Skip navigation

A Deeper Dive into Racial Disparities in Policing in Vermont, Stephanie Seguino and Nancy Brooks, 2018

Download original document:
Brief thumbnail
This text is machine-read, and may contain errors. Check the original document to verify accuracy.
A Deeper Dive into Racial Disparities in Policing in Vermont

Stephanie Seguino
Professor
Department of Economics &
Fellow, Gund Institute for the Environment
University of Vermont
Burlington, VT 05401
stephanie.seguino@uvm.edu
and
Nancy Brooks
Visiting Associate Professor
Dept. of City and Regional Planning
Cornell University
Ithaca, NY USA 14853
nb275@cornell.edu
March 26, 2018
Acknowledgements: We would like to thank Jianxiong Huang for research assistance, and
Colchester Police Chief Jennifer Morrison, Vergennes Police Chief George Merkel, and
Vermont State Police Lt. Garry Scott and Major Ingrid Jonas for helpful comments that
informed this report.

A Deeper Dive into Racial Disparities in Policing in Vermont
EXECUTIVE SUMMARY
In 2014, the Vermont legislature passed a bill requiring all Vermont law enforcement
agencies to collect traffic stop data so as to make it possible to identify and track any racial
disparities in policing. The first round of data became available in 2016, and “Driving While
Black and Brown in Vermont” (Seguino and Brooks 2017), an analysis of that data was
released in January 2017.
The report has generated wide-ranging conversations about the role of race in policing.
Several Vermont law enforcement agencies have responded by taking the initiative to invest
in training to address potential implicit bias in policing and to improve the quality of their
data. At the same time, there have been questions raised by some observers about the quality
of the data and methodology used in the 2017 study, and a concern that the study does not
account for the context of traffic stops that may justify the racial disparities found in the
original report.
In this brief, we address these questions and report results from a logistic regression analysis
that accounts for other factors beyond race that may influence the probability of being
searched and of contraband being found. We also present new results of an analysis of race
and the types of contraband found in 2016 Vermont State Police vehicle searches.
A brief summary of our findings is as follows:
• Data quality. A concern has been raised that the data are of poor quality and therefore
should not be used to make inferences. Control for data quality lies with law
enforcement agencies. There is room for improvement to ensure complete data and
uniform methods of coding. These efforts appear to be underway. That said, we did
not find evidence of serious miscoding or overt efforts to manipulate the data in the
dataset law enforcement agencies provided.
• Home state of driver and vehicle registration, and passenger data. There is a view that racial
disparities in stops and post-stop outcomes would be smaller if the data analysis
controlled for out-of-state drivers, vehicles, and passengers. That is possible, but law
enforcement agencies have not provided that data to the public—nor does the
legislation require it. Once that is made available, those variables can be included in
the analysis.
• The benchmarking issue. A perennial question in race data analysis in policing across the
country is what is the appropriate denominator to use in calculating stop rates by
race? The best practice nationally is to use data on the race of not-at-fault drivers in
accidents, which we have done in our 2017 study. Our analysis also includes
indicators based on post-stop outcomes, where the race of the driver is known (or,
more precisely, at which time the officer has formed a perception of the race of the
driver). By basing conclusions from analysis of multiple indicators instead of only
one (such as the stop rates), we avoided the problem of “cherry-picking” indicators.
This approach provides a more accurate assessment of the data in the event there are
mismeasurement problems in any one of the indicators. Our assessment of racial
disparities in policing then relies not on any one indicator, but the patterns across all
	
  

1	
  

•

•

	
  

indicators. Insofar as those patterns are consistent across indicators, more robust
conclusions can be drawn about the degree of racial disparities.
The role of context in explaining racial disparities. In this report, we conducted logistic
regression analysis to account for all the contextual factors that are available in the
data that might influence the probability of being searched and of contraband being
found. Controlling for these factors, we find that race continues to be a statistically
significant factor in an officer’s decision to search a vehicle and in the probability of
finding contraband. Wide racial disparities persist. Specifically, Black and Hispanic
drivers continue to be roughly 2.5 to 4.0 times more likely to be searched that White
drivers, and 30 to 50 percent less likely to be found with contraband subsequent to a
search than White drivers. These findings indicate probable oversearching of Black
and Hispanic drivers compared to White drivers.
The analysis of Vermont State Police contraband data, resulting from 440 searches in
2016, provides an initial detailed racial analysis of the types of contraband found in
searches based on probable cause or reasonable suspicion. Two striking features of
these data are: 1) over 70 % of all contraband found is marijuana with 13%
comprised on heroin, cocaine, or opioids, and 2) only White drivers were found with
heroin, cocaine, and/or opioids. No drivers of any other race with found with this
contraband. This suggests that, at least for this dataset for 2016, and for these types
of searches, assumptions held by the public and law enforcement about the race of
drivers carrying this type of contraband should be revisited.

2	
  

A Deeper Dive into Racial Disparities in Policing in Vermont
I.

INTRODUCTION

In early 2017, a report on racial disparities in Vermont policing was issued by the
authors (“Driving While Black and Brown in Vermont”). Our goal was to assess the extent
of racial differences in the experience of drivers on Vermont roads. The study included data
from 29 law enforcement agencies, policing roughly 78% of the Vermont population.1 We
utilized a methodology that is easily replicable so as to provide a format for data analysis that
community groups, law enforcement, and policymakers could all use to monitor the
evolution of racial disparities in policing.
We reported on seven indicators that measure racial disparities: arrest and search rates by
race, “hit” rates2 by race, racial shares of traffic stops by agency and officer as compared to
racial shares of the driving population, racial differences in ticket rates, male shares of stops
by race, and officer stop rates by race. In our analysis, we explored the data for a pattern across
several indicators. The state-wide data showed that there were racial disparities across a number
of indicators (disparities differed by agency, however), with Black and Hispanic drivers more
likely to be stopped and arrested than White or Asian drivers. Of particular note, they were
also significantly more likely to be searched but less likely to be found with contraband than
White or Asian drivers. For reference, Table 1 reports several post-stop outcomes from the
2017 report for all Vermont law enforcement agencies combined. Panel A shows results for
all years for which we had data and Panel B is for 2015 only.
By basing conclusions drawn from the data analysis on multiple indicators instead of only
one (such as the stop rates), researchers avoid the problem of “cherry-picking” those
indicators that are consistent with their prior assumptions. This methodological approach,
which is widely accepted in social science research, also provides a more accurate assessment
of the data in the event there are mismeasurement problems in any one of the indicators.
Mismeasurement may be due to miscoding, misunderstanding of the concept of
discretionary stops (and other variables), missing data, or simple mistakes in recording.

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
1	
  Numerous agencies did not submit their data by September 2016 and so could not be included in our original
analysis.	
  	
  
2	
  A “hit” is defined as a search of a vehicle that results in contraband being found. The hit rate then is the
percentage of searches that are “productive,” that is, the percentage of searches in which contraband is found.

	
  

3	
  

Table 1. Post-Stop Outcomes, All Agencies
Panel A. All Years
White

Black

Asian

Hispanic

Arrest rate

1.1%

2.0%

0.8%

1.5%

Search rate

1.0%

3.9%*

0.6%

3.4%*

Hit rate (includes all outcomes)

76.3%

62.6%*

77.8%

64.3%*

Hit rate (outcome = arrest/ticket)

60.0%

49.9%*

79.4%

46.6%*

White

Black

Asian

Hispanic

Arrest rate

1.2%

2.1%*

1.1%

1.3%

Search rate

0.9%

3.6%*

0.5%

2.6%*

79.4%

72.8%

88.9%

75.0%

Hit rate (% of searches with contraband)

Panel B. 2015 only

Hit rate (% of searches with contraband)
Hit rate (includes all outcomes)

56.1%*

Hit rate (outcome = arrest/ticket)
67.0%
88.9%
60.7%
Note: Total stops for all years is 425,388 and for 2015, there were 107,497 stops. These data exclude externally
generated stops and searches based on a warrant. Asterisks on arrest, search, and hit rates indicate statistically
significant differences of a racial group as compared to White drivers at the 5 percent level.

