An Exploratory Multilevel Analysis of Pedestrian Frisks in Philadelphia

Abstract

The City of Philadelphia has faced significant litigation related to racial and ethnic disparities in stop-and-frisk practices. The Philadelphia Police Department has made much of its stop-and-frisk data publicly available in the name of transparency and to facilitate independent investigation (the data describe over 350,000 pedestrian stops with over 45,000 pedestrian frisks for 2014–2015). The current analysis made use of this public data set to explore whether the individual-level relationship between Black racial classification and being subjected to a frisk can be explained by associated neighborhood-level factors such as the violent crime rate. Additionally, the present analysis examined whether variation in the violent crime rate is similarly related to the likelihood of being frisked in predominantly Black versus non-Black areas and whether area racial composition affects the likelihood that an officer’s decision to frisk will be supported with uncovered contraband. The results were consistent with theories of neighborhood racial stigma. In particular, the violent crime rate was a significantly weaker predictor of being frisked in Black areas, and, net of a variety of factors at the individual and neighborhood levels, Black citizens and Black places experienced a disproportionate amount of frisks where no contraband was found or arrest made.

Keywords

race and policing criminal profiling police legitimacy racial profiling use of force criminalblackman criminological theories minority group threat neighborhood stigma spatial analysis

The City of Philadelphia recently settled a class-action lawsuit alleging illegal targeting of Black and Latino citizens in the city’s stop-and-frisk policy.¹ The settlement (2011) involved a consent decree where the police department would continuously collect data that would be analyzed to ensure that (1) investigative stops were based on articulable and reasonable suspicions that crime is afoot and (2) frisks were based on articulable and reasonable concerns about an armed and dangerous suspect. The hope was that monitoring the general lawfulness of stop-and-frisk practices would significantly reduce impermissible biases associated with racial profiling.

Two of the lead plaintiffs in the case, Mahari Bailey and Jewell Williams, have publicly relayed the personal stories that drove them to action. In an article in the New York Times (Goode, 2012), Bailey recounted how police officers repeatedly stopped him:

“It just became too much,” Mr. Bailey said. In one instance, Mr. Bailey said, he was standing with friends outside a house in West Philadelphia when an unmarked car screeched up and two men in plain clothes jumped out, guns drawn, and told them to put up their hands. “We thought we were being robbed,” he said. Two other cars arrived. Mr. Bailey was handcuffed and placed spread-eagle against a police car. When Mr. Bailey said he was a lawyer and asked why he had been stopped, he got no answer, he said. But the officers threatened to call his employer “and say I was hanging with drug dealers.”

Pennsylvania state legislator and retired Temple University police officer Jewell Williams was driving home through North Philadelphia when he came upon the scene that led to his involvement in the class-action lawsuit. Williams described the circumstances in an interview published in Newsweek (Saul, 2016):

[Williams] noticed that three police officers had stopped a car with two elderly black men inside. The officers frisked the driver, and when a cop set his cash down on the trunk of the car, the loose paper bills fluttered away in the wind and were scooped up by patrons from the nearby Red Top bar. “So I got out of the car, and I was yelling to the people, ‘Yo, leave that man’s money alone!’”…The cops got angry. They swore at and handcuffed the driver and passenger, a city employee and a retired tailor, and put them both in a patrol car. Williams, standing next to his black Chrysler with legislative tags, pulled out his state representative ID and even pointed to his home, which was on the other side of an overgrown lot. “Get back in your fucking car before I give you a bunch of tickets,” one cop told him, according to court papers.

Stories such as these have increasingly found their way into the public’s view and many U.S. cities have moved to reform the way they do policing across a variety of dimensions.

The movement to reform stop-and-frisk practices has been especially prominent, but it has also drawn considerable resistance. Critics of efforts to curb the use of stop and frisk, including President Trump (Ehrenfreund, 2016), have worried that overregulating the practice will ultimately lead to rising violent crime rates, especially in Black and Latino neighborhoods. A common refrain among law enforcement officials responding to accusations of racial profiling in stop-and-frisk cases is that geography, not race, is driving the racial disparities in raw statistics. For example, in defending New York’s now significantly curtailed stop-and-frisk practices former Police Commissioner Raymond Kelly argued, “We went where the crime was, whatever color the perpetrators turned out to be” (Lipsky, 2015). Social commentators wrestling with the limits of what constitutes “racial profiling” have expressed similar sentiments. For instance, referring to former Philadelphia Mayor Michael Nutter’s call to use stop-and-frisk practices in “targeted enforcement areas,” one columnist noted that “Since ‘high crime’ and ‘high violence’ areas are in low-income and minority neighborhoods, cries of racism were inevitable…I see it as targeting a crime zone rather than a skin tone, but others disagree” (Bykofsky, 2015).

Race scholars might add to the discussion by noting that even if the observed racial disparities in raw statistics were entirely explained by geographic profiling, that would not necessarily mean the absence of racial bias, as individual officers might stigmatize entire neighborhoods based on racial predominance rather than react to the actual frequency of crime and disorder (Meehan & Ponder, 2002; Sampson & Raudenbush, 2004). Fundamentally, those arguing that variation in neighborhood dangerousness explains racial disparities in police scrutiny should seek evidence that (1) the individual-level relationship between race and police intrusiveness disappears when holding constant the level of neighborhood crime and violence and (2) the prevalence of neighborhood crime and violence similarly determines how police officers treat individuals in predominantly White and predominantly non-White areas.

The current analysis explored these issues using publicly available stop-and-frisk data for the City of Philadelphia in 2014 and 2015. Results from multilevel logistic regression analyses indicated that while violent crime at the census tract level is significantly positively related to the likelihood that someone stopped will be frisked for weapons, tract violent crime does not significantly mediate the individual-level relationship between race and the odds of being frisked, especially being frisked unproductively (no found contraband or arrest made at the conclusion of the investigatory stop). Additionally, closer inspection of the data revealed that the overall significant relationship between an area’s violent crime and the likelihood of being frisked was not present for Philadelphia’s many tracts that were almost entirely Black. Social psychological perspectives on racialized neighborhood stigma may help explain why violent crime appears to be significantly less of a predictor of variation in police defensive behavior in Black versus non-Black areas. Moreover, these same perspectives may help elucidate why frisks are associated with exceptionally low contraband recovery rates in predominantly Black places.

Racial Composition, Neighborhood Stigma, and the Police

Building on a research literature demonstrating that the perceived racial composition of an area is strongly associated with fear of criminal victimization (Bursik & Grasmick, 1993; Chiricos, Hogan, & Gertz, 1997), Quillian and Pager (2001) examine the influence of racial composition on residents’ subjective assessment of neighborhood crime, controlling for multiple objective indicators of local crime incidence. Using multilevel data for three large U.S. cities, Quillian and Pager (2001) find that the percentage Black in an area matters much more than actual crime rates in driving the general perception of neighborhood crime. They further argue that this has implications for residential segregation patterns. While Quillian and Pager (2001) focus on how stereotypes of criminal propensity at the neighborhood level may explain continued hypersegregation in the United States, especially between Black and non-Black residents, they note that it is important to consider how such racialized perceptions can supersede objective indicators for other social dynamics.

Following Quillian and Pager’s (2001) call for more research, Sampson and Raudenbush (2004) conducted a multilevel analysis of the degree to which objectively assessed cues of neighborhood disorder (counts of the presence of graffiti, garbage, vacant houses, etc.) influence survey respondent perceptions of neighborhood disorderliness relative to factors associated with group stereotypes. Similar to Quillian and Pager’s study (2001), the results from Sampson and Raudenbush’s (2004) analysis of Chicago data indicated that the racial composition of a neighborhood mattered more than objective signs of disarray in predicting perceptions of broken windows disorderliness. Sampson and Raudenbush (2004, p. 320) argued that “dark skin is an easily observable trait that has become a statistical marker in American society, one imbued with meanings about crime and disorder that stigmatize not only people but also the places in which they are concentrated” [emphasis added]. Thus, one of the key contributions of Sampson and Raudenbush’s (2004) multilevel research was how it highlighted the notion of “neighborhood stigma” as a sui generis phenomenon (Besbris, Faber, Rich, & Sharkey, 2015; Wacquant, 2008).

