The Boys in Blue Are Watching You: The Shifting Metropolitan Landscape and Big Data Police Surveillance in the United States

Abstract

Despite decades of crime decline, police surveillance has continued to expand through a range of tactics oriented towards policing social disadvantage. Yet, despite attention to the linkages between residential inequality and policing, few studies have accounted for two intertwined structural developments since the turn of the 21st century: (1) the shift away from spatially concentrated patterns of racial segregation within urban centers towards sprawling patterns of economic segregation and (2) the turn from reactive policing towards proactive surveillance. Using the case of big data policing, we create a new measure of big data surveillance in metropolitan areas to examine how changes in segregation have affected the expansion of proactive police surveillance. In contrast to theoretical accounts emphasizing the role of police surveillance in governing economic inequality and perpetuating racial segregation, we do not find evidence that racial segregation or income inequality increase big data surveillance. Instead, much of the recent rise in big data policing is explained by increases in sprawling patterns of income segregation. These results provide new insight into the linkages between policing and residential inequality and reveal how changes in metropolitan segregation influence criminal justice surveillance in the era of big data.

Policing plays a central role in creating and perpetuating social inequality. Police contact has been linked to worsened mental health (Sewell, Jefferson, and Lee 2016; Sugie and Tuney 2017), poorer educational outcomes (Bernburg and Krohn 2003; Legewie and Fagan 2019), and risk of future incarceration (Bernburg et al. 2006). The recent spate of highly publicized police violence against young black men has invigorated critical attention to police oversight and social inequality (Edwards, Lee, and Esposito 2019). As Bell (2020a: 133 – 162) describes, police help to carve racial enclaves by criminalizing communities of color, patrolling residential boundaries, and constructing neighborhood reputations. These intertwined processes create a “deadly symbiosis” between criminal justice and residential inequality, where constant policing ensures that crime and poverty stay in communities of color (Wacquant 2001).

Although a long line of research explores how residential segregation influences policing (e.g., Bell 2020a, 2020b; Carmichael and Kent 2014; Chamlin and Liska 1992; Goffman 2009; Kent and Jacobs 2005; Liska and Chamlin 1984; Stults and Baumer 2007; Wacquant 2009), much of this research conceptualizes the urban environment through a 20^th century lens, where white flight in the mid-20^th century created “chocolate cities and vanilla suburbs” (Farley et al. 1978). Yet, since the turn of the 21^st century, policing and residential segregation have both changed in significant ways. On the one hand, policing strategies have moved away from reactive interventions that emphasize rapid responses to crime towards proactive strategies that utilize constant surveillance to predict outbreaks of crime and displace visible disorder (Fagan and Ash 2017; Herbert, Beckett, and Stuart 2017). On the other, economic segregation between inner cities and surrounding suburbs has steadily increased while racial segregation within city centers has declined (Lacy 2016; Massey and Tannen 2017; Reardon et al. 2018). Consequently, whereas 20^th century segregation was characterized by pronounced within-municipality racial segregation, 21^st century segregation is characterized by “moderate racial-ethnic segregation and rising class segregation” (Massey, Rothwell, and Domina 2009: 74), especially class segregation between urban centers and suburban municipalities (Frey 2015; Lacy 2016).

Despite these shifts in policing philosophy and metropolitan segregation, few studies have analyzed how changes to the residential landscape affect police surveillance in the 21^st century. Consequently, although past research describes how new policing tactics can reinforce and deepen residential inequalities, few studies have examined why proactive policing has grown more common in recent years or whether the rise of economic segregation has contributed to the influx of police surveillance. This is an important omission, as key theoretical claims about the relationship between criminal justice oversight and inequality hinge on assumptions about the black-and-white nature of residential segregation and its attendant consequences for differential exposure to the criminal justice system (e.g., Blalock 1967; Herbert et al. 2017; Wacquant 2001).

In this study, we examine why proactive police surveillance has expanded through the case of big data policing, a particularly recent policing innovation. Big data policing refers to a range of technological practices used by law enforcement to gather and process surveillance data on people, places, and populations and to forecast where and when crime is likely to occur. Although data-driven tactics offer to reduce racial biases and promote efficiency (Ridgeway 2018), a growing body of scholarship notes that, in practice, big data tactics tend to produce similar inequalities as are observed in traditional policing (Brayne 2020; Ferguson 2017). For instance, in China, big data police surveillance has increased political suppression (Xu 2021). Scholarship also documents stark negative consequences of institutional surveillance outside of as well as within the criminal justice system, including decreased interaction with hospitals, schools, and labor markets (Brayne 2014; Haskins and Jacobsen 2017), worsened mental health (Sewell et al. 2016; Sugie and Tuney 2017), and economic precarity (Hughes 2020). Thus, big data surveillance not only carries theoretical implications for scholarship on segregation and policing, but for stratification and state surveillance as well.

Leveraging novel data from the Atlas of Surveillance (AOS) to create a new measure of big data police surveillance in metropolitan areas, we evaluate whether, how, and to what extent changes in metropolitan segregation have influenced the expansion of big data surveillance between 2009 and 2019. In doing so, we make two primary theoretical and empirical contributions. First, while there is growing interest in the use of big data in the criminal justice system (Brayne and Christin 2020; Ferguson 2017; Harcourt 2007; Lageson and Maruna 2018), few studies have measured big data surveillance and, consequently, inquiries into the broader patterning of big data policing have been stunted. By developing a new measure, our analysis provides the first look into the prevalence of big data police surveillance in the U.S.

Second, although prior research on segregation and policing focuses on racial segregation (e.g., Kent and Jacobs 2005; Liska and Chamlin 1984; Stults and Baumer 2007), few studies have accounted for the growth of between-municipality economic segregation in recent decades. Yet, to demonstrate that new styles of segregation produce policing adaptations, as claimed in theoretical accounts (Bell 2020a; Fagan and Ash 2017), it is necessary to examine the most relevant type of segregation at the level of analysis where it is most meaningful. We advance this literature by considering how sprawling patterns of economic segregation may have a unique influence over police surveillance beyond what can be explained by racial segregation or income inequality at large. By advancing and evaluating a new explanation for the policing-segregation relationship that focuses on macro economic segregation, we move past prior research’s focus on racial segregation to provide new insight to how shifts in the type and scale of segregation impacts policing in the 21^st century.

SEGREGATION AND POLICE SURVEILLANCE

In the 1990s, law enforcement underwent a fundamental shift away from a focus on felony offenses and towards “broken windows” philosophies that relentlessly targeted minor forms of disorderly behavior (Kelling and Coles 1996; Zimring 2011). The number of police officers per capita in major U.S. cities increased by 13% between 1992 and 1999, while police expenditures increased by 15% (McCarty et al. 2012). Since the 1990s, declining crime rates have forced law enforcement to reconceptualize policing. Instead of rapid responses to ongoing crimes, many departments now regard order maintenance to be their primary function (Herbert et al. 2017; Vitale 2008). Police increasingly respond to minor complaint calls (Herring 2019), oversee business districts (Beckett and Herbert 2009), and monitor poor neighborhoods for visible disorder (Herbert et al. 2017; Stuart 2016).

