Policing the Homeless: An Evaluation ofEfforts to Reduce Homeless-Related Crime

Richard BerkDepartment of Statistics

Department of CriminologyUniversity of Pennsylvania

John MacDonaldDepartment of CriminologyUniversity of Pennsylvania

September 11, 2009

Abstract

Summary — Police officials across the United States are increas-ingly relying on place-based approaches for crime prevention. In thisarticle, we examine the Safer Cities Initiative, a widely publicizedplace-based policing intervention implemented in Los Angeles’s “SkidRow” and focused on crime and disorder associated with homelessencampments. Crime reduction was the goal. The police division inwhich the program was undertaken provides 8 years of times seriesdata serving as the observations for the treatment condition. Fouradjacent police divisions in which the program was not undertakenprovide 8 years time series data serving as the observations for thecomparison condition. The data are analyzed using a generalized ad-ditive model. On balance, we find that this place-based interventionis associated with meaningful reductions in violent, property, and nui-sance street crimes. There is no evidence of crime displacement.

Policy Implications —This study provides further evidence that ge-ographically targeted police interventions can lead to significant crime

1

prevention benefits, with no evidence that crime is simply displaced toother areas. Criminologists and the media have given the Los AngelesPolice (LAPD) little credit for major reductions in crime that haveoccurred over the past five years following a number of major policyreforms. We suggest that researchers should look more closely at thetargeted interventions the LAPD has undertaken for evidence-basedexamples of effective policing. Importantly, this work suggests thatcrime associated with homeless encampments can meaningfully re-duced with targeted police actions. However, law enforcement actionsdo not address the roots of homelessness nor most of its consequences.Getting tough on the homeless should not be confused with policiesor programs that respond fundamentally to the social and personalproblems that homelessness presents.

1 Introduction

A growing body of evaluation research supports the practical benefits ofplaced-based policing. Yet, scholars are often critical of such approaches inpart because place-based interventions can shift the burden of crime to otherareas. Perhaps most directly, crime displacement can result from geograph-ically targeted police-crackdowns (Sherman, 1990), although displacementappears to be less problematic when place-based policing is attuned to thelocal criminal environment (Weisburd et al., 2006). As Weisburd and col-leagues (2006) note, crime does not just move around the corner. Morefundamentally, there is little reason to expect that police interventions alonechange underlying social and economic factors that contribute to crime. Ef-fective police interventions can make a short-term difference in a given locale(Skogan, 1990), but on larger spatial and temporal scales, there will usuallybe little change.

These issues are particularly noteworthy for this paper. For years, down-town Los Angeles has been characterized in part by a sizable homeless pop-ulation (Berk et al., 2008) concentrated in an area commonly referred to as“Skid Row” (Magnano and Blasi, 2007). Crime and disorder have been astable feature of Skid Row life. Open-air drug markets, prostitution, nightlyrobberies, drug overdoses, theft, and vandalism have been a common. Lo-cal merchants have long lobbied for a tougher response from the LAPD (seeHarcourt, 2005 for some history). In the fall of 2005, they got their wish.

The majority of Los Angeles City services for the homeless are located

2

in downtown Los Angeles. These providers and other stakeholders speakwith many different voices about the needs of the homeless and how best torespond (Koegel et al., 1988; Wenzel et al., 2000). There are also, not sur-prisingly, some basic disagreements and battles over turf. But no one wouldargue that a crack down on the homeless could be more than a temporary andpartial measure, and some complain that law enforcement resources might bebetter spent providing additional homeless services (Blasi, 2007). As impor-tant as they debates may be, they neglect three logically prior key questions.Did the LAPD interventions intended to reduce the crime associated withthe Skid Row encampments really reduce crime? If they did, how large werethe reductions and were they true reductions or merely displacements? Inthis paper, we attempt to provide answers.

2 Background

That crime and place are closely linked is indisputable. Over a century ofempirical research in the United States and Europe has demonstrated thatcrime is geographically concentrated (see Weisburd, Bruinsma, and Bernasco,2009 for an historical overview). Theoretical explanations at various spatiallevels have followed (Sampson, 1995; Taylor, 2001; Welsh and Hoshi, 2002).Practical applications of this knowledge have also been developed. Perhapsmost notable is a focus of police resources on crime hot spots or areas ofparticularly high crime concentrations (Sherman et al., 1989). While policeactions directed at specific persons or crimes have produced mixed-results(Sherman, 1990), a number of studies point to consistent effects from space-based police interventions (see Braga, 2001; Weisburd and Eck, 2004 forreviews).

A common criticism for place-based policing strategies is that they aremyopic. The underlying causes of crime are not addressed, although severalnotable policing scholars have argued that the police can in principle affectsystemic issues under certain circumstances (Eck and Spellman, 1987; Gold-stein, 1990). Perhaps more to the point, reductions in crime are a desirablegoal and one for which police are well suited. Also, there is nothing aboutpolice initiated attempts to reduce crime that necessarily preclude more fun-damental interventions by other agencies. In short, there is little that thepolice can do to “solve” the problem of homelessness. But perhaps well-designed policing strategies can reduce some of its undesirable consequences.

3

Whether in this case the resources allocated to the police might be betterused to address root problems (National Law Center on Homelessness andPoverty, 2009) is an important issue, but not one that can be appropriatelyaddressed here.

There is a marked increase in victimization when individuals find them-selves in settings where excessive drinking of alcohol, drug use, and otherrisky behaviors are prevalent (Felson, 2002). It is not surprising, therefore,that homeless individuals have significantly higher rates of criminal victim-izations than individuals who have a place to live (Koegel et al., 1988; Kushel,Evan, et al., 2003). It is also not surprising that homeless encampments canbe especially problematic for the homeless and others because there is a con-centration of prospective perpetrators and victims (Meithe and Meyer, 1990;Wenzel et al., 2000; Sampson and Lauritsen, 1990). A key implication forcrime that is probably more important than the number of homeless in ajurisdiction is their spatial density.

3 The Safer Cities Initiative

Los Angeles (LA) County has the largest number of homeless individuals ofany county in the United States (Berk et al., 2008). Consistent with histor-ical trends (Rossi, 1989), the population of homeless is very heterogeneous.According to LA Almanac, (2009) the County homeless population:

1. has an average age of about 40;

2. nearly 50% with a high school education;

3. is at least 33% female;

4. is at least 20% composed of families;

5. has about 20% with disabilities;

6. has at least 16% employed;

7. has about 25% mentally ill; and

8. has at least 33% with substance abuse problems.