Citing Werthman and Piliavin’s (1967) classic work on policing and their formulation of the concept of “ecological contamination,” Sampson and Raudenbush (2004, p. 321) noted that police officers categorize the areas they patrol with racialized distinctions between regular and “bad” neighborhoods, and these neighborhood-level moral judgments supplant considerations of individual attributes and circumstances in police–resident interactions. Similarly, Meehan and Ponder (2002) have argued that such neighborhood-level categorizations are best understood as emanating from general social currents rather than a deviant police subculture. That is, police understandings of place are colored by the same racial stereotypes underlying prejudice outside law enforcement; these racial attitudes are ultimately more important than what is taught and learned in the academy and in more informal occupational settings.

In sum, existing research suggests that, for both the general public and police officers, urban neighborhoods with a predominance of Black residents are more likely to be vilified as crime ridden regardless of the actual crime rate. What is less clear from existing research is whether non-Black urban neighborhoods are stereotypically perceived as safe places where suspicion is unwarranted regardless of the actual crime rate. That is, are non-Black areas as positively stereotyped as Black areas are negatively stigmatized? If not, this suggests a larger role for the actual crime rate in determining fear and suspicion in non-Black urban areas. Moreover, previous research has not assessed whether neighborhood racial composition conditions how much the local crime rate matters for provoking police defensive actions. The present study helps addresses this gap in the literature by examining the interaction of neighborhood racial composition and violent crime on police officers’ decisions to frisk pedestrians in hypersegregated Philadelphia.

Insights From Previous Stop-and-Frisk Studies

Especially coinciding with the release of public data for New York City, there have been dozens of studies examining racial disparities in stop-and-frisk occurrences as well as the degree to which the practice might reduce crime (Abrams, 2014; Ayres, 2002; Ayres & Borowsky, 2008; Geller & Fagan, 2010; Gelman, Fagan, & Kiss, 2007; Gutman, 2016; Harris, 2002; Kubrin, Messner, Deane, McGeever, & Stucky, 2010; Lacoe & Sharkey, 2016; Levchak, 2017; Rosenfeld & Fornango, 2017; White & Fradella, 2016). Although a full review is beyond the scope of this article, most of these studies have concluded that racial bias likely plays an important role, particularly in an officer’s decision to frisk, and that maximizing stop and frisk is not an effective crime reduction strategy. As recently argued by White and Fradella (2016), attempts to maximize stop and frisk for general deterrence purposes are misaligned with the original intention of this policing tactic and such initiatives dramatically increase the likelihood that an officer’s discretionary powers will be abused.

While researchers want to know if the totality of a stop-and-frisk encounter reflects racial bias, examining some aspects of stop-and-frisk incidents face steeper methodological hurdles than others in terms of controlling for spurious influences. In particular, while it is important to examine racial differences in officer decisions to stop certain pedestrians, it is difficult to adjust these statistics to take into account legally relevant racial differences in criminal victimization and offending. While many studies compare the percentage of a particular group in the general population to their proportional representation in stops, this type of external benchmarking is disputed as significantly overestimating bias since it does not account for racial differences in targeted criminal behavior (Walker, 2001). Other studies benchmark race-specific stop rates to race-specific arrest rates, but this is disputed as significantly underestimating bias, as the decision to arrest may reflect an important part of the prejudice that researchers are ultimately trying to measure (Fridell, 2004). Furthermore, officer decisions to stop individuals may reflect different types of cognitive biases than decisions to frisk, with the former sometimes driven by a sense that one is firmly in control and the latter potentially driven by a sense that one is not.

Additional complications arise when considering that some stops are falsely recorded as investigatory detentions when they are better characterized as voluntary “mere encounters” and other stops are so highly structured that they leave little room for investigative discretion (such as when an officer immediately recognizes, stops, and searches a known fugitive as standard procedure for arrest). As Gutman (2016) has argued, racial bias can be masked when improperly recorded stops such as these are included in investigatory stop-and-frisk data.

Due in part to these methodological difficulties, rather than focusing on racial disparities in stops, researchers have increasingly turned their attention to variation in practices and outcomes once a stop has occurred (Carroll & Gonzalez, 2014; Gutman, 2016). Although still susceptible to a variety of methodological problems (such as selection bias), an advantage of this general strategy is that a consensus external benchmark is not required for the statistics to provide meaningful insights. An important component of the current analysis explorers how various individual- and neighborhood-level factors influence the likelihood that a stop will involve a frisk (the term “stop-and-frisk” is more accurately “stop-and-maybe-frisk,” as typically most stops do not involve frisks). The decision to focus on the practice of frisking, the patting down of outer clothing, reflects two defining features of this police behavior: (1) the widely agreed upon lawful purpose of a frisk is to remove a perceived threat to officer or citizen safety associated with a potential weapon in reach of a dangerous individual and (2) unlike searches, which require probable cause and often occur incident to arrest, frisks are more discretionary and require a lower threshold of “reasonable suspicion” that the detainee is armed and dangerous. Additionally, recent research has demonstrated that, beyond just the number of stops a person experiences, the level of police intrusiveness during stops appears to be especially related to decreased police legitimacy (Tyler, Fagan, & Geller, 2014).

After exploring variation in the decision to frisk, the current analysis proceeds to examine the race-specific productivity of frisks in terms of found contraband. One benefit to examining the contraband recovery rate for stops with frisks, as opposed to all stops, is that a stop with frisk is more unquestionably an investigatory detention (Gutman, 2016). The idea underlying this type of analysis is that it can get at whether certain groups are more likely to be stopped and frisked due to stereotypes of dangerousness and criminality, rather than empirically justified cause. A lower contraband hit rate for racial and ethnic minorities than Whites is a red flag for racial bias in the threshold officers use to operationalize reasonable suspicion in frisks (Fridell, 2004). Although the use of hit rate statistics has been criticized as an oversimplified test of racial bias in stop-and-frisk evaluations (Anwar & Fang, 2006), its relative simplicity has made it an increasingly popular measure. Moreover, at various times for various locations, advocates and opponents of stop-and-frisk implementations have both cited versions of contraband recovery statistics to support their position (MacDonald, 2001; New York Civil Liberties Union, 2011).

While recognizing the potential utility of racial disparities in contraband hit rates as an indicator of bias, Ridgeway (2007) argued that raw hit rate statistics could be misleading in that they do not take into account racial differences in the suspected crime motivating the officer’s decision to detain someone. In a project sponsored by the New York City Police Foundation, Ridgeway (2007) analyzed a large data set for New York City. Looking at the top seven most commonly suspected crimes prior to a frisk, Ridgeway (2007) noted that although all seven exhibited the general pattern of higher rates of overall contraband for Whites relative to Blacks, there was allegation-specific variation in the disparity (e.g., for suspected drug-related crime there was a 11.1 contraband recovery rate for Blacks vs. 16.7 for Whites, while for suspected trespassing it was 8.1 versus 10.3 and for suspected burglary it was 2.7 vs. 3.1). Ridgeway pointed out that if you control for the fact that officers suspected Whites and Blacks of committing different types of crimes, the observed hit rate disparity shrinks to about 80% of its original size. Ridgeway (p. 11) interpreted this result as implying that “part of the gap between the 3.3 percent recovery rate for black suspects and the 6.4 percent recovery rate for white suspects is due not to race but rather to differences in suspected crimes.”

