Background and Theory

A natural consequence of segregation are boundary zones, which sociologists describe as, “interstitial or transitional areas sandwiched between two homogeneous communities” (Legewie and Schaeffer Reference Legewie and Schaeffer2016, 124). As the spaces where different neighborhoods meet, boundary zones are revealing of interracial dynamics and local power struggles. Moreover, to the degree that racial and class-based segregation circumscribe the distribution of municipal resources, this can be seen most starkly at boundary zones, where “objective forms of social differences [manifest] in unequal access to and unequal distribution of resources” (Dean et al. Reference Dean, Dong, Piekut and Pryce2019, 272).

A lengthy literature identifying boundary zones as deeply political places develops from scholarly attention to borders between countries and the conditions under which those borders become more (or less) regulated (Braun and Kienitz Reference Braun and Kienitz2022). Notably, defense of resources drives the establishment of physical fortification, and scholars have likewise observed that municipal security often develops to protect private wealth—this is particularly stark when states bifurcate access to citizenship along ethno-racial lines and fail to adequately provide for public safety (Spocter Reference Spocter2022; Simmons and Kenwick Reference Simmons and Kenwick2022; Robins Reference Robins2002; Davies Reference Davies1981; Clarno and Murray Reference Clarno and Murray2013). The state has an interest in responding to the demands of wealthy and powerful constituents, but it also has an interest in promoting economic development within cities. In the United States, law enforcement have been deployed for this exact purpose by regulating access to and movement between invested and disinvested spaces (Desmond and Valdez Reference Desmond and Valdez2013; Beck Reference Beck2020; Collins, Stuart, and Janulis Reference Collins, Stuart and Janulis2022; Laniyonu Reference Laniyonu2018). Strategies of order maintenance policing, also referred to as “broken windows policing”, provide the institutional basis by which this occurs. The logic underlying order maintenance policing is that law enforcement can enhance feelings of public safety through regulating how people interact with space (like preventing loitering) and dictates that they should pursue low level infractions associated with misdemeanor charges in order to prevent more serious crime (Herbert and Beckett Reference Herbert and Beckett2017).

Such practices provide a legal mechanism for law enforcement to use discretion when making decisions about who to stop and where, but the deployment of these tactics are also at least in part in response to citizen’s demands, vis-à-vis 311 and 911 calls (Harris, Rigolon and Fernandez Reference Harris, Rigolon and Fernandez2020). As one scholar put it, “the history of planning could be rewritten as the obsession with managing fear in the city” (Sandercock Reference Sandercock2000, 205)—where public space is increasingly organized around consumption, law enforcement are deployed to create an environment supportive of consumption-related activities and are highly responsive to wealthy citizens and business owners (Summers and Howell Reference Summers and Howell2019). Empirically, scholars have focused on the extent to which race and class boundaries foster intergroup conflict and in turn, criminogenic behavior (Kim and Hipp Reference Kim and Hipp2022; Legewie Reference Legewie2018; Dean et al. Reference Dean, Dong, Piekut and Pryce2019). Evidence from New York city suggests that non-whites are at higher risk of arrest when on the white side of the line in boundary zones—at least in part because white residents are especially prone to requesting police intervention in their neighborhoods (Neil and Legewie Reference Neil and Legewie2023; Legewie and Schaeffer Reference Legewie and Schaeffer2016).

Intergroup conflict among citizens no doubt yields heightened police presence in racial boundary zones. Yet civilian demands for police intervention that spring from intergroup conflict work alongside top-down decisions about where and how to allocate police resources. Ethnographic work provides a thick account of police regulation of disadvantaged spaces and boundaries (Gordon Reference Gordon2022). For example, in Gordon’s description, the construction of administrative boundaries that circumscribe police work (like police beats) themselves provide a means for the unequal distribution of law enforcement resources. Law enforcement began with beliefs about the utility of a community policing approach, but filtered through notions about coherent neighborhoods, themselves embedded with race and class biases about criminal pathology and value. Thus, comparatively white, wealthy, and economically developed spaces received different resources than poorer neighborhoods with larger racial minority populations. Describing these dynamics, Gordon (Reference Gordon2022) writes:

Gordon’s work on the differential distribution of law enforcement resources relative to patterns of development is important because it explicates the role police play in maintaining geographic racial divisions by managing the flow of people around invested districts and in how they respond to requests for service. Scholars elsewhere find that adjacency to a minority neighborhood is a stronger predictor of use-of-force incidents than is neighborhood composition, providing evidence that police engage racial boundary zones differently than other parts of the city (Omori, Lautenschlager, and Stoler Reference Omori, Lautenschlager and Stoler2022). It is not simply that police provide better services to invested spaces than to disinvested ones. Citizen-driven and organizational dynamics cohere to produce differential policing in and out of the borders between said spaces, and likely do so in ways that diverge from crime. This produces our primary hypothesis:

Complementing empirical work on the citizen-side factors that yield extra policing in boundary zones, and following the theoretical intuitions developed by Gordon (Reference Gordon2022), we critically engage the organizational factors that promote the differential regulation of space. Research around how police aid gentrification identifies order maintenance policing practices as essential tools to regulate poverty, consumption, and investment (Desmond and Valdez Reference Desmond and Valdez2013; Laniyonu Reference Laniyonu2018; Beck, Reference Beck2020; Collins, Stuart, and Janulis Reference Collins, Stuart and Janulis2022; Muñiz Reference Muñiz2015). To wit, under the height of the stop-and-frisk program in New York City, stops overall were more prevalent and less efficacious in boundary zones than in more racially homogeneous spaces (Legewie and Schaeffer Reference Legewie and Schaeffer2016). As focused as order-maintenance policing strategies are on preempting rather than reacting to crime, these practices are endogenous to the kinds of crimes that are ultimately reported (low-level activity likely to result in a misdemeanor charge, for example; Desmond and Valdez Reference Desmond and Valdez2013; Herring Reference Herring2019; Harris, Rigolon, and Fernandez Reference Harris, Rigolon and Fernandez2020; Huff Reference Huff2021). On the other hand, arrest rates for violent felonies are less likely to be driven by geographical considerations (i.e., they should prompt a police response wherever they occur). We therefore expect that to the extent police give attention to boundary zones in excess of crime, this should be most readily observable in police activity addressing low-level, non-violent activity:

We are further interested in the racial dynamics of policing in racial boundary zones. On one hand, we might expect that police heavily target non-white people in non-white spaces in order to protect white privilege (Muñiz Reference Muñiz2015), or that heightened policing mostly occurs on the white side of the line as white people call on police to regulate their space (Neil and Legewie Reference Neil and Legewie2023). Yet Gordon directs our attention to how institutional practices can yield racialized outcomes without implicating individual explicit, racialized attitudes, writing that “[officers] could generate racial disparities without racial intent: they did so just by following their official duties within the circumscribed work environment of the police district. Racial outcomes in policing were facilitated by organizational structures like district boundaries, resource allocations, deployment decisions, formal policies and legal standards” (2022, 183). To the extent law enforcement views boundary zones as in need of additional regulation, law enforcement resources should be distributed in boundary zones in ways that are commensurate on both sides of the boundary, but elevated relative to co-ethnic, homogeneous spaces:

Police make further decisions about how, when, and with whom to intervene. When such decisions are made with little evidence of actualized crime, officers rely on other kinds of heuristics to determine potential criminality, which in turn can reify social hierarchies rooted in said heuristics (Gordon Reference Gordon2022; Epp, Maynard-Moody, and Haider-Markel Reference Epp, Maynard-Moody and Haider-Markel2014; Harris, Rigolon, and Fernandez Reference Harris, Rigolon and Fernandez2020; Gaston Reference Gaston2019; Kane Reference Kane2006). Regarding segregation, Gordon writes that it “created a series of associations that shaped police action, which, in turn, reified the realities of segregation … racial and spatial segregation was so entrenched and naturalized that the police problematized any boundary crossing,” (2022, 163).

