U.S. Department of Justice 
Office of Community Oriented Policing Services 

Image of United States Department of Justice seal

Communty Oriented Policing Services logo

INNOVATIONS 
A Closer Look

A Suggested Approach to Analyzing Racial Profiling: 
Sample Templates for Analyzing Car-Stop Data 
Image of police car and officer

Authors: 
Joyce McMahon, Ph.D., 
Amanda Kraus, Ph.D., 
The CNA Corporation 

www.cops.usdoj.gov 


A Suggested Approach to Analyzing Racial Profiling: 
Sample Templates for Analyzing Car-Stop Data 

By 
Joyce McMahon, Ph.D., 
Amanda Kraus, Ph.D., 
The CNA Corporation 


This project was supported by cooperative agreement #2002-HS-WX-0061 
by the Office of Community Oriented Policing Services, U.S. Department of 
Justice. The opinions contained herein are those of the author(s) and do not 
necessarily represent the official position of the U.S. Department of Justice. 
References to specific companies, products, or services should not be 
considered an endorsement of the product by the author(s) or the U.S. 
Department of Justice. Rather, the references are illustrations to supplement 
discussion of the issues. 


Introduction 

Because of community concerns about the use of race as a significant 
factor in police decision-making, many police departments have collected 
and used data from car stops to investigate the issue of racial profiling, or 
bias-based policing. Many more departments are considering doing so. 
Most existing evaluations based on car-stop data have relied on aggregate 
comparisons between the racial composition of stops in a city and the 
racial composition of the city’s total population. Unfortunately, this simple, 
aggregate approach can be dangerous. By failing to include important 
information on police operational procedures, specific city circumstances, 
and socio-demographic characteristics of neighborhoods, the aggregate 
approach is too simplistic and may generate misleading results. 

Decisions regarding the merits of racial profiling concerns are important 
and should not be based on either anecdotal evidence or incomplete 
analysis. Evaluating the extent and nature of police profiling patterns 
may lead to decisions regarding proper training and appropriate police 
tactics. It is crucial that such evaluations rely on appropriate methodological 
approaches, objectively obtained data, and appropriate benchmarks or 
comparison guidelines. 

The Justice Department’s Office of Community Oriented Policing 
Services (COPS) asked The CNA Corporation (CNAC) to work with 
two police departments to address data collection and evaluation issues. 
This effort represents a follow-on to a previous CNAC study for the 
COPS Office, which produced the report, How to Correctly Collect 
and Analyze Racial Profiling Data: Your Reputation Depends On It! 
(available on the COPS web site).1 The purpose of the new work was to 
apply the tools and methods recommended in the previous report by 
creating partnerships with two police departments and helping them use 
rigorous analytical methods that go beyond the standard practice. The 
approach relies on blending analytical research methods with operational 
police insights. 

1 How to Correctly Collect and Analyze Racial Profiling Data, 2002. Report available in hard copy from the 
COPS Office (800-421-6770) or at the COPS web site, http://www.cops.usdoj.gov/default.asp?Item=770 


This pamphlet does not include any actual data or specific findings from 
the two departments. Instead, it describes the general approaches used, 
and illustrates them with sample templates of the analytical output. 
These templates represent examples of how to display and evaluate 
results from various methods of analysis. They were selected to demonstrate 
the range of analytical questions that can be addressed by these 
techniques. 

The traditional (standard) approach to data analysis–and its limitations 

The typical approach for using police administrative data to identify the 
presence of racial profiling practices is to compare the racial breakdown 
of vehicle stops to that of the city’s population. Consider an example for 
a hypothetical city in which 33 percent of the people who are stopped 
belong to a minority race group and 67 percent belong to a majority race 
group. If the demographic data for the city show that the racial breakdown 
of the population is the same–33 percent minority and 67 percent non-
minority–the standard conclusion that would be drawn from the fact of 
equal racial shares is that there is no racial profiling; all is well. 