However, a potential problem with this conclusion is that it assumes that race plays no role in the type of crimes Black people are falsely suspected of committing relative to Whites. Ridgeway’s (2007) data showed that the majority of Black detainees were stopped due to suspected weapon contraband, while only 28% of White detainees were initially suspected of this violation. Ridgeway’s data also indicated that the relative racial difference in contraband recovery was greatest for this particular suspected offense (weapon related). Thus, the statistically adjusted hit rate disparity presented by Ridgeway comparing Blacks to “similarly situated Whites” assumes that the higher prevalence of Black citizens erroneously suspected of criminal possession of a weapon has nothing to do with race and racial stereotypes. If one allows for the possibility that racial stereotypes may influence an officer’s initial perception/misperception of who is armed and dangerous, then it is not surprising that controlling for a measure potentially reflecting racial profiling in police citizen interactions (type of suspected crime justifying a stop) would account for some of the variation in another potential indicator of racial profiling in police citizen interactions (the hit rate after a frisk). In this sense, while undercontrolling is an important issue for researchers to contemplate, particularly regarding analyses of stop rates without a consensus benchmark, overcontrolling may be an equally important issue, especially for stop-outcome studies focusing on frisks, contraband acquisitions, and arrests.

Analyzing Philadelphia stop-and-frisk data for the plaintiffs in the class-action suit mentioned earlier (Bailey v. City of Philadelphia), Rudovsky, Messing, Roper, and Kreimer (2016) reported that Black citizens who were frisked had an overall contraband recovery rate of 5.1 while Whites who were frisked had a rate of 8.3. They noted that the racial disparity in hit rates continued to be significant in regression analyses utilizing a large sample and controlling for a variety of factors, such as individual age and sex as well as the police service area crime rate. Employed as a consultant with Rudovsky et al. for the plaintiffs, Abrams (2014) noted that one particular analysis of the Philadelphia data stood out: In about 40% of the cases sampled, there was no legal basis for the stop and frisk. Common examples were cases where the officer justified the stop because the suspect was “loitering” or “obstructing the sidewalk,” or the officer justified the frisk based on “narcotics investigation.” While illuminating, the analyses conducted by Rudovsky et al. (2016; as well as Abrams, 2014) were based on a portion of the full data and were primarily descriptive. Incorporating multilevel modeling techniques to disentangle individual from place-level associations, the present study utilizes a public data set covering all recorded pedestrian stop and frisks in Philadelphia from 2014 to 2015.

Data

Like the New York Police Department, the Philadelphia Police Department (2016) has made much of its stop-and-frisk data publicly available in the name of transparency and to facilitate independent analysis. The current project makes use of the Philadelphia public data set (a total of 389,005 pedestrian stops with 55,598 pedestrian frisks for 2014–2015) to investigate the individual- and neighborhood-level correlates of the odds of frisking a stopped pedestrian and the odds of finding contraband. For the entirety of 2014 and 2015, the data were collected while Charles Ramsey served as Police Commissioner and were derived from an officer’s reporting of investigatory and custodial detentions on Form 75-48a. Officer instructions specifically noted that the form was not to be used for ordinary consensual interactions: commonly referred to as mere encounters.

There is undoubtedly a considerable amount of error in officer decisions to classify an interaction as a mere encounter versus an investigatory stop, especially in comparison to the terms justifying a frisk, which are more clearly limited to observed and articulable threats to officer safety. A benefit to focusing on stops with frisks is that such encounters are more clearly investigatory detentions than stops without frisks (Gutman, 2016). However, even the frisk data are subject to some reporting error. As Rudovsky et al. (2016) have pointed out, there are a number of cases in the Philadelphia data set where the suspected crime justifying the stop involved a weapon, and yet there was no recorded frisk. Although it is possible that a frisk was unnecessary in these cases due to unspecified circumstances, it seems likely that a frisk occurred but was undocumented. Assuming that such reporting error is random, the overall impact for the present analyses would be to reduce the likelihood of statistically significant frisk predictors.

As Carroll and Gonzalez (2014) have argued, it is important to try to restrict analyses of potential racial bias in officer behavior on police judgments that are truly discretionary in nature, such as the decision to frisk a detainee for weapon contraband based on the perception of a suspicious bulge. A frisk that occurs in lieu of an incident-to-arrest search is arguably an example of routinized police work with limited discretion. In line with Carroll and Gonzalez’s (2014) call, the present analyses focus on stops where if an arrest occurs, it is associated with contraband uncovered by a police officer’s reasonable suspicion. This filter for potential incident-to-arrest cases removes 7% of all stops, but still leaves 362,237 observations for analysis. Although important conceptually, the results were practically identical without the filter.

In addition to the publicly available data covering stop-and-frisk incidents, the Philadelphia Police Department (2016) provides data on crime occurrences with x and y coordinates matched to the street block address. These data were merged with the stop-and-frisk data set, which also contained x and y coordinates, by 2010 census tract identifier. American Community Survey 5-year population estimates (2009–2013) for racial composition and various socioeconomic characteristics of Philadelphia’s 384 census tracts were downloaded from the Census Bureau (2016). These were then appended to the merged data from the Philadelphia Police Department. Although census tracts are not truly socially delineated neighborhoods, they are widely considered useful proxies. In Philadelphia, the average census tract has about 4,000 people and is about one-quarter of a square mile. To ensure the reliable calculation of residential population measures, the current analysis focuses on the 367 areas with at least 1,000 residents where no more than 33% were institutionalized (in a nursing home, jail, etc.). Each of these 367 residential tracts saw at least some occurrences of both stops and frisks in 2014 or 2015.

Estimates from the most recent census indicate that the two largest racial/ethnic groups in Philadelphia are non-Hispanic Whites (37%) and non-Hispanic Blacks (42%). As a number of studies have documented, neighborhood segregation, especially Black–White residential segregation, is very high in Philadelphia (Massey & Denton, 1993). For example, Census and American Community Survey data reveal that the majority of non-Hispanic Black Philadelphians live in census tracts that are at least 85% non-Hispanic Black. Hispanic residents (about 12% of the city’s total) are significantly less segregated. Less than 3% of the city’s Hispanic population lives in a census tract that is at least 85% Hispanic.

Method

Disentangling potential racial bias at the individual and place levels is conceptually and methodologically challenging. However, contemporary research employing multilevel modeling techniques has supported earlier theories (e.g., Smith, 1986) emphasizing that area-level characteristics can strongly predict policing practices independent of individual-level attributes (Ross, 2015). Such supporting evidence cannot be reliably obtained without attention to the nested structure of the data. A variety of social science research demonstrates that failing to account for nesting in the data can distort estimated variances (Moerbeek, 2004), effect sizes (Wampold & Serlin, 2000), and overall substantive interpretation of the results (Raudenbush & Bryk, 2002).

Because the current analysis utilizes nonnormally distributed binary outcome measures, a hierarchical generalized linear model (HGLM) is preferred over a traditional hierarchical linear model. The HGLM used here employs a Bernoulli distribution with a logit link and quasi-likelihood Laplace estimation (Ene, Leighton, Blue, & Bell, 2015). For the sake of parsimony, all of the equations were random intercept only. In cases where zero would not be a substantively meaningful value, the variables were grand-mean centered before they were entered into the equations.