Prior research describes how the influx of police surveillance has helped to reify and perpetuate social inequality. As Kramer and Remster (2021) describe, contemporary policing enacts “slow violence” by utilizing practices and tactics that decrease academic achievement, worsen health, and generate cultural trauma in overpoliced neighborhoods. Although historically reactive policing created such inequalities by espousing policies that encouraged aggressive encounters with suspects for petty crimes (Zimring 2011), contemporary policing strategies produce social inequalities by increasing police oversight and contact (Herring 2019; Legewie and Fagan 2019). For instance, as Beckett and Herbert (2009) describe, efforts to limit reactive policing in Seattle have resulted in aggressive patrols, where officers utilize constant surveillance to detect and displace poor and homeless populations from public view.

The increasing dependence on police surveillance for regulating social marginality has directed scholarly attention towards segregation as a characterizing feature of highly policed areas (Bell 2020a; Fagan and Ash 2017; Goffman 2009; Herbert et al. 2017; Wacquant 2001). Two explanations have been proposed for how segregation can impact police surveillance. First, racial threat theorists argue that segregation is inversely associated with police oversight (Blalock 1967: 121 – 126). To racial threat theorists, a primary function of the criminal justice system is to control the economic and political influence of racial minorities. Within this vein of theory, segregation acts a social control mechanism that suppresses minority group competitive power (Kent and Jacobs 2005; Liska 1992; Stults and Baumer 2007). By concentrating minorities in disadvantaged neighborhoods, inequalities resulting from segregation are sufficient to protect white economic and political interests. Hence, when segregation is high, police surveillance should decrease as there is no need for additional social control mechanisms.

A second hypothesis comes from criminological theory on the expansion of the carceral state. The carceral state hypothesis argues that segregation should increase a range of crime control tactics, including police surveillance. Here, neoliberal economic policies produce high levels of poverty in racially segregated areas (Wacquant 2009). The resulting critical mass of disadvantage in predominantly black neighborhoods creates a “deadly symbiosis” between the criminal justice system and urban ghettos, where criminal justice interventions act as the primary apparatus used to regulate social precarity (Brydolf-Horwitz and Beckett 2021; Wacquant 2001: 92 – 93). Police surveillance acts as one among many criminal justice interventions intended to control and monitor inequality in racially segregated areas. Echoing this reasoning, Bell (2020a: 140), argues that police reinforce residential inequality by patrolling poor neighborhoods of color and by monitoring the boundaries of predominantly white and affluent neighborhoods.

Despite the central role of segregation in theoretical accounts of policing, evidence that segregation influences policing outcomes is inconsistent across study designs and levels of analysis. Ethnographies find that police presence is typically greater in black neighborhoods and that people of color are disproportionately exposed to police contact (Goffman 2009; Herring 2019; Rios 2011; Stuart 2016). Tract-level studies similarly report positive effects of racial concentration on policing (Zhao, Yang, and Messner 2018). For instance, 311 complaint calls are more common in tracts where otherwise segregated ethnic groups come into contact (Legewie and Schaefer 2016). Studies conducted at the county-level, however, do not find an association between segregation and police force size (Stults and Baumer 2007). At the city-level, Carmichael and Kent (2014) report a positive association between segregation and police force size, while Chamlin and Liska (1992) report a negative association. Kent and Jacobs (2005) report regional differences, where segregation is negatively associated with police force size in southern cities but positively associated elsewhere. Beck and Goldstein (2018), alternatively, find no association between segregation and police expenditures.

Thus, although theoretical accounts provide divergent expectations on the relationship between segregation and policing, evidence has been mixed. One possible reason for these inconsistencies may be that prior studies have not accounted for the most relevant form of segregation at the unit of analysis where it is most impactful. Although 20^th century segregation was characterized by black-white segregation within urban centers (Massey and Denton 1993), the 21^st century is characterized by rising economic segregation between cities and surrounding suburbs. In this article, we examine how and why the recent increase in economic segregation may influence police surveillance beyond the independent contributions of racial segregation and income inequality that have been highlighted in prior theoretical accounts.

(SUB)URBAN INEQUALITY AND POLICE SURVEILLANCE IN THE 21^st CENTURY

Two recent changes in residential inequality may affect the segregation-policing relationship. First, decades of suburbanization have caused the dominant form of segregation to be between inner-cities and suburban municipalities, as opposed to the neighborhoods that lie within city centers (Lacy 2016; Light and Thomas 2019; Massey and Tannen 2017). Second, economic segregation has increased while racial segregation has declined (Massey et al. 2009; Reardon and Bischoff 2011). Between 1970 and 2010, extreme racial segregation decreased by 40% (Massey and Tannen 2015), while income segregation rose by 25% (Reardon et al. 2018). While historically suburbanization was associated with white flight from inner cities (Logan et al. 2004), the simultaneous decrease in racial segregation and increase in income segregation since 2000 has been driven by minority households moving into mostly white, suburban areas (Frey 2015; Lacy 2016; Reardon et al. 2018). Whereas only 18% of metropolitan black households lived in suburban areas in 1970, 40% lived in the suburbs by 2010 (Massey and Tannen 2017).

Despite these changes to residential inequality, few studies have accounted for the impact of between-municipality economic segregation on police surveillance. Yet, there are several theoretical pathways through which macro economic segregation can shape policing. First, income segregation can increase reliance on police to stop crime and poverty from spilling into suburban municipalities. As Bell (2020b: 943 – 946) describes, while poor families tend to regard police as a nuisance, wealthy families tend to regard police as an amenity, particularly because police presence helps to exclude poor families from middle class suburbs. For example, although crime rates are far higher in inner cities, arrest rates in city centers are comparable to suburban areas (Beck 2019). Homeowners also tend to respond more positively to police presence than renters (Reisig and Parks 2000), and some research finds that racial differences in attitudes towards the police can be explained by differences in economic conditions (Ashcroft, Daniels, and Hart 2003; Schuck et al. 2008).

Second, income segregation can increase police surveillance by increasing the salience of perceived victimization risk when middle-class families travel to city centers. Although between-municipality segregation provides a means for high income households to distance themselves from inner cities, commuters still travel to downtown districts for work and leisure where they typically encounter visible disorder (Vitale 2008). Beck (2020), for instance, finds that growth in the number of middle-class households in gentrifying neighborhoods is associated with increases in 311 complaint calls.¹ Third, income segregation can increase police surveillance by increasing dependence on police for economic growth. As Herbert et al. (2017, 1497) describe, “business interests increasingly call on police to… revitaliz[e] downtowns, maximizing the profitability of retail spaces and providing boosts to the tourism and convention industry.” Because urban revitalization relies on middle-class consumption (Hochstebach and Musterd 2018), the safety concerns that originate among middle-class suburbanites can exert political pressure on officials who seek to promote shopping, dining, and business in downtown districts.

In sum, prior research provides several hypotheses for how metropolitan residential inequality can influence police surveillance. First, the threat and carceral state hypotheses offer competing claims that racial segregation will either be positively or negatively associated with police surveillance. Second, the carceral state hypothesis suggests that there will be a positive interaction between racial segregation and economic inequality, where a critical mass of inequality in segregated areas increases police surveillance. Third, we argued that economic segregation can have an independent positive effect on police surveillance beyond what can be explained by racial segregation or economic inequality at large. We now introduce our case of big data policing before describing our data and measurement strategy.