4

A large fraction of the LA County’s homeless can be found in the City ofLos Angeles, with the highest concentration by far on Skid Row. It is likelythat the Skid Row population has many features in common with the Countypopulation, with perhaps a greater density of disadvantage. Crime problemsthat go well beyond mere nuisance are rampant. Open-air drug markets,prostitution, robbery gangs, theft, and vandalism are common. Until re-cently, these Skid Row conditions went largely unaddressed.

Starting in 2004, media and public policy attention began to highlightthe conditions surrounding homelessness in downtown LA’s Skid Row (LosAngeles Times, 2009). Notable stories in the Los Angeles Times describeddirty needles littering the streets, prostitution conducted in public porta-potties placed on Skid Row, robbery gangs targeting homeless individuals fortheir disability checks, drug-related overdoses, and frequent violence (Lopez,2005). Local merchants and private developers were particularly vocal aboutthe these and a variety of related problems (Harcourt, 2005).

In September 2005, the Los Angeles Police Department pilot tested aneffort to clean up the area. The intervention was officially named the “SaferCities Initiative” (SCI) and was a hallmark of Chief William Bratton’s “bro-ken windows” approach (Bratton, 1989; Wilson and Kelling, 1982) targetingwell-defined defined geographical locations. Under the direction of CaptainJodi Wakefield, a pilot program called the “Main Street Project” was in-tended to reduce the density of homeless encampments through fines andcitations. Encampments located in the “historic core” section downtown LA(S. Main Street between 4th and 7th streets) were deemed to be a publichealth nuisance and sanctions were authorized under LA city statute 1

The LAPD also cracked down on crimes like public intoxication, druguse, and prostitution that made the area congenial for criminal predators.Starting in October, the LAPD placed 4 to 5 officers on foot in the historiccore section of downtown, who were to focus exclusively on general nuisancecrimes and basic order maintenance. There followed a further increase inpolice presence with the introduction of one of the LAPD’s mobile policecommand stations, commonly referred to a “Big Blue.” In addition, under-cover vice teams were placed in the areas known for open air drug markets

1The LA municipal code notes, “No person shall sit, lie or sleep in or upon any street,sidewalk or other public way.” (L.A., CAL., MUN. CODE 41.18(d). The use of thismunicipal code to explicitly arrest homeless persons was deemed a violation of the 8thamendment by the 9th Circuit Court of Appeals (Jones v. City of Los Angeles, 444 F.3d1118, 1120 (9th Cir. 2006)

5

and for prostitution. Finally, a special undercover squad focused on localrobberies (Wakefield, 2009).

After the pilot phase was implemented largely as designed and with anec-dotal evidence of success, the LAPD officially launched the full-scale versionof the SCI on September 17th, 2006 by placing 50 full-time officers on thestreet in downtown LA. Just as in the Main Street Project, the officers weredeployed only to the historic core of downtown LA. The officers worked east-ward through the Skid Row section, breaking up homeless encampments,issuing citations and making arrests for violations of the law. The plan wasto clear out specific street areas, maintain a visible police presence for atleast a week, and then move onto other parts of downtown. The dedicatedofficers also worked closely with two vice units (Narcotics Buy Team andField Enforcement Section).

The immediate goals of the SCI were demonstrably achieved. The SkidRow homeless encampments were cleared. The concentration of homelessindividuals was dispersed. The debris they left behind was removed. Butwhat about crime in the downtown LA? Was there a reduction? Accord-ing to LAPD internal documents and media reports (see Los Angeles Po-lice, 2008a,b), homeless-related drug overdoses, murders, and reported crimesdropped the year after the intervention.

However, the program was not independently and rigorously evaluated,and there were many skeptics (Blasi, 2007). In this paper, we report theresults of an independent, empirical evaluation. We consider the possiblecrime reduction impact of the Main Street Project, which served as the pilot,and of the SCI. Three broad kinds of crime are examined: (1) nuisance crime,(2) violent crime, and (3) property crime. We also consider possible effectson four areas adjacent to downtown.

4 Research Methods

4.1 Research Design

The unit of observation and analysis for this study is the Police Division.The Los Angeles Police Department is organized into four Police Bureaus andwithin them, nineteen divisions. The LAPD intervention was implementedin the Central Division, which is part of the Central Bureau. The CentralDivision contains much of the downtown area, including large office and city

6

government buildings, their support services, light industry and a mix ofresidential housing. The Central Division also contains several square blocksof shops patronized by people of modest means living near the downtown.And the division contains the LA’s Skid Row. The Central Division is theexperimental unit.

As shown in Figure 1, the Central Division is bordered by four otherCentral Bureau Police Divisions: Northeast, Rampart, Hollenbeck and New-ton. These divisions contain a mix of moderate to low income householdsand a mix of industrial and commercial establishments. The four divisionscontiguous to the Central will serve as the comparisons units because oftheir proximity to the downtown, their broad similarities with many impor-tant spatial and demographic features with the Central Division, and theirshared command structure under the Central Bureau.

For each of the divisions, we have three time series of crime counts byweek: the number of violent crimes, the number of property crimes, and thenumber of nuisance crimes. Nuisance crimes often are associated with thehomeless.2 We consider different kinds of crimes because although commonlore has the homeless as perpetrators of a wide variety of nuisance crimes,they can also be perpetrators and victims of very serious crimes. An impor-tant question, therefore, is whether all crimes can be affected by the inter-ventions or just nuisance crimes. Although the immediate goal of dispersingthe homeless was achieved, the homeless are hardly the only crime victimsand perpetrators in downtown Los Angeles. Even if homeless-related crimesare affected, other factors may dominate any local trends for certain kindsof crimes.

The three time series for the Central Division serve as three outcomevariables for the experimental group. The three time series for each of thefour contiguous divisions serve as three “non-equivalent no-treatment con-trol group time series” in a quasi-experimental design (Shadish et al., 2002:181-184). Crime patterns over time that the control divisions share with theexperimental division should not be a direct result of interventions that wereintroduced only in the experimental division. But if ignored, they risk beingconfounded with any effects of the interventions. In principle, the confound-

2Violent crime included aggravated assault (57.7%), homicide (1.4%), rape (2.3%),robbery (38.6%). Property crimes included burglary (12.8%), theft from vehicles (25.3%),attempted burglary (25.7%), grant theft auto (21.2 %), and larceny theft (14.5 %). Nui-sance crime included disturbance (0.8 %), lewd conduct (1.3 %), petty theft (40.4 %),pickpocketing (1.8 %), trespass (4.6 %), and vandalism (51.1 %).