Our report has since generated numerous conversations and responses. On the one hand,
some law enforcement agencies, communities, and legislators have used the results of our
paper and methodology to further conversations about racial disparities and potential racial
bias in policing. Some agencies (Vermont State Police and Burlington Police Department, in
particular) have invested heavily in training to address potential implicit bias in policing and
to improve the quality of their data. Other agencies have raised questions about the results
of our study, concerned it does not account for the context of traffic stops that may justify
the racial disparities founds in our original report. Worries have also been voiced about the
quality of the data on which our study was based.
This report addresses those concerns. We discuss the issue of data quality first. We then
present additional analysis of the data, using logistic regression analysis to further explore the
role of race by controlling for contextual factors that may be correlated with the race of the
driver. Logistic regression analysis is a method for measuring the odds a particular outcome
will occur, such as the search of a vehicle subsequent to a stop. This type of analysis allows
us to identify how various factors (such as reason for the stop, and age and gender of the
driver) contribute to those odds. In this way, we can control for the context of the event,
and avoid misattributing to race the reason for a search (or other outcome). Finally, we take
this opportunity to discuss results of a preliminary analysis of 2016 Vermont State Police
data on types of contraband found in searches that sheds light on racial outcomes by type of
contraband.

	
  

4	
  

II.

DATA ISSUES

Several concerns have been raised about the quality of the data our analysis was
based on. Questions about how the data were analyzed have also been indicated. We take
this opportunity to respond to those questions.
1. Data quality and missing data: Police leaders have expressed concern that the data are of
poor quality in the sense that some data is incomplete, and in some categories, there
is not a uniform agreement across agencies (or officers) about what actual policing
situations the data captured. (For example, one agency may record searches of
passengers while another may only record searches of drivers). Police leadership sees
the lack of uniform agreement across departments about the circumstances to be
captured as an important issue, undermining their confidence in any analysis of those
data.
In addition to quality of the data, a major problem is that the law requiring traffic
data collection by race has not been followed. The law (20 V.S.A. § 2366) requires
that by or before September 1, 2014, every state, county, and municipal law
enforcement agency is required to collect roadside stop data consisting of the
following: 1) age, gender, and race of driver, 2) reason for the stop, 3) type of search
conducted, if any, 4) evidence located, if any, and 5) outcome of the stop. All
roadside stop data, as well as reports and analysis of roadside stop data, are to be
made public. The law further required agencies by September 1, 2016 and annually
thereafter, to provide the data collected to the vendor chosen by the Criminal Justice
Training Council.
A number of agencies did not report their data by the September 2016 due date,
although some agencies subsequently submitted their data after that time. Table A1
in Appendix A lists agencies for which data was available and those whose data had
not been submitted in time for this analysis. For those that did submit data, a
number of required categories of data were missing from some agencies. A table of
missing data by agency is provided in Table A2 in Appendix A.
The question of whether racial disparities would disappear if quality were improved
is an empirical one. It will be incumbent on agencies to provide accurate data and
only then will we know the effect on racial disparities. Further, it should be noted
that all data sets have some problems of quality, whether it is educational data,
unemployment data, measures of GDP, or many others. This problem is not unique
to the Vermont traffic policing dataset. (That said, of course, the negative impact of
inaccurate data on perceptions of police are of great concern).
Absent evidence of serious miscoding or overt efforts to manipulate the data, social
science researchers use existing data to answer questions, and then typically offer
advice on how to improve data quality. Were researchers to wait for perfect datasets,
no analysis would ever be conducted. It is our professional opinion with over 50
years of combined experience in statistical analysis, that the size of the racial
disparities found in our dataset should lead to deeper reflection about the role of
race in policing in Vermont. A number of Vermont law enforcement agencies are
	
  

5	
  

now actively engaged in improving their data quality. Over time, as quality improves,
the task will be to analyze that data and to compare to earlier years to observe trends
over time. An instructive example is the Vermont State Police, which has invested
significant resources into improving data quality. Their analysis of their 2016 data
reflects this improvement, and also shows a substantial decline in missing data.3 It is
instructive that in 2016, racial disparities in stop, arrest, and search rates (notably,
between Black and White drivers) continue to be large, although search rate
disparities have narrowed from 2015.
Another important issue is the inability of a number of Vermont law enforcement
agencies to produce all the data required by law. Data reporting is new for a number
of agencies that find themselves under-resourced and without the skilled staff to
perform this work. Moreover, the legislation (cited above) requires law enforcement
agencies to work with the Criminal Justice Training Council with the goals of
collecting uniform data, adopting uniform storage methods and periods, and
ensuring that data can be analyzed. That work appears to be underway, although
funding challenges to address this problem exist. As a result, analyzing traffic stop
data has until now been an arduous process because coding is not uniform and due
to numerous instances of missing data. We are hopeful some of the quality issues are
in the process of being resolved.
2. Duplicate records: A question was raised about whether our dataset, obtained from law
enforcement agencies, included duplicate incidents. More specifically, the concern is
that in cases of more than one outcome of a stop (for example, a driver may have
been issued more than one ticket), the incident is recorded multiple times, one for
each outcome. This could skew the results and erroneously contribute to data
suggesting wider racial disparities if drivers of color were more likely to have more
than one outcome per stop. We addressed that issue in our original report.
Specifically, we “de-duplicated” our data through a complex programming
procedure, relying on incident numbers, and matching of age, gender, date of birth,
and time of day to identify duplicates.4
3. Home state of driver and vehicle registration: There is a view that racial disparities in stops
and post-stop outcomes would be smaller if the data analysis controlled for out-ofstate drivers and vehicles. That is possible, but law enforcement agencies have not
provided that data to the public—nor does the legislation require it. Once that is
made available, that variable can be included in the analysis.
4. Search and contraband data may be inaccurate because data does not differentiate between passenger
and driver. Some agencies have indicated that there is confusion about whether data
on passengers should be recorded or only on the driver. This may affect results if the
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
3	
  Vermont State Police traffic policing data can be found on their website:
http://vsp.vermont.gov/communityaffairs/trafficstops	
  
4	
  Some agencies did not provide incident numbers, making the de-duplication process more complex. In the
future, such data should be reported with incident numbers so that agencies can more easily monitor their own
traffic stop data for racial disparities. 	
  

	
  

6	
  

driver and passenger are of different races. We concur and have proposed that
agencies begin to collect passenger data as well.
5. Small sample sizes. A concern has been raised about small sample sizes for some
agencies, inhibiting the ability to make sound statistical inferences. The sample size in
our 2017 study for our full dataset was over 400,000. For some analyses, we
restricted our sample to data for 2015, yielding a sample size of over 107,000. These
sample sizes are more than adequate to make reliable statistical inferences. We made
no inferences on data from agencies if sample sizes were too small. We also
conducted tests of statistical significance of all of our results.
6. Lack of controls for context. Concerns have been raised that were we to have controlled
for driver age, gender, and reason for the stop, racial disparities would be smaller if
not non-existent. For example, younger drivers are typically more likely to exhibit
poorer quality driving than older drivers. If stopped Black or Hispanic drivers on
average are younger than White or Asian drivers, failure to control for age may result
in overstating the role of race in traffic policing. We address this concern in the next
section where we show that even when these variables are controlled for, our original
results stand: Black and Hispanic drivers are searched at a higher rate than White
drivers, but are less likely to be found with contraband.
7. The benchmarking issue. A perennial concern with traffic data analysis is that researchers
do not have an accurate measure of the driving population. In the absence of precise
data (which would be very costly to obtain), many studies use the Census population
as an estimate of the driving population. The best practice for measuring the driving
population in calculating stop rates by race, however, is to use racial shares of not-atfault drivers in accident reports as the denominator in the stop rate calculation
(Alpert, et al., 2004; McLean and Rojeck 2016). This is because not-at-fault drivers in
accidents are considered to be a representative sample of the driving population.5 In
our 2017 study, we use both the Census data and not-at-fault driver data by race from
the Department of Motor Vehicles. The accident data capture the fact that the
driving population in towns and on highways is at least partly comprised of
transients (that is, non-residents traveling through a city or town). Using both
measures, racial disparities exist to varying degrees in many of agencies in our
sample. There is a good deal of variation in stop rates by race across agencies, and
this in itself is notable.
8. Racial disparities and racial bias. There is a concern that evidence of racial disparities
may be interpreted by the public as racial bias. The two concepts have different
meanings. Racial disparities are the statistically significant differences in racial
outcomes of policing. They may be due to a variety of factors, including contextual
factors as well as implicit or unconscious officer bias. Therefore, evidence of racial
disparities should signal to law enforcement agencies the importance of digging
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
5

Another method used to identify racial shares of the driving population is an observational traffic study, where
surveyors stand on street corners, recording the perceived race of drivers. This approach is rarely used, in part
because it is very expensive and also because it too has problems of accuracy due to changing driving
populations over time.