THE RISE OF BIG DATA SURVEILLANCE

Although prior research describes law enforcements’ growing reliance on proactive surveillance (Bell 2020a; Fagan and Ash 2017; Goffman 2009; Herbert et al. 2017), few studies have examined the expansion of police surveillance in recent decades. Research on policing tactics using post-1990s data is either ethnographic, documenting the experiences of the targets of police surveillance and the surveillance practices of officers (Goffman 2009; Herring 2019; Rios 2011; Stuart 2016), or only measures police surveillance indirectly, for instance, by using stop-and-frisks and police force size (e.g., Carmichael and Kent 2014; Sewell et al. 2016). While important for analyzing officer allocation and behavior, these measures are unable to disentangle the proactive surveillance tactics central to theoretical accounts of contemporary policing from police contact more generally. We address this limitation by focusing on the case of big data policing, a particularly recent innovation in proactive surveillance.

Big data policing entails the use of digital technologies to collect, aggregate, and analyze surveillance data on populations in a manner that improves departments’ (perceived) ability to control crime (Ridgeway 2018). Surveillance data are compiled by triangulating information from a diverse range of sources, processed using specialized algorithms, and circulated through automated alert systems that direct officer presence. A growing number of scholars have detailed how officers incorporate surveillance data into their routine behaviors (Brayne 2020; Rios, Prieto, and Ibarra 2020; Sausdal 2020). Rios et al. (2020), for example, find that police blend traditional and big data surveillance methods to monitor criminal suspects, requesting social media information during stop-and-frisk searches to link to online databases.

A central theme to emerge from this literature is that big data surveillance enables police to monitor public spaces and populations far beyond the neighborhoods that officers patrol. A core feature of big data policing is that law enforcement collects fine-grained data on mobility patterns and social activities en masse (Brayne 2020: 37 – 54). Data are not obtained selectively to surveil individuals suspected of crime. Instead, surveillance technologies are coordinated to provide a stream of information that can be monitored for suspicious activity in metro areas at large (Didier 2018). As Brayne (2020: 23 – 24) describes, “The tendency for data or systems initially intended for one purpose to be used for another … is a fundamental component of the big data [policing] landscape.” This type of “function creep” means that even technologies that offer to improve organizational efficiency and increase police oversight can contribute to big data surveillance. For instance, in Los Angeles, the implementation of automated license plate readers (ALPRs) required creating queryable databases to store and access the glut of image data. Although ALPRs are a non-intrusive technology, the image data from ALPRs is often linked to video analytic software to identify unregistered license plates (Brayne 2020). Similarly, although offering to increase police accountability, body-worn cameras (BWCs) provide officers with live-stream footage of citizen encounters and are often linked to facial recognition software to identify suspects with outstanding warrants (Hood 2020). Such technologies enable police surveillance to become more “programmatic, ongoing, cumulative, … and embedded into [law enforcement’s] routine activity” (Brayne 2020: 49).

While prior studies detail how police increasingly gather and analyze surveillance data, the literature on big data policing is limited in two important ways. First, prior research has been ethnographic, directing attention to how officers react to, make use of, and navigate the range of technological practices within their routine patrols (Brayne 2020; Rios et al. 2020; Sausdal 2020). While providing important insight to the organizational aspects of big data, we know less about the prevalence of big data surveillance or why it has increased. Second, because prior research has focused on organizational incentives, this literature has developed independently from theory and research that ties criminal justice expansion to residential inequality (Herbert et al. 2017; Wacquant 2009). We advance this literature by evaluating whether and how changes in metropolitan segregation influence the recent expansion of big data police surveillance.

CURRENT STUDY

Although prior research examines the relationship between traditional police behaviors and segregation, few have accounted for how segregation has changed in recent decades or its effects on contemporary policing. Through the case of big data policing, we evaluate this relationship to make several contributions to the literatures on segregation, policing, and state surveillance. First, we move past prior studies’ emphasis on racial segregation within city centers to examine how between-municipality income segregation may have a distinct effect on police surveillance. In doing so, we draw attention to how the type and scale of segregation can impact conclusions about the segregation-policing relationship. Second, we develop a new measure of big data police surveillance that allows us to empirically test our hypotheses on income segregation, racial segregation, and income inequality. In doing so, we provide the first insight into aggregate trends and the prevalence of big data police surveillance.

DATA SOURCES

We utilize balanced panel data on 381 metropolitan statistical areas (MSAs) observed annually between 2009 and 2019 (N = 4,191). As described in prior research (Jargowsky and Park 2009; Light and Thomas 2019), MSAs incorporate both suburban areas and city centers and are therefore the most appropriate unit of analysis for examining how the suburban-inner city divide influences police surveillance. While racial segregation increased in the mid-20^th century due to white flight from city centers (Light and Thomas 2019), suburban diversification since the turn of the 21^st century has contributed to a simultaneous decline in racial segregation and an increase in economic segregation (Lacy 2016). The timeframe is thus appropriate for examining racial integration, economic segregation, and their respective effects on big data policing.

Measuring Big Data Police Surveillance

Our primary data source is the Atlas of Surveillance (AOS). The AOS data seeks to provide agency-level information on surveillance data collection and processing strategies for every police department in the U.S., including municipal, county, and state agencies. The AOS data are collected by a team of over 20 data journalists and scholars as part of the Electronic Frontiers Foundation. Each entry is obtained from public records, including federal documents, individual police departments, freedom of information act requests, official state audits, nonprofit organizations, existing databases, press releases, and sales records of technology vendors. This procedure reflects a growing trend towards measuring police behavior using public records to subvert the biases that emerge from official statistics (Edwards et al. 2019).²

Since the AOS data are newly released, we have taken several steps to validate the coverage of the data. We first assessed whether there was error in how measures obtained from existing databases were entered. We compared the measures of drone usage that are aggregated from the Bard college dataset (Gettinger 2020) and measures of fusion centers obtained from the Department of Homeland Security (DHS 2020) to evaluate any inconsistencies. Both had 100% correspondence. Next, we compared measures to alternative databases where available. First, we correlated the use of BWCs with the Law Enforcement Management and Administrative Statistics (LEMAS). Although the LEMAS data is known to suffer from measurement error due to the voluntary nature of the program, the two measures correlate at the MSA-level at .73. Second, we compared the measures of real-time crime centers to the RAND data (Hollywood et al. 2019, 2 – 4). While the RAND data is not comprehensive, we identified no cases that appeared in the RAND data and did not appear in the AOS data. Collectively, these results indicate that the AOS data is consistent with existing measures where available and exhibits reasonable amounts of divergence where errors are known to exist for alternative datasets.

Big data surveillance encompasses a range of tactics used to “scrutinize individuals, groups, and contexts through the use of technical means to extract [and analyze] information” (Marx 2016: 20). It relies on integrating distinct surveillance technologies to increase the “volume, veracity, and variety” of data available to monitor public areas and population behavior (Ridgeway 2018: 402 – 403). A measure of big data policing must therefore account for a range of surveillance technologies that increase the size and diversity of surveillance data. The measure must also include surveillance tactics that span multiple departments, as a primary goal of big data policing is to improve information sharing and coordination among law enforcement at multiple levels of government and in different municipalities (Bloss 2007; Didier 2018). For instance, drones, real-time crime centers, and fusion centers represent powerful surveillance tools, but are rarely owned by a single department, instead being shared by multiple departments, often in different jurisdictions (Bloss 2007; Gettinger 2020; Hollywood 2019). Finally, a measure of big data surveillance must account for “function creep,” where technologies that may not be adopted with the goal of enhancing surveillance are routinely used to increase surveillance once the technology becomes available (Brayne 2020: 23 – 24).