7

Figure 1: Spatial Organization of the LAPD and Central Bureau Divisions

8

ing can be removed by regression-based covariance adjustments. Preciselyhow we will do this is discussed shortly.

There are 419 weeks of data starting on January 1st of 2000 and endingon December 31st of 2007. The first and last weeks are dropped because theywere not a full seven days. The timing of the LAPD Main Street Project forSkid Row and the SCI was determined through several conversations withthe LAPD captain in charge of both, who in turn consulted staff activity logs(Wakefield, 2009).

The formal start of the SCI was easy to document. It began in week351 and continued through the end of 2007. The start of the pilot programwas far more ambiguous. On September 27, 2005 the Main Street Projectofficially began (week 300). Skid Row porta-potties were removed beginningon October 26 (week 304). On November 16, 2005 (week 307), the LAPDadded 43 recruit officers to walk beats en mass in Skid Row. However, severalweeks before the official announcement, the proposed pilot program was beingwidely discussed by the media. Moreover, a variety of stakeholders werealready involved so that by the middle of August 2005 it was widely knownthat a police initiative to clear Skid Row of the homeless was immanent. Itmay well make sense, therefore, to make week 294 (or so) the starting datefor the pilot program. Week 294 is six weeks before the Main Street Projectbegan officially.

4.2 The Statistical Model

Because we do not know the functions by which the crime time series forthe comparison divisions may be related to crime time series for the exper-imental division, we apply the generalized additive model. The generalizedadditive model (GAM) allows functional forms for the relationship betweenquantitative regressors and the response to be inductively determined fromthe data (Hastie and Tibshirani, 1990).3 The generalized additive model inthis application can take the following form separately for each of three crimecategories.

yc,t = e[αc,0+αc,1I1,t+αc,2I2,t+P4

d=1 fc,d(xc,d,t)] + εc,t, (1)

where

• c is an index for one of the three kinds of crime;

3There are no functional form issues for the categorical predictors because they are notquantitative. However, transformations of the response are on the table.

9

• t is an index for week;

• d is an index for one of the four comparison divisions,

• yc,t is the response crime count by week for one of the three kinds ofcrime,

• I1,t an indicator variable coded as “1” for the period in which the MainStreet Project was in place, and “0” otherwise;

• I2,t an indicator variable coded as “1” for the period in which the SaferCities Initiative was in place, and “0” otherwise;

• fc,d(xc,d,t) is an inductively generated relationship between the crimetime series c for comparison division d and the response crime series c;and

• εc,t is a random Poisson disturbance meeting the usual regression as-sumptions.

The parameter αc,0 is the usual intercept for crime type c. The parameterαc,1 will capture the direction and size of any average shift up or down inyc,t associated with the introduction of the Main Street Project. As such,it can provide an estimate of the pilot program’s treatment effect for crimetype c. The parameter αc,2 will capture the direction and size of any averageshift up or down in yc,t associated with the introduction of the SCI. As such,it can provide an estimate of the SCI’s treatment effect for crime type c.More complicated treatment functions can be formulated and if necessary,time-related patterns that are unique to the Central Division, but not as-sociated with the intervention, can be removed through additional terms inequation 1. The systematic part of equation 1 is exponentiated consistentwith the conventional log-link function for Poisson regression (Hastie andTibsharini, 1990: 139).

There can be no regression coefficients associated with the functions ofquantitative covariates. Any information in a single regression coefficientwould be automatically incorporated in the inductively constructed func-tional form.4 Moreover, for nonlinear functions, there is not one regressioncoefficient but many; indeed a limitless number under many circumstances.The derivative of the function at each predictor value is a slope.

4Formally speaking, they are not identified.

10

In practice, the inductive functions are likely to be rather smooth. Con-sequently, the covariance adjustments will control for relatively smooth tem-poral patterns in crime that the experimental division shares with the com-parison divisions. Sharp discontinuities are usually not well captured or canbe missed altogether. We will return to this issue later.

Key elements of equation 1, as in any regression model, are the propertiesof εt. In particular, there is here the real possibility of dependence withinthe disturbances of the response time series. Such dependence can underminestatistical inference. Equation 1 assumes implicitly that such dependence isremoved by covariance adjustments using the four comparison time series.Insofar as the temporal dependence for the crime counts in any of the com-parison divisions is similar to the temporal dependence for the crime countsin the experimental time series, it will be eliminated. But the possibility ofany dependence remaining is an empirical matter we will address below.

4.3 Estimation Procedures

Obtaining estimates of the regression parameters of equation 1 while alsoobtaining estimates of the functional forms for the covariates complicatesmatters. In the same spirit as the generalized linear model, a form of itera-tively reweighted least squares is used. However, the usual regression sum ofsquares is “penalized” (Wood, 2008: 500).5 A penalty is introduced so thatan appropriate degree is smoothing is achieved for the functional forms to beestimated. The intent is to balance the bias that comes from not smooth-ing enough against the variance that comes from smoothing too much. Themore weight given to the penalty, the smoother the function. The less weightgiven to the penalty, the rougher the function. The weight arrived at is usu-ally determined by some measure of out-of-sample performance, such as thecross-validation statistic.

5The function to be minimized is ||W(z−Xβ)||2+∑

j λjβT Sjβ, where W is a diagonal

weight matrix as usual, z is the adjusted response for the given iteration, X is a matrix ofregressors including the design matrix for the splines associated with each of the smoothingfunctions, β is a vector of coefficients whose values are to be estimated, λj is a weightgiven to the penalty function for the jth function that is also to be estimated, and Sj isa matrix of constants representing the basis functions. The term on the left is just theusual weighted sum of squares. The term on the right imposes a penalty that increasesas the functions to be estimated become more complex. It is the second term that makesestimation different from iteratively reweighted least squares applied to the generalizedlinear model.

11

To summarize, separately for each crime type c we are seeking the val-ues of αc,0 through αc,2, and the four functions of crime in the comparisondivisions that iteratively minimize a penalized sum of squares. A formaldiscussion penalized regression splines can be found in Green and Silverman(1994), in Wood (2006), and in Hastie and his colleagues (2009: Section5.4). The GAM implementation we used is discussed in Wood (2008). Berk(2008: Chapter 2) provides a very accessible introduction to penalized fittingfunctions.

5 Results

The analysis was undertaken in several steps. These steps are presented insome detail for the nuisance crime outcome to help make our methods moretransparent. We proceed much more rapidly through the results for violentcrimes and property crimes.

