citygiven the city’s high density, and thus the comparison group of blockfaces with a Censustract is limited and potentially affected by a given foreclosure too (which may bias ourestimated coefficients towards zero). Thus, we generally show both sets of results.10Finally, to further test the direction of causality, following Schuetz et al (2008)and Campbell, Giglio, and Pathak (2011), we estimate the relationship between pastforeclosure starts on the blockface and crime while also controlling for futureforeclosures on the blockface (a count of the number of foreclosure starts in the 18months following the quarter for which we measure crime). Specifically, we estimate:Foreclosure notices issued 18 months in the future (Xbnt 1) should pick up unobservedtrends in the blockface that might be correlated with both foreclosures and crime, but nothave any direct effect on criminal activity today (ybnt).10The full set of results is available from the authors upon request. To estimate models with more than onehigh-dimensional fixed effect we use the Stata command “reg2hdfe” developed by Guimaraes & Portugal(2010).13

Additional AnalysisOur paper addresses five secondary questions as well. First, we examine whattypes of crime are most sensitive to foreclosures by re-estimating the above equationseparately for violent, property, and public order crimes. As noted above, we expectlarger effects on public order and violent crimes than on property crimes. Second, wetest whether impacts on crime vary depending on the outcomes of the foreclosures.Specifically, we compare results of regressions using several distinct measures offoreclosures ranging from a simple count of foreclosure notices issued to properties onthe blockface in the past 18 months to a count of properties headed to foreclosureauctions or in bank ownership. These measures are discussed in greater detail below.Third, we explore whether effects of foreclosure extend beyond the blockface on whichthe foreclosure occurs (blockface i). To do so, in modeling crime on blockface j, weinclude both a variable capturing a count of recent foreclosure notices issued onblockface i and a variable capturing the number of foreclosure notices recently issued onthe six blockfaces that are contiguous to blockface i. Fourth, we explore whether impactsdepend on the initial level of crime in the neighborhood by estimating models separatelyin the lowest and highest crime quartiles of police precincts. Finally, we examinewhether there is a threshold level of foreclosures that triggers crime by modelingforeclosure counts using a set of categorical variables to capture the intensity offoreclosure activity.Negative Binomial ModelThe fine-grained nature of our data presents some potential pitfalls for estimation.Specifically, because blockfaces are such small geographic units, there are a substantial14

number of time periods when no crimes in our chosen categories occur. As shown inFigure 2, this skews the distribution of crime levels towards zero, violating the normaldistribution assumption and making it likely that an ordinary least squares regression willbe a poor fit for the data, especially for models of violent crime. We address this issue byestimating the above relationship using a negative binomial model (following Sampson,Raudenbush, and Earls (1997), and Osgood (2000) in their studies of neighborhoodcrime). These models estimate the relationship between foreclosure activity and crimeusing a different distributional assumption for the underlying empirical relationship thanthe standard linear model, and arguably provide a better fit for our data. (See theAppendix for more detail on estimation.)IV.Data SourcesCrime DataUnder an agreement with the New York City Police Department, we haveobtained point-specific data on all crimes initiated by victim or bystander complaints andreported in New York City between 2004 and 2008.11 This detailed dataset includes thespatial coordinates of each reported crime, along with its date, time, and offense category(shown in Table 1). We used GIS procedures to assign each crime to various levels ofgeography, including police precincts, Census tracts, and blockfaces.12 Many of the X/Y11The NYPD assigns each incident a unique identification number, therefore there are not duplicates basedon multiple reports of the same incident.12There are 76 police precincts, 2,246 census tracts, 36,601 census blocks, and 96,933 blockfaces in NewYork City. We limit our sample of blockfaces to those that have at least 1 building, and that are able to bematched to the New York City Department of Finance’s Real Property Assessment Database (RPAD) dataabout property characteristics, resulting in a sample of approximately 89,000.15

coordinates of crimes are geo-coded to the middle of the street, or literally on the borderof two Census blocks (and often two Census tracts). These crimes do not pose a problemfor our analysis, because they clearly occur on a single blockface. We assign the 20percent of crimes that take place at intersections to multiple blockfaces, as they could beaffected by conditions on all adjoining blockfaces. We randomly assign the 19 percent ofcrimes that take place on a blockface on the border of two Census tracts to one tract orthe other. Although we have the exact date of both crimes and foreclosure notices, weaggregate crimes to quarters, as sample sizes do not permit shorter time periods forblockfaces.Mortgage Foreclosure DataIn New York State, a mortgage foreclosure is initiated when the foreclosing partyfiles a legal document, called a lis pendens, in the county court.13 We use foreclosurefiling data from 2003-2010 obtained from a private vendor, the Public DataCorporation.14 Note that many properties that receive foreclosure notices do not go toauction; owners are able to cure those foreclosures through paying back arrears, receivinga modification, or selling quickly to a new owner. Despite this, most researchers simplyuse data

unclear whether elevated foreclosures lead to crime or whether some changes in unobserved conditions lead to increases in both foreclosures and crime. Similarly, Kirk and Hyra (2011) study the connection between foreclosures and crime in community areas in Chicago, which are relatively large neighborhoods comprised of about 10 Census tracts each.