	
  

7	
  

deeper into police practices to understand the source of those disparities. The
existence of racial disparities therefore does not in and of itself provide evidence of
racial bias (defined as the case where the race of the driver influences officer
decisions about whom to stop, ticket, arrest, or search). Evidence of disparities,
therefore, should be interpreted with caution. That said, the “gold standard” used in
the race and traffic policing literature for identifying racial bias is racial differences is
hit rates (the percentage of searches that yield contraband) [Persico and Todd 2008].
Evidence that the Black or Hispanic hit rate is significantly lower than that of White
drivers is interpreted as an indication of inefficient policing at best, and is consistent
with a claim of biased policing. We found such evidence using the statewide data and
for several agencies.
In sum, higher quality data are always welcome in the world of social science research. But in
the real world, data are imperfect. This does not preclude such data being used unless it is
found to be so flawed as to be misleading, which is not our assessment of the the data we
analyzed. In our 2017 report, we highlighted our concerns with quality of data and urged
agencies and the state to take steps to address any lacuna or inconsistencies. A few agencies
are taking steps to address this issue, with the most significant efforts being made by the
Vermont State Police followed by Burlington Police Department.
III.

DO RACIAL DISPARITIES IN SEARCH RATES AND CONTRABAND
FOUND DISAPPEAR WHEN WE CONTROL FOR CONTEXT?

A. Background
One of the apprehensions about the results of our 2017 report is that it does not
take into account the context of the stop and of post-stop outcomes. That is, it is argued
that if we controlled for time of day, gender, age, reason for stop, and other relevant
variables, racial disparities would likely disappear. Our original report did not include such an
analysis because our goal was to produce a data analysis methodology that is easily replicable
and thus a useful tool for law enforcement agencies and for community groups who wish to
monitor racial disparities in policing. It is worthwhile to note that our methodology reflects
guidelines by a number of national experts on race data analysis (Fridell 2004).
It is, however, reasonable to question whether context can explain racial disparities in
traffic stop policing in Vermont. Some driving behaviors and circumstances may co-vary
with race, and may be the dominant reason behind a search rather than the race of the
driver. For example, perhaps people of color are more likely to be younger drivers, a
group that typically exhibits poorer quality driving than older drivers. Age of the driver
may contribute to the likelihood of a search for drivers of any race. If drivers of color are
more likely to be young, compared to White drivers, then they would experience higher
search rates on average. Failing to control for such factors risks misattributing search
rate differences to race rather than the explicit behavior of the driver. If, even after
controlling for factors like gender, age, reason for stop, and time of day, we still find that
race is a statistically significant predictor of a search and a predictor of a lower likelihood
that contraband is found, then that provides evidence that the race of the driver,
independent of other factors, influences traffic policing in Vermont.
	
  

8	
  

The way that researchers address the possibility that other factors may influence the
probability of a driver being searched is to conduct multivariate logistic regression analysis.
The goal in this methodology is to determine whether there are statistically significant
differences in the probability of drivers being searched by race.6 Such analyses have been done for
other states and for Vermont agencies (Seguino, Brooks, and Mitofsky 2012; Baumgartner et
al., 2016; Pierson, et al., 2017), and have found that once context is controlled for, Black
drivers (and in some cases, Hispanic drivers) are still more likely to be searched than White
drivers.
In this section, we report the results of the logistic regression analysis on our Vermont-wide
data to respond to this concern. In our analysis, we use the data produced by the law
enforcement agencies themselves upon which our 2017 study was based.7 8
B. Probability of Search Methodology
Multivariate logistic regression analysis assesses the differential probability of being
searched for Black, Hispanic and Asian drivers as compared to White drivers. The full
model takes this general form:
Probability of Search = β0 + βb*Black + βa*Asian βna*Native American + + βh*Hispanic +
βi*Day of Weeki + βj*Reason for Stopj + βm*Male + βage*Age +
βk*Time of Dayk + βl*Agencyl + Residual.
Dummy variables for each racial group are included, with White the excluded category.
In our model, the coefficients for each of the driver race variables can be interpreted as
the odds of a search for a driver of that race as compared to the odds for White drivers
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  

Researchers cannot conduct this type of analysis on the probability of being stopped because we do not have
data on who is not stopped. Although this analysis could theoretically be conducted on the probability of being
arrested, we lack sufficient information about arrest reasons to make this analysis meaningful. Hence, studies
that use this method to assess racial disparities in policing typically analyze the probability of being searched.
We add to this an analysis of the probability of contraband being found, subsequent to a search. The reason for
this is that the “gold standard” for identifying racial disparities and potential bias in traffic policing is a
comparison of search and “hit” rates (Persico and Todd 2008). In particular, we would expect that the
probability of finding contraband is equal across all racial groups, regardless of differences in search rates. The
reason for this is that law enforcement agencies presumably have an interest in efficient use of their resources.
If agencies are over searching a group such that it has a lower “hit” rate than other racial groups, we would
expect that officers recognize the higher percentage of unproductive searches of that racial group and they
would adjust their assessment of the criteria on which they base a search.
7	
  Numerous agencies did not submit their data by the September 2106 deadline. Some have subsequently
submitted their data, but it is not included in this analysis because we want to assess the extent to which racial
disparities in our 2017 report can be explained by other contextual factors.	
  
8	
  Additional variables might have been useful to include in our analysis (discussed above), but Vermont law
enforcement agencies are not required to report to the public data on: 1) the state in which the stopped vehicle
is registered, 2) the year the vehicle was made (which might offer information on potential “poverty profiling”),
3) officer information (the age, race, gender, and job tenure of officers, for example), or 4) the specific type of
contraband found during a search. Based on other studies done in the US that control for these variables, it is
unlikely that our results would differ substantially, even if these data were available. Nevertheless, it would be
useful for agencies to also report these data so as to better assess the role of race in policing. 	
  