We take a latent variable approach that utilizes an index of surveillance technologies that covary with a single underlying construct. Following the above criteria, we included technologies that span a range of intrusive (gunshot sensors, cell-site simulators, facial recognition, predictive policing) and unintrusive (ALPRs, BWCs, video analytics) techniques, that are often shared between levels of government and different municipalities (real-time crime centers, drones, fusion centers), and that are routinely integrated with data systems to increase surveillance, even when those technologies may not be initially designed for those goals (ALPRs, BWCs). We consider a total of 10 technologies to create our measure: ALPRs, BWCs, cell-site simulators, drones, facial recognition software, fusion centers, gunshot sensors, predictive policing software, real-time crime centers, and video analytics. Descriptions of each technology are provided in Table 1. Each technology has been highlighted in prior research as reflecting or contributing to big data policing, meaning that our measure is consistent with theoretical conceptualizations of big data surveillance that permeate the literature (Bloss 2007; Brayne 2020; Gettinger 2020; Hood 2020; Lyon 2013; Marx 2016).

Table 1.

Description of Surveillance Technologies in 381 Metropolitan Areas, 2009 – 2019.

Surveillance Technologies	Correlation with surveillance indexa	Number (%) of technology-using police departments	Description
Body-worn Cameras	.79	971 (40.4%)	Portable cameras attached to the chests of police officers. Some cameras stream live and can be combined with facial recognition software.
Drones	.82	733 (30.5%)	Unmanned aerial vehicles used to monitor crowds when on-the-ground surveillance is difficult or dangerous.
Real Time Crime Center	.62	50 (2.1%)	Hubs where police analyze surveillance data in real time. Analysts make use of a variety of data and processing strategies, including automated license plate readers, predictive policing, face recognition, cell-site simulators, and gunshot detection sensors.
Gunshot Detection Sensors	.80	74 (3.1%)	Technology used to track loud, gunshot-like noises. Sensors are typically mounted on streetlights or on the sides of building. When triggered, the sensor triangulates the estimated location of the gunfire and sends the information to law enforcement.
Video Analytics	.60	20 (.8%)	Software applied to video feeds to detect movement patterns that might be indicative of a crime or terrorist event.
Cell-site Simulator	.30	56 (2.3%)	Devices that behave like cellphone towers, connecting to cellphones to collect personal data.
Fusion Centerb	.26	59 (2.5%)	Hubs that share intelligence and surveillance data between state, triable, territorial, and federal agencies in real time. They are similar to real-time crime centers but focus on non-local matters and are funded by the US Department of Homeland Security.
Automated License Plate Reader	.79	348 (14.5%)	Cameras attached to fixed locations or police patrol cars that capture every license plate that passes and upload the picture to searchable databases.
Facial Recognition Software	.61	59 (2.5%)	Software to automatically identify an individual from a picture, video, or in real-time using facial features. Law enforcement may also use mobile face recognition software to identify people during police stops in real time.
Predictive Policing	.64	32 (1.3%)	Artificial intelligence used to help police decide which neighborhoods, individuals, or blocks they should devote resources to during specific time periods. Law enforcement feed surveillance data such as police stops, past records of violence, and calls for service, as well as crime data into the algorithms to help predict where and when crime will occur, and who is likely to be involved.
Total		2,402

^aCorrelation is calculated after excluding the focal surveillance technology from the index.

^bDates for the construction of new fusion centers were recorded by examining fusion center websites, executive orders, and official documentation from the DHS National Network of Fusion Center 2011 – 2019 reports.

To create the index, we aggregated the number of surveillance technologies in each police department to the MSA-level by linking the department-level identifiers to their respective metro areas. Then, we divided the sum by the number of departments in each MSA to correct for exposure (i.e., MSAs with a greater number of departments tend to use a greater number of surveillance technologies). The surveillance index can be interpreted as the average number of surveillance technologies in use per department. The index is highly reliable with a Cronbach’s α of .85. Confirmatory factor analysis is also consistent with the existence of a single latent construct. Factor loadings range from .478 to .709 and are all significant at p < .001. Model fit is strong, with a root mean squared error of approximation of .037, standardized root mean residual of .031, Tucker-Lewis fit index of .921, and confirmatory fit index of .948. Further, each of the surveillance technologies positively correlates with the surveillance index when omitted from the measure (see Table 1). This means that the variation in the surveillance index is not unduly dependent on a single technology.³ Collectively, these results indicate that the surveillance index has high reliability, face validity, and construct validity. We provide additional discussion of our measurement strategy in the Supplementary Materials along with sensitivity analyses.

Segregation Measures

Our focal independent variables are income segregation, racial segregation, and income inequality. Many income segregation measures are defined, in part, by income inequality metrics, meaning that it is difficult to disentangle income segregation from income inequality. To overcome this issue, we use the entropy-based rank-ordered information theory index (Reardon and Bischoff 2011). The benefit of the measure is that it ranks the share of the population that fall above and below a given income threshold. Hence, the measure is conceptually distinct from income inequality (Reardon and Bischoff 2011: 1108 – 1110). The measure obtains a maximum of 1 when all households within a tract have the same income but household income differs between tracts and obtains a minimum of 0 when all households within a tract have different incomes.⁴ We measure racial segregation with the black-white dissimilarity index. The dissimilarity index ranges between 0 and 1 and captures the proportion of the black or white population that would have to relocate to achieve uniform dispersion. We measure income inequality using the Gini income inequality index.

We use the American Community Survey (ACS) 5-year estimates to calculate each measure.⁵ The rank-ordered information theory index can be biased when constructed from ACS estimates due to low sampling rates in some tracts (Logan et al. 2018). We correct the rank-ordered information theory index for sampling bias using the method proposed by Reardon et al. (2018). The method involves calculating the approximate bias in the rank-ordered information theory index using ACS tract sampling rates and estimated population sizes and then subtracting the bias from the segregation index (Reardon et al. 2018: 2135 – 2137). Because segregation measures can be sensitive to population size in small MSAs (Logan et al. 2018; Napierala and Denton 2018), we conducted sensitivity analyses only examining MSAs with populations greater than 100,000 and 200,000. Results are consistent with those reported below.

Control Variables

Public statements on the organizational benefits of big data surveillance emphasize thwarting violent crime and terrorism (e.g., Bloss 2007). We measure the violent crime rate per 10,000 capita and drug arrest rate per 10,000 capita using data from the Uniform Crime Reports. We also include a measure of whether an MSA was the target of recent terrorist violence using data from the START Global Terrorism Database (e.g., Ramey and Steidley 2018). We use a three-year rolling time-window because we expect that terrorist attacks will have lasting effects on surveillance efforts. Prior research suggests that highly policed cities tend to endorse big data tactics (Brayne 2020; Didier 2018). We account for police force size by controlling for the number of sworn-in officers per 10,000 capita. We also control for police expenditures using data from the Annual Survey of State and Local Governmental Finances. The measure is the amount spent on police protections among constituent cities, counties, and townships in tens of thousands of inflation-adjusted dollars (e.g., Beck and Goldstein 2018).