Variables

Table 1 displays descriptive statistics for all of the variables in the analyses. Panel A of Table 1 lists the relevant measures for the analysis of the likelihood of a stop and frisk between 2014 and 2015. The data indicate that about 13% of pedestrian stops in Philadelphia resulted in a frisk. This percentage is quite different from New York’s, which traditionally has hovered around 55% (Levchak, 2017). Despite the apparent greater selectivity in frisking in Philadelphia, the average contraband recovery rate for stops was practically identical to that of New York’s for the same time period: about 3%. Unfortunately, the public data for Philadelphia do not include a breakdown of contraband type, which would be particularly useful for the current analysis of frisks. Existing contraband-specific research for New York and other cities strongly suggests that illicit drugs are uncovered much more frequently than weapons. Guns in particular are retrieved at extremely low rates (Abrams, 2014). Still, this existing research also suggests that similar racial disparities in hit rates exist for both weapon and nonweapon recoveries (Goel, Rao, & Shroff, 2016; Levchak, 2017).

Table 1.

Descriptive Statistics.

Variable	Mean	SD	Minimum	Maximum
Panel A. Characteristics of investigative stops (N = 362,237)
Pedestrian frisked	0.13	0.34	0.00	1.00
Contraband recovered	0.03	0.18	0.00	1.00
Pedestrian Black	0.69	0.46	0.00	1.00
Pedestrian Latino/Latina	0.09	0.29	0.00	1.00
Pedestrian age	33.19	13.38	0.00	115.00
Pedestrian male	0.83	0.37	0.00	1.00
Occurred at night	0.43	0.50	0.00	1.00
Tract poverty rate	36.22	16.42	0.40	84.70
Tract percent Black	58.69	34.65	0.00	99.70
Tract violent crime rate	32.45	16.12	0.76	105.93
Tract stop rate	60.38	56.51	1.42	250.26
Panel B. Characteristics of investigative frisks (N = 47,538)
Contraband recovered	0.11	0.32	0.00	1.00
Pedestrian Black	0.78	0.41	0.00	1.00
Pedestrian Latino/Latina	0.10	0.31	0.00	1.00
Pedestrian age	28.76	11.10	0.00	115.00
Pedestrian male	0.96	0.20	0.00	1.00
Occurred at night	0.46	0.50	0.00	1.00
Tract poverty rate	37.56	15.52	0.40	84.70
Tract percent Black	64.79	31.98	0.00	99.70
Tract violent crime rate	34.22	15.28	0.76	105.93
Tract frisk rate	16.65	5.75	1.20	30.32

Note. The citywide tract averages are 45.60 for the percent Black, 26.15 for the poverty rate, and 20.92 for the violent crime rate (N = 367).

The selection of independent variables was based on previous research demonstrating that younger individuals, males, Blacks, and Latino/Latina residents were all more likely to be frisked (Rudovsky, Messing, Roper, & Kreimer, 2016) and that, net of other factors, frisks may be more common at night (Levchak, 2017). The independent variables were fairly standard in their construction. Data for pedestrian sex (male or female), race (Black, White, Asian, Native American, or Unknown), and ethnicity (Latino/Latina, non-Latino/Latina, or Unknown) were collected by officer observation. The violent crime rate was defined as the total number of murders, robberies, rapes, and aggravated assaults for 2014 and 2015, divided by the 2010 census tract population, and multiplied by 1,000. The nighttime hours were delineated as being between 7 p.m. and 5 a.m.: About 43% of stops occurred during this period. The means for the detainee demographics indicated that the typical stop in Philadelphia involved a non-Latino/Latina (91%), male (83%), Black pedestrian (69%), typically in his early 30s (mean = 33 years old).² The means for the geographic context variables suggested that, relative to the overall city averages, stops were more common in neighborhoods with a higher poverty rate (36% vs. 26%), a greater percentage Black (59% vs. 46%), and a higher violent crime rate (32 vs. 21 violent crimes per 1,000 residents).

Panel B of Table 1 lists the relevant variables for the analysis of the likelihood of a pedestrian frisk producing contraband. On average, 11% of investigatory frisks uncovered contraband. Relative to the demographic characteristics of all those stopped, those stopped and frisked were more likely to be Black (78% vs. 69%), male (96% vs. 83%), and younger (mean = 29 years old vs. 33 years old). Stop and frisks were also somewhat more likely to involve a Latino/Latina pedestrian (10% vs. 9%) and to occur at night (46% vs. 43%). Similarly, for the geographic context variables, stops involving frisks were more common in neighborhoods with a higher percentage Black (65% vs. 59%) and were just slightly more likely in places with higher poverty (38% vs. 36%) and violent crime rates (34 vs. 32 violent crimes per 1,000 residents). There was substantial variability at the tract level in both the number of pedestrian stops relative to the resident population and the percentage of pedestrian stops incorporating frisks. The highest stop-with-frisk rates tended to be clustered in North Philadelphia, but there was also a considerable amount of spatial dispersion (see Figure 1).

Figure 1.

Percentage of pedestrian stops with frisks in Philadelphia Census Tracts 2014–2015.

Results

The Odds of Being Frisked

Table 2 displays the results from the first set of multilevel logistic regression analyses. These analyses were directed at addressing the following questions: (1) Is the neighborhood violent crime rate significantly associated with the level of physical intrusiveness of investigative stops? (2) Is the reported higher average frisk rate for Black pedestrians due to higher average violent crime rates in the communities where Black citizens reside? and (3) Is the association between neighborhood violent crime rates and the likelihood of being frisked contingent on the racial composition of an area?

Table 2.

Multilevel Estimates for the Odds of a Stop Leading to a Frisk.

Item	Model 1	Model 2	Model 3
Fixed effects
Intercept	−2.0991* (0.0339) [0.1226]	−4.1100* (0.0384) [0.0164]	−4.2124* (0.0469) [0.0148]
Pedestrian Black		0.6383* (0.0180) [1.8933]	0.6179* (0.0181) [1.8550]
Pedestrian Latino/Latina		0.4998* (0.0227) [1.6484]	0.4881* (0.0227) [1.6292]
Pedestrian age		−0.0289* (0.0004) [0.9715]	−0.0290* (0.0004) [0.9714]
Pedestrian male		1.6322* (0.0242) [5.1151]	1.5995* (0.0238) [4.9506]
Stop occurred at night		0.0756* (0.0104) [1.0785]	0.0716* (0.0104) [1.0742]
Tract poverty rate			0.0032 (0.0019) [1.0032]
Tract violent crime rate			0.0192* (0.0025) [1.0194]
Tract percent Black			0.0041* (0.0007) [1.0041]
Percent Black × Crime rate			−0.0003* (0.0001) [0.9997]
Tract stop rate			−0.0073* (0.0010) [0.9927]
Error variance
Level-2 intercept	0.3857* (0.0320)	0.2461* (0.0214)	0.1480* (0.0130)
Model fit
−2 Log Likelihood	272758	258306^a	258142^b

Note. N = 362,237. Level-2 variables are italicized (N = 367 census tracts). Cells show parameter estimates on the log odds scale with standard errors in parentheses and odds ratios in brackets. Estimation method = Laplace.

^aModel 2 fits the data better than Model 1, χ²(5) = 14452, p < .05. ^bModel 3 fits the data better than Model 2, χ²(5) = 164, p < .05.

*p < .05.

Regarding Question 1, the results indicate that a significant amount of variability in the likelihood of being physically frisked exists at the neighborhood level (intraclass correlation coefficient [ICC] = .11, p < .001) and that, on average and net of a variety of factors, the neighborhood violent crime rate is positively associated with the odds of a frisk occurring during an investigative stop. Notably, this relationship was independent of the tract stop rate, which was introduced to control for any confounding influence of police officers spending more investigatory time in violent neighborhoods.³

The individual-level variables in Model 3 in Table 2 all exhibited statistically significant associations with the likelihood of a frisk that were in the expected direction (given findings in previous research and the descriptive statistics presented in Table 1). The coefficients for individual classification as Black, Latino/Latina, and male were all positive, as was the coefficient for the stop occurring during nighttime hours. The coefficient for the age variable was strongly negative.