Figure 2 shows the three crime time series for the Central Division. Avertical line for the start of each intervention is overlaid. It is clear that allthree time series varied a lot over the study period. Property crime (in green)ranged from a low of 21 crimes in a week to a high 181 crimes in a week. Thereis a very large drop in all three crimes about the time when the (pilot) MainStreet Program began and a much smaller downward shift about the timethe SCI was introduced. The three time series for the Central Division havebroadly similar temporal patterns with between series correlations rangingfrom 0.50 to 0.59.

One might be tempted to conclude that the Main Street Project andthe SCI had important effects on all the three time crime types. In otherwork (Berk and MacDonald, 2009), however, we have found some patternscitywide that are also evident in the four comparison divisions. For example,Figure 3 shows the comparable plot for the Hollenbeck Division. Note thatthe large increase around week 200 and the large decrease around week 300corresponds substantially to the dramatic rise and subsequent fall in Figure 2.

5.1 Results for Nuisance Crime

Figure 4 plots the nuisance crime time series by week and overlays a lowesssmoother. There is again a dramatic increase several months before any ofthe interventions are in place followed by a significant drop in crime associ-

12

Crime Multiple Times Series For The Central Division

Week

Crim

e C

ount

0 100 200 300 400

50100

150 Safer Cities

Main Street Project

Figure 2: Crime Counts for The Central Division: Green = Property Crime,Black = Nuisance Crime, Red = Violent Crime

Crime Multiple Times Series For The Hollenbeck Division

Week

Crim

e C

ount

0 100 200 300 400

050

100

150

200

250

300

Safer Cities

Main Street Project

Figure 3: Crime Counts for the Hollenbeck Division: Green = PropertyCrime, Black = Nuisance Crime, Red = Violent Crime

13

ated with the Main Street Project and and perhaps a small drop in crimeassociated with the SCI. However, as already suggested, much of the varia-tion over time is shared with the four adjacent police divisions and even thecity overall. Were that variation removed, the patterns might differ substan-tially. For the requisite covariance adjustments, we turn to the generalizedadditive model. The nuisance crime counts by week for each of the four com-parison divisions serve as the covariates. The intervention variables are notyet included.

0 100 200 300 400

2040

6080

100

Central District Nuisance Crime Over Time

Week

Cen

tral D

istri

ct N

uisa

nce

Crim

e C

ount

Main Street Project

Safer Cities

Figure 4: Nuisance Crime Counts for Central Division With No CovariacneAdjustments And An Overlaid Lowess Smoother

Figure 5 shows for didactic purposes the residualized nuisance crime pro-duced by the model. Once again, a lowess smoother is overlaid. The dif-ference between Figure 5 and Figure 4 is striking. For example, the largeshift upwards around week 260 and the large shift downwards around week300 are both gone. That pattern was shared with other police divisions inthe Central Bureau and has been removed. Visually at least, the covarianceadjustments appear to be effective, and there are suggestions of relativelysmall intervention effects.

Yet, there is also some evidence for a gradual downward trend uniqueto the Central Division. Were that not taken into account, the average of

14

0 100 200 300 400

-20

020

4060

Residualized Central District Nuisance Crime Over Time

Week

Cen

tral D

istri

ct N

uisa

nce

Crim

e C

ount

Main Street Project

Safer Cities

Figure 5: Nuisance Crime Counts for Central Division After Covariance Ad-justments With An Overlaid Lowess Smoother

0 100 200 300 400

-20

020

4060

Residualized Central District Nuisance Crime Over Time

Week

Cen

tral D

istri

ct N

uisa

nce

Crim

e C

ount

Main Street Project

Safer Cities

Figure 6: Nuisance Crime Counts for Central Division After Covariance Ad-justments Include Week With An Overlaid Lowess Smoother

15

the number of nuisance crimes before the two interventions would be higherthan the average number of crimes after the two interventions because of thetrend, not because of the intervention. Prudence dictates addressing thattime trend, which apparently is unique to the Central Division. Again fordidactic purposes, Figure 6 is the result when a counter for week is includedas an additional covariate whose functional relationship with the responseis, as before, determined empirically. It seems now that the time series isvery well behaved. Large scale trends, even highly nonlinear trends, thatthe experimental Central Division shares with the four adjacent comparisondivisions have been removed. So has the overall downward trend that isunique to the Central Division. We are now is a good position to considerany treatment effects characterized by change in level.

We do this by applying the generalized additive model much as in equa-tion 1. The residualizing process just described goes on the behind the scenesso that covariance adjusted estimates of any treatment effects can be prop-erly obtained. The model includes two treatment indictor variables, and fivecovariates, one for the weekly totals for nuisance crimes in each of the fourcontrol police divisions and one that is a counter for week. The fit is con-structed from the contributions of all seven regressors, only two of which areintended to capture any intervention effects. Overall, the fit is quite good,at least to the eye, and 64% of the deviance is account for by the model.Figure 7 is a plot of the time series of the number of nuisance crimes in theCentral Division with the GAM fitted values overlaid.

Table 1 contains the key parameter estimates. As noted earlier, a pe-nalized regression spline smoother (represented by an “s” in the table) wasused for each covariate as part of the fitting process. Conditional Poissondisturbances were assumed along with the canonical log-link function.

Consider first the five covariates. They are of little substantive interest.Their role is to remove crime trends that the Central Division shares withany of the four adjacent divisions and overall time trends that are uniqueto the Central Division. Moreover, their relationships with the response arevery difficult to interpret because they are the result of several covarianceadjustments. How would one think about, for example, the relationship be-tween the number of nuisance crimes in the Central Division and the numberof nuisance crimes in the Hollenbeck Division with the crime trends in theother three comparison divisions held constant?

Rather, the issues are methodological. All but one have effective degrees

16

0 100 200 300 400

2040

6080

100

Central District Nuisance Crime Analysis

Week

Cen

tral D

istri

ct N

uisa

nce

Crim

e C

ount

Main Street Project

Safer Cities

Figure 7: Nuisance Crime Counts for Central Division With Fitted ValuesOverlaid

Variable Estimate EDF P-ValueConstant 3.74 1.0 >.001

Main Street Pilot -0.27 1.0 .002Safer Cities -0.36 1.0 .004

s(Week) — 6.5 >.001s(Hollenbeck) — 4.0 0.159s(NorthEast) — 2.2 .002

s(Newton) — 1.0 0.089s(Rampart) — 7.9 .014

Table 1: GAM Resuls for Nuisance Crime in the Central Division: Log linkand Poisson Disturbances (N = 417, 64% of the Deviance Accounted For)

17

of freedom (EDF) greater than 1.0.6 This indicates that all of the partialresponse functions except for the Newton Division are nonlinear, some dra-matically so.7 Had a conventional parametric form of regression been used,such as Poisson regression, the assumed functions for the covariates wouldlikely be very wrong. Incomplete covariance adjustments would have followedalong with badly biased estimates of any treatment effects.