By using this approach, people are effectively allowing the demographic 
mix of a city’s total population to act as an estimate of the demographic 
mix of the observed violator population. More specifically, they are 
assuming that members of each race group violate traffic laws at the 
same rates and that officers observe members of each race group 
committing these violations in the same proportions. Whether the first 
assumption holds has been a subject of much debate, and is virtually 
impossible to prove or disprove. The second assumption is generally 
not expected to hold because it is well understood that deployed police 
officers, drivers of different races, and problem traffic areas are all 
dispersed unevenly across a city’s geography. 


Finding the right comparison group 

A more general characterization of the main conceptual problem with 
the standard approach is that city population data don’t provide the right 
comparison groups. Figure 1 shows how members of a city’s driving 
population–whether they are residents of the city or not–become part 
of the stopped and cited populations. The ideal approach to determining 
whether members of one race group are inappropriately stopped at 
higher rates than members of other groups would be (a) to compare the 
race breakdown of the total violator population with that of the observed 
violator population, (b) to compare the race breakdown of the observed 
violator population with that of the stopped population, and (c) to 
consider reasons why variations between the groups exist. Unfortunately, 
it is virtually impossible to determine the racial make-up of either the 
total violator population or the observed violator population, so researchers 
and police analysts have usually made comparisons between the race 
breakdowns of the stopped and the resident populations. The trouble is 
that the resident population is several steps removed from both violator 
populations and is likely to be quite different for reasons that are 
independent of police activity as well as for reasons that are due to 
police policies and behavior. 

Reasons for differences in the racial compositions of the various 
populations in the figure can be considered in more detail. Starting 
from left to right in Figure 1, it can first be seen that the racial composition 
of the available driving population may be different from the racial 
composition of the city’s population for two reasons. First, the city’s 
own driving population may not have the same racial composition as the 
city’s total population. This will be true if race breakdowns vary by 
driving age or car ownership. Second, if the non-resident driving 
population is sufficiently large, then differences between the racial 
compositions of non-resident drivers and the city’s population will also 
cause the driving population to be racially different from the city’s 
population. If the race mix of the driving population is different from 
that of the city’s population for either of these reasons, then it is likely 
that all the populations that are “downstream” from the driving population 
will also be racially different from the city’s population. Since these 
differences will occur independent of any action on the part of police 
officers, they cannot be the result of racial profiling. 


Figure 1. How members of a city’s total population become members of the 
city’s driving and stopped populations 

Figure showing reasons for differences in the racial compositions of the various populations. Starting from left to right in Figure 1, it can first be seen that the racial composition of the available driving population may be difference from the racial composition of the city's population.

Next, consider sources of differences between the racial compositions of 
the driving population and the violator population. 
is whether people of different races commit infractions at the same rates. 
If they do, then the race distributions of the two groups will match; if 
they don’t then the race distributions will not match. 
difference will occur independent of police activity.

Moving from the violator population to the observed violator population 
is the first place where police department policies can contribute to 
differences between the stopped population and the city population. 
Specifically, the racial make-up of the observed violator population may 
differ from that of the total violator population due to variations in 
police deployments across geographic areas and times of daydifferences are legitimate if deployment patterns are determined without 
respect to race. Next, differences in the race breakdowns of the observed 
violator and stopped populations can arise due to the interaction between 
racial differences in violation rates on the one hand, and police policies 
and procedures regarding which violations should be singled out for 
stops, on the other. Again, if such policies are determined without respect 
to race, then differences between populations can be considered legitimate. 
Therefore, an important corollary to recommendations on effectively 
using administrative data to show whether a department practices racial 
profiling is that departments need to be able to document the reasons 
behind existing deployment patterns and other operational decisions. 


Finally, individual officers will use their own judgment and understanding 
of department policies to decide whom to stop, and then whom to cite 
or search among those they have stopped. Racial differences between 
the observed violator and stopped populations and racial differences 
between the stopped and cited (or searched) populations are the result 
of the behaviors of both the officer involved and the stopped citizen, 
therefore, they can be the result of racial profiling or not. 