For the tract-level variables in Model 3 in Table 2, only the poverty rate was statistically insignificant. Holding constant the other variables in the model, the tract stop rate was significantly negatively associated with the likelihood of a frisk occurring.⁴ Consistent with previous research (e.g., Levchak, 2017), controlling for individual racial classification, the percentage Black in an area was positively associated with the likelihood that a stop would incorporate a frisk. As mentioned, the violent crime rate also exhibited a statistically significant average positive effect on the likelihood of a frisk occurring. The coefficient for the violent crime rate (b = .0192, p < .001) indicates that a 1-unit increase in violent crime is expected to increase the odds of a frisk by approximately, .0194 exp (.0192) = 1.0194. Alternatively stated, for stops in places that are 1 standard deviation (15) above the average citywide violent crime rate, the odds of a frisk occurring are estimated to be about 29% greater than for stops in places with mean levels of violent crime. Although statistically significant and in the expected direction, the average effect of the tract violent crime rate was not especially large (evaluated at the mean of percent Black).

Concerning Question 2, the results reveal that controlling for area violent crime and other factors, Black racial classification for pedestrians is strongly and significantly associated with the odds of a frisk. The coefficient for Black racial classification in Model 3 in Table 2 (b = .6179, p < .001) suggests that individuals classified as Black are, on average, approximately 1.86 times more likely to be frisked than those not classified as Black, exp (.6179) = 1.86. This strong association was fundamentally unaffected by the addition of the violent crime rate into the equation.

The answer to Question 3 regarding how neighborhood racial composition conditions the relationship between violent crime and the odds of a frisk requires considerable elaboration. The coefficient for the product term of the percent Black and violent crime rate in Model 3 in Table 2 (b = −.0003, p < .001) suggests that the positive effect of area violent crime on the odds of a frisk decreases as the percentage Black increases. Alternatively stated, racial composition and violent crime appear to interact such that variation in violent crime matters more for explaining variation in the odds of being frisked in places with fewer Black residents.⁵ The coefficients for the product and component terms suggest that the effect of violent crime on frisking in areas that are 1 standard deviation below the citywide average in percentage Black is more than 3.5 times larger than that in areas 1 standard deviation above the average of percent Black. This interaction effect was also evident using an alternative split-sample method for assessing the conditioning impact of racial composition on the relationship between violent crime and the likelihood of a frisk. Estimating the full models separately for majority Black and non-Black tracts, the effect of violent crime in predominantly non-Black neighborhoods was similarly estimated to be about 2.5 times larger than that for majority Black neighborhoods in Philadelphia.⁶ Additionally, the relatively weak relationship between violent crime and frisking in predominantly Black areas was also observable in a simple regression utilizing the tract-level data only (see Figure 2).

Figure 2.

A weak relationship between violent crime and frisk rates in majority Black tracts. The simple linear regression line was fitted for all residential tracts in Philadelphia that were greater than 50% non-Hispanic Black (N = 150). The correlation was −.01. For comparison, the correlation using data for all other tracts was .42 (N = 217, P < .001). The source of the 2010 population data was the U.S. Census Bureau. Violent crime data were collected by the Philadelphia Police Department and were totaled for 2014 and 2015.

Regardless of how it is estimated and illustrated, the interpretation of the apparent interaction between neighborhood violent crime and racial composition is complicated by the need to address potentially confounding nonlinearity and race-specific distributions in rates of violence (Hannon & Knapp, 2003; McNulty, 2001). That is, it is possible that the effect of violent crime on police response significantly decreases at higher levels of violence and that predominantly Black and non-Black neighborhoods rarely experience similar levels of criminal victimization. However, in the case of the Philadelphia data used here, there is a sufficient amount of commonality in the violent crime distributions that can be exploited to help differentiate nonlinearity from an interaction effect. Figure 3 displays the overlay in the violent crime rate distributions for majority Black and nonmajority Black areas. As can be seen, while majority Black neighborhoods tend to have significantly higher levels of violence than other places, there are still many predominantly Black neighborhoods with relatively low violent crime rates. Supplemental analyses reveal that the significant interaction between the percent Black and violent crime rate holds when a quadratic term for violent crime is introduced into the equation for the full sample and when the sample is restricted solely to stops in tracts with violent crime rates 25 and below.⁷

Figure 3.

Overlap of violent crime rate distributions for <50% Black versus >50% Black tracts. The dashed kernel density curve is for residential tracts that are greater than 50% non-Hispanic Black (N = 150) and the solid kernel density curve is for all other tracts (N = 217). The source of the 2010 population data is the U.S. Census Bureau. Violent crime data come from the Philadelphia Police Department and are totaled for 2014 and 2015.

The Probability of Finding Contraband on Frisked Detainees

As noted earlier, previous research has examined some of the individual-level Philadelphia data used here and reported that while Black detainees experience frisks more often than White detainees, contraband is actually less likely to be uncovered from frisks of Black individuals than Whites (Rudovsky et al., 2016). These patterns can also be seen at the tract level in hypersegregated Philadelphia. Despite the fact that predominantly Black areas saw nearly 70% more frisks than non-Black areas, the absolute amount of contraband recovered from frisks was actually greater in non-Black places where frisks appeared to be conducted more selectively (see Figure 4).

Figure 4.

Frisks were more common, yet less productive in majority Black tracts. Black tracts were those having a resident population greater than 50% non-Hispanic Black. There were 150 Black and 217 Non-Black tracts. The source of the 2010 population data was the U.S. Census Bureau. Frisk and contraband recovery data were provided by the Philadelphia Police Department and were totaled for 2014 and 2015.

Table 3 displays the results from the set of multilevel logistic regression analyses directed at addressing whether both pedestrian’s Black racial classification and the neighborhood’s percentage Black are independently associated with decreased odds of a frisk uncovering contraband. Examination of the ICC suggests that about 7% of the variability in contraband recovery is accounted for by the census tracts in the sample (ICC = .07). While the ICC here is relatively low (suggesting that variability is overwhelmingly at the individual level), the coefficient was larger than that cited in other comparable multilevel analyses (e.g., Levchak, 2017; Terrill & Reisig, 2003) and the associated value for the between-neighborhood variance was highly statistically significant (p < .001).

Table 3.

Multilevel Estimates for the Odds of a Frisk Uncovering Contraband.

Item	Model 1	Model 2	Model 3
Fixed effects
Intercept	−2.0908* (0.0340) [0.1236]	−1.0291* (0.0711) [0.3573]	−0.9977* (0.0755) [0.3687]
Pedestrian Black		−0.3293* (0.0463) [0.7194]	−0.2166* (0.0472) [0.8053]
Pedestrian Latino/Latina		−0.1097 (0.0564) [0.8961]	−0.0662 (0.0568) [0.9360]
Pedestrian age		0.0017 (0.0013) [1.0017]	0.0014 (0.0013) [1.0014]
Pedestrian male		−0.7244* (0.0593) [0.4846]	−0.7488* (0.0588) [0.4729]
Stop occurred at night		−0.2676* (0.0304) [0.7652]	−0.2681* (0.0304) [0.7648]
Tract poverty rate			−0.0011(0.0023) [0.9989]
Tract violent crime rate			0.0039 (0.0028) [1.0039]
Tract percent Black			−0.0071* (0.0009) [0.9930]
Tract frisk rate			−0.0172* (0.0053) [0.9830]
Tract stop rate			0.0003 (0.0011) [1.0003]
Error variance
Level-2 intercept	0.2416* (0.0297)	0.1869* (0.0252)	0.1131* (0.0178)
Model fit
−2 Log Likelihood	32739	32446^a	32359^b

Note. N = 47,538. Level-2 variables are italicized (N = 367 census tracts); Cells show parameter estimates on the log odds scale with standard errors in parentheses and odds ratios in brackets. Estimation method = Laplace.

