Despite numerous obstacles to data collection, criminologists have long sought to shed light on the relationships between race and crime (Walker et al., 2007). In that attempt, they have collected data that fall, roughly, into three categories: (1) crime data, (2) officer data, and (3) public opinion data. These three categories will serve as an outline for our review of both “what we know” about bias in law enforcement, and “what we do not yet know.”
Criminologists use aggregate crime data far more than any other kind of quantitative data. Broadly, crime data refer to data collected by municipal law enforcement on the types of crimes that occur within a jurisdiction. Municipal-level data are often crucial for understanding the ways in which racial disparities are created and maintained—and for correcting those biases. Litigating claims of racial discrimination, for instance, would be far more difficult on a national level. Individuals who are targets of biased policing are also more likely to make convincing arguments based on the behavior of local law enforcement, as opposed to national-level trends.
Still, national data are crucial for studying the factors that produce and moderate racial disparities in policing. Without the ability to examine variations in demographics, policies, and outcomes, it is difficult to develop principles and theoretical frameworks with which to predict racial disparities and diagnose racial discrimination. Consequently, the lack of national data on racial disparities is a significant obstacle to the study of bias in policing. What national data sets are available result from municipal data that are gathered in one (or more) of four possible ways: Universal Crime Report format, National Incident-Based Reporting System (NIBRS) format, National Crime Victimization Survey (NCVS) format, and department-specific formats. Unfortunately, each has severe limitations with regard to diagnosing even racial disparities. We will briefly describe each format below.
Types of crime data and their limitations
Municipal law enforcement collects data about types of arrests and reports them annually, in the aggregate, to the Federal Bureau of Investigation (FBI). These data are themselves aggregated annually by the FBI and published as the Uniform Crime Report (UCR; Federal Bureau of Investigation, 2009; Maltz, 1977; Robinson, 1911). As a result of the federal statute that requires the collection of these data, many police departments use the UCR format for the maintenance of their own data. This practice means that, if someone is seeking data from a particular department, they are likely to receive these in a format consistent with UCR reporting standards (Gabbidon & Greene, 2005). Consequently, it is common to refer to these data as UCR data, whether they are at the level of a given municipality or at the national level.
These data have provided significant insights into national-level racial disparities. For instance, UCR data have revealed that Blacks and Latinos are significantly more likely to be arrested for violent and drug-related crimes (Federal Bureau of Investigation, 2003; Sidanius & Pratto, 1999; Walker et al., 2007). Despite the utility of having a relatively uniform reporting system, researchers and practitioners alike acknowledge a number of shortcomings of UCR data, chief among them that (1) they fail to collect information on police contacts that do not fall easily into existing UCR categories or (2) on victims.
To address this first concern, the FBI endorses a new model for collecting crime-related data that would focus on incidents of crime, rather than arrests. The NIBRS requires municipal law enforcement to change the way they collect and report criminal data, with all calls for service and recorded criminal events receiving a unique entry, thereby changing the unit of analysis from the arrestee to the arresting incident instead. Though the FBI and the International Association of Chiefs of Police (IACP) both endorsed this new structure in 1988, to date, only 21 states use the NIBRS format, severely limiting the effectiveness of the innovation (Finkelhor & Ormrod, 2000; Gabbidon & Greene, 2005). Law enforcement's resistance to using NIBRS, in addition to the significant logistical hassle of changing to a new data management system, stems from a concern that, if they switch from UCR to NIBRS reporting then the headline in the paper is likely to be “crime goes up,” despite observational evidence to the contrary (Rantala & Edwards, 2000). Consequently, whatever the potential benefits of NIBRS-style reporting might be, they are mitigated by the perceived costs to law enforcement of changing systems, making them less comprehensive than UCR statistics.
In much the same way that NIBRS addresses concerns of underreporting, the NCVS attempts to address concerns that victims of crime are not considered in UCR data (Mosher, Miethe, & Phillips, 2002). These data have revealed important racial disparities in the rates of criminal victimization, providing the evidence that Blacks and Latinos are more likely to be the targets of violent and property crimes than are Whites (Bureau of Justice Statistics, 2001a; Walker et al., 2007). However, while this data set is the best available, it is neither comprehensive (as it is based on sampling estimates) nor is it linked to NIBRS or UCR data. As a result, researchers must analyze data about victimization separately from data about perpetrators.