6

	
  

9	
  

with the same other characteristics. This is called the odds ratio, because it is the ratio of
the odds of a non-White driver being searched over the odds that a White driver is
searched.9 An odds ratio of 1 indicates equal probabilities of being searched. A ratio that
is greater than one indicates a group is more likely to be searched than the omitted (or
benchmark) group. Finally, an odds ratio that is less than 1 is indicative of a lower
probability of a group being searched relative to the omitted group.
We add driver age, measured in years, as an explanatory variable. The coefficient on Male
indicates the odds a male driver will be searched as compared to the odds a female driver
will be searched. The excluded category for the Reason for Stop set of variables is motor
vehicle violation (such as speeding). The coefficients on the Reason for Stop variables
indicate the odds of being searched for each reason for stop as compared to odds of
being searched due to motor vehicle violations where the reason is one of the following:
investigatory stop, suspicion of driving while under the influence (DWI), vehicle
equipment, and for reasons unknown (that is, the reason was not stipulated in the
incident report). This control can help to eliminate misattribution of race to search
disparities, if for example, people of color are more likely to drive older vehicles with
equipment problems, or are more likely to be DWI. Externally generated stops are
excluded from this analysis because we want to focus on searches based on officer
discretion.10
We also control for the day of the week and the time of day, with the excluded
categories, respectively, Friday and afternoon. And finally, we control for law
enforcement agencies. This helps to address the potential problem of unequal sample
sizes by agency as well as the fact that agencies are located in different parts of the state
with varying conditions that might influence search rates. Controlling for all of these
factors allows us to interpret the race variable, net of the impact of the other control
variables.11
We report results on four variations of our basic model. The need to run the regression with
more than one specification is due to the missing data problem. Because some agencies did
not report data on control variables such as gender or age, we lose those observations when
we include those independent variables in our regression.
In Model 1 in Table 2, we start with a basic model, in which race of the driver is our only
explanatory variable (this specification is analogous to our crosstab analysis in our 2017
paper). The results show that, compared to White drivers, Black drivers are 3.994 times more
likely to be searched than White drivers. (This represents the ratio of the odds of a Black
driver being search compared to the odds of a White driver being searched). In contrast,
Asian drivers are a little more than half has likely to be searched as White drivers. Native
Americans are 2.83 times more likely to be searched, and Hispanic drivers 3.49 times more
likely.
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
9	
  Appendix 2 provides a brief explanation of how to interpret odds ratios.	
  
10	
  For this reason, we also exclude searches based on a warrant.	
  	
  
11	
  Omitted variable bias is still possible if race is correlated with variables that we are not able to include in
the model. 	
  
	
  

10	
  

In Model 2, adding controls for reason for stop and agency, we find that investigatory stops
are more than 6.5 times as likely to lead to a search as a stop initiated due to a moving
violation.12 Stops based on suspicion of DWI are almost 10 times more likely to lead to a
search compared to a moving violation. Stops based on vehicle equipment are about 60%
more likely to lead to a search than stops due to a moving violation. These results are
statistically significant and not surprising. It is notable that the odds ratio by racial group
changes very little with the addition of controls for reason for stop and agency. In particular,
for example, the Black/White odds ratio changes just a small amount from 3.994 to 3.796.
Asian, Native American, and Hispanic odds ratios are similarly stable when we control for
reason for stop and agency. Coefficients on all racial variables continue to be statistically
significant at the one percent level. That is, we can reject the null hypothesis that there is no
difference in search rates by race with a high degree of certainty.
In Model 3, we incorporate all available controls, adding gender and age of driver, time of
day, and day of week to the independent variables in Model 2. Since not all agencies report
these variables, the number of observations has decreased from roughly 409,000 to about
367,000. (If all of the agencies had reported gender, age, etc., we would not have bothered
running the second specification). The odds ratio on age is below 1.0, meaning that the older
the driver, the lower the probability of being searched. The odds ratio on the gender of the
driver indicates that male drivers are more than twice as likely to be searched as female
drivers. The time of day results show that the probability of being searched is lower in the
morning than in the afternoon and about 28% greater at night than in the afternoon. The
addition of these controls slightly reduces the odds ratio of Black drivers being searched
compared to White drives to 3.2. For Hispanic drivers, the odds ratio also falls marginally to
3.022.

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
12	
  Law enforcement agencies that have reviewed their data entry practices, such as VSP, have indicated that
some searches may not have been recorded. This may have happened if, for example, a driver was arrested for
DWI, careless and negligent driving, or excessive speed. Subsequent to those arrests, a search may have been
conducted. There may therefore be an undercounting of searches in the data up to this point, although there
are efforts underway to rectify this issue. Because those were not discretionary searches, their omission may not
have had a significant impact on the data, however.	
  

	
  

11	
  

Table 2. Probability of a Search, All Years
Model 1
VARIABLES

Race only
Black

Race

Asian
Native American
Hispanic
Day of Week

Saturday

3.994***
(0.208)
0.559***
(0.094)
2.829***
(0.909)
3.498***
(0.322)

3.790***
(0.203)
0.562***
(0.095)
3.100***
(1.003)
3.326***
(0.310)
1.065
(0.056)
1.050
(0.058)
0.920
(0.051)
0.963
(0.0529)
0.947
(0.052)
1.031
(0.055)
6.564***
(0.390)
9.434***
(1.219)
3.508***
(0.391)
1.599***
(0.057)

0.010***
(0.001)
409,390

0.004***
(0.001)
408,872

Sunday
Monday
Tuesday
Wednesday
Thursday
Reason for Stop

Investigatory Stop
Suspicion of DWI
Unknown
Vehicle Equipment

Gender

Male

Age

Age

Time of Day

Morning (4AM - Noon)

Model 2
Race and
partial
controls

Night (8PM - 4AM)
Constant
Observations

Model 3
Race and all
controls
3.210***
(0.189)
0.449***
(0.086)
3.427***
(1.123)
3.022***
(0.294)
0.979
(0.055)
0.950
(0.056)
0.947
(0.056)
0.981
(0.0579)
1.010
(0.059)
1.075
(0.060)
5.978***
(0.389)
8.637***
(1.175)
5.004***
(0.689)
1.432***
(0.054)
2.075***
(0.081)
0.942***
(0.002)
0.611***
(0.031)
1.281***
(0.046)
0.023***
(0.005)
367,045

Model 4
Race and all
controls except
agency
3.213***
(0.186)
0.478***
(0.091)
3.193***
(1.044)
3.387***
(0.327)
0.932
(0.052)
0.938
(0.055)
0.978
(0.057)
1.007
(0.0591)
1.038
(0.060)
1.110*
(0.062)
5.735***
(0.359)
6.686***
(0.890)
4.491***
(0.564)
1.479***
(0.054)
2.088***
(0.081)
0.943***
(0.002)
0.649***
(0.032)
1.273***
(0.045)
0.034***
(0.002)
367,679

Note: Standard errors are in parentheses. * p<0.10, **p<0.05, ***p<0.01.

Finally, in Model 4, the regression results include all controls except for law enforcement
agency. As can be observed, this has very little effect on the odds ratio by racial group,
except for Hispanics. There, we observe the odds of being searched rises to 3.387 for
Hispanic drivers compared to White drivers.
In Table 3, we report results from restricting our sample only to nighttime searches. The
odds ratio of a Black driver being searched as compared to a White driver is 3.107, lower
than for the full sample. The odds ratio for all other racial groups compared to White drivers
falls as well. It is notable that adding the remainder of our controls (Models 2-4) has virtually
no effect on the odds ratios by race.

	
  

12	
  

Table 3. Probability of Search, Nighttime Only
Model 1
VARIABLES
Race

Race
Black
Asian
Native American
Hispanic
Saturday
Sunday
Monday
Tuesday
Wednesday
Thursday

Reason for Stop

Model 2
Race and
partial
controls

Investigatory Stop
Suspicion of DWI
Unknown
Vehicle Equipment

Age

Age

Gender

Male
Constant
Observations

3.011***
(0.247)
0.403***
(0.113)
2.559*
(1.329)
2.890***
(0.449)
0.941
(0.077)
1.174*
(0.099)
1.236**
(0.111)
1.049
(0.010)
1.076
(0.010)
1.237**
(0.106)
5.734***
(0.569)
6.488***
(0.971)
2.817***
(0.542)
1.297***
(0.071)
0.947***
(0.002)

3.122***
(0.238)
0.441***
(0.108)
2.223
(1.144)
2.491***
(0.370)

0.0657***
(0.006)
101,748

0.0173***
(0.003)
112,247

Model 3
Race and all
controls

7.333***
(0.680)
6.455***
(0.959)
2.159***
(0.334)
1.241***
(0.066)

3.011***
(0.260)
0.385***
(0.113)
2.635*
(1.373)
2.551***
(0.406)
0.962
(0.081)
1.165*
(0.101)
1.174*
(0.109)
1.006
(0.098)
1.07
(0.101)
1.223**
(0.107)
6.034***
(0.638)
7.976***
(1.233)
2.981***
(0.667)
1.274***
(0.073)
0.942***
(0.003)
1.780***
(0.106
0.0267***
(0.005)
99,156

Model 4
Race and all
controls except
agency
2.911***
(0.246)
0.389***
(0.114)
2.585*
(1.345)
2.758***
(0.434)
0.946
(0.079)
1.171*
(0.101)
1.215**
(0.112)
1.025
(0.010)
1.093
(0.102)
1.259***
(0.110)
5.701***
(0.585)
6.594***
(0.991)
2.872***
(0.595)
1.331***
(0.074)
0.945***
(0.002)
1.793***
(0.106
0.0249***
(0.004)
99,350

Note: Standard errors are in parentheses. * p<0.10, **p<0.05, ***p<0.01.