Following the flow from the city’s total population to the stopped and 
searched or cited populations clearly shows the problems associated 
with using comparisons between the racial compositions of these groups 
to prove or disprove that a police department’s officers are engaged in 
racial profiling. Therefore, such comparisons should be interpreted with 
extreme care. 

Matching definitions of race across data sources 

Another problem with making simple comparisons between the racial 
compositions of the stopped populations and racial compositions of 
cities’ populations is the fact that measuring race is not always 
straightforward. The U.S. Census Bureau publishes self-reported race 
data by one race alone, two races, and three or more races. Census data 
also indicate self-reported Hispanic ethnicity. In contrast, police-
collected information on a citizen’s race typically reflects an officer’s 
assessment of the individual’s racial background based on how the 
person looks. Since it’s difficult for police officers to capture the nuances 
of mixed races, stop or citation data usually just have one category for 
each race and don’t capture multiple races. Therefore, in many cases, 
U.S. Census race definitions may not match race definitions in police 
data even when they have the same labels. In addition, in some 
instances, stop or citation data are recorded as Caucasian or African-
American, omitting all other races and ethnic categories (e.g., Hispanic). 


Controlling for many factors at one time 

Finally, simplistically comparing race data for stops against race data for 
total populations can’t capture multiple correlations between different 
factors associated with stops and violations. Therefore, a better approach 
is to use multivariate techniques that allow researchers to hold constant 
the effects of several factors all at the same time. For example, it is 
likely that more police officers are deployed and making stops in high-
crime, low-income neighborhoods. Unfortunately, in the United States, 
such neighborhoods are also likely to be minority neighborhoods. 
Therefore, it is possible that disproportionately high stop rates for 
minorities may be the result of police deployment patterns rather than 
police bias. 

One way to explore the combined impact of multiple factors, and 
investigate the direction of causality, is by the use of nested tables. 
Samples of these types of tables are provided later on (e.g., figures 3 and 
4). For example, suppose that an investigator is examining stop rates by 
minority classification. Rather than just look at stop rates associated 
with various race/ethnic groups, it is important to look at stop rates for 
these groups within specific types of car stops (e.g., driving related and 
vehicle equipment), for different times of day (e.g., rush hour, daytime, 
evening), or within different neighborhoods (e.g., central city district, 
residential areas, shopping areas). Examining the basic measure of car 
stops in this way has the potential to provide substantially more 
information about the multiple influences that may affect car stops. The 
tables can be set up to explore these different influences by explicitly 
calculating rates for each pertinent cell defined (e.g., stop rate by race/ 
ethnic group for driving related incidents at night in the central city). 

In addition, other multivariate techniques (such as regression analysis) 
that include neighborhood-specific data on income, race, and the number 
of deployed police officers can measure the separate effects of each of 
these potentially correlated factors. One specific example relates to 
controlling for type of stop along with race. Research has shown that 
Black drivers are less likely than drivers of other races to wear seat 
belts.2 Therefore, if a large percentage of stops were related to failure to 
wear a seat belt, Black drivers would be likely to have a disproportionately 

2 See: “Achieving a Credible Health and Safety Approach to Increasing Seat Belt Use Among African 
Americans,” Department of Occupational and Preventive Medicine, Meharry Medical College, May 1999 and 
“Blue Ribbon Panel to Increase Seat Belt Use Among African Americans: A Report to the Nation,” December 
2000, p. 11, DOT HS 809 185. 


high share of the total. Thus, it would be important to control for type of 
violation or reason for stop in a racial profiling study based on data from 
car stops. Regression analysis is well suited for this type of analytical 
investigation. 