A prominent explanation for the recent expansion in big data surveillance is that digital technologies are now more widely available in all sectors of public and private life (Bloss 2007; Brayne 2020). We include several variables to account for technological determinism. First, if big data policing is increasing because there is a universal expansion in the availability of surveillance technology, this would imply a uniform trend towards big data surveillance. We control for this by including a vector of year fixed effects, which holds all uniform trends constant. This specification also offers the benefit of controlling for all possible period effects, the most important of which for our purposes is the nationwide push for BWCs after the 2014 killing of Michael Brown.⁶ While year fixed effects control for uniform trending, it is possible that market forces may make technology hubs, like Silicon Valley, epicenters for big data surveillance. We control for local market forces by including a binary variable equal to 1 if a technology vendor is headquartered in an MSA. Finally, the effect of technological availability may vary temporally, as technologies hit local markets at different points in time. We address this by including interactions between the technology vendor variable and the year fixed effects.⁷

In addition to controls for technological determinism, we account for demographic factors and labor market conditions that might influence police surveillance. Threat theory argues that ethnic, racial, and economic threats lead to increased social control practices (e.g., Blalock 1967; Liska 1992). We account for racial, ethnic, and economic threat by including measures of the percent black population, percent Latino population, the poverty rate, and the unemployment rate. We also control for population density using the population per square mile to account for the possibility that MSAs with diffuse populations rely more so on big data surveillance than MSAs with dense populations (e.g., Fagan and Ash 2017). Finally, we control for socioeconomic and political factors by including the median household income, proportion of the population that holds a baccalaureate degree, and the proportion of voters who voted for a Republican presidential candidate in the most recent election. These controls account for the possibility that areas that are more affluent, more highly educated, or more conservative may be more inclined to expand police surveillance through big data tactics. We also include a vector of region indicator variables to account for regional differences. Descriptive statistics for all variables are provided in Table 2.

Table 2.

Descriptive Statistics for 4,191 MSA-years (381 MSAs).

Variable	Mean (SD) or %	Range
Surveillance index	.116 (.222)	.00 to 2.00
Rank-order information theory index	.079 (.029)	.014 to .165
Black-white dissimilarity index	.498 (.097)	.243 to .797
Gini inequality index	.452 (.024)	.388 to .544
Violent crime rate per 10,000 capita	32.936 (19.692)	.026 to 128.740
Drug arrest rate per 10,000 capita	49.720 (59.490)	.011 to 2,888.091
Terrorist attack in the last 3 years	8.66%	0 to 1
Police officers per 10,000 capita	5.970 (3.861)	.053 to 47.015
Police expenditures in tens of thousands of dollars (inflation adjusted)	18.859 (61.965)	.013 to 1026.864
Technology vendor in MSA	10.79%	0 to 1
Percent black	11.643 (10.870)	.563 to 55.162
Percent Latino	13.022 (15.663)	.718 to 95.731a
Percent 15 – 24 year-old male	7.516 (1.944)	2.280 to 19.706
Percent with a baccalaureate degree	10.519 (4.700)	3.200 to 34.000
Median household income in thousands of dollars (inflation adjusted)	52.163 (99.230)	31.264 to 122.478
Unemployment rate	4.765 (1.433)	.800 to 10.800
Percent living in poverty	10.756 (3.573)	3.500 to 31.000
Population per square mile	242.887 (255.139)	7.152 to 2,329.380
Percent Republican voters	51.919 (12.422)	16.63 to 88.284
N of MSAs	381
T	11
NT of MSA-years	4,191

^aThe maxima for the percent Latino metric is from Laredo Texas, an MSA with a population of roughly 235,000 on the Mexican border.

Analytic Strategy

The surveillance index is overdispersed (dispersion = 12.361, p <.001). Log-transforming such a variable and fitting a linear model would yield biased coefficients (Osgood 2002). We address overdispersion by modeling the MSA-level count of surveillance technologies in a negative binomial regression and specifying the natural logarithm of the number of police departments as an offset parameter. This is algebraically equivalent to modeling the number of surveillance technologies in use per department, but coefficients are not biased by bimodality (Osgood 2002). Hence, the exponentiated coefficients can be interpreted as increasing/decreasing the number of surveillance technologies in use per police department.

Given our interest in both between- and within-MSA variation in big data surveillance, we use a between-within model (Allison 2009).⁸ The between-within model decomposes the variance in the independent variables into separate between and within coefficients. The within coefficients are identical to a fixed effects model in that they control for all time invariant confounders and reflect within-MSA change. The between coefficients reflect the effects of time invariant MSA averages on time invariant between-MSA differences. In practice, the model is estimated as a multilevel model with an MSA-level random intercept. Note that this specification controls for unobserved heterogeneity by holding MSA-level differences constant (for the within coefficients) and by including an MSA-level random intercept (for the between coefficients). Table S1 of the Supplementary Materials presents supplementary analyses that address the possibility of reverse causation. No variance inflation factor is above 4 in our analysis, indicating that multicollinearity is not a problem.

RESULTS

Figure 1 plots the geospatial distribution of the surveillance index. While the mean of the surveillance index is .116, the range is .00 to 2.00, reflecting substantial geospatial variation. Consistent with expectations, the most highly surveilled areas tend to be segregated along class lines. For instance, two of the most heavily surveilled MSAs, Stockton CA and Trenton-Ewing NJ, score .102 and .153 on the rank-ordered information theory index (above the national average), and 1.71 and 1.40 on the surveillance index, respectively. Figure 2 reveals that big data surveillance has increased both steadily and sharply since approximately 2012. While the mean of the surveillance index was .011 in 2009, it was .348 by 2019—a 32-fold increase in 10 years. Consistent with expectations, Figure 2 reveals a close temporal association between the rank-ordered information theory index and the surveillance index (r = .82). Consistent with the racial threat hypothesis, but inconsistent with the carceral state hypothesis, there is an inverse relationship between the black-white dissimilarity index and the surveillance index.

Figure 1.

Surveillance Index in MSAs using 2019 values. Rounded to nearest quintile.

Figure 2.

Over-time Trends in Surveillance Index.

Descriptive results align with the expected association between income segregation and big data surveillance and with the threat explanation for racial segregation. We now turn to between-within models to provide a formal assessment. Models 1 – 3 present bivariate results from regressing the surveillance index on segregation and inequality indices (Table 3). In line with expectations, the rank-ordered information theory index is positively associated with both between- and within-MSA differences in the surveillance index. However, neither the black-white dissimilarity index nor Gini index within coefficients are significant. The between coefficient for the Gini index is positive, indicating that MSAs with, on average, higher income inequality tend to, on average, score higher on the surveillance index. Consistent with threat theory, the between coefficient for the black-white dissimilarity index is negative.

Table 3.

Restricted Negative Binomial Between-within Models of Surveillance Index with Variable Exposure and Inequality and Segregation Measures (N = 4,191).