It follows from the institutional perspective on policing, where individual prejudices held by either police or civilians are less important than are the routinized policies and practices that police are trained to engage in, that individuals are likely to be stopped when in neighborhoods where they appear to be out of place, irrespective of civilian race. Indeed, courts have ruled that the fish-out-of-water heuristic is a justifiable component of reasonable suspicion (Novak and Chamlin Reference Novak and Chamlin2012; Gaston Reference Gaston2019; Hannon, Neal, and Gustafson Reference Hannon, Neal and Gustafson2021). Scholars have paid special attention to its application to non-white bodies in predominantly white spaces. However, the institutional perspective suggests that police are patrolling racialized spatial boundaries, and as such the fish-out-of-water heuristic should be operative in reference to both white and non-white civilians when they are out of place in boundary zones (Bell Reference Bell2020).

This is not to say that white and non-white people face equal risk of arrest when out of place. Scholars have demonstrated that the fish-out-of-water heuristic is unequally applied writ large, without attention to boundary or non-boundary zones (Hannon, Neal, and Gustafson Reference Hannon, Neal and Gustafson2021; Gaston Reference Gaston2019). Instead, the fish-out-of-water heuristic provides an institutional mechanism by which racialized policing in boundary zones might occur in ways that exceed civilian requests for intervention. If we observe that both white and non-white people face heightened risk of arrest when in boundary zones and are racially out of place relative to when in homogeneous zones, this provides evidence that institutional mechanisms promote the policing of racial boundaries, in addition to civilian demands that develop from intergroup conflict and feelings of out-group threat. The key comparison is to other members of one’s racial subgroup. We do not attempt to adjudicate between citizen and organizational factors that promote heightened policing in racial boundary zones. Instead, this hypothesis and the attendant empirical test provide a means to evaluate whether organizational factors can be observed to shape the politics of racial boundaries in American cities. Our final hypothesis, then, is as follows:

In sum, we argue that racial boundary zones are spaces that command special attention from municipal police departments. Existing empirical accounts spotlight the demand-side factors that yield heightened police activity in racial boundary zones. Our goal is to develop complementary evidence for the supply-side factors that generate not just more activity, but attention that exceeds that given to crime in homogeneous spaces. We leverage Gordon (Reference Gordon2022)’s ethnographic account of how boundaries organize law enforcement behavior to develop a set of expectations around the relationship between policing in boundary zones and crime, and the unique character policing takes in said spaces. In what follows, we leverage Legewie’s novel, data-driven metric that allows us to identify organic demographic transitional zones (Legewie and Schaeffer Reference Legewie and Schaeffer2016; Legewie Reference Legewie2018; Neil and Legewie Reference Neil and Legewie2023) and evaluate policing in racial boundary zones across seven cities.

Data and Design

Our measure of police activity is arrests. Prior evidence suggests arrests measure policing activity net of crime, since many arrests are discretionary and motivated by officer biases (Stashko Reference Stashko2018). We collected incident-level arrest data across several cities: Atlanta, GA; Austin, TX; Boston, MA; Chicago, IL; Milwaukee, WI; and Seattle, WA. We also collected data for Louisville, KY. For this particular locality, we were only able to obtain records for felony arrests. However, we argue that low-level infractions most likely to result in a misdemeanor offense result from processes that are most reflective of officer discretion, and as such, that misdemeanor arrests are the most sensitive measure of excessive policing. This logic gives rise to hypothesis 2. In the case of Louisville, we distinguish between felony arrests for crimes against persons, property and society, where we are likely to see effects stemming from crimes against property and society, relative to crimes against persons. For the sake of parsimony, we present the findings from the six cities where we have records of misdemeanor arrests and discuss the findings from Louisville (located in the appendix) where appropriate. While we set out to incorporate more cities, these seven provided us with the complete range of data necessary for our analysis. For instance, some cities for which we had incident-level arrest data such as El Paso, Oklahoma City, and Idaho Falls lacked incident-level crime data, while others such as Baltimore, had crime data but largely lacked addresses of arrests. Table 1 displays the kind of data we were able to attain for each city.