^aModel 2 fits the data better than Model 1, χ²(5) = 293, p < .05. ^bModel 3 fits the data better than Model 2, χ²(5) = 87, p < .05.

*p<.05.

Only three of the five individual-level variables in Table 3, Model 3 exhibited a statistically significant association with the likelihood of a productive frisk. Despite the large sample size (N = 47,538 frisks), the effects of individual ethnic classification as Latino/Latina and pedestrian age were not statistically significant at conventional levels. All three of the statistically significant coefficients were negative in sign: for individual classification as Black, individual classification as male, and the nighttime frisk indicator. Recalling that the legal justification for a frisk is to uncover something that a potentially dangerous suspect could use as a weapon, one substantive interpretation of these results is that they imply a greater level of perceived but unsubstantiated threat to officer safety in cases where the detainee is Black, male, and encountered at night. Holding constant the other factors at their mean, the estimates from Table 3, Model 3 suggest that the predicted probability that contraband will be found on a non-Latino Black male selected to be frisked at night is 9.7%, while the estimated probability of contraband recovery for a non-Latina non-Black female selected to be frisked during the day is 26.9%. This suggests that there is systematic variation in reasonable suspicion thresholds by race, gender, and time of day. Such variation may be driven by deeply internalized normative expectations. That is, for example, a plausible explanation for why women selected to be frisked are more likely to have contraband is that women are generally perceived to be less capable of harmful criminal action than men, and so it takes a more obvious telltale sign that a woman is criminally dangerous to trigger a defensive frisk.

Some neighborhood-level factors also appear to affect the probability of contraband recovery. Several social critics have suggested that the high level of stop and frisk in predominantly Black communities has produced a deterrent effect where residents in these communities do not carry weapons or other contraband due to the fear of getting caught. For example, Butler (2012) argues that

In heavily policed neighborhoods like Brownsville, where the average young man is seized and searched five times a year, according to that data, you would be crazy to walk around with a firearm. So you just give the gun to your girlfriend to carry, or you keep it at home.

To help account for the possibility that any observed effect of racial composition would be spurious due to this dynamic, Model 3 in Table 3 includes controls for the rate of stops relative to the resident population size and the percentage of stops with frisks. While the coefficient for the tract stop rate was statistically insignificant, higher tract frisk rates did significantly reduce the odds of a frisk uncovering illicit goods. However, even controlling for the frisk rate, increased Black presence was significantly associated with decreased likelihood of a productive frisk (b = −.0071, p < .001).⁸ Thus, both the individual pedestrian’s Black racial classification and the neighborhood’s percentage Black were independently associated with significantly reduced odds of a frisk yielding contraband.⁹

Robustness Tests Common to Both Sets of Analyses

For each set of multilevel analyses (predicting frisks and predicting contraband recovery), a number of robustness tests were implemented. First, the violent crime rate was replaced with another indicator of neighborhood dangerousness: the number of shooting victims.¹⁰ These data were also made publicly available by the Philadelphia Police Department and geocoded with x and y block coordinates. The gun violence measure produced substantively very similar results in the main models: Percent Black and the number of shooting victims significantly interacted in their effects on the likelihood of a frisk, while the measure was not significantly related to contraband recovery.

Second, spatial lags of tract violent crime were incorporated into the models to help account for spatial clustering and the possibility that tract boundaries are more meaningfully defined for some groups than others. The spatial lags were calculated using the k-nearest neighbor method and can be thought of as the average amount of violent crime in the areas surrounding a particular tract (five nearest neighbors in this case). Inclusion of the spatial lags did not alter the key findings in either set of results. Independent of the spatial lag term, the impact of tract violent crime rates on the likelihood of a frisk was significantly conditioned by tract racial composition. Both the tract violent crime rate and the spatial lag were not significantly related to contraband recovery.

Third, and relatedly, to help further alleviate concerns that census tracts may differently reflect sociopolitical boundaries for minority and majority populations, the analyses were conducted using police service areas (PSAs) as the Level-2 grouping variable rather than census tracts. The Philadelphia Police Department notes that PSA boundaries were drawn in the spirit of community policing principles. PSAs are much larger than census tracts, but they are considerably smaller administrative units than police districts. A potential advantage to using these spatial boundaries is that relevant police deployment decisions are made at this level (N = 65). Geocoded PSA data were available for the violent crime rate and percent Black, but not the other Level-2 variables. Using the available data, the main models were reestimated and the basic findings reemerged. In particular, the PSA violent crime rate and percentage Black significantly interacted in their effects on the odds of a frisk occurring and both individual Black racial classification and the PSA’s percentage Black were independently associated with decreased odds of finding contraband on frisked detainees.

Conclusion

Existing pedestrian stop-and-frisk research has mostly focused on publicly available data for New York City. The current analysis extended this line of inquiry to Philadelphia. This is important for several reasons. First and most obviously, what is true about the defining characteristics of stop-and-frisk practices in New York City need not be true everywhere. Second, while New York City’s reported use of stop and frisk has declined to historically low levels after high profile litigation, Philadelphia’s use has remained elevated, and recent reports suggest that relatively little progress has been made in terms of reducing racial disparities (Rudovsky et al., 2016). Third, different from New York City, the City of Philadelphia chose not to aggressively fight stop and frisk litigation in court but instead entered into a consent decree to more closely monitor investigatory stops. As Abrams (2014) has argued, the cooperative/consent decree approach, which is used in other cities as well, may ultimately slow progress in reforming police practices relative to more adversarial court battles (which tend to receive considerable press coverage). Finally, Philadelphia’s violent crime rate is nearly twice that of New York City’s. This may affect the relative degree to which stop and frisk is retracted or reformed.

One basic result from the current analysis that was consistent with research focused on New York City was a lower contraband recovery rate for frisked Black individuals relative to Whites. Evidence of this disparity, which was robust to a variety of individual- and tract-level controls, suggests that police officers may be using a lower bar for reasonable suspicion in their decisions to frisk Black residents, potentially due to stereotypes and what Goffman (1963) termed “tribal stigma.” Moreover, independent of the individual’s racial classification, the proportion Black in an area was significantly associated with heightened odds of a person being subjected to an unproductive frisk; thus, there is also evidence in line with the concept of “neighborhood stigma” (Besbris et al., 2015). Another potential indicator of neighborhood racial stigma uncovered in the present analyses was that the relationship between neighborhood violent crime and the likelihood of a stop-with-frisk incident was considerably stronger in areas that were not predominantly Black. For Black tracts, the relationship between variation in violent crime and frisking activity was weak, which implies that in Black places perceptions of dangerousness and criminality are uniquely driven by something other than objectively measured threat.

If one assumes that racial socialization for police officers is similar to that of the general public, then it is not surprising that area racial composition would be meaningfully tied to police defensive actions regardless of objective danger (Quillian & Pager, 2001; Sampson & Raudenbush, 2004; Smith & Alpert, 2007). However, police officers have considerably more factual information about local variation in serious crime rates than the general public. That is, while the public increasingly can monitor current neighborhood crime rates online, the police actually respond to the calls behind those statistics and are routinely briefed about specific incidents at specific locations. Therefore, one might reasonably expect a stronger association between the objective rate of violence and perceived violence for police officers relative to the general public. Future research might focus on the differences and similarities between the public and the police in what generates fear for one’s safety.