As an example, consider the partial response function for the number ofnuisance crimes in the Rampart Division, which is shown Figure 8. The solidline is the estimated function, and the dotted lines are 95% error bands.8

There is a rug plot at the base of the the figure to show roughly how theregressor is distributed. The vertical axis metric is that of the fitted values incentered log units. The label indicates that a smoother has been applied tothe Rampart covariate and that the function uses up 7.9 degrees of freedom.

Taking the error bands in account, Figure 8 shows that for the RampartDivision the partial response function is approximately flat until about 60nuisance crimes per week. Between 60 and nearly 100 crimes per week, thefunction is increasing. There are so few weeks with more than 100 nuisancecrimes that, as the very wide error band indicates, the apparent declineshould not be taken seriously.

Figure 9 shows the partial response function for week, which captures anyoverall time trends in the Central Division with temporal crime patterns inthe comparison divisions held constant. The trend is down or flat, exceptfor the period between about week 250 and week 320. The sharp downwardtrend beginning around the time that the pilot study was fully in place mayactually represent a treatment effect. An upward trend in nuisance crimeswas reversed. However, our model assumes that any treatment effects arechanges in level, not slope, and we hesitate to respecify after looking at theresults. That would be data snooping (Freedman, 2005: Section 4.9). Shouldthe change in slope be a real treatment effect, the likely consequence for our

6Effective degrees of freedom (EDF) is a generalization of the usual degrees of freedomconcept. Loosely speaking, the EDF indicates how many degrees of freedom are being“used up” by the estimated function. The EDF does not have to be a whole number.

7When the EDF is 1.0, the functional form is linear.8The error bands are constructed point by point. They are an attempted to represent

approximately the interval in which 95% of the fitted values would fall were it possible torepeat the study, independent of all other studies, a very large number of times. It is nota 95% confidence interval because there is likely to be some unknown amount of bias inthe fitted values.

18

20 40 60 80 100

-0.4

-0.2

0.0

0.2

0.4

Partial Response Plot for Rampart Division

Number of Nuisance Crimes

s(Rampart,7.9)

Figure 8: Partial Response Function for Nuisance Crimes in the RampartDivision

0 100 200 300 400

-0.4

-0.2

0.0

0.2

0.4

Partial Response Plot for Week

Week

s(Week,6.51)

Figure 9: Partial Response Function for Week

19

model is to underestimate any beneficial effects for the two interventions.The top three rows of Table 1 contain results for the constant and the

two treatment indicators. The parameter estimates for the interventions canbe interpreted just as one would for Poisson regression within the generalizedlinear model. The Main Street Project regression coefficient of -.27 translatesinto a count multiplier of .76. With the introduction of the pilot program, thenumber of nuisance crimes in the Central Division is on the average about75% of what it had been before the pilot program was introduced. Theassociated p-value is .002. The SCI regression coefficient of -.36 translatesinto a count multiplier of .70. With the introduction of SCI, the numberof nuisance crimes is on the average about 70% of what it had been beforeeither of the interventions had been introduced. The associated p-value is.004.

Recall that the starting date for the Main Street Project was difficult toprecisely define and perhaps could have been moved several weeks in eitherdirection. Not surprisingly, the estimated effects of both programs dependon the starting date for Main Street Project. As the starting date is movedtoward the starting date for the Safer Cities Initiative (SCI), the variance ofthe Main Street Project indicator is necessarily reduced. A reduction in aregressor’s variance can introduce greater instability into any treatment effectestimates and reduces statistical power. Moreover, because in this case allof the regressors are correlated, changing one estimate affects all estimates.Suffice it to say, moving the starting date of the Main Street Project a week ortwo in either direction makes no important difference. This is good becauseour chosen start date is probably accurate within plus or minus a week ortwo. But moving the starting date several weeks forward or back can changethe story. Indeed, moving the starting date for the Main Street Project sixweeks forward in time can essentially remove any nuisance crime treatmenteffects for either intervention. But in so doing one would have to assume thatthe Main Street Project was sprung on the Central Division with no advancewarning whatsoever and without any prior media coverage and related publiccontroversy. Facts to the contrary are readily available.

Interpreting the treatment effect p-values also requires some caution. Asnoted earlier, statistical inference for GAM can be difficult to justify becausethe model is arrived at inductively (Leeb and Potscher, 2005; 2006; 2008;Berk et al., 2009).9 However, the p-values for the two interventions are so

9There is also a bit of dependence in the disturbances. The estimate correlation at a

20

small that even if they biased downward by a factor of 10, one would stillreject the conventional null hypothesis.10 For these, the conclusions from thestatistical tests are perhaps sound. Moreover, one gets the effectively thesame results assuming a normal generalized additive model.11 The analysesto which we now turn provide futher support for our procedures.

5.2 Results for Violent Crime

Having gone through the analysis of nuisance crimes in considerable depth,we can move very quickly through the analyses for violent crime and propertycrime. Applying the same steps and same analysis procedures as used fornuisance crimes, Table 2 shows the results for violent crimes. All of thepreliminary and intermediate results anticipating Table 2 broadly replicatethe preliminary and intermediate steps for nuisance crime. And as before,the story is to be found in the regression coefficients for the Main StreetProject and the SCI.

Variable Estimate DF or EOF P-ValueConstant 3.70 1.0 >.001

Main Street Pilot -0.17 1.0 .023Safer Cities -0.50 1.0 >.001

s(Week) — 5.2 >.001s(Hollenbeck) — 5.4 >.001s(NorthEast) — 6.9 >.001

s(Newton) — 1.7 0.102s(Rampart) — 2.6 0.227

Table 2: GAM Results for Violent Crime in the Central Division: Log linkand Poisson Disturbances (N = 417, 68% of the Deviance Accounted For)

The average treatment effect estimates of roughly the same as found fornuisance crime. The Main Street Pilot regression coefficients of -.17 translates

lag of one week is .06. This is small by most any standard and will not have much impacton statistical tests.

10For the two interventions, a one tailed test is used.11For large counts as we have here, a conditional Poisson distribution and a conditional

normal distribution are much the same.