Table 1 Data availability for seven cities

Table 2 Summary of results by hypothesis and city

These cities range widely in geographic location, racial composition, and histories with policing. Evaluating policing patterns across multiple cities allows us to assess the generalizability of our theory and account for departmental idiosyncrasies across multiple urban contexts. While we sought the broadest time window possible from each department, we succeeded in obtaining complete data on arrests between 2012–2014 for every city. To prepare the data for analysis, we 1) geocode each arrest location and 2) geolocate each arrest inside its respective 2010 Census block (U.S. Census Bureau 2010). We omit warrant arrests as police may exercise less discretion during warrant arrests and they may occur in locations that are distinct from offense locations. We summarize the number of total arrests and arrests by category (misdemeanor and felony) between 2012–2014 for each block and log the outcome (plus 1 to ensure identification) to account for the right-skewed outcome distribution.

We chose to model arrests for two reasons: 1) arrests are the particular metric of police behavior for which we were able to obtain the broadest collection of U.S. cities; and 2) because arrests are more likely to occur with evidence of criminal activity, they are a hard test of our theory and biased against finding any differences between boundary and non-boundary zones. We think that felony arrests are suitable for analysis since many do not lead to a criminal conviction and arise from the same processes that yield misdemeanor arrests, but overall, we think felony arrests are less reflective of discretionary processes than are arrests for misdemeanor infractions. Misdemeanor arrests are correlated with our different measures of crime at the census block-level across cities between ρ = 0.3 - 0.73, a moderate-to-strong correlation (appendix table 4), implying arrests mostly occur where crime occurs, but the relationship is not 1-to-1, consistent with Hypothesis 1. One may argue that an alternative metric such as stops are more appropriate because they reflect discretionary police activity that may or may not reveal wrongdoing by the civilian, even as arrests occur when evidence of criminal behavior is apparent. We anticipate that estimates derived from arrests are conservative. We were able to obtain stop data for Austin, Chicago. and Milwaukee (indicated in table 1). To validate our central findings, we evaluate this data concurrently and reference it where appropriate. It becomes particularly important when we evaluate hypothesis four, which concerns the race of the civilian stopped. We prepare stop data in the same fashion applied to arrest data.

Our key independent variable is boundaries as defined by the level of racial difference between adjacent neighborhoods. Sociologists have developed innovative methods for detecting boundaries, to assess how spatial areas marked by deep demographic differences may lead to heightened conflict or decreased social cohesion (Legewie and Schaeffer Reference Legewie and Schaeffer2016; Kramer Reference Kramer2017; Dean et al. Reference Dean, Dong, Piekut and Pryce2019). To this end, we use an areal wombling technique, which detects boundaries based on differences in the values of spatial data derived from areal units (e.g., a census block or block group). We calculate racial boundaries at the census block level because they offer the granularity necessary to identify and analyze racial variation that may be diluted at the block-group level.

To measure racial boundaries we identify a given census block within a city context. We then identify its adjacent blocks. Next, we calculate the absolute value difference between the proportion of the block’s population that is white and the proportion white in its adjacent blocks (using 2010 decennial census data). We characterize racial boundaries using the white/non-white composition because it allows for consistent measurement across cities, even as the racial composition of each city varies. If we consistently identify an association between white/non-white boundaries and arrests even given this variation, then we can be confident that our theory travels across a wide array of ethno-racial contexts and, moreover, that a primary way in which racial boundaries operate is on the basis of whiteness. Following Legewie (Reference Legewie2018), we take the maximum difference and use that as the racial boundary value. This helps to capture drastic geographic boundaries between predominantly white and non-white blocks.Footnote ¹ The equation to calculate the racial boundary measure is:

$$ {racialboundary}_i={max}\left(|\%{white}_i-\%{white}_{ij}|\right) $$

Where i represents census block i. j are neighboring blocks to i, where j = {1…n_i} in n neighbors to i. The boundary value is scaled between 0–1, where 0 represents no difference in the proportion of the population that is white between adjacent areas, and 1 represents a maximum difference. As the demographic maps in Figure 1 demonstrate visually, racial boundaries capture the relative racial concentration of a given subgroup at the block level, although the quantities of interest represent points of inequality rather than homogeneity. Appendix table 1 provides further information about the measure demonstrating that there are no large outliers in the measure across all cities and comparing the boundary measure to the level of racial segregation in each city context. We use this continuous measure to evaluate hypotheses 1 and 2, which posit that police will give more attention to crime in boundary zones than non-boundary zones, and that this will be more apparent among arrests leading to a misdemeanor conviction. We are also concerned with the moderating effect of crime and racial composition on policing in boundary zones. Following the observation by Legewie and Schaeffer (Reference Legewie and Schaeffer2016) that boundaries characterized by the steepest racial differences also feature the starkest differences in public safety outcomes, we dichotomize the boundary metric such that those census blocks with a boundary score that falls in the top quartile of racial difference are coded as one and those falling below the top quartile of racial difference are coded as zero.

Figure 1 Racial boundaries in the city of Milwaukee

Note: The map on the left displays the percentage of the population that is white by block using 2010 decennial census data. The map on the right displays the white racial boundary value by census block

Previous work has found greater levels of crime at racial boundaries (Legewie and Schaeffer Reference Legewie and Schaeffer2016; Kramer Reference Kramer2017; Kim and Hipp Reference Kim and Hipp2022; Dean et al. Reference Dean, Dong, Piekut and Pryce2019; Legewie Reference Legewie2018). Our central claim in this paper is that police activity is distributed in ways that reflect additional interests and power arrangements and does not strictly map onto crime. To account for criminal activity, we gather incident-level crime data for each city. First, we subset the crime data to the same date range as the corresponding arrest data. Second, we distinguish between violent and property crimes, which will serve as two distinct crime types for which we adjust.Footnote ² Third, we geolocate each incident into its respective census block. Finally, we sum the total crimes for each type, for each census block, during 2012–2014. We evaluate whether we observe excess police activity in boundary zones relative to crime with two tests: we include control variables for crime type and perform a sensitivity analysis to evaluate the robustness of the relationship between boundaries and policing after conditioning on crime; and we assess whether crime moderates the relationship between arrests and boundary zones. This allows us to evaluate whether crime in boundary zones receives more attention than crime outside boundary zones.Footnote ³

We adjust for several block-level characteristics that may jointly explain police activity and racial boundaries. We adjust for population (logged), the proportion of the male population that is 15–35, the proportion of the population that is white, the level of commercial activity in the block, the presence of physical boundaries, and ethno-racial diversity across the four major ethno-racial groups in the U.S. (non-Latinx white, non-Latinx Black, Latinx, non-Latinx Asian) using the Herfindahl index. We leverage percent white to evaluate Hypothesis 3. Hypothesis 3 expects that heightened arrests in boundary zones will be distributed on both the white and the non-white side of the boundary. To test this, we employ a moderation analysis between the racial boundary metric and percent white. We leverage a similar strategy to evaluate Hypothesis 4, which concerns the race of the civilian stopped and where. To assess this, we repeat the moderation analysis, but we subset by race of civilian among cities where we have this information (Austin, Chicago, and Milwaukee).