For purposes of comparison with our 2017 study, we repeat the logistic regression analysis,
limiting our sample to data for 2015 only, the year for which we have full data from all
agencies. Those results are shown in Table 4. In Model 1, where we control only for race,
Black drivers are 3.871 times more likely to be searched than White drivers. The odds of
being searched for Asian relative to White drivers are slightly more than one half (they are
half as likely to be searched, that is). And Hispanic drivers are 2.95 times more likely to be
searched, compared to White drivers. Models 2-4 show that the odds ratio of a Black driver
being searched falls slightly but is still high, ranging from 2.6 times to 3.5 times, depending
on the model’s controls. The odds of Asian drivers being searched compared to White
drivers falls even more with the added controls, while the odds ratio of a Hispanic compared
to White driver ranges from 3.0 to 3.5 times more likely, depending on the model used. (The
number of Native American drivers stopped and searched in 2015 is too small to be able to
make reliable inferences). The odds ratios reported here for each racial group can be
compared to our 2017 results in Table 1.

	
  

13	
  

Table 4. Probability of Search, 2015
Model 1
VARIABLES
Race

Model 3
Race and
all
controls

Model 4
Race and all
controls except
agency

3.871***
(0.385)
0.529*
(0.177)
2.950***
(0.555)

3.455***
(0.357)
0.531*
(0.179)
2.607***
(0.498)
1.272**
(0.133)
1.226*
(0.138)
1.007
(0.116)
1.148
(0.127)
0.935
(0.107)
1.034
(0.114)
7.618***
(0.831)
8.087***
(2.107)
5.672***
(0.940)
1.253***
(0.0964)

0.010***
(0.0003)
104,596

0.003***
(0.001)
100,916

2.651***
(0.327)
0.310***
(0.140)
2.493***
(0.510)
1.178
(0.138)
1.149
(0.144)
1.075
(0.139)
1.106
(0.142)
1.019
(0.131)
1.136
(0.139)
6.472***
(0.795)
7.178***
(2.042)
6.294***
(1.247)
1.074
(0.0947)
1.721***
(0.136)
0.944***
(0.00313)
0.436***
(0.0530)
1.298***
(0.0980)
0.018***
(0.007)
81,772

2.755***
(0.333)
0.352**
(0.159)
3.025***
(0.612)
1.126
(0.131)
1.141
(0.141)
1.089
(0.139)
1.130
(0.143)
1.038
(0.132)
1.177
(0.142)
6.318***
(0.729)
6.442***
(1.794)
7.828***
(1.394)
1.153*
(0.0969)
1.734***
(0.136)
0.945***
(0.00309)
0.454***
(0.0539)
1.325***
(0.0961)
0.034***
(0.005)
85,440

Race only
Black
Asian
Hispanic

Day of Week

Model 2
Race and
partial
controls

Saturday
Sunday
Monday
Tuesday
Wednesday
Thursday

Reason for Stop

Investigatory Stop
Suspicion of DWI
Unknown
Vehicle Equipment

Gender

Male

Age

Age

Time of Day

Morning (4AM - Noon)
Night (8PM - 4AM)
Constant
Observations

Note: Standard errors are in parentheses. * p<0.10, **p<0.05, ***p<0.01.

Figure 1 visually represents the relative likelihood of a search occurring by racial group as
compared to White drivers. Panel A shows results from Table 2 for all years and Panel B
shows results for 2015 only (from Table 4).

	
  

14	
  

Figure 1. Relative Odds of Search by Race
Panel A. All Years
4.5

4.0

3.8

4.0

3.5

3.3

3.5

3.2

3.4

3.2

3.0

3.0
2.5
2.0
1.5
1.0

0.6

0.6

0.4

0.5

Race only

Race, some
controls

Race, all
controls

Race, all
controls except
agency

0.5
0.0

Black

Asian

Hispanic

Same odds as White drivers

Panel B. 2015 Only
4.5
4.0

3.9
3.5

3.5

3.0

3.0

3.0

2.6

2.8

2.7

2.5

2.5
2.0
1.5
1.0

0.5

0.5

0.5
0.0
Race only

Black

Race, some
controls

Asian

0.3

0.4

Race, all controls

Race, all controls
except agency

Hispanics

Same odds as White drivers

Note: See Tables 2 and 4 for exact odds-ratios. The horizontal (purple) line indicates the situation of
a non-White racial group having the same odds of being searched as Whites. Columns that exceed
that bar indicate higher odds for a racial group being searched compared to White drivers, and below
the bar indicates a lower probability of a search occurring for that racial group compared to White
drivers.

	
  

15	
  

Taken together, this evidence suggests that racial disparities in search rates are extremely
robust at the state level, regardless of the contextual factors controlled for. (As we noted
earlier, the ability to make inferences about search rates disparities at the agency level
depends on the sample size. Only four agencies have sufficient data on which to make
statistical inferences about search rates disparities—Burlington, Rutland, South Burlington,
and VSP). Moreover, the levels of disparity indicated by the logistic regressions are
very similar to those suggested by the search rate ratios in our original 2017 study
(see Table 1). The use of more rigorous statistical techniques does not in any
meaningful way change the nature of our 2017 findings. They simply reinforce the
stark racial differences we reported in our earlier study.
C. The Probability of Finding Contraband
We conduct logistic regression analysis to assess the role of race in the probability of finding
contraband, subsequent to a search. In our 2017 study, we found that Black and Hispanic
drivers were less likely to be found with contraband than White or Asian drivers. As in the
analysis of search rates, we control for other factors that may influence the probability of
contraband being found to avoid erroneously attributing to race the effect of other factors.
Again, we exclude externally generated stops and searches based on a warrant. The equation
we estimate is as follows:
Probability of Finding Contraband = β0 + βb*Black + βa*Asian βna*Native American +
βh*Hispanic + βi*Day of Weeki + βj*Reason for Stopj +
βm*Male + βage*Age + βk*Time of Dayk + βl*Agencyl +
Residual.
Table 5 reports the results of the probability of contraband found for searches for any
outcome of the stop and search (that is, in which the result was a warning, a citation, or an
arrest) for all years for which we have data.13 The results shown for Model 1, where the only
explanatory variable is race of the driver, indicate that Black drivers are less than half as likely
to be found with contraband, subsequent to a search, as White drivers. There is no
statistically significant difference in the probability of Asian drivers being found with
contraband compared to White drivers, while Hispanic drivers’ likelihood of being found
with contraband is roughly 50% less than White drivers’. The addition of controls in Models
2-4 does not in any meaningful way alter the odds ratios of finding contraband in searches of
Black, Asian, and Hispanic as compared to White drivers.

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
13	
  Results for 2015 from our 2017 study indicated that hit rate disparities were slightly smaller and were
statistically significant only for contraband resulting in an arrest or ticket. 	
  