Another aspect of controlling for many factors is to consider which 
factors are likely to be important in a given jurisdiction. Different cities 
are likely to have different problems, geographic considerations, and 
histories with respect to citizen-police interactions over time. One 
approach is to survey the citizens of a jurisdiction to determine their 
perceptions (e.g., of the police force, crime, police initiatives) and to 
understand citizens’ views concerning various issues (e.g., racial 
profiling, neighborhood crime concerns, traffic congestion, illegal 
parking). Using surveys is a productive way to determine the range 
of factors that are relevant for a particular city or jurisdiction.3 

Data collection 

What kind of data can be collected? 

There are at least four general categories of data, as described on the 
next page. This should not be considered a definitive or exhaustive list 
of data elements. Depending on the questions that are considered to be 
important, and the particular characteristics of a department and the 
surrounding city or county, some of these measures may be irrelevant. 
There may also be data elements that would be important to consider 
that are not included. 

Although this may appear to be a lot of information, most of it is already 
collected–often as part of standard practices. For example, officers 
making a traffic stop will usually record the year of the car and the date 
of stop. With this information, it is possible to create a variable to 
describe the age of the car. In addition, characteristics of the officer 
making the stop need not to be collected for each incident. Instead, with 
a coded officer identifier for each stop, stop data can be merged with data 
from personnel files. Finally, information about the stop location can be 
derived from U.S. Census data and supplementary operational data. 

3 See Deborah Weisel, Conducting Community Surveys: A Practical Guide for Law Enforcement Agencies, NCJ 
178246, 1999, Washington, DC: U.S. Department of Justice, Office of Justice Programs. Weisel provides a 
practical guide for law enforcement agencies, which includes a description of how surveys can be applied to 
improve policing services, in addition to techniques for identifying survey goals and procedures for survey 
administration and analysis. The guide reviews many practical aspects of using surveys, including the use of 
telephone surveys, tailoring surveys to fit different needs, sample size, probability sampling, analytical techniques, 
and costs.


Categories of Data 

Citizen characteristics         Officer characteristics 
. Demographics                  . Demographics 
. Car age/type                  . Career and assignment 
. Address                       . Behavior 
. Behavior 
Stop information                Site information 
. Location                      . Demographics 
. Time of day                   . Urban/rural, business/residential 
. Duration                      . Socio-economic data 
. Reason                        . Traffic flow 
. Outcome                       . Officer deployment 
                                . Calls for service 

Deciding what data to collect 

Figure 2 provides a flow diagram of a process that can be used to guide 
data collection decisions. It is motivated by the most basic consideration– 
what do the department and the community want to learn? Or put 
another way, which questions are analysts attempting to answer? The 
ability to articulate what is to be investigated should precede the process 
of exploring data elements and selecting an analytical approach. 

Following this process will allow departments to evaluate and categorize 
the data elements they are collecting in tandem with the methods 
available to them. One of the important aspects of this method is that 
it allows investigators to explicitly consider what questions can be 
answered using alternative approaches. For example, given a certain set 
of available data elements, there will be some questions that can be 
answered, but there are likely to also be some questions that can’t be 
addressed by that data. The investigators need to know this and consider 
it in advance, so that they can determine whether it is important to 
obtain additional data, or alternatively so that they can explain that 
certain questions will not be answered using the current data elements. 


Figure 2. Steps to Guide Data Collection 

Image of a flow chart showing the steps to guide data collection

Depending on what analytical method is eventually selected, it may be 
advisable for police departments to consider the merits of a partnership 
with an analytically trained social scientist. 

CNAC research partnerships results 

General lessons learned 

CNAC established research partnerships with two law enforcement 
agencies that differed substantially in terms of characteristics of the 
surrounding communities and the data being collected. In each case, 
there were separate challenges to evaluating the data; surprisingly, 
however, some lessons learned were the same for both departments in 
that similar patterns showed up in both data sets. Specifically, four 
findings held across both communities: 

. Nonresident drivers affect outcomes. 

. The race share of stops/citations varies significantly by time 

of day. 

. Race and gender interact with type of stop/citation. 

. The race mix of stops/citations varies with the race mix of the 

stop location’s population. 