It is sometimes suggested that rather than reflecting officer defensiveness associated with racial profiling, higher frisk rates for Black residents could reflect a greater tendency to question a police officer’s authoritative commands. The Philadelphia public data set does not provide much useful information in this regard. However, the more commonly used public data for New York City can offer some illuminating evidence (New York City Police Department, 2017). In 2006, when stop-and-frisk policy was still widely implemented in New York, a sizable 16.6% of frisks involving non-Hispanic Black individuals were described by uniformed officers as necessary due to “Refusal to comply with officer’s direction(s) leading to reasonable fear for safety.” However, the relevant percentage for frisked non-Hispanic Whites was statistically indistinguishable: 16.1% (p > .05, N = 81,690 for Black detainees and N = 9,525 for White detainees). This suggests that racial differences in the likelihood of being frisked are not fundamentally about varying levels of obedience to immediately recognized police authorities.¹¹

Although from a legal perspective an officer’s decision to frisk should be based on an articulable threat of an armed and dangerous individual, it is possible that the general level of perceived community support partially influences frisk decisions. Officers may feel more isolated and threatened in predominantly Black and Latino/Latina areas because they perceive lower levels of trust. Indeed, some classic research examining attitudes among Philadelphia police officers suggests that perceptions of low levels of community support are correlated with heightened concerns with on-the-job security (Greene, 1989). Assuming that repeated frisking can increase community members’ mistrust of the police (Carr, Napolitano, & Keating, 2007; Rios, 2011; Tyler et al., 2014), stop-and-frisk practices may potentially create a positive feedback loop where frisking in Black and Latina/Latino areas begets more frisking in Black and Latina/Latino areas. Growing attention to issues of procedural justice in police–citizen interactions, such as ensuring that the reason for the stop is respectfully explained and that the detainee is given ample opportunity to state his or her case, might help disrupt this self-reinforcing dynamic. Various community policing initiatives where officers meet residents outside the context of reasonable suspicion, such as Police Community Days, might also help.

Like earlier New York–based studies, the present investigation is limited in its generalizability. Further research for other cities and time periods is needed, especially research employing observational data from independent sources, rather than information solely supplied by the police. Moreover, as Rosenfeld and Fornango (2014, 2017) have emphasized, in the absence of formal experimental conditions, it is important not to draw strong causal conclusions from data sets like the ones complied for New York and Philadelphia. Additionally, it is important to keep in mind that these secondary data sets, created by official paperwork, cannot adequately capture dynamics where informal rules contradict, and override, formal directives. For example, while frisks are formally about checking for weapons, informally, officers may feel pressure to utilize the less-than-probable-cause threshold of frisks to catch certain offenders with drugs or other nonweapon contraband that would still constitute admissible/prosecutable evidence. Indeed, previous research suggests that officers frequently conflate searches with frisks, despite their very different legal standards (Carroll & Gonzalez, 2014).¹²

Regardless of these important limitations, future research on racial disparities in police activity will hopefully be moved to (1) consider the possibility that racial profiling can simultaneously occur at the individual and neighborhood levels and (2) explore whether the relationship between local crime and police activity varies by area racial composition. In a sense, the second proposal is essential to extend the logic underlying standard contraband recovery rate comparisons to race-specific contextual influences. There is a conceptual congruence between studying whether a Black individual’s actual innocence is less relevant for avoiding contraband-seeking frisks and examining whether a Black community is able to translate a low crime rate into predictably lower levels of policing intrusiveness.¹³

Although replication was not the central purpose of the current analysis, several key descriptive findings regarding Rudovsky et al.’s (2016) report for the plaintiffs in Philadelphia’s Bailey et al. case (2011) were supported (e.g., significant racial disparities in frisk and contraband recovery rates). Whether reacting to litigation or proactively seeking public trust, police departments across the country have increasingly made their stop data publicly available for independent analysis. This is undoubtedly a positive development and those police departments voluntarily participating in open data sharing deserve appropriate recognition for their efforts (including Philadelphia). Still, given the complex structure of much of these data, it is likely that average readers will find it difficult to meaningfully parse the vast amounts of information released. Social scientists, and criminologists in particular, can play an important role in actualizing meaningful transparency in police behavior. Such transparency, along with evidence-based efforts to minimize racial bias, has the potential to enhance the legitimacy and crime-control efficacy of the police.

This study attempted to synthesize insights from multilevel research on racialized perceptions of crime and disorder (e.g., Quillian & Pager, 2001; Sampson & Raudenbush, 2004) with research focused on racial disparities in stop-and-frisk practices (e.g., Levchak, 2017; White & Fradella, 2016). In terms of future theory development in the racial profiling literature, the present analysis suggests that the notion of racialized neighborhood stigma should play a more important role. This means considering not only how the racial composition of a neighborhood can matter beyond individual-level racial dynamics (additively) but also how neighborhood racial composition might condition the meaningfulness of factors commonly assumed to be race neutral (multiplicatively). Likewise, it is important for policy evaluations of the effectiveness of stop-and-frisk practices to consider threshold effects and nonlinear relationships (Goel et al., 2016). In particular, the current results suggest that increasing the number of frisks does not necessarily translate into more uncovered contraband and related arrests. Given existing research on how unwarranted frisking can undermine police legitimacy and promote disregard for law (Tyler et al., 2014), this is not a simple matter of ineffectiveness—it is an issue of potentially harmful policing.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Notes

References

Abrams

D. S.

(2014). The law and economics of stop-and-frisk. Faculty Scholarship. Retrieved from http://scholarship.law.upenn.edu/faculty_scholarship/1592

Anwar

Fang

(2006). An alternative test of racial prejudice in motor vehicle searches: Theory and evidence. American Economic Review, 96, 127–151.

Ayres

(2002). Outcome tests of racial disparities in police practices. Justice Research and Policy, 4, 131–142.

Ayres

Borowsky

(2008). A study of racially disparate outcomes in the Los Angeles. Police Department. Retrieved from the ACLU of Southern California Website: https://www.aclusocal.org/sites/default/files/wp-content/uploads/2015/09/11837125-LAPD-Racial-Profiling-Report-ACLU.pdf

Bailey v. City of Philadelphia. (2010). No. 2:10-cv-05925, U.S. District Court for the Eastern District of Pennsylvania.

Bailey v. City of Philadelphia. (2011). Settlement Agreement, Class Certification and Consent Decree. No. 10-cv-05952, U.S. District Court for the Eastern District of Pennsylvania.

Baumer

(2002). Neighborhood disadvantage and police notification by victims of violence. Criminology, 40, 579–616.

Besbris

Faber

Rich

Sharkey

(2015). The effect of neighborhood stigma on economic transactions. Proceedings of the National Academy of Sciences of the United States of America, 112, 4994–4998.

Bursik

Grasmick

H. G.

(1993). Neighborhoods and crime: The dimensions of effective community control. New York, NY: Lexington Books.

10.

Butler

(2012, 9 20). Don’t antagonize those who could help fight crime. The New York Times. Retrieved from https://www.nytimes.com/roomfordebate/2012/07/17/does-stop-and-frisk-reduce-crime/dont-antagonize-those-who-could-help-fight-crime

11.

Bykofsky

(2015, 2 27). Falling out of love with stop-and-frisk. Philadelphia Daily News. Retrieved from http://www.philly.com/philly/columnists/stu_bykofsky/20150227_Falling_out_of_love_with_stop-and-frisk.html

12.

Carr

P. J.

Napolitano

Keating

(2007). We never call the cops and here is why: A qualitative examination of legal cynicism in three Philadelphia neighborhoods. Criminology, 45, 445–480.

13.

Carroll

Gonzalez

M. L.

(2014). Out of place: Racial stereotypes and the ecology of frisks and searches following traffic stops. Journal of Research in Crime and Delinquency, 51, 559–584.

14.

Chiricos

Hogan

Gertz

(1997). Racial composition of neighborhood and fear of crime. Criminology, 35, 107–131.

15.

Ehrenfreund

(2016, 9 22). Donald Trump claims New York’s stop-and-frisk policy reduced crime. The data disagree. The Washington Post. Retrieved from https://www.washingtonpost.com/news/wonk/wp/2016/09/22/donald-trump-claims-new-yorks-stop-and-frisk-policy-reduced-crime-the-data-disagree/

16.