	
  

16	
  

Table 5. Probability of Contraband, All Outcomes
	
  	
  	
  

Model 1

Black
Asian
Native American
Hispanic

	
  
Day of Week

Model 3
Race and
all
controls

Model 4
Race and all
controls except
agency

0.484***
(0.058)
0.993
(0.449)
0.319*
(0.206)
0.492***
(0.102)

0.524***
(0.066)
0.894
(0.412)
0.306*
(0.201)
0.442***
(0.096)
0.667***
(0.010)
0.872
(0.141)
0.631***
(0.097)
0.579***
(0.088)
0.636***
(0.097)
0.718**
(0.108)
1.195
(0.190)
0.995
(0.360)
1.098
(0.299)
1.057
(0.010)

4.700***
(0.204)
4,197

3.696***
(0.964)
4,189

0.468***
(0.064)
0.686
(0.350)
0.385
(0.264)
0.409***
(0.091)
0.668**
(0.109)
0.844
(0.149)
0.623***
(0.103)
0.598***
(0.099)
0.611***
(0.099)
0.682**
(0.109)
1.284
(0.232)
1.034
(0.377)
0.669
(0.250)
1.080
(0.109)
1.253**
(0.133)
0.971***
(0.00408)
0.774*
(0.103)
0.939
8.588***
(4.480)
3,715

0.451***
(0.059)
0.764
(0.384)
0.399
(0.273)
0.459***
(0.098)
0.745*
(0.117)
0.869
(0.148)
0.689**
(0.111)
0.625***
(0.100)
0.656***
(0.103)
0.737**
(0.115)
1.218
(0.210)
0.854
(0.277)
0.423***
(0.137)
1.016
(0.099)
1.259**
(0.129)
0.972***
(0.004)
0.687***
(0.0871)
0.869
13.71***
(2.683)
3,721

Race only

VARIABLES	
  
Race

Model 2
Race and
partial
controls

Saturday
Sunday
Monday
Tuesday
Wednesday
Thursday

	
  
Reason for Stop

Investigatory Stop
Suspicion of DWI
Unknown
Vehicle Equipment

	
  
Gender

Male

Age

Age

Time of Day

Morning (4AM - Noon)

	
  
	
  
	
  	
  

Night (8PM - 4AM)
Constant
Observations

Note: Standard errors are in parentheses. * p<0.10, **p<0.05, ***p<0.01.

We re-ran the regressions, focusing on contraband that resulted in issuance of a citation or in
an arrest. Minor cases of contraband, such as an open container or a juvenile in possession
of cigarettes, were coded as no contraband. The results, shown in Table 6, indicate that by
recoding warnings as no contraband, the probability of Black and Hispanic drivers being
found with contraband is similar to when all types of contraband are included. In this
regression, we found that searches following investigatory stops were more correlated with
serious contraband being found compared to stops due to moving violations. Searches
following stops due to vehicle equipment problems were less likely to result in contraband
being found that triggers a ticket or arrest.

	
  

17	
  

Table 6. Probability of Contraband, Citations and Arrests Only

VARIABLES
Race

Black
Asian
Native American
Hispanic

	
  
Day of Week

Saturday

Model 1

Model 2

Model 3

Race only

Race and
partial
controls

Race and
all
controls

0.612***
(0.070)
2.077*
(0.903)
0.503
(0.322)
0.498***
(0.100)
0.842
(0.101)
0.971
(0.124)
0.817
(0.103)
0.713***
(0.0891)
0.791*
(0.0971)
0.726***
(0.086)

0.550***
(0.069)
1.574
(0.750)
0.563
(0.367)
0.429***
(0.090)
0.902
(0.116)
1.018
(0.140)
0.832
(0.110)
0.722**
(0.0966)
0.754**
(0.0975)
0.707***
(0.088)

0.508***
(0.061)
1.580
(0.745)
0.574
(0.371)
0.422***
(0.084)
0.982
(0.122)
1.071
(0.142)
0.908
(0.117)
0.737**
(0.0951)
0.765**
(0.0958)
0.733**
(0.089)

1.781***

1.754***

1.651***

(0.251)
1.087
(0.341)
0.863
(0.205)
0.656***
(0.050)

(0.272)
1.029
(0.325)
0.575
(0.195)
0.645***
(0.052)
1.316***
(0.115)
0.988***
(0.004')
0.839
(0.0942)
0.988
(0.0801)
3.787***
(1.823)
3,721

(0.242)
0.984
(0.275)
0.458***
(0.134)
0.626***
(0.049)
1.339***
(0.113)
0.988***
(0.004)
0.729***
(0.0777)
0.933
(0.0713)
2.921***
(0.455)
3,721

0.538***
(0.058)
2.087*
(0.888)
0.541
(0.343)
0.471***
(0.089)

Sunday
Monday
Tuesday
Wednesday
Thursday
	
  
Reason for
Stop

Investigatory Stop
Suspicion of DWI
Unknown

	
  
Gender

Vehicle Equipment
Male

Age

Age

Time of Day

Morning (4AM - Noon)
Night (8PM - 4AM)

	
  

Constant
Observations

1.848***
(0.064)
4,197

0.666*
(0.161)
4,195

Model 4
Race and
all
controls
except
agency

Note: Standard errors are in parentheses. * p<0.10, **p<0.05, ***p<0.01.

Using the data reported in Tables 5 and 6, Figure 2 visually portrays these odds ratios. Panel
A gives the odds ratios of being found with contraband for Blacks, Asians, and Hispanics
compared to White drivers for all outcomes of a stop and search, and Panel B reports odds
ratios for those cases where the outcome was a citation or an arrest only. In both panels, it
can be observed that Black and Hispanic drivers are significantly less likely to be found with
contraband than White drivers. The Asian/White odds ratio differs, however, depending on
the controls.

	
  

18	
  

Figure 2. Relative Odds of Contraband Found by Race, All Years
Panel A. All Contraband Found
1.2
1.0

1.0

0.9

0.8
0.6

0.8

0.7
0.5

0.5

0.5

0.4

0.5

0.4

0.5

0.5

0.4
0.2
0.0
Race only
Black

Race, some
controls
Asian

Race, all
controls
Hispanic

Race, all
controls except
agency
Same odds as White drivers

Panel B. Contraband Leading to Citation or Arrest
1.4

1.22

1.14

1.2
1.0

0.85

0.8
0.6

0.44

0.42

0.49
0.39

0.4

0.43

0.45 0.47
0.34

0.38

0.2
0.0
Race only
Black

Race, some
controls
Asian

Race, all
controls
Hispanic

Race, all
controls except
agency
Sames odds as White drivers

Note: See Tables 5 and 6 for exact odds ratios. The horizontal (purple) line indicates the situation of
the racial group having the same odds of being found with contraband as Whites. Columns that
exceed that bar indicate higher odds for a racial group to be found with contraband compared to
White drivers, and below the bar indicates a lower probability of a contraband being found compared
to White hit rates. The data in Panel A are for all outcomes of the search, while in Panel B, data are for
contraband that led to a citation or an arrest. Externally generated stops are excluded as are searches
on warrant and arrests on warrant.

To sum up the results of the logistic regressions, adding controls for a variety of contextual
factors has little effect on racial disparities in the probability of being searched and of
contraband being found during a search. This is not to say that the controls were not
	
  

19	
  

meaningful or significant. Searches and the likelihood of finding contraband are more likely
to happen under some conditions as compared to others (e.g., during investigatory stops as
compared to motor vehicle stops). But even controlling for these factors, race continues
to be a statistically significant factor in an officer’s decision to search a vehicle.
Moreover, and with regard to the question of racial bias as an explanation for such
disparities, the analysis shows that Black and Hispanic drivers are less likely to be
found with contraband, a finding that is consistent with oversearching of those
groups. These results are consistent with those reported in our 2017 analysis.
IV.