We also adjust for several socio-economic characteristics, including percent homeowners, logged median household income, percent in poverty, percent unemployed, and percent college-educated.Footnote ⁴ However, the Census provides these variables at the block group level, while our racial boundary metric is calculated at the block level. Socio-economic differences, moreover, likely coincide with racial differences and are an alternative explanation for any disparate police activity in and out of boundary zones we may observe. In order to craft a measure of socio-economic boundaries that is commensurate with our racial boundary metric, we followed methods developed by Wang et al. (Reference Wang, Qian, Kats, Kontokosta and Sobolevsky2017) and leverage an alternative data source to measure socio-economic status at more granular levels than is available via the Census: incident-level 311 call data. Briefly, we identify call types associated with more affluent neighborhoods, calculate the proportion of calls overall for which those categories account, and use areal wombling to identify points of socio-economic difference at the block level. The result is an SES boundary measure at the same level of aggregation as our racial boundary measure.Footnote ⁵

We use the following linear model for each city to test our hypotheses:

$$ Log\left({Arrests}_i+1\right)={\displaystyle \begin{array}{l}\alpha +{\beta}_1{racialboundary}_i\\ {}+\hskip2px {\beta}_2{econboundary}_i\\ {}+\hskip2px \sum \limits_{k=1}^k{\beta}_{k+2}{X}_{ib}^k+{\varepsilon}_{ib}\end{array}} $$

where Arrests_i is the number of arrests in census block i, racialboundary_i is the racial boundary measure scaled between 0–1 at the block-level (i), and econboundary_i is the socioeconomic boundary measure scaled between 0–1 at the block-level (i). $ {\sum}_{k=1}^k{X}_{ib}^k $ are k block and block group-level covariates. ε_ib are block group-clustered errors.

Assessing Racialized Policing in Boundary Zones

Hypotheses 3 and 4 concern the racialized distribution of police activity in racial boundary zones. Hypothesis 3 posits that heightened police activity observed in racial boundary zones will be apparent on both sides of the boundary. On one hand, we can imagine that policing might predominate on the non-white side of the line in an effort to control non-white populations, and that this reflects over-policing of non-white people in general. On the other hand, researchers have demonstrated that white individuals living in boundary zones make heightened requests for government intervention out of a sense of threat (Legewie Reference Legewie2018). We do not contest the citizen-driven factors that contribute to over-policing in boundary zones, but we do contend that top-down processes are essential to understanding how boundary zones organize police behavior. It should by now be clear that police activity does not simply follow crime, but that it is distributed in ways that reflect beliefs about how best to maintain social order and which spaces are vulnerable to the breakdown of order. Those beliefs are baked into the institutional policies and practices that govern, constrain, and organize police activity. As such, the disparate application of policing, whether racial, spatial, or otherwise, can occur in the absence of individual prejudices. We therefore expect that heightened police activity in boundary zones will be apparent on both sides of the boundary. We argue that the racialized nature of policing, in this case, manifests in heightened attention paid to racial boundary zones themselves, relative to racially homogeneous spaces.

In order to assess where arrests are most likely to occur, we interact percent white with the racial boundary metric. This allows us to assess whether arrests are more likely to occur in or out of boundary zones, conditional on percent white. The results are displayed in figure 6 (associated regression estimates are located in appendix table 10 and appendix table 11). In Boston and Chicago, it appears that in non-white spaces, there is no difference in the likelihood of arrest in boundary zones relative to non-boundary zones, but in white spaces risk of arrest is heightened in boundary zones. On balance, however, it appears that rather than moderating the effect of living in a non-white space on the risk of arrest, there is an intercept effect whereby one is more likely to face arrest in boundaries than in non-boundaries, even as non-white spaces—in or out of boundary zones—experience more policing overall than do white ones.Footnote ⁸ We repeat this analysis in Louisville among felony arrests (displayed in appendix figure 8 and appendix table 14), and in Austin, Chicago, and Milwaukee using stop data (located in appendix figure 10 and appendix table 17). The substantive takeaway, that heightened risk of arrest is higher in boundary zones for both white and non-white spaces, holds.