General approach to interpreting data 

To investigate the presence or absence of racial profiling within a force, 
data can be analyzed at three levels: department or citywide, by census 
tract, and by individual officer. Generally, our approach was to use the 
data to identify patterns in the race mix of stops and stop outcomes that 
might call for further investigation. More specifically, rather than using 
the standard technique of making simplistic comparisons between the 
race mix of stops and the race mix of the cities, we focused on within-
sample comparisons across race groups, and used both bivariate and 
multivariate analysis techniques. This approach allowed us to avoid 
some of the pitfalls associated with the standard technique. 

Although the focus of this project was on using data to analyze racial 
profiling issues, the same data can be used to help departments analyze 
operational efficiency and to learn more about what practices are more 
effective in achieving operational goals (and what practices are less 
effective). In the process of trying to understand why outliers exist, a 
department can learn about many aspects of its operations–not just racial 
profiling. 

Sample standard templates to evaluate stop and citation data 

In the following samples of standard templates, we present table 
structures that include a selection of important variables and a suggested 
structure of analysis. These are examples of how investigators can use 
templates, either in table or regression formats, to investigate stop and 
citation data. In addition, these approaches can be used to investigate 
associated search data, as well as officer efficiency metrics. In these 
tables, we do not present actual numbers–while we show the structure of 
the table, all numerical entries are represented by XX. We do not show 
actual numbers, in part because we do not want to imply that our 
examples support findings that might indicate the possibility of racial 
profiling, or appear to negate the existence of racial profiling. 


If we were to use actual numbers in our tables, it would be tempting 
for readers to try to interpret the meaning of the results. This would 
be inappropriate for a number of reasons. First, we are not using full 
models – just template examples. In the real world, there might be other 
influences that should be considered, and these factors will vary for 
different jurisdictions. In addition, if we used numerical examples, we 
would need to describe how all of the variables are measured (their 
scale), discuss the variables’ observed variations (standard deviations) 
and correlations with each other, and motivate a discussion of hypothesis 
testing and statistical significance. It is not our intent to address the 
overall topic of how to do analytical modeling or assess tests of 
significance–these can be complicated issues and are best addressed 
in the context of social science estimation or prediction techniques. 

We begin our examples of data templates by considering ways to address 
stop data. The most general way to expand the traditional comparison of 
stop data by race is to create a table structure that supplies information 
about additional factors, such as gender, time of day, or type of stop. 
Figure 3 displays a template for disaggregating data by type of stop, 
controlling for minority status. In this example, stop data are evaluated 
separately for driving-related stops and vehicle equipment stops, 
although other types of stops could be considered as well. This type of 
analysis has the advantage of being very easy to display and understand, 
but has the disadvantage that only a limited number of factors can be 
displayed in a single table. 

Figure 3. Disaggregate Data by Type of Stop 

Image of a table displaying a template for disaggregating data by type of stop, controlling for minority status. 

.Ways to disaggregate: by gender, time of stop, location, etc. 
.This list of variables informs what data to collect. 


When looking at aggregate stop data, operational experts and researchers 
alike begin to think about other ways to consider what the data might be 
telling them. For example, it may be that most officers are deployed in 
minority neighborhoods, or that minorities are more likely to live in 
urban neighborhoods with more traffic. These and a host of similar 
considerations can be used to motivate the collection of additional data 
elements that can be used to control for other influences and illuminate 
the data findings. 

One of the most common considerations is the influence of the stop 
location, which can be mapped to a distinct neighborhood or census 
tract. Figure 4 offers a sample template of one way to examine data by 
stop location. Narrowing down to census tract generally yields a better 
idea of the characteristics of the neighborhood and, theoretically, gets 
closer to the real candidate population that is available to be stopped. 
However, it is still possible that many nonresidents may be stopped, 
depending on the nature of the neighborhood and the type of highways 
or streets that run through it. 

Figure 4. Analyze Data by Stop Location–Neighborhood or Census Tract 

Image of a table offering a sample template of one way to examine data by stop location. 