2016 VERMONT STATE POLICE CONTRABAND DATA

In traffic stop reports made available to the public, Vermont law enforcement
agencies identify searches that result in contraband being found, but do not indicate what
that contraband is. Vermont State Police (VSP) generously agreed, however, to provide
details on the type of contraband found in 2016 searches as well as the degree of the offense
(civil or criminal). The dataset provided by the VSP includes the gender, race, and age of the
driver, the reason for the stop, and outcome of the stop.14
These data are useful to better understand both the types of contraband found in searches
and the racial composition of drivers found with various types of contraband. This is
especially important since Vermont police officers frequently state that racial disparities in
policing (and searches) are linked to concerns about drug trafficking, especially along the
borders of Vermont with concerns about drugs entering from people of color from out-ofstate.
The VSP dataset is comprised of 369 unique incidents in which contraband was found
during a search.15 The total number of searches was 440. We coded the contraband data into
8 groups. Those groups are summarized in Table 7 along with the percentage of searches
with each type of contraband. The data include arrests on warrant in order to provide a full
spectrum of the types and relative importance of various types of contraband found in
discretionary searches.16 (In total, there were six incidents with an arrest based on a warrant,
all of White drivers).

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
14	
  Data on city of the stop, agency, stop date, day, and time were also provided, although we do not analyze
those data here.	
  	
  
15 The data exclude externally generated stops and searches based on warrants. 	
  
16	
  By discretionary, we mean searches in which troopers made a decision to search, based on their assessment
of the evidence.

	
  

20	
  

Table 7. Types of Contraband
Number of
incidents

Percentage of Total

Cigarettes

1

0.3%

Cash/Counterfeit money

1

0.3%

Alcohol

23

6.2%

Drug Paraphernalia

39

10.6%

Opioids

9

2.4%

Marijuana (includes + drug
paraphernalia)

264

71.5%

Marijuana + opioids, LSD,
mushrooms, alcohol in any
combination

5

1.4%

Cocaine/Heroin

28

7.6%

Total

369

100.0%

Type of Contraband

Marijuana is the most frequently identified type of contraband, comprising 71.5% of all
searches in which contraband was found. This is followed by drug paraphernalia17 (10.6%),
and then cocaine/heroin (7.6%). Opioid contraband comprised 2.4% of incidents.
It is useful to clarify Vermont law on marijuana possession in effect in 2016. Vermont
decriminalized possession of marijuana in 2013, removing criminal penalties on small
amounts and replacing them with civil fines. Possession of more than an ounce of marijuana
remains a criminal offense.18 It is, however, still common practice for police across Vermont
to request consent to search or apply for a warrant when they smell the odor of or see
evidence of marijuana use in plain view. This is because driving under the influence
continues to be an offense and therefore is considered a public safety matter. Further, it
remains a crime to sell marijuana, and an officer may conduct a search to determine how
much marijuana is contained in a vehicle.
Table 8 shows racial shares of all contraband cases and degree of offense (civil vs.
criminal).19 White drivers comprised 91.3% of all drivers with contraband in 2016, and a
higher percentage of all drivers with criminal quantities and types of contraband (96.9%).
Black drivers were 5.9% of all drivers with contraband and 3.1% of drivers with criminal
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  

Drug paraphernalia is not automatically considered contraband. If it contains residue, it would be illegal. On
the other hand, a new glass blown pipe unused is not illegal and would not be counted as contraband.
18 The law has since changed. In January 2018, Vermont legalized recreational marijuana, allowing adults over
the age of 21 to possess an ounce of marijuana and to possess up to two mature plants and four immature
plants in their home. The law stipulates that people convicted of possessing more than one ounce of marijuana,
or more than two mature and four immature plants, can be imprisoned up to six months and fined $500 unless
they participate in a court diversion program. Impaired driving remains illegal under the new law, and neither
drivers nor passengers are allowed to use marijuana in a vehicle. Anyone with marijuana in a vehicle can be
fined $200. 	
  
19	
  Criminal and civil offenses differ in terms of their punishment. Criminal cases may have jail time as
punishment, whereas civil cases generally only result in monetary damages or a requirement to do or not do
something.
17

	
  
	
  

21	
  

amounts and types of contraband. No Asian or Hispanic drivers were found with
contraband that led to a criminal offense. It should be noted that the sample sizes for
Asians, Blacks, and Hispanics are too small to make statistical inferences. These numbers are
therefore illustrative; more data would be required to identify statistically significant
differences in racial shares of incidents and criminal incidents.
Table 8. Racial Shares of Incidents by Degree
Number

Civil

Criminal

Share of
all
incidents

Asian

2

2

0

0.5%

Share of
criminal
incidents
0.0%

Black

22

20

2

5.9%

3.1%

Hispanic

8

8

0

2.2%

0.0%

White

337

274

63

91.3%

96.9%

Total
369
304
65
100.0%
Note: Data do not include searches with arrests on warrant.

100.0%

Table 9 provides data on racial shares of three key types of contraband: marijuana,
heroin/cocaine, and opioids. The majority of marijuana incidents were civil offenses
(93.8%). White drivers comprised 93.3% of the incidents in which criminal quantities of
marijuana were found, and Black drivers were 6.7%. Again, it is worth noting that the small
sample sizes, particularly for Blacks, indicates these data should be interpreted with caution.
More data would be required to make sound statistical inferences from these data.
In the case of heroin/cocaine and opioids, all incidents resulted in criminal offenses, and all
were of White drivers. No Black, Asian, or Hispanic drivers were found to be in possession
of heroin/cocaine or opioids in the searches conducted by Vermont State Police in 2016.20
Table 9. Racial Shares of Marijuana, Heroin/Cocaine, and Opioid Contraband
Marijuana
Heroin/Cocaine
Opioids
Civil

Criminal

Racial
Shares of
Criminal

Civil

Criminal

Racial
Shares of
Criminal

Civil

Criminal

Racial
Shares of
Criminal

Asian

1

0

0.0%

0

0

0%

0

0

0%

Black

17

1

6.7%

0

0

0%

0

0

0%

Hispanic

4

0

0.0%

0

0

0%

0

0%

White
228
14
93.3%
0
Note: These data do not include arrests on warrant.

31

100%

0

0
9

Race

100%

In sum, the 2016 VSP contraband data are best understood as illustrative. More data would
be needed (and from more agencies) to draw statistically sound inferences with regard to the
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
20	
  Of the five incidents in which marijuana was found in some combination with opioids, LSD, mushrooms,
alcohol in any combination, four incidents involved White drivers and one a Black driver. 	
  
	
  

22	
  

relationship between race and contraband. That said, the data offer a means to compare and
contrast assumptions held by officers in some local law enforcement agencies (though not
VSP) with regard to contraband and race. The data presented here suggest that some of
those assumptions may be inaccurate, at least with regard to contraband that turns up during
discretionary searches.
The notable highlights of these data are as follows:
• Marijuana is overwhelmingly the most frequently identified contraband in VSP
searches in 2016, comprising 71.5% of contraband found subsequent to a search. Of
those incidents, only 5.6% were criminal cases. The remainder were civil.
• Heroin/cocaine and opioids combined were found in 10% of all searches yielding
contraband. White drivers comprised 100% of those found with these types of
contraband.
V.