These disjoint national data sets force researchers—particularly ones interested in racial disparities—to focus primarily on locally collected data. Though individual departmental records can vary widely, they often have incident as well as victim data tied together in offense reports. Importantly, departmental data often also have information about the number of individuals who are stopped, detained, or searched by police—the most common interactions civilians are likely to have with law enforcement—none of which are captured in any of the national data sets (Bureau of Justice Statistics, 2002; Walker et al., 2007). In other words, only departmental data capture information that would reveal racial profiling.
Similarly, departmental data can be connected to individual officers, allowing researchers to determine whether or not the race, sex, age, or experience of individual officers plays a role in aggregate outcomes—something not possible with the existing national data sets. To the extent that scientists and researchers can obtain access to them, these departmental data have been the most useful in producing new knowledge about issues such as racial profiling and racial disparities in use of force (Goff et al., 2010; Ridgeway & MacDonald, 2010). In fact, at this stage, departmental data offer the best hope of investigating the police biases that scholars, concerned communities, and progressive law enforcement deem most critical (Fridell, 2004; Wilson, Dunham, & Alpert, 2004). For instance, voluntarily reported departmental data on use of force have revealed that Blacks are presently four times more likely than Whites to be targeted for use of police force, down from eight to one a quarter century ago (Bureau of Justice Statistics, 2001b; Walker et al., 2007).
However, despite the persistent racial disparities that all levels of analyses and types of data reveal, it is not as simple in the domain of law enforcement to conclude when and where discrimination enters. If Latinos are arrested at twice their representation in a given population, does that mean that there are too many or too few officers in their neighborhoods? Similarly, if Blacks are stopped at twice their representation in a given population, is that because they are committing more crimes (as those who face discrimination in employment, housing, health care, wealth accruement, and education might), or because the police are biased against them? Alternatively, disparities may arise because those who pass laws and direct police enforcement efforts (i.e., municipal executives) direct law enforcement to engage in functionally discriminatory behavior, or because chiefs and sheriffs choose to deploy their officers more aggressively in Black and Latino neighborhoods.
Although it would be naïve to imagine that officers and departmental policies play no role in the creation of racial disparities, it is quite difficult to distinguish between racial disparities in policing and racial discrimination at the individual officer, departmental, and national levels. That is, is racial discrimination in law enforcement the cause of racial disparities or are those disparities a symptom of racial discrimination in other domains? Many criminologists remain agnostic regarding these questions and some of the best-intentioned professionals are left ill-equipped to identify bias where it occurs. Despite these limitations, creative criminologists (and, increasingly, economists and sociologists) have found ways to sidestep some of the issues that we have outlined earlier. For these scholars, departmental data have offered the most promise of identifying departmental bias. Below, we will review the different ways in which scholars have approached the difficult issue of measuring racial profiling, all of which uses crime data gathered from departments.
The problem of measuring racial profiling
To review the specific methodologies that others have used for analyzing police bias, we return again to the example of racial profiling. As detailed earlier, despite repeated efforts to pass federal legislation that would mandate a national database on racial profiling (e.g., The Traffic Stop Statistics Act, 1997; End Racial Profiling Act of 2010: HR 5748, originally introduced in 2001), so-called “racial profiling data” tend to be kept at the municipal or state level, with 25 states enacting some form of racial profiling data collection (Racial Profiling Data Collection Resource Center, n.d.). Consequently, analyses of racial disparities in stops must take on the idiosyncrasies of the jurisdiction under study. While this narrow scope can be problematic, this limitation is not the largest barrier to quality analyses of racial bias in police stops.
Rather, the largest barrier to an accurate accounting of racial bias in police stops is the difficulty scholars have identifying the appropriate way to analyze the data that are collected. More specifically, while some departments keep racial demographic information on vehicle and pedestrian stops, these data only permit an analysis of racial disparities—not racial discrimination (Banks, 2003; Blank et al., 2004; Goff et al., 2010; Ridgeway & MacDonald, 2010). Again, the difference between observed racial disparities versus racial discrimination as the cause is critical. As described earlier, the central difficulty in measuring racial profiling is that, if one believes a police department is engaged in racial profiling, it is reasonable to assume that they are stopping too many Blacks and/or Latinos. The question then becomes, how does one measure “too many”? How does one know if observed disparities are truly due to officer racial discrimination, as opposed to a plethora of other potential causes?