Figure 6 Heterogenous influence of racial boundary on arrests by white racial context

Notes: X-axis is the percent white measure. Y-axis is the predicted logged number of arrests (+1 to ensure identification), by arrest type. Each row characterizes a different city sample. Estimates from fully-specified models with controls at means. 95% CIs displayed from HC2 robust block group clustered SEs.

Hypothesis 4 concerns the race of civilian subject to police intervention. We have argued that top-down policies and procedures that determine the distribution of police resources are an important mechanism underlying the correlation between racial boundary zones and police activity. Other factors that also contribute to this relationship include demand-side factors, where white citizens request police intervention at higher rates in diverse spaces than in white, homogenous ones out of feelings of threat. Indeed, one study observed that in boundary zones non-white individuals were stopped more often on the white side of the boundary. However, again as earlier, police do not only respond to requests for service, they are first and foremost provided training around how to identify signs of potential criminality and given directives about where and how to patrol. One such heuristic that officers may draw on to identify the likelihood that an individual intends to commit crime is the fish-out-of-water heuristic. A fish out of water is an individual who appears to be in a neighborhood where they do not belong. The supply-side model of racially disparate policing predicts that individuals in boundary zones should be at greatest risk for police intervention when they are on the side of the boundary that contrasts with their race, and further that this should be true not just for non-white individuals in white spaces, but for all individuals when their race mismatches their surroundings. That is not to say that non-white people who are out of place face risk of arrest that is equal to white people who are out of place—non-white people are likely to face greater risk of arrest overall than are their white counterparts. Instead, both white and non-white people will face heightened risk of arrest when out of place in boundary zones than when in place and in boundary zones. The critical comparison is to other members of one’s racial group.

Only our data on police stops in Austin, Chicago, and Milwaukee includes the race of the civilian involved in a given interaction with police. We therefore leverage this data to evaluate the likelihood of being stopped among subsamples of white and non-white civilians, conditional on boundary status and racial composition of neighborhood.Footnote ⁹ The results of this analysis are displayed in figure 7 (the associated regression estimates can be found in appendix table 18). In Austin, both white and non-white individuals are more likely to be stopped when in boundary zones irrespective of which side of the line they are on. In Chicago and Milwaukee we observe that the fish-out-of-water heuristic is operative for both white and non-white individuals as within boundary zones, they face a heightened risk of arrest when they appear out of place on the basis of race. On balance, we take this as supportive evidence for our fourth hypothesis.

Figure 7 Heterogenous influence of racial boundary on police stops by race of civilian

Notes: X-axis is the percent white measure. Y-axis is the predicted logged number of police stops, by race of civilian stopped. Each row characterizes a different city sample. Estimates from fully-specified models with controls at means. 95% CIs displayed from HC2 robust block group-clustered SEs.

There is variation across cities in terms of how racialized policing expresses itself vis-à-vis boundaries. Hypothesis 3 anticipates that the probability of arrests will be higher in boundary zones than in non-boundary zones irrespective of which side of the boundary one is on. We do observe some differences in risk of arrest between boundary and non-boundary zones conditional on percent white, but which side of the line receives more attention from law enforcement differs by city. With respect to Hypothesis 4, we consistently observe that both white and non-white people are more likely to be stopped when they are racially out-of-place in boundary zones than when in non-boundary zones. Here, we leverage more sensitive data in the form of stops of civilians, but we only have this data for three of six cities. For this reason, it is difficult to make sense of inter-city heterogeneity on these points. Setting data differences aside, these heterogeneous patterns highlight that policing is fundamentally a local phenomenon. We have argued that police departments broadly embrace a style of policing that gives officers discretion, instructs them how to preempt crime, that doing so is a fundamental task of police work, and that policing racial boundaries is part-and-parcel to the task of maintaining social order. We think we have offered evidence that is sufficiently consistent to support that argument. Nevertheless, that there is some heterogeneity across contexts highlights the importance of cross-city approaches (as we have done here), and the ongoing need for scholars to verify and validate their theories outside of a single city, an approach that continues to define much of modern policing scholarship.