Ene

Leighton

E. A.

Blue

G. L.

Bell

B. A.

(2015). Multilevel models for categorical data using SAS PROC GLIMMIX: The basics. SAS Global Forum Proceedings. Retrieved from https://support.sas.com/resources/papers/proceedings15/3430-2015.pdf

17.

Fridell

(2004). By the numbers: A guide for analyzing race data from vehicle stops. Washington, DC: Police executive Research Forum.

18.

Geller

Fagan

(2010). Pot as pretext: Marijuana, race, and the new disorder in New York City street policing. Journal of Empirical Legal Studies, 7, 591–633.

19.

Gelman

Fagan

Kiss

(2007). An analysis of the New York City Police Department’s “stop-and-frisk” policy in the context of claims of racial bias. Journal of the American Statistical Association, 102, 813–823.

20.

Goel

Rao

J. M.

Shroff

(2016). Precinct or prejudice? Understanding racial disparities in New York City’s stop-and-frisk policy. The Annals of Applied Statistics, 10, 365–394.

21.

Goffman

(1963). Stigma: Notes on the management of a spoiled identity. New York, NY: Simon & Shuster.

22.

Goode

(2012, 7 12). Philadelphia defends policy on frisking, with limits. The New York Times, p. A11.

23.

Gutman

(2016). Stop and frisk redux: Analysis of racial bias in New York City. CUNY Academic Works. Retrieved from http://academicworks.cuny.edu/hc_sas_etds/45

24.

Greene

Jack R.

(1989). Police officer job satisfaction and community perceptions: Implications for community policing. Journal of Research in Crime and Delinquency, 26, 168–183.

25.

Hannon

Knapp

(2003). Reassessing nonlinearity in the urban disadvantage/violent crime relationship: An example of methodological bias from log transformation. Criminology, 41, 1427–1488.

26.

Harris

(2002). Profiles in injustice: Why racial profiling cannot work. New York, NY: The New Press.

27.

Kubrin

C. E.

Messner

S. F.

Deane

McGeever

Stucky

T. D.

(2010). Proactive policing and robbery rates across U.S. cities. Criminology, 48, 57–97.

28.

Lacoe

Sharkey

(2016). Life in a crime scene: Stop, question, and frisk activity in New York City neighborhoods in the aftermath of homicides. Sociological Science, 3, 116–134.

29.

Levchak

P. J.

(2017). Do precinct characteristics influence stop-and-frisk in New York City? A multi-level analysis of post-stop outcomes. Justice Quarterly, 34, 377–406.

30.

Lipsky

(2015, 9 9). Ray Kelly offers defense of stop-and-frisk the courts never heard. New York Post. Retrieved from http://nypost.com/2015/09/09/ray-kelly-offers-defense-of-stop-and-frisk-the-courts-never-heard/

31.

MacDonald

(2001). The myth of racial profiling. City Journal, 11, 2–5.

32.

Massey

Denton

(1993). American apartheid. Cambridge, MA: Harvard University Press.

33.

McNulty

(2001). Assessing the race–violence relationship at the macro level: The assumption of racial invariance and the problem of restricted distributions. Criminology, 39, 467–488.

34.

Meehan

A. J.

Ponder

M. C.

(2002). Race and place: The ecology of racial profiling African American motorists. Justice Quarterly, 19, 399–430.

35.

Moerbeek

(2004). The consequences of ignoring a level of nesting in multilevel analysis. Multivariate Behavioral Research, 39, 129–149.

36.

New York Civil Liberties Union. (2011). Stop and frisk practices. Retrieved from http://www.nyclu.org/issues/racial-justice/stop-and-frisk-practices

37.

New York Police Department. (2017). Stop, question and frisk report database [Data file]. Retrieved from http://www1.nyc.gov/site/nypd/stats/reports-analysis/stopfrisk.page

38.

Philadelphia Police Department. (2016). Vehicle and pedestrian investigations [Data file]. Retrieved from https://www.opendataphilly.org/dataset/vehicle-pedestrian-investigations

39.

Quillian

Pager

(2001). Black neighbors, higher crime? The role of racial stereotypes. American Journal of Sociology, 107, 717–767.

40.

Raudenbush

S. W.

Bryk

A. S.

(2002). Hierarchical linear models: Applications and data analysis methods (2nd ed.). Thousand Oaks, CA: Sage.

41.

Ridgeway

(2007). Analysis of racial disparities in the New York Police Department’s stop, question, and frisk practices. Santa Monica, CA: Rand Corporation.

42.

Rios

(2011). Punished: Policing the lives of Black and Latino boys. New York: New York University Press.

43.

Rosenfeld

Fornango

(2014). The impact of police stops on precinct robbery and burglary rates in New York City, 2003–2010. Justice Quarterly, 31, 96–122.

44.

Rosenfeld

Fornango

(2017). The relationship between crime and stop, question, and frisk rates in New York City neighborhoods. Justice Quarterly, 1–21. doi:10.1080/07418825.2016.1275748

45.

Ross

C. T.

(2015). A multi-level Bayesian analysis of racial bias in police shootings at the county-level in the United States, 2011–2014. PLoS One, 10, 1–34.

46.

Rudovsky

Messing

Roper

M. C.

Kreimer

(2016). Plaintiffs’ sixth report to court and monitor on stop and frisk practices: Fourth amendment issues. Bailey v. City of Philadelphia. Retrieved from https://www.aclupa.org/download_file/view_inline/2674/198/

47.

Sampson

R. J.

Raudenbush

S. W.

(2004). Seeing disorder: Neighborhood stigma and the social construction of “broken windows.” Social Psychology Quarterly, 67, 319–342.

48.

Saul

(2016, 5 18). America has a stop-and-frisk problem. Just look at Philadelphia. Newsweek. Retrieved from http://www.newsweek.com/2016/06/10/stop-and-frisk-philadelphia-crisis-reform-police-460951.html

49.

Smith

D. A.

(1986). The neighborhood context of police behavior. In Reiss

A. J.

Tonry

(Eds.), Crime and justice: An annual review of research (pp. 313–342). Chicago, IL: University of Chicago Press.

50.

Smith

M. R.

Alpert

G. P.

(2007). Explaining police bias: A theory of social conditioning and illusory correlation. Criminal Justice and Behavior, 34, 1262–1283.

51.

Terrill

Reisig

M. D.

(2003). Neighborhood context and use of force. Journal of Research in Crime and Delinquency, 40, 291–321.

52.

Tyler

T. R.

Fagan

Geller

(2014). Street stops and police legitimacy: Teachable moments in young urban men’s legal socialization. Journal of Empirical Legal Studies, 11, 751–785.

53.

U.S. Census Bureau. (2016). Profile of general population and housing characteristics: 2010 [Data file]. Retrieved from https://factfinder.census.gov/faces/nav/jsf/pages/index.xhtml

54.

Wacquant

(2008). Urban outcasts a comparative sociology of advanced marginality. Cambridge, MA: Polity Press.

55.

Walker

(2001). Searching for the denominator: Problems with police traffic stop data and an early warning system solution. Justice Research and Policy, 3, 63–95.

56.

Wampold

B. E.

Serlin

R. C.

(2000). The consequences of ignoring a nested factor on measures of effect size in analysis of variance. Psychological Methods, 5, 425–433.

57.

Werthman

Piliavin

(1967). Gang members and the police. In Bordua

(Ed.), The police: Six sociological essays (pp. 56–98). New York, NY: Wiley.

58.

White

M. D.

Fradella

H. F.

(2016). Stop and frisk: The Use and abuse of a controversial policing tactic. New York: New York University Press.