21

into a count multiplier of .84. With the introduction of the pilot program, thenumber of violent crimes in the Central Division is on the average about 84%of what it had been before the pilot program was introduced. The associatedp-value is .02. The Safer Cities Initiative (SCI) regression coefficients of -.50translates into a count multiplier of of .61. With the introduction of SCI,the number of violent crimes is on the average about 61% of what it hadbeen before either of the interventions had been introduced. The associatedp-value is substantially less than .001. However, the same cautions noted fornuisance crimes still apply.12

5.3 Results for Property Crime

Table 3 shows the results for property crime. Again, all of the preliminaryand intermediate results anticipating Table 3 broadly replicate the prelimi-nary and intermediate steps for property crimes. The Main Street Projectregression coefficients of -.25. The count multiplier is .78. With the intro-duction of the pilot program, the number of property crimes in the CentralDivision is on the average about 78% of what it had been before the pilotprogram was introduced. The associated p-value is .005. The SCI regressioncoefficients is -.43, which translates into a count multiplier of .65. With theintroduction of SCI, the number of property crimes is on the average about65% of what it had been before either of the interventions had been intro-duced. The associated p-value is less than .001. In short, for all three kindsof crime the estimated treatment effects are rather similar.

5.4 Spillover Spatial Effects

Finding treatment effects in the Central Division leaves unaddressed anyimpact the Main Street Project and the SCI may have had on the fouradjacent police divisions. There are three clear possibilities. First, theremay have been no impact whatsoever outside of the Central Division.

Second, there may have been crime displacement manifested by crime in-creases in the four adjacent police divisions about the same time the LAPDintervened in the Central Division. Before looking at the data, we weresomewhat skeptical of the displacement hypothesis because of our view that

12The estimate correlation between the residuals a week apart was less than .10 and farto small to be of any real concern.

22

Variable Estimate DF or EOF P-ValueConstant 4.57 1.0 >.001

Main Street Pilot -0.25 1.0 .005Safer Cities -0.43 1.0 .001

s(Week) — 8.4 > .001s(Hollenbeck) — 1.0 .222s(NorthEast) — 8.0 >.001

s(Newton) — 6.8 >.01s(Rampart) — 6.1 .319

Table 3: GAM Results for Property Crime in the Central Division: Log linkand Poisson Disturbances (N = 417, 81% of the Deviance Accounted For)

crime associated with homelessness depended substantially on homeless den-sity, and there were apparently no new, large homeless encampments in thecomparison divisions after the interventions. There was certainly nothing onthe scale of Skid Row.

Third, the interventions in the Central Division may have had spillovereffects manifested by crime reductions in the four adjacent police divisionsabout the same time as the LAPD intervened in the Central Division. Al-though there were no large scale encampments on the adjacent divisions,there were areas where homeless individuals tended to congregate. Perhapsthese would areas would be affected. Before looking at the data, spillovereffects seemed like a real possibility because word of the police crackdownwould spread quickly among the homeless and related stakeholders. In addi-tion, police officers in the comparison divisions might have been empoweredand motivated to aggressively intervene with their own homeless populations.

To address these three possibilities, we proceeded much as before. Theresponse variables were the three kind of crime counts in each the four com-parison divisions. The same two intervention variables were defined. Weekand the corresponding crime count in the Central Division were the covari-ates. We essentially reversed the logic of the previous analyses. A total oftwelve analyses followed (three kinds of crime by four comparison divisions).

In general, the same sort of intervention effects were found. There was ab-solutely no evidence for crime displacement. There was evidence for spillovereffects. As a summary, Table 4 uses the sum of the three kinds of crime in

23

Variable Estimate DF or EOF P-ValueConstant 6.91 1.0 >.001

Main Street Pilot -0.26 1.0 >.001Safer Cities -0.23 1.0 >.001

s(Week) — 8.62 > .001s(Central Total Crime) — 8.51 > .001

Table 4: GAM Results For Total Crime For The Comparison Divisions UsingTotal Crime in the Central Division and Week as Regressors: Log link andPoisson Disturbances (N = 417, 87% of the Deviance Accounted For)

the four comparison divisions as the response variable. The sum of the threekinds of crime for the Central Division and week are to two regressors. Thetwo treatment effect estimates imply multipliers of about .75, about what wefound previously.

There were no Main Street Project of SCI interventions in any otherdivision but the Central. According to the LAPD, there were also no otherlaw enforcement interventions in those divisions that could be confoundedwith what was going on in the Central. Nor were there apparently any othersorts of interventions in the four comparison divisions starting around week300. It is important to stress that for the crime reductions in the comparisondivisions to be other than indirect spillover effects from the Central Division,the timing would have to be very similar. There would need to be a sharpbeginning to any confounded intervention around week 300.

6 Discussion

We have shown that broad time trends in certain crimes that the CentralDivision shares with its immediate neighbors can be effectively removed. It isthen possible in principle to isolate processes unique to the Central Division.This is at least a good start. It follows that treatment effect estimates areconsistent with meaningful but modest crime reductions for both the MainStreet Project and the SCI for nuisance, violent, and property crimes.

Consider first the overall effects in the Central Division. Everything de-pends on the quality of the covariance adjustments that GAM implements.Highly nonlinear relationships with the response variable can be effectively

24

taken into account even when they are unknown before the data are exam-ined. Therefore, the key vulnerability is that there are one or more con-founders having highly local temporal impacts on the amount of crime inthe Central Division and that the local effects just happen to materializearound the time when both police initiatives were introduced. For example,had the City of Los Angeles instituted significant new ways to provide forthe homeless at the same time that the police were trying to disperse them,it would be difficult to determine which interventions, if any, were reducingcrime. Under such circumstances, the bias in our treatment effect estimatescould be positive or negative. We might attribute to either of the interven-tions crime reduction effects that were not real or fail to find interventioneffects that were. We know of no such confounders, nor do our informantswithin the LAPD. However, we cannot rule them out categorically. Whatis clear is that crime in the Central Division dropped very soon after bothpolice interventions were introduced.

The apparent spillover effects complicate matters. One interpretationis that the Main Street Project and the SCI implemented in the CentralDivision led to more widespread but unofficial changes in police practices. Itis also plausible that media and word-of-mouth descriptions of events in theCentral Division dispersed collections of homeless individuals in at least theadjacent police divisions. These are true spillover effects that one might callindirect.

Another interpretation has already been anticipated. In all five policedivisions in the Central Bureau or in the comparison divisions alone, one ormore other interventions may have been introduced coinciding in time withthe Main Street Project and the SCI. It is important to emphasize that suchinterventions would need to be be sharply implemented around the sametime as the two known police actions. The impact of any interventions withmore gradual effects over longer time spans would likely be absorbed in thecovariance adjustments. Again, we can find no evidence of interventions thatthe covariance adjustment would not be able to handle.