A seemingly common sense approach to the racial profiling question would involve comparing the racial demographics of the stops to the racial demographics of a population. In other words, one wants to create a fraction, with the percentage of Latinos (or Blacks, etc.) stopped as the numerator and the percentage of Latinos in the population as the denominator. Using this analytic technique, also known as population benchmarking, researchers hypothesize that a municipality with a 25% Latino population will produce vehicle stops that are also around 25% Latino. Any deviation from this ratio of population demographics to police stop rates is assumed to be due to police racial bias or profiling. However, this metric is flawed for several reasons, which we detail later.
First, as argued earlier, if racial discrimination exists in all other important social institutions (i.e., education, employment, health care, housing, and wealth accruement), then it is highly probable that these racial inequalities will produce a disproportionate incentive to commit crime among targeted populations of non-Whites? In other words, if racial discrimination encourages a group to engage in criminal behavior, then it is likely that racial disparities in stops can also be a symptom of wider discrimination, rather than a product of police biases (Goff et al., 2010).
Second, using the general population—or even the residential population of a given area—as a benchmark is problematic for a variety of reasons. For example, when assessing car stops, it is not clear that the residents of a given municipality are represented among the driving population in proportion to their racial demographics or among the pedestrian population in the case of pedestrian stops. Similarly, it is often the case that large urban areas are business and social hubs for surrounding municipalities, meaning that the foot and vehicle traffic in a large city is likely to include large numbers of nonresidents. Moreover, in areas with significant undocumented populations, census data are unlikely to reflect the actual racial makeup of the city. Therefore, the population demographic number used to compare the racial demographics in stops against is itself a flawed comparison metric.
Despite these and other flaws, collecting “racial profiling data” usually means keeping track of the racial demographics of stops, and both municipal and federal “racial profiling analyses” frequently use population benchmarking as a standard technique. Thankfully, because population benchmarking is so imprecise, scholars and police have searched for a replacement to populations as the relevant benchmark—or denominator. This search is also known as the “denominator problem” (Walker, 2001) or the benchmarking problem.
Many researchers have wrestled with the “denominator problem,” and have found creative alternatives to simple population benchmarking. Specifically, six methods, with a range of popularity and effectiveness, use departmental crime data to examine racial profiling and have become popular in recent years. They are (1) adjusted neighborhood benchmarking, (2) arrest data, (3) DMV and vehicle registrations data, (4) so-called “blind enforcement mechanics,” (5) observational data, and 6) consent search or “outcomes tests” analyses. Below, we will outline each of these methodologies, their advantages and disadvantages, and summarize their effects on how scholars approach analyses of racial bias in law enforcement.
Adjusted neighborhood benchmarking
This approach encompasses a number of analytical techniques that attempt to solve the denominator problem by more accurately quantifying the number of residents who might be stopped. This method frequently involves benchmarking stops within smaller geographic areas and using the neighborhood or census tract demographics rather than municipal demographics to produce appropriate benchmarks (Fridell, 2004; Goff et al., 2010; Ridgeway & MacDonald, 2010). This technique is designed to reduce the disproportional impact of targeted enforcement techniques on the racial demographics of stops data. That is, when departments choose to engage in increased vehicle stops within a given neighborhood (often as a crime reduction tool), it is likely to drive up the number of stops in that neighborhood. If the neighborhood is majority Black or Latino, then, even if stops in that neighborhood are racially proportional to the population, the enforcement pattern will result in a higher proportion of Blacks or Latinos stopped citywide than their representation in the population.
An adjusted neighborhood benchmarking approach alleviates this concern by matching stops to neighborhoods rather than to an entire city or county. Therefore, by narrowing the unit of analysis to smaller areas, researchers are better able to account for the racial compositions of the area, yielding a more precise “denominator.” This technique, used in the infamous RAND report on New York City's stop and frisk practices (Ridgeway, 2006), however has many of the same drawbacks as less nuanced population benchmarking techniques. First, it does not take into account commuter or undocumented populations, which are again likely to change the “denominator” used in population benchmarking. Further, neighborhood benchmarking is also unable to distinguish between police bias or previous discrimination as an explanation of the data. That is, if it is established that racial bias are responsible for observed disparities in stop rates, population benchmarking can never answer the question of whether officer racial bias caused the disparities. And, perhaps most importantly, this technique is unable to determine whether targeted enforcement patterns might be motivated by the racial demographics of certain areas—potentially obviating police departments for policies that produce racial harmful results. That is, if department policies dispatch unequal numbers of patrol units to particular racial areas, observed disparities in stops are likely to ensue, but they would not necessarily be indicative of individual officer bias.