.Is the difference between the stops’ minority share and the population’s 
 minority share statistically significant? 

This type of tabular result allows a researcher to compare the minority 
share of stops in a specific census tract with the minority share of the 
population in that census tract. If there are differences, two questions to 
ask are how large are the differences and would the community consider 
them to be meaningful? Again, most evaluators would also explore other 
census tract characteristics, such as the median income for the tract, major 
through streets, presence or absence of shopping centers, and so on. 


The effort to control for many factors simultaneously will often lead to 
using regression analysis to investigate the issues. Multiple regressions 
have the advantage of being able to estimate the separate influences of 
several factors at once. This allows the independent effect of one factor 
to be considered while holding constant, or taking into consideration, the 
effect of a range of other factors. Figure 5 is a sample template of the 
results of a multiple regression to evaluate stop rates by census tract, 
holding constant several important factors like the minority share of the 
tract’s population. (The variables or factors controlled for in figure 5 are 
illustrative of the estimation technique, but should not be considered to 
exhaust the possible list of variables.) 

Figure 5. Explaining tract-specific stop rates as a function of tract 
characteristics 

Image of a table showing a sample template of the results of a multiple regression to evaluate stop rates by census tract, holding constant several important factors like the minority share of the tract's population. 

.Controlling for other tract characteristics, is the minority population share 
 a significant determinant of stop activity? 

This regression approach can offer compelling information because it is 
often argued that differences between the minority share of stops and the 
minority population share at the aggregate level are due to unequal 
distributions of officers across areas/neighborhoods in the city. If this 
is the case, evaluating at the census tract level should yield useful 
information regarding how stops differ across these areas and why. In 
particular, it is also often argued that the correlation between police 
presence in a neighborhood and the race mix of the neighborhood is 
really a result of the correlation between race and income/poverty status 


in U.S. cities. If this is true, then it is important to include both income 
and race in the analysis (in our terminology, hold constant the separate 
effects of income and race). 

Another way to use the information shown in figure 5 is to use the 
estimated model to generate predictions of the amount of stop activity 
that would be expected in each census tract, given the levels of the 
variables or factors observed in each tract. Figure 6 shows comparisons 
between predicted stop activity and actual stop activity by census tract. 
In this example, the expected rates are used as benchmarks for actual 
rates. 

When the actual rates are substantially different from the predicted 
rates (i.e., when there are outliers), it may be useful to consider what 
supplementary information might explain the differences. For example, 
if actual stop rates are substantially lower than the predicted rates, it 
might be that red light cameras are being used to monitor intersections 
and that, consequently, officers are turning their efforts in other directions. 

Figure 6. Use Predicted Stop Rates for Benchmark Comparisons 

Image of a table showing the comparisons between predicted stop activity and actual stop activity by census tract. 

.Are actual numbers of stops and stop rates significantly different from 
predicted? 
. Can supplemental data explain differences? 


Another way to disentangle stop data is to look for patterns in stop 
outcomes, such as conducting a search or issuing a citation. For example, 
to analyze search outcomes, the first step is to explore whether, given 
that the stop has occurred, minorities are more likely than Whites to be 
searched. In addition, it is interesting to compare the results of searches 
by race (and possibly by other factors as well). 

Looking at the search results can indicate how effective, or productive, 
searches are across racial subcategories, and can also indicate what type 
of illegal product was most frequently discovered (e.g., weapons, drugs). 
A sample template for this type of analysis can be seen in figure 7. 

Figure 7. Use Data on Search to Analyze Profiling as Well as Search Yields 

Image of a table showing a sample template showing how effective, or productive searches are across racial subcategories, and also indicates what type of illegal product was most frequently discovered. 

.Conditional on being stopped, are minorities significantly more likely to be 
 searched? 
. Conditional on being searched, are minorities significantly more likely to 
 have contraband? 
. Do searches have high overall yield rates? 