	1	2	3	4		5	6	7
Within MSA effects
Rank-ordered information theory index	9.366** (3.251)			10.268** (3.239)			9.790** (3.367)	10.600** (3.350)
Black-white dissimilarity index		−.624 (.844)		−.708 (.854)		−.636 (.846)		−.707 (.854)
Gini inequality index			−.332 (1.354)			.078 (1.349)	−1.595 (1.142)	−1.333 (1.401)
Between MSA effects
Rank-ordered information theory index	10.079***  (2.131)			13.340***  (2.127)			9.932***  (2.364)	13.117***  (2.345)
Black-white dissimilarity index		−2.192***  (.650)		−3.547***  (.657)		−2.523***  (.637)		−3.488***  (.654)
Gini inequality index			5.734*  (2.905)			5.481*  (2.825)	4.078 (2.901)	3.113 (2.764)
Intercept	−6.930***  (.266)	−5.005***  (.382)	−11.143***  (1.295)	−5.393***  (.378)		−10.395***  (1.262)	−9.688***  (1.367)	−8.018***  (1.333)
MSA frailty term	1.203	1.178	1.195	1.068		1.105	1.187	1.054
Year fixed effects?	Yes.	Yes.	Yes.	Yes.	Yes.		Yes.	Yes.
Log-likelihood	−4249	−4256	−4253	−4235	−4245		−4246	−4232
AIC	8529	8542	8536	8505	8525		8525	8501
BIC	8624	8638	8632	8613	8633		8633	8622

^***p <.001

^**p<.01

^*p <.05.

Two-tailed tests. All models include the natural logarithm of the number of police departments as an offset parameter.

In Models 4 – 7 we evaluate confounding between each pair of segregation and inequality variables. Although the black-white dissimilarity index correlates with the rank-ordered information theory index (r = .29), it is a suppressing effect. Once the black-white dissimilarity index is controlled, the rank-ordered information theory index within coefficient increases by approximately 11% and the between coefficient increases by roughly 32%. Consistent with prior research (Reardon et al. 2018), these findings indicate that although racial and income segregation are intertwined, the positive effect of income segregation is independent from, and suppressed by, the negative effect of racial segregation. Moreover, the Gini index between effect becomes insignificant once the rank-ordered information theory index is controlled (Models 6 – 7). Consistent with expectations, this result indicates that income segregation is a powerful determinant of the surveillance index. In contrast to the carceral state hypothesis, we do not find evidence that either the Gini index or dissimilarity index are positively associated with the surveillance index. Instead, consistent with threat theory, the negative between coefficient for the black-white dissimilarity index is robust.

We now include control variables to assess the robustness of these results to alternative explanations. Models 8 – 14 tell a story consistent with Models 1 – 7. The dissimilarity index and Gini index within coefficients are not significant in any model, nor is the Gini between coefficient. The dissimilarity index between coefficient is again negative. As above, within-MSA increases in the rank-ordered information theory index are associated with increases in the surveillance index. And the rank-ordered information theory index between coefficient is positive and significant in all models where the black-white dissimilarity index is controlled. Turning to controls, we find mixed support for public safety explanations. On one hand, neither the between nor within coefficients are significant for either violent crime rates or drug arrest rates. On the other hand, the within coefficient for recent terrorist attacks is positive, indicating that terrorist events tend to increase big data surveillance. Consistent with expectations that traditional and big data policing are correlated (Brayne 2020; Rios et al. 2020), Models 8 – 14 uncover positive effects from the number of police officers per capita. Results also reveal positive within-MSA effects from Republican voting, which is consistent with prior work on police force size and militarization (Kent and Jacobs 2005; Ramey and Steidley 2018). The positive between coefficient for the percent Latino population is consistent with the ethnic threat hypothesis and with research finding that Latino men are disproportionately targeted for police surveillance (Blalock 1967; Rios et al. 2020). Collectively, results for controls replicate findings from prior research that report positive effects of terrorist violence, organizational capacity, conservative politics, and ethnic threats on police behavior.

The carceral state hypothesis suggests that the effect of racial segregation should be compounded by economic inequality. We test this by including interactions between each of the segregation and inequality indices. Models 15 – 18 reveal that none of the interactions are significant (Table 5). Nor does including interactions influence substantive conclusions about the rank-ordered information theory index, black-white dissimilarity index, or Gini index. Also of note is the sharp increase in AIC and BIC with respect to primary models, indicating worse model fit when interactions are included. In contrast to the carceral state thesis, but consistent with threat theory, this result indicates that the negative effect of racial segregation on big data surveillance cannot be attributed to an interaction with economic disadvantage.

Table 5.

Negative Binomial Between-within Models of Surveillance Index including Interactions (N = 4,191).

	15	16	17	18
Within MSA effects
Rank-ordered information theory index	9.993** (3.047)	9.702** (2.933)	11.710*** (3.006)	11.790*** (4.127)
Black-white dissimilarity index	−1.289 (.802)	−1.198 (.831)	−1.353 (.777)	−.1.285 (.851)
Gini inequality index	−5.037 (2.836)	−5.069 (2.835)	−5.504 (2.826)	−5.470 (2.827)
Information theory index x Dissimilarity index	30.320 (150.080)			17.270 (151.100)
Dissimilarity index x Gini index		−37.330 (84.350)		−16.110 (84.950)
Information theory index x Gini index			−93.940 (60.260)	−93.020 (62.560)
Between MSA effects
Rank-ordered information theory index	7.890*  (3.018)	7.254*  (3.208)	7.854*  (3.208)	7.600*  (3.354)
Black-white dissimilarity index	−.659 (2.123)	9.285 (13.460)	−2.387** (8.630)	6.588 (15.740)
Gini inequality index	6.307 (4.487)	11.918 (15.610)	7.574 (8.813)	13.780 (18.58)
Information theory index x Dissimilarity index	−20.040 (22.350)			−13.970 (26.800)
Dissimilarity index x Gini index		−25.780 (29.640)		−17.160 (37.120)
Information theory index x Gini index			−11.502 (86.100)	13.660 (92.000)
Intercept	−9.014 (10.910)	−14.720 (14.060)	−7.149 (11.420)	−12.750 (14.410)
MSA frailty term	.981	.981	.980	.989
Controls included?	Yes.	Yes.	Yes.	Yes.
Year fixed effects?	Yes.	Yes.	Yes.	Yes.
Year x Vendor in MSA fixed effects?	Yes.	Yes.	Yes.	Yes.
Log-likelihood	−4478	−4477	−4475	−4474
AIC	9054	9054	9050	9057
BIC	9371	9371	9367	9399

^***p <.001

^**p<.01

^*p <.05.

Two-tailed tests. All models include the natural logarithm of the number of police departments as an offset parameter. All variables were within-transformed before calculating interactions.

To facilitate substantive interpretation of our results, we now direct attention to effect size. On average, a one percentage point within-MSA increase in the rank-ordered information theory index produces a 12% (exp(.112)=1.12) increase in the surveillance index. Similarly, a one percentage point increase in the MSA-average of the black-white dissimilarity index correlates with an 2.3% (exp (−.023) =.977) decrease in the MSA-average of the surveillance index, and a one percentage point increase in the MSA-average of the information theory index correlates with a 7.8% (exp(.075)=1.078) increase in the MSA-average of the surveillance index. These results illustrate that the within-MSA change in income segregation has a particularly large effect on the recent rise of big data surveillance. We can assess what levels of surveillance might look like in 2019 if income segregation had not increased by holding the information theory index at its 2009 value and calculating the change in expectation. An MSA like Denver-Aurora-Bloomfield, CO would experience a 57.1% decrease in the surveillance index had income segregation not increased since 2009, while an MSA like Los Angeles-Long Beach-Anaheim would expect a 29.8% decrease (Figure 3). Nationally, we would expect a 29.0% mean decrease in the surveillance index in 2019 had income segregation not increased since 2009.

Figure 3.