The use of arrest data as the denominator for racial profiling statistics was mostly popular among conservative journalists and pundits for a short time, and not with more rigorous scholars (Harris, 2002). Comparing the racial demographics of stops to the racial demographics of arrest data may seem reasonable if one is searching for the ideal denominator—the demographics of individuals who commit crimes. But upon even cursory inspection, this technique violates multiple rules of logic and statistical inference. Since a police “stop” is often a precursor to an arrest, arrest records may be racially skewed. That is, if stops are biased, then arrests may be biased. Consequently, using (street-level) arrest data to benchmark stops is highly suspect.
DMV and vehicle registration data
Some researchers have attempted to solve the denominator problem by restricting the benchmark population to those who are licensed to drive or who own a vehicle. While this seems a reasonable modification for analyses of vehicle stops, it does no better than unadjusted population benchmarking at identifying the appropriate commuter population (i.e., those who live outside a jurisdiction but regularly drive through it), and suffers from similar problems with regard to causal attributions for observed disparities. Of course, this technique cannot be used to address disparities in pedestrian stops, another arena of potential racial profiling by police officers.
“Blind enforcement mechanics
” This technique is a far more promising approach than the ones previous reviewed. By using functionally race-blind mechanisms to estimate the relevant population, it would seem possible to create a better benchmark. Examples of so-called “blind enforcement mechanics” include comparing traffic stops made in the daylight (when officers might be able to tell the race of the suspect) to ones made at night (when this is, ostensibly, harder to infer race of suspects). Using radars, airplanes, and video traps to estimate the racial demographics of those speeding is another example. And, perhaps most promising of all, using no-fault accidents and red-light cameras to estimate the appropriate racial benchmarks have become increasingly popular.
Using “no fault” accident information, for instance, has the advantage of providing a reasonably unbiased account of the demographics of those who are on the road, since “no fault” accidents are, ostensibly, equally likely to affect everyone. Red-light cameras have the ability to record vehicle information from heavily traveled areas which is linked to vehicle owner demographics, thus providing an estimate of benchmark demographics by providing racial information on those who actually break at least one traffic rule.
Again, however, these techniques are limited. “No fault” accidents are likely to underrecord individuals who are in the country illegally or civilians with criminal records who are wary of engaging with law enforcement. Individuals without insurance are also less likely to report accidents, which is likely to skew the racial demographics “no fault” accidents further. Similarly, because red-light cameras are expensive, they tend to be placed in high-traffic areas and dangerous intersections rather than in quieter residential neighborhoods or suburbs, both of which tend to have fewer traffic lights in general (Walker et al., 2007). Consequently, using red-light cameras is likely to produce artificially high numbers of Blacks and Latinos who are concentrated in inner cities (Jargowsky, 1998; Massey & Denton, 1993; Wilson, 1996). These techniques are also ill-equipped to address the issue of pedestrian stops (with street-corner cameras suffering from even more severe limitations than red-light cameras) and suffer from the same inability to distinguish between police discrimination and broader racial disparities.
Using a technique that is similar, though more robust, than recordings from red-light cameras, a few dedicated researchers have attempted to estimate the population of potential vehicular law violators by simply observing them—in extraordinary numbers (Engel & Calnon, 2004; Ridgeway & MacDonald, 2010; Walker et al., 2007). If observers are well trained, this approach has the benefits of better location sampling and can be adapted (with some difficulty) to pedestrian contexts. However, creating a benchmark from observational data is exceptionally time-consuming and prohibitively expensive.
Consent search data
One notable exception to many of the limitations found with most crime data analyses of racial profiling is research using consent search analyses or “outcomes testing.” These techniques are considered the contemporary gold standard in both internal benchmarking and aggregate benchmarking analyses. The first of these approaches is “consent search” analyses that focus on officer or departmental “hit rates.” The logic behind these searches is as follows: since consent searches are at the discretion of the officer, they provide a unique insight into potential officer biases. If officers are policing in an unbiased manner, then one would expect that the percentage of searches that produce contraband among Black or Latino targets—the “hit rate”—should be similar to the percentage of searches among White targets. If, on the other hand, hit rates for Black suspects are significantly lower than for White suspects, this ratio might be an indication of officer or departmental bias (Dominitz & Knowles, 2006; Knowles, Persico, & Todd, 2001; Persico & Todd, 2006; Sanga, 2009; Smith & Petrocelli, 2001).