Some many find the treatment effects unsurprising. The large concentra-tion of homeless individuals in downtown Los Angeles, and especially on SkidRow, contained a substantial pool of potential crime victims who were easytargets. Likewise, there was a substantial pool potential crime perpetrators.The area was also a magnate for individuals with lawless inclinations. Dis-persing the homeless population would seem on its face an obvious way toreduce crime in downtown Los Angeles. As already noted, more important

25

for crime than the number of homeless in a city may be how much they arespatially concentrated. Areas characterized by a high density of homelesspeople may overwhelm local resourses. It is one thing, for example, for ashopkeeper to ask a single homeless person not to camp right in front of theentrance to his or her establishment, and quite another to ask that of severalindividuals at once many times over the course of a working day. In the firstcase, there may be little disruption of customer traffic. In the second case,the disruption could be substantial. Therefore, one can view the two LAPDinterventions primarily as a way to reduce the spatial density homeless indi-viduals, and one could have walked through Skid Row soon after the policeintervened and plainly seen that the homeless encampment was no more.

Unaddressed, however, is whether it is good public policy more generallyto have police break up concentrations of homeless individuals living on thestreets. In our view, the wisdom of such police action depends on whatservices are made available for the homeless. There might be little objectionto aggressive policing in the short term, for example, if there were a sufficientnumber of adequate shelters where homeless individuals might safely eat andsleep. Getting the homeless off the streets and into shelters is probably asensible stop-gap approach. In practice, however, there will often be toofew shelter beds. Then, the appropriate public policy will require difficulttradeoffs. There is also the matter of costs. Police resources reallocatedto Skid Row are not available elsewhere. In addition, there are secondarycosts that come from processing homeless individuals who are arrested andhomeless individuals who are incarcerated. Far more is involved than thewages of patrol officers reassigned to Skid Row.

7 Conclusions

On balance, there is evidence in the Central Division for modest but mean-ingful reductions across a wide range of crimes that can attributed to theMain Street Project and the SCI. The analysis does not have the “gold stan-dard” internal validity of a randomized experiment, but our non-equivalentno-treatment control group time series design has real merit, and we canfind no alternative explanations for our main findings. It also helps thatthere is unequivocal evidence that the Skid Row homeless encampment wascompletely cleared. A key intervening requirement for crime reduction wasdemonstrable fulfilled.

26

The evidence for beneficial spillover effects in the comparison divisions isless compelling. The estimated effects are roughly comparable, but in con-trast to the interventions in the Central Division, we have no data whatsoeveron what the police may have done in the comparison divisions. We can onlyspeculate on what may have happened. Although our post hoc account hassome plausibility, it is no substitute for real measures of police practices inthe comparison divisions and of responses of the homeless across the entireCentral Bureau to interventions in Central Division.

At the same time, the usefulness of our study depends on the knowledgebase from which the Main Street Project and the SCI were implemented.The scientific gold standard is not the only yardstick. Another yardstick isthe evidence about any treatment effects that would otherwise be available.By that measure, our finding are certainly not definitive, but still a stepforward.

Still, there are some strong caveats to keep in mind. The modest crimereductions can imply that much of the crime, perhaps even most crime, in theCentral Division was not driven by the Skid Row homeless encampment. Ifthe goal is crime reduction overall, dispersing the homeless is at best a start.In addition, whether the findings can be usefully generalized to other citieswith their own kinds of homeless problems is an empirical question. Forexample, the interventions in the Central Division were premised on largehomeless encampments thought to be a key factor in local crime. Finally,there is the matter of cost-effectiveness. Even if one only considers the costsof the intervention, the arrests, the prosecutions, and the incarcerations, it isnot apparent that clearing the Skid Row homeless encampment was a gooduse of police resources. It may have been, but that case remains to be madein real economic terms.

Crime reduction is no doubt an important policy goal. The high victim-ization rates for homeless individuals implies that the Main Street Projectand the SCI probably had some direct benefits for the people formerly livingon Skid Row streets. There were also the benefits for local shopkeepers, theircustomers, and residents in dwellings near by. We emphasize again, however,that police interventions of the sort undertaken by the Main Street Projectand the Safer Cities Initiative do not solve the problem of homelessness. Atbest, they can only address one of its possible manifestations.

27

References

Berk, Richard A. (2008) Statistical Learning from a Regression Perspective.New York: Springer.

Berk, Richard A., Brown, Lawrence., and L. Zhao. (2009). “Statistical In-ference After Model Selection.” Working Paper, Department of Statis-tics, University of Pennsylvania (under review).

Berk, Richard A., Kriegler, Brian, and D. Ylvisaker. (2008). “Counting theHomeless in Los Angeles County.” In Probability and Statistics: Essaysin Honor of David A. Freedman, Monograph Series for the Institute ofMathematical Statistics, D. Nolan and S. Speed (eds.),

Berk, Richard A., and John M. MacDonald. (2008). “Overdispersion andPoisson Regression.” Journal of Quantitative Criminology 24(3): 269-285.

Berk,Rrichard. A. and John M. MacDonald. (2009). “The Dynamics ofCrime Regimes.” Criminology, forthcoming.

Blasi, Gary. (2007). Policing our way out of homelessness?: The first yearof the Safer Cities Initiative on Skid Row. Los Angeles: The UCLASchool of Law Fact Investigation Clinic.

Braga, Anthony A. (2001). ”The Effects of Hot Spots Policing on Crime.”Annals of the American Academy 578: 104-125.

Brantingham, Patricia L. and Paul J. Brantingham. (1999). ”A TheoreticalModel of Crime Hot Spot Generation.” Studies on Crime and CrimePrevention 8: 7-26.

Bratton, William (with Peter Knobler). (1998). Turnaround:How AmericasTop Cop Reversed the Crime Epidemic. New York: Random House.

Cohen, Lawrence and Marcus Felson. (1979). ”Social Change and CrimeRate Trends: A Routine Activity Approach.” American SociologicalReview 44: 588-608.

Eck, John E., and William Spelman. (1987). Problem solving: Problemoriented policing in Newport News. Washington, DC: Police ExecutiveResearch Forum.

28

Felson, Marcus. (2002). Crime and Everyday Life. 3rd Edition. ThousandOaks, CA: Pine Forge Press.

Goldstein, Herman. (1990). Problem-oriented Policing. New York: McGraw-Hill.

Green, P.J. and B.W. Silverman (1994) Nonparametric regression and Gen-eralized Linear Models. New York: Chapman & Hall.