Some authors have focused especially on stop outcomes because data 
on stop outcomes are considered to be cleaner than data on stops alone. 
Specifically, since it is usually not possible to identify the observed 
violator population (see figure 1), analysts don’t have any information 
about drivers who could have been stopped, but weren’t. In contrast, 
since stop outcomes are conditional on the occurrence of a stop, it is 
possible to identify who experienced each outcome, and equally 
important, who did not. In this case, therefore, there is no ambiguity 
about having identified the denominator correctly. 


Knowing which drivers experienced a specific outcome and which did 
not also enables researchers to use regression analysis to estimate the 
likelihood that a specific event will occur as a function of the drivers’ 
characteristics and characteristics of the stop. In other words, regression 
techniques allow us to look at a number of different variables that may 
be influencing an outcome, and figure out which of the variables 
actually have an important and substantial impact. 

Figure 8 shows a template for reporting the results from a multiple 
regression technique that estimates the separate impacts of various 
factors that influence whether a person who has been stopped is 
subsequently searched.4 The citizen race variable in the list of independent 
(or influencing) variables is the one of interest with respect to racial 
profiling questions. In this case, the model estimates the effect of the 
driver’s race on the likelihood of being searched, holding constant, or 
separately taking into account, the influence of other factors, such as 
citizen gender, car age, reason for stop, time of day, and various officer 
characteristics. 

Figure 8. Control for Multiple Factors Using Regression Analysis of Search 

Image of a table showing a template for reporting the results from a multiple regression technique that estimates the separate impacts of various factors that influence whether a person who has been stopped is subsequently searched.

. Is citizen race significant even when controlling for other aspects of the 
stop? 

4 The statistical model used here is a logit model. A logit is a non-linear regression model that is frequently used 
in analyses when the outcome variable being modeled is a probability (such as the probability that a stopped cited 
driver was searched or not), rather than a continuous variable (such as the number of citations issued in a tract). 
Logit models are better than linear models for predicting probabilities because the predictions generated by a 
logit always fall between zero and one. Logit models are often used for dependent variables with two possible 
outcomes (denoted as 0 or 1 outcomes), but can also be used for more complex dependent variables. For 
technical notes on the logit model, see Peter Kennedy, A Guide to Econometrics, MIT Press, 5th edition, 2003; 
William H. Greene, Econometric Analysis, Prentice Hall, 5th edition, 2003; and David Knoke and Peter J. Burke, 
“Log Linear Models (Quantitative Applications in the Social Sciences), Sage Publications, 1980. 


In figure 8, officer characteristics were included in the list of explanatory 
variables as a suggestion rather than as an example of typical control 
variables. Although most departments seem certain that the stopping 
officer’s race does not influence the race of the stopped (or cited or 
searched) citizen, very few departments have actually put this to the test. 
However, it is important to consider using variables that control for 
officer characteristics. 

For example, minority officers might do racial profiling of minorities, 
or they might do “reverse” profiling. The key is that different results on 
officer race and experience level may tell different stories. In popular 
terms, this notion is framed by asking whether profiling is a question of 
White vs. Black or a question of Blue vs. others. The primary reasons 
that officer data have not been included in analyses are concerns about 
privacy and liability. Therefore, the suggested method for including 
officer information is to use identification codes so that data can be 
matched to specific officers without identifying them by name. 

Another way officer identification can be used to analyze racial profiling 
is to examine officers’ stops and compare them with some benchmark. A 
benchmark can be externally provided, by taking a standard set by other 
departments, for example. Another approach is to create an internal, 
department-specific benchmark by calculating an average behavior 
among a department’s own officers, and then comparing individual 
officer data with the average for the department, unit, branch, precinct, 
or neighborhood, depending on what exact behavior is being evaluated. 

Using hypothetical data, Figure 9 shows an example of the type of 
analysis that can be performed. This example allows an evaluator to 
consider whether there are officers who stop significantly more (fewer) 
minorities than the unit’s overall average. It is important to stress that, 
although this approach identifies officers who are performing above or 
below an expected average range, this does not necessarily imply 
whether they are doing anything wrong. There may be many factors that 
would easily explain why a particular officer had an unusually high or 
low minority share of stops over a given period of time, such as a 
month. However, looking at such a metric allows department managers 
to consider these factors, and determine whether any unusual patterns 
might merit additional investigation or consideration in the future. 