This technique has the potential benefit of providing both initial evidence of an unacceptable disparity and simultaneously demonstrating ways in which police efficiency can be improved. Similarly, other outcomes-based analyses of post-stops data (i.e., number of arrests that result from stops, quality of the interaction, etc.) have a greater degree of accuracy without having to engage with the denominator problem. However, there are still significant limitations of outcomes testing.
Some scholars still bemoan the fact that consent search data (and other outcomes data) often neglects geographic variation in searches (Sanga, 2009). Others argue that racial differences in hit rate can occur for spurious statistical reasons (Bjerk, 2007). Still others suggest that there are concerns with regard to racial differences in when civilians give consent.
For instance, some have suggested that there are racial differences in community awareness that consent searches also require the consent of civilians (Sklansky, 1997). If this is true, then Whites who are hiding contraband may feel more comfortable refusing police search requests than Blacks or Latinos, thus escalating the likely hit rate for non-Whites. Conversely, if officers know that Blacks and Latinos are less likely to refuse searches, this could increase incentives to engage in so-called “pretextual stops,” the practice of stopping someone for a minor infraction in hopes of finding something more substantive during the interaction. This practice, of course, would reduce the hit rate for non-Whites.
Importantly, the strength of this analytic technique is that it does not focus on the decision to stop an individual but, rather, on something that happens after a stop. Consequently, while it may be a superior metric of the racial biases of officers and departments, it does not answer the question of whether or not individuals are stopped because of their race. Similarly, though outcomes testing measurements do not suffer from the same denominator problem as the previously discussed approaches, there are still difficulties interpreting consent search analyses. For instance, while a departmental policy may result in more searches, and proportionally fewer “hits,” within a particular community, this pattern may be the result of a deliberate enforcement strategy rather than an indication of ineffective policing. Similarly, emerging research suggests that non-Whites may feel anxious during police encounters in response to the fear that they will be labeled as criminals—even if they know they are not criminals (Najdowski & Goff, 2011). This apprehension can result in behaviors that would appear to observers as if they are guilty of something—likely resulting in the desire by an officer to search the individual.
Summary of popular approaches to racial profiling analyses
Each of these approaches seeks to approximate the racial demographics of the criminal population—and each does so imperfectly. Yet, despite having so many options available, no scholar nor practitioner has suggested that there is a “one size fits all” solution to the problem. In fact, every significant review of racial profiling analytic approaches has stressed both the need for agency-specific approaches and for the need to look beyond simple stops data to ensure equity in law enforcement (Fridell, 2004; Goff et al., 2010; Harcourt, 2006; Harris, 2002; Ridgeway & MacDonald, 2010; Walker et al., 2007). This assertion represents a scholarly acknowledgment of the methodological imperfections of existing measurement techniques.
It is important to note, also, that each of the above techniques assumes that the proper level of analysis is the level of the institution (Harcourt, 2006).1 Of course, institutions and departments are an important level of analysis. It is much easier for communities to seek redress from a department with a demonstrated pattern of racially disparate treatment than for any individual to demonstrate that a given stop was motivated by racial prejudice. Still, however reasonable it is to suspect that some police departments engage in aggregate discrimination, it is at least as reasonable to suspect that police departments (of a suitable size) contain individual officers who engage in racially biased policing. Yet, none of the above techniques are well positioned to address this possibility. This assertion is not to say that criminologists have not acknowledged both the existence of institutional-level bias and the need to study both “rotten apples” (i.e., biased officers) and “rotten apple barrels” (i.e., biases in police culture and/or policy; Walker, 2001; 2005; Walker & Alpert, 2000). Still, the metrics for studying racial bias in law enforcement are poorly fitted for a quantitative analysis of any given department's level of bias—much less a comparison between departments. In the next section, therefore, we address the relatively smaller literature that addresses the possibility of individual-level biases: research on officer attitudes.
Perhaps due to criminology's origins in clinical psychology, there are numerous personological approaches to policing (Adlam, 1982; Balch, 1972; Bennett & Greenstein, 1975; Evans, Coman, & Stanley, 1992; Fenster & Locke, 1973; Hanewicz, 1978; Hogan & Kurtines, 1975; Lester, Babcock, Cassisi, Genz, & Butler, 1980; McNamara, 1967; Mills & Bohannon, 1980; Niederhoffer, 1967; Sherman, 1980; Sidanius & Pratto, 1999; Skolnick, 1977; Toby, 2000; Walker, 1992). Yet, despite the extensive literature on police personality, there is relatively less literature on individual officer-level biases as opposed to the above literature on profiling at a higher institutional level. What research there is tends to fall in one of two categories: officer racial attitudes research and internal benchmarking analyses.