Harcourt, Bernard E. (2005). Policing L.A.s Skid Row: Crime and RealEstate Development in Downtown Los Angeles (An Experiment in RealTime). University of Chicago Legal Forum: 2005.

Hastie, T.J., and R.J. Tibshirani (1990) Generalized Additive Models NewYork: Chapman and Hall.

Hastie, T., Tibshirani, R. and J. Friedman (2009) The Elements of Statis-tical Learning, Second Edition. New York: Springer.

Koegel, Paul, Audrey Burnam, and Rodger K. Farr. (1988). ”The Preva-lence of Specific Psychiatric Disorders Among Homeless Individualsin the Inner-city of Los Angeles.” Archives of General Psychiatry 45:1085-1092.

Kushel, Margot B., Jennifer L. Evans, Sharon Perry, Marjorie J. Robert-son, and Andrew R. Moss. (2003). ”No Door to Lock: VictimizationAmong Homeless and Marginally Housed Persons.” Arch Intern Med163(20):2492-2499.

Leeb, H., B.M. Potscher. (2005). “Model Selection and Inference: Factsand Fiction,” Econometric Theory 21: 21–59.

Leeb, H., B.M. Potscher. (2006). “Can one Estimate the ConditionalDistri-bution of Post-Model-Selection Estimators?” The Annals of Statistics34(5): 2554–2591.

Leeb, H., B.M. Potscher. (2008). “Model Selection,” in T.G. Anderson,R.A. Davis, J.-P. Kreib, and T. Mikosch (eds.), The Handbook of Fi-nancial Time Series, New York, Springer: 785–821.

Lopez, Steven. (2005). ”Demons Are Winning on Skid Row.” Los AngelesTimes, October 16.

29

Los Angeles Times. (2009). ”Full Coverage: Homeless in America.” avail-able at: www.latimes.com/la-homeless-pkg,0,5295266.special

Los Angeles Almanac (2009) www.laalmanac.com/social/so14.htm

Los Angeles Police Department. (2008a). Presentation at the ManhattanInstitute/Milken Institute Policing Skid Row Conference. Los Angeles,CA, January 17. available at: ww.lapdonline.org (3735)

Los Angeles Police Department. (2008b). News Release (38591). availableat: www.lapdonline.org

Mangano, P.F., and G. Blasi. (2007). “Stuck on Skid Row.” Los AngelesTimes. October 29, 2007, Opinion Section (articles.latimes.com/2007/oct/29/news/oe-mangano29)

McMorris, Emily N. 2006. Jones V. City of Los Angeles: A DangerousExpansion of Eighth Amendment Protections Stifles Efforts to Cleanup Skid Row. Loyola of Los Angeles Law Review, 40,1149-1168.

Miethe, Terry D. and Robert F. Meier. (1990). ”Opportunity, Choice, andCriminal Victimization: A Test of a Theoretical Model.” Journal ofResearch in Crime and Delinquency 27: 243-266.

National Law Center on Homelessness & Poverty (2009) Homes not Hand-cuffs: The Criminalization of Homelessness in U.S. Cities. Wash-ingtion, D.C., The National Law Center on Homelessness & Poverty.www.nlchp.org

Rossi, P.H. (1989) Down and Out in America: The Origins of Homelessness.Chicago: University of Chicago Press.

Sampson, Robert J. The Community, in James Q. Wilson and Joan Pe-tersilia, eds., Crime, San Francisco, Calif.: Institute for ContemporaryStudies, 1995, pp. 193-216.

Sampson, Robert J., and Janet L. Lauritsen. (1990). ”Deviant Lifestyles,Proximity to Crime, and the Offender-Victim Link in Personal Vio-lence.” Journal of Research in Crime and Delinquency 27(2): 110-139.

30

Shadish, W., Cook, D.D., and D.T. Campbell (2002) Experimental andQuasi-Experimental Designs for Generalized Causal Inference. NewYork: Houghton Mifflin.

Sherman, Lawrence W., Patrick R. Gartin, and Michael E. Buerger. (1989).”Hot Spots of Predatory Crime: Routine Activities and the Criminol-ogy of Place.” Criminology 27: 27-55.

Sherman, Lawrence W. (1990). ”Police Crackdowns: Initial and ResidualDeterrence.” In Crime and justice: A Review of Research. Vol. 12,edited by Michael Tonry and Norval Morris. Chicago: University ofChicago Press.

Skogan, Wesley G. (1990). Disorder and Decline: Crime and the Spiral ofDecay in American Neighborhoods. Berkeley, CA: University of Cali-fornia Press.

Taylor, Ralph B. (2001). Breaking Away from Broken Windows: BaltimoreNeighborhoods and the Nationwide Fight Against Crime, Grime, Fear,and Decline. Boulder, Colo.: Westview Press.

Wakefield, Joan. (2009). Personal communications with John MacDonald.

Weisburd, David and John Eck. (2004). ”What Can Police Do to Re-duce Crime, Disorder and Fear?” Annals of the American Academy ofPolitical and Social Science 593: 42-65.

Weisburd, David, Laura A. Wyckoff, Justin Ready, John E. Eck, Joshua C.Hinkle, and Frank Gajewski. (2006). ”Does crime just move aroundthe corner? A Controlled Study of Spatial Displacement and Diffusionof Crime Control Benefits.” Criminology 44: 549-592.

Weisburd, David, Gerben J.N. Bruinsma, and Wim Bernasco. (2009).”Units of Analysis in Geographic Criminology: Historical Development,Critical Issues, and Open Questions.” (pp. 3-31). In D. Weisburd etal. (eds.), /textitPutting Crime in its Place. New York: Springer.

Wenzel, Susan L., Paul Koegel, and Lillian Gelberg. (2000). ”Antecedentsof Physical and Sexual Victimization Among Homeless Women: AComparison to Homeless men.” American Journal of Community Psy-chology 28: 367-390.

31

Welsh, Brandon C., and Akemi Hoshi. (2002). Communities and Crime Pre-vention (pp. 165-197). In Lawrence W. Sherman, David P. Farrington,Brandon C. Welsh, and D. Layton MacKenzie, eds., Evidence-BasedCrime Prevention, New York: Routledge.

Wilson, James Q., and George L. Kelling (1982). Broken Windows: ThePolice and Neighborhood Safety, Atlantic Monthly, March, 29-38.

Wood, Simon N. (2006) Generealized Additive Models: An Introduction withR. New York: Chapman & Hall.

Wood, Simon N. (2008) “Fast Stable Direct Fitting and Smoothness Selec-tion for Generalized Additive Models.” Journal of the Royal StatisticalSociety, Series B, 70(3):495-518.

32