Figure 9. Analyze Data at the Officer Level and Compare with Benchmark 

Image of a bar graph using hypothetical data to show how to analyze data at the officer level and compare with benchmark

.Which officers stop significantly more minorities than the unit’
.Can supplemental data explain differences? 

The key to this type of evaluation is picking the right benchmark to use 
as a comparison, and then identifying outliers and using other data to 
examine and explain them. For example, the race mix of stops for each 
officer is likely to depend on the area he or she patrols, and, within each 
area, on the time of day during which he or she is on dutythe benchmark should incorporate as many relevant factors as possible. 

It is also possible to use officer identification codes to examine issues 
other than racial profiling. Figure 10 provides a very generalized 
example of the potential use of stop data to explore ofFor a group of hypothetical officers, the figure compares the rates at 
which each officer issued citations after making stops to the group’average citation rate. Thus, in this case the group average is the 
designated internal benchmark. 


Figure 10. Using Stop Data to Analyze Officer Efficiency 

Image of a bar graph providing a very generalized example of the potential use of stop data to explore officer efficiency

Which officers give fewer citations than the unit average? 
What is the relevant benchmark? It should reflect department and unit 
goals. 

Again, the key is to select the correct benchmark and then identify 
outliers and use other data to explain them. Officers who have unusual 
citation rates may or may not have excellent reasons that explain why 
their rates are unusually high or low. For example, different units may 
have different operational mandates. The traffic unit is goaled to monitor 
traffic and ensure traffic safety. So, officers in this unit are more likely 
to be looking for specific types of violations that are associated with 
danger and may be more likely to issue tickets. In contrast, officers in 
other units will make car stops and traffic stops more opportunistically, 
and potentially for different reasons. These and many other factors (e.g., 
time of day officer is on duty) must be taken into consideration when 
comparing across officers. 


Summary 

The methods of analysis shown in this report are only a sampling of the 
types of analysis that can be performed in the context of racial profiling 
or officer evaluations. As indicated before, the type of analysis that 
should be done depends on the nature of the department, the characteristics 
of the surrounding population, and the specific questions that the 
department and citizen groups may be posing. The questions being 
asked will determine what data should be collected and how it should 
be analyzed–both with respect to modeling technique and the level of 
detail that is maintained (e.g., stops, searches, officer identification). 
Generally speaking, multivariate analysis will require careful attention to 
properly control for the various factors that influence stop rates, citations, 
and searches. A productive way to approach this type of analysis is 
through a partnership between local police operational experts and social 
science researchers versed in applying analytical methods. This type of 
partnership can greatly enhance the processes of data collection, 
analysis, and interpretation. 


Additional Resources 

Fridell, Lorie. By the Numbers: A Guide for Analyzing Race Data from 
Vehicle Stops. Washington, DC: Police Executive Research Forum, 2004. 

Fridell, Lorie, Robert Lunney, Drew Diamond and Bruce Kubu with 
Michael Scott and Colleen Laing. Racially Biased Policing: A 
Principled Response. Washington, DC: Police Executive Research 
Forum, 2001. 

McMahon, Joyce, Joel Garner, Ronald Davis, and Amanda Kraus. How 
to Correctly Collect and Analyze Racial Profiling Data: Your Reputation 
Depends On It! Washington, DC: Government Printing Office, 2002. 


Image of United States Department of Justice seal

FOR MORE INFORMATION 

U.S. Department of Justice 
Office of Community Oriented Policing Services 
1100 Vermont Avenue, N.W. 
Washington D.C. 20530 

To obtain details on COPS programs, call the 
COPS Office Response Center at 800.421.6770 
Visit the COPS internet address listed below. 

Created March 2005 e03180002