The first of these approaches involves simply measuring the racial attitudes of officers. Though few and far between, these data have created a consensus that law enforcement in the United States shares the racial biases of civilians, though there is a tendency for law enforcement to be slightly more racially prejudiced than the population at large (Bayley & Mendelsohn, 1969; Eberhardt, Goff, Purdie-Vaughns, & Davies, 2004; Sidanius & Pratto, 1999; but, see Correll et al., 2007 for an exception). This finding would seem to indicate that officers are prone to some level of bias-based policing. But many researchers and practitioners are quick to point out that there is a difference between biased attitudes and discriminatory behavior (Correll et al., 2007; Eberhardt et al., 2004; Ogloff, 2000; Walker et al., 2007).
This is a distinction that is also well known—though too often forgotten—within social psychology, with attitudes traditionally predicting less than 10% of the variance of both behaviors in general and racially discriminatory behaviors in particular (Dovidio, 2001; LaPierre, 1934; Wicker, 1969). Additionally, even if racially biased attitudes produce some level of biased behavior, it is not at all clear how much biased behavior they produce. A lack of real-world behavioral metrics, therefore, prevents the officer racial attitude research from having a more significant effect on the science or policies surrounding racial bias in policing.
The second significant category of research on officers is research on internal benchmarking (Fridell, 2004; Ridgeway & MacDonald, 2010; Walker, 2001, 2005; Walker et al., 2007; Wilson et al., 2004). Internal benchmarking analyses eschew population denominators for officer performance denominators, essentially comparing one officer's in the field performance to her or his peers. These analyses have the advantage of identifying individual officers who are more prone to stop or use force against Blacks or Latinos as opposed to Whites than their peers, effectively permitting researchers to avoid questions of suspect criminality in their analyses by analyzing variance among officers from the same patrols, same neighborhoods, and within the same department.
This technique is essential if departments wish to identify the officers whose behavior needs correction. In fact, many believe that so-called early warning systems—that predict future behavior based on past behavior—are the best hope to ensure equity and effectiveness in policing (Walker, 2005). However, this technique tends not to include predictors (i.e., attitudes) nor to link officer data to institutional-level variables, making it difficult to target an intervention or discern the role that the institution might be playing in the production of the bad behavior. Additionally, because early warning systems vary so widely and cannot be validated without considerable aggregate data, they remain limited in their capacity to predict behavior across departments.
Taken together, while officer data represent an essential level of analysis, researchers have yet to connect predictors (such as attitudes) to real-world behaviors. Consequently, as with crime data, there is scant research that uses officer data to produce advances in our understanding of biased police behaviors.
Public Opinion Data
As criminology evolved from the study of the criminally insane to the study of how crime functions in society (Garland, 1985; Ogloff, 2000; Rafter, 2008), there emerged a growing interest in how the public felt about public safety (Walker et al., 2007). This recognition of the importance of public perception about crime has translated into a contemporary interest in public opinion data, particularly data gathered from national polls.
Public opinion on public safety is an essential part of the policing puzzle, since a department cannot gain public cooperation if the communities being policed feel that law enforcement engage in racially biased policing (Tyler & Huo, 2002; Walker et al., 2007; Weitzer & Tuch, 2006). That said, opinion data are fundamentally about perceptions of racial inequality in policing and not about the realities of racial inequality in policing. In addition, because journalists and/or professional pollsters frequently gather these data, rather than scholars (cf. Walker et al., 2007), there are often questions by rigorous scientists about inappropriate methodologies and an inability to track opinions over time. For the same reason, it also tends to be the case that some large cities (e.g., Washington DC) are overrepresented in public opinion data on race and policing.
Were it tied to data on police behavior, researchers might reveal important relationships between public perceptions and the behavior of officers. However, as this technique has yet to be applied to the issue of race and policing, this relationship remains a question for empirical exploration. As it stands, while public opinion data are crucial information for scholars and practitioners who want to understand how race and policing is lived, it does little to determine whether police are actually engaging in biased practices.