Proportional Change in Predicted Value of Surveillance Index in 2019 when Rank-ordered Information Theory Index is Held at 2009 Value. Rounded to nearest quartile.

In sum, while theoretical accounts of the expanding carceral state argue that police surveillance is used to maintain and monitor residential boundaries in racially segregated areas, we find that income segregation, rather than racial segregation or income inequality at large, has contributed to much of the growth of big data surveillance since 2009. We do find partial support for the threat hypothesis, where time invariant differences in levels of racial segregation are inversely associated with time invariant differences in big data surveillance. However, we do not find the same support from the within coefficients, meaning that threat processes cannot account for the change in big data policing since 2009. Collectively, these findings reveal that changes in metropolitan segregation carry important implications for policing in the 21^st century.

DISCUSSION

Although theoretical accounts emphasize the important role of residential inequality in policing, few studies have accounted for how changes to metropolitan segregation in recent decades have influenced contemporary policing tactics. Drawing on threat theory and research on the expanding carceral state, we evaluated several explanations for the expansion of big data policing and advanced a new explanation that focuses on macro income segregation. Constructing a new index measure of big data police surveillance, we found partial support for the racial threat hypothesis. Metropolitan areas that are, on average, less racially segregated tend to, on average, have higher levels of big data surveillance. Alternatively, we did not find support for the carceral state hypothesis, which predicts that racial segregation and income inequality at large increase criminal justice oversight. Instead, our results lend the strongest support to the income segregation hypothesis, indicating that the recent expansion of big data surveillance is particularly sensitive to recent increases in spatial income inequalities.

Collectively, our findings scaffold arguments that segregation is central to understanding police behavior (Bell 2020a; Blalock 1967; Herbert et al. 2017; Wacquant 2001). As our analysis demonstrates, big data surveillance in the least economically segregated areas is often a fraction of the size of the most economically segregated areas. In this regard, our analysis reinforces claims that shared social space maps onto shared living experiences. In our case, exposure to police surveillance increases in areas where there is greater class-based segregation. And, it is important to note that because exposure to police contact increases risk of future incarceration (Bernberg et al. 2006), this means that more economically segregated areas may also have higher rates of collateral consequences from criminal justice contact (e.g., Sugie and Tuney 2017).

Our results also highlight the significance of changes in residential inequality for explaining the shift from reactive policing tactics to proactive surveillance. As Fagan and Ash (2017) reason, new styles of segregation beget new styles of policing. Although prior research suggests that many changes to policing tactics are an organizational response to oversized police departments left over from an era of high crime (Beckett and Herbert 2009; Herbert et al. 2017; Herring 2019), it is difficult to explain why novel policing innovations, such as big data surveillance, would continue to expand into the 21^st century when crime has not risen. Our results run counter to this intuition, suggesting that innovations in police surveillance—in our case, big data—are partly influenced by sprawling patterns of metropolitan segregation that concentrate affluence in the suburbs and poverty in inner cities.

By evaluating the impact of segregation on big data surveillance, our findings align with the racial threat hypothesis and update expectations from the carceral state hypothesis to congeal with 21^st century residential inequality. First, threat theorists reason that increases in racial segregation should depress police surveillance as segregation is sufficient to stunt black economic and political influence (Blalock 1967; Liska 1992). Although prior studies find mixed evidence of an association between racial segregation and policing across units of analysis (e.g., Carmichael and Kent 2014; Kent and Jacobs 2005; Stults and Baumer 2007), by accounting for sprawling patterns of segregation in metro areas, we find an inverse relationship between racial segregation and big data surveillance consistent with threat theory. In this regard, our study highlights the importance of the unit of analysis for studies of policing and segregation. If whites self-segregate by moving into the suburbs (e.g., Massey and Denton 1993; Farley et al. 1978), then segregation fails as a social control mechanism when the suburbs diversify, as has occurred in recent decades (Lacy 2016; Massey and Tannen 2017). In this regard, future research must account for this broader unit of analysis when examining how racial integration and economic cleavages influence policing.

Second, by advancing income segregation as a focal explanatory variable, we update expectations from the carceral state hypothesis. Although theoretical accounts of the expanding carceral state and changes in policing philosophy direct substantial attention to racial segregation and economic inequality at large (Herbert et al. 2017; Wacquant 2001), the core claim from these arguments is that pronounced inequality increases social control and police oversight. Wacquant (2009), for instance, argues that people of color bear the brunt of carceral expansion due to their historical role as the underclass in American racial capitalism. This reasoning aligns with scholarship on segregation which notes that racial segregation is largely consequential because it overlays with patterns of economic segregation that disadvantage communities of color (Massey and Denton 1993; Peterson and Krivo 2010). Thus, we believe that our results for income segregation align with the core reasoning of the carceral state thesis.

Our findings also underscore the importance of income segregation for understanding how residential inequalities influence policing and social control. Although income segregation has drawn attention from demographers (e.g., Logan et al. 2018; Reardon and Bischoff 2011), policing scholars have been relatively silent on this increasingly prevalent form of inequality. Hence, although past work has emphasized racial segregation and its effects on police oversight throughout the 20^th century (Blalock 1967; Liska 1992; Wacquant 2001), our results imply that income segregation may be more relevant for understanding contemporary police surveillance and its attendant consequences. For instance, a growing body of scholarship reports negative effects of income segregation on educational resources and life-course outcomes (e.g., Owens 2018). It is possible that police surveillance plays a mediating role in these processes by increasing risk of criminal justice contact in economically segregated areas.

We also offer an important methodological and empirical contribution by developing a new measure of big data police surveillance. Our measure accounts for core features of big data surveillance, including technological integration, data size and diversity, “function creep,” and interdepartmental coordination. Using this measure, we provide a first look at the prevalence of big data policing—a topic that has drawn substantial scholarly attention (Brayne 2020; Ferguson 2017; Ridgeway 2018). Our results reveal that big data surveillance is unevenly spatially distributed among metro areas and has expanded in recent years, increasing 30-times over since 2009. In line with arguments from surveillance studies that characterize the contemporary U.S. as a “surveillance society” (Lyon 2013; Marx 2016), our results reveal that big data police surveillance is already common in many metropolitan areas and is rapidly expanding in others.

As scholarship on residential stratification continues to boom, a central task is to explain the perpetuation of residential segregation over time (Light and Thomas 2019). Our study takes an important step towards this goal by exploring how segregation contributes to contemporary police surveillance. As Bell (2020a) and others have described (Fagan and Ash 2017; Wacquant 2001), the simultaneous under- and over-policing of disadvantaged neighborhoods creates a perpetual cycle of social precarity, where disadvantaged populations in segregated areas are shuffled between urban ghettos and the criminal justice system. Our findings illustrate that this feedback loop may continue to exist even though the residential environment and police philosophies have both changed. By adopting sprawling big data tactics that enable police to monitor large areas beyond where officers can physically patrol, police are able to surveil the increasingly population-diffuse patterns of suburban and exurban residential inequality that characterize the contemporary metropolitan environment.

Table 4.

Negative Binomial Between-within Models of Surveillance Index with Variable Exposure including Controls (N = 4,191).

	8	9	10	11	12	13	14
Within MSA effects
Rank-ordered information theory index	9.205** (3.295)			10.567** (3.293)		9.864** (3.369)	11.177*** (3.457)
Black-white dissimilarity index		−.527 (.841)		−.617 (.844)	−.485 (.841)		−.621 (.847)
Gini inequality index			−2.311 (1.411)		−2.150 (1.428)	−2.743 (1.535)	−2.616 (1.430)
Between MSA effects
Rank-ordered information theory index	4.423 (2.863)			7.509*  (3.053)		3.921 (2.903)	7.498** (3.008)
Black-white dissimilarity index		−1.416 (.741)		−2.172** (.797)	−1.649* (.752)		−2.283** (.795)
Gini inequality index			5.486 (4.035)		7.296 (4.186)	4.707 (4.216)	6.999 (3.999)
Controls
Within MSA effects
Violent crime rate per 10,000 capita	.002 (.002)	.002 (.002)	.002 (.002)	.002 (.002)	.002 (.002)	.002 (.002)	.002 (.002)
Drug arrest rate per 10,000 capita (logged)	−.006 (.007)	−.006 (.007)	−.006 (.007)	−.006 (.007)	−.006 (.007)	−.006 (.007)	−.006 (.007)
Terrorist attack in last 3 years	.286*** (.038)	.286*** (.038)	.288*** (.038)	.285*** (.038)	.288*** (.038)	.289*** (.038)	.296*** (.040)
Percent black population (logged)	−.777 (.444)	−.808 (.441)	−.813 (.440)	−.785 (.443)	−.816 (.439)	.782 (.443)	−.978* (.443)
Percent Latino population (logged)	−.736 (.623)	−.619 (.619)	−.591 (.618)	−.800 (.623)	−.635 (.618)	−.775 (.623)	.050 (.644)
Percent 15 – 24 year old male population	−.152 (.120)	−.126 (.120)	−.115 (.120)	−.150 (.120)	−.106 (.120)	−.139 (.121)	−.014 (.122)
Police officers per 10,000 capita (logged)	.087** (.033)	.088** (.033)	.090** (.033)	.087** (.033)	.089** (.033)	.088** (.033)	.085** (.035)
Police expenditures in tens of thousands of inflation adjusted dollars (logged)	.012 (.035)	.015 (.033)	.015 (.033)	.012 (.037)	.016 (.034)	.012 (.037)	.011 (.040)
Percent Republican voters	.027*** (.006)	.027*** (.006)	.028*** (.006)	.027*** (.006)	.028*** (.006)	.028*** (.006)	.023*** (.006)
Median household income in thousands of inflation adjusted dollars (logged)	.088 (.103)	.091 (.104)	−.148 (.179)	.090 (.103)	−.145 (.178)	−.148 (.179)	−.148 (.179)
Unemployment rate	.039 (.042)	.023 (.042)	.014 (.042)	.038 (.042)	.012 (.042)	.027 (.043)	−.010 (.240)
Poverty rate	−.065* (.027)	−.061* (.027)	−.045* (.029)	−.066* (.027)	−.046 (.029)	−.046 (.029)	−.039 (.031)
Percent with baccalaureate degree	.000 (.010)	.001 (.010)	.000 (.010)	.000 (.010)	.001 (.010)	−.001 (.010)	.002 (.010)
Population per sq. mile (logged)	1.045 (.877)	.795 (.874)	.868 (.873)	.852 (.878)	.725 (.873)	.0995 (.877)	.471 (.896)
Between MSA effects
Violent crime rate per 10,000 capita	.001 (.005)	.003 (.005)	.001 (.005)	.002 (.005)	.001 (.005)	.000 (.006)	−.001 (.005)
Drug arrest rate per 10,000 capita	.104 (.089)	.118 (.089)	.125 (.090)	.102 (.088)	.132 (.089)	.114 (.090)	.081 (.090)
Terrorist attack in last 3 years	.762 (.411)	.951*  (.399)	.879*  (.402)	.742 (.411)	.931*  (.398)	.759 (.411)	.666 (.393)
Technology vendor in MSA	−.165 (.405)	−.184 (.405)	−.181 (.350)	−.197 (.357)	−.190 (.357)	−.193 (.357)	−.186 (.533)
Percent black population (logged)	.042 (.098)	.068 (.095)	.035 (.095)	−.010 (.099)	.068 (.094)	.485 (.099)	.040 (.084)
Percent Latino population (logged)	.335*** (.077)	.321*** (.078)	.355*** (.077)	.267*** (.080)	.312*** (.078)	.336*** (.075)	.240** (.076)
Percent 15 – 24 year old male population	.003 (.038)	.004 (.040)	.004 (.038)	−.015 (.041)	.007 (.040)	.032 (.038)	−.002 (.040)
Police officers per 10,000 capita (logged)	.342*** (.099)	.255*** (.099)	.292*** (.097)	.289** (.100)	.229*  (.100)	.326** (.100)	.270** (.100)
Police expenditures in tens of thousands of inflation adjusted dollars (logged)	−.002 (.099)	.103 (.093)	.078 (.094)	−.002 (.102)	.074 (.094)	−.002 (.102)	−.002 (.103)
Percent Republican voters	−.012 (.007)	−.013 (.007)	−.013 (.007)	−.014* (.007)	−.015* (.007)	−.013 (.007)	−.017* (.007)
Median household income in thousands of inflation adjusted dollars (logged)	−.307 (.526)	−.307 (.527)	.039 (.560)	−.308 (.526)	.039 (.560)	.039 (.560)	.040 (.561)
Unemployment rate	−.011 (.089)	−.033 (.088)	−.019 (.089)	−.017 (.089)	−.029 (.088)	−.008 (.089)	−.068 (.076)
Poverty rate	−.026 (.026)	−.013 (.025)	−.026 (.026)	−.020 (.026)	−.044 (.031)	−.047 (.032)	−.043 (.031)
Percent with baccalaureate degree	.029 (.026)	.045 (.025)	.017 (.031)	.026 (.026)	.014 (.031)	.011 (.031)	−.010 (.031)
Population per sq. mile (logged)	−.143 (.103)	−.102 (.101)	−.133 (.102)	−.130 (.101)	−.113 (.101)	−.152 (.103)	−.106 (.092)
Region (vs. Northeast)
Midwest	−.003 (.099)	.031 (.099)	.020 (.099)	.008 (.098)	.041 (.097)	.004 (.099)	.009 (.098)
South	.017 (.068)	.015 (.068)	.018 (.068)	.009 (.068)	.013 (.067)	.017 (.068)	.002 (.068)
West	−.208 (.241)	−.205 (.241)	−.136 (.241)	−.274 (.238)	−.155 (.239)	−.169 (.243)	−.231 (.240)
Intercept	−7.078*** (.983)	−6.213*** (1.095)	−9.113*** (.983)	−5.551*** (1.122)	−8.589*** (1.174)	−.8.725*** (1.772)	−7.969*** (1.39)
MSA frailty term	.983	.965	.984	.952	.959	.984	.972
Year fixed effects?	Yes.	Yes.	Yes.	Yes.	Yes.	Yes.	Yes.
Year x Technology Vendor fixed effects?	Yes.	Yes.	Yes.	Yes.	Yes.	Yes.	Yes.
Log-likelihood	−4171	−4172	−4173	−4167	−4170	−4169	−4020
AIC	8426	8429	8429	8422	8428	8426	8422
BIC	8693	8696	8696	8702	8708	8706	8715

^***p <.001

^**p<.01

^*p <.05.

Two-tailed tests. All models include the natural logarithm of the number of police departments as an offset parameter.