CONCLUSION

Vermont has embarked on a long-term project of using data to expand awareness of traffic
policing and race. Because traffic stops are the most frequent interaction people have with
the police, combined with the large number of traffic stops in any given year, data on stops
can be a useful tool for understanding the extent of racial disparities in these interactions.
They are, in other words, a way of holding up a mirror to ourselves.
Though data often and usually are imperfect, that does not preclude their usefulness.
In this report, we have discussed an array of concerns with traffic stop data quality, many of
which have been shared with us by police. Efforts to improve data quality are important and
should continue to be pursued. It is clear that there are a number of agencies pursuing that
goal. In the interim, however, the data we do have are useful at gauging racial disparities in
policing and give no evidence of being so systematically flawed that they are unusable.
In this report, we provide details on a statistical analysis that controls for other factors
that may influence the probability of being searched or of contraband being found
during a search. Those results demonstrate that while other factors also contribute to the
likelihood of either of those outcomes, racial disparities continue to exist when those
factors are controlled for. In particular, Black and Hispanic drivers in Vermont are
substantially more likely to be searched than White or Asian drivers, and are less likely to
be found with contraband. The levels of disparity indicated by the logistic regressions are
very similar to the search and hit rate ratios in our original 2017 study. The use of more
rigorous statistical techniques therefore does not alter the nature of our 2017 findings.
These disparities should be of great concern to law enforcement agencies, communities,
and legislators. While the disparities in no way suggest that agencies are intentionally
profiling people of color, they do indicate the necessity for law enforcement to be selfreflective about their policing practices and to interrogate the role of implicit bias in
decision-making. Research shows that implicit racial bias is evident in numerous
domains, not just policing. As its name suggests, it is often unconscious rather than
intentional. Several agencies have planned or are planning implicit bias trainings, a
positive step to work toward fair and unbiased policing in Vermont. The Vermont State

	
  

23	
  

Police has gone beyond this to rigorously examine a wide array of practices, procedures,
and policies to ensure fair and impartial policing at every level.
Finally, with regard to the descriptive analysis of 2016 VSP contraband, it is instructive
that for searches turning up heroin, cocaine, and opioids, drugs that are so much in the
Vermont news of late, only White drivers were found with such contraband. There may
be other aspects of drug trafficking in Vermont not reflected in these data. But the data
tell us that in terms of discretionary searches in the course of traffic policing, the
stereotype, held by society as a whole, that people of color are more likely to be drug
traffickers is erroneous.

	
  

24	
  

APPENDIX A
Table A1. Agencies Represented in 2017 Study
Agencies Included in 2017 Study
Addison County Sheriff

Agencies Not Included in 2017 Study
Bellows Falls

Pittsford

Barre City

Bennington County Sheriff

Richmond

Barre Town

Berlin

Royalton

Bennington

Bradford

Rutland Town

Brandon

Brighton

Shelburne

Brattleboro

Caledonia County Sheriff

Stowe

Bristol

Canaan

Swanton

Burlington

Castleton

Thetford

Colchester

Chester

VT Dept. of Fish and Wildlife

Essex

Chittenden County Sheriff

VT Dept. Of Motor Vehicles

Grand Isle County Sheriff

Dover

Washington County Sheriff

Hinesburg

Essex County Sheriff

Waterbury

Manchester

Fair Haven

Weathersfield

Middlebury

Fairlee

Wells Constabulary

Milton

Franklin County Sheriff

Wilmington

Montpelier

Hardwick

Windham County Sheriff

Northfield

Hartford

Windsor County Sheriff

Randolph

Killington

Winhall

Rutland

Lamoille County Sheriff

Woodstock

Rutland County Sheriff

Ludlow

Springfield

Lyndonville

St. Albans

Mendon Constabulary

St. Johnsbury

Morristown

UVM

Newport

Vergennes

Norwich

VSP

Orange County Sheriff
Williston
Orleans County Sheriff
Note: Those agencies in italics provided data but it was incomplete or in a format that was unusable.

	
  

25	
  

Table A2. Missing Values in Data Submitted by Agencies
Agency
Addison County
Sheriff
Barre City

Race

Reason
for Stop

Outcome

Search

Search
Outcome

Gender

Age

Time of
Day

6,020

17.4%

0.3%

0.3%

0.2%

100.0%

0.2%

17.5%

0.0%

602

2.8%

4.3%

3.5%

10.8%

11.3%

0.0%

0.0%

100.0%

Barre Town

2,852

0.1%

0.5%

0.0%

0.0%

0.0%

0.1%

0.1%

0.0%

Bennington

5,213

3.0%

0.4%

0.4%

0.4%

2.7%

1.2%

97.5%

0.0%

Brandon

2,010

10.0%

15.8%

5.7%

7.3%

7.4%

6.2%

5.8%

0.0%

Brattleboro

7,945

11.5%

0.3%

1.4%

1.4%

1.6%

0.2%

0.3%

0.0%

Bristol

576

8.7%

4.2%

3.0%

0.0%

0.0%

1.0%

2.6%

0.0%

Burlington

24,850

4.8%

3.2%

1.1%

1.8%

1.8%

2.1%

1.7%

0.0%

Colchester

7,116

3.2%

1.5%

0.4%

0.0%

0.0%

0.1%

90.4%

0.0%

Essex

4,987

3.1%

2.1%

0.5%

1.9%

2.0%

2.1%

100.0%

0.0%

Grand Isle

0.0%

0.0%

0.0%

0.0%

0.0%

0.0%

0.0%

4,481

0.0%

Hinesburg

918

0.0%

0.0%

0.0%

0.0%

0.0%

0.0%

0.0%

0.0%

Manchester

2,885

3.8%

15.7%

14.3%

19.0%

20.1%

0.5%

0.4%

0.0%

Middlebury

1,465

0.8%

3.7%

0.1%

0.1%

1.6%

0.1%

0.0%

0.0%

Milton

5,672

4.2%

2.2%

1.9%

2.2%

2.2%

3.7%

3.6%

0.0%

Montpelier

4,147

7.6%

5.7%

1.4%

1.3%

1.8%

0.8%

0.5%

0.0%

Northfield

617

0.8%

1.3%

1.5%

1.3%

1.0%

0.7%

100.0%

0.0%

Randolph

292

1.4%

1.7%

0.0%

7.2%

7.2%

0.0%

0.0%

100.0%

Rutland

13,523

2.6%

0.0%

0.0%

0.0%

0.0%

0.4%

4.1%

0.0%

Rutland County
Sheriff

4,384

3.7%

8.0%

1.5%

100.0%

100.0%

2.4%

0.5%

0.0%

S. Burlington

11,644

0.5%

0.7%

0.1%

0.2%

0.2%

100.0%

100.0%

0.0%

Springfield

6,356

0.0%

0.0%

0.0%

0.0%

0.0%

0.0%

0.0%

0.0%

St. Albans

7,657

28.9%

6.2%

3.5%

3.9%

100.0%

3.7%

1.3%

0.0%

St. Johnsbury

4,175

2.5%

1.7%

3.0%

1.7%

1.5%

0.5%

0.3%

0.0%

UVM

8,851

1.3%

28.8%

1.7%

1.8%

1.9%

62.1%

62.2%

0.0%

Vergennes

1,808

2.2%

0.8%

0.1%

0.1%

0.6%

0.0%

4.6%

0.0%

Williston

11,912

1.0%

0.8%

1.8%

0.9%

1.0%

0.5%

4.9%

0.0%

Winooski

5,283

5.4%

10.8%

8.0%

8.6%

9.1%

100.0%

3.2%

0.0%

283,285

1.7%

0.8%

0.8%

0.8%

1.0%

0.6%

0.4%

0.0%

2.9%

2.0%

1.1%

2.1%

5.4%

5.8%

8.8%

0.2%

VSP
Percent
missing data

	
  

Total
Incidents

26	
  

APPENDIX B
A Primer on Interpreting Odds and Odds-ratios
The first column of Table 5 with the odds ratios is copied below with an additional column
showing the difference in the probability of finding contraband for each race group relative
to White drivers. These are the numbers used in the example below.
The Odds of an outcome occurring as a result of a police stop is defined as the ratio of the
probability of that outcome occurring over the probability of that outcome not occurring.
In the case of contraband being found subsequent to a search, the probability of finding
contraband when the driver is White is 0.8113 and thus the probability of not finding
contraband is 0.19. The odds of finding contraband when the driver is White are then
0.81/0.19 = 4.3 to 1. If the driver is Black, the probability of finding contraband is 13
percentage points lower (0.132 from table below), so the probability of finding contraband is
0.68 for a black driver and the probability of not finding contraband is 0.32. The odds of
finding contraband when the driver is Black are 0.68/0.32 = 2.125 to 1.
The Odds Ratios reported in the regression equations are the ratios of odds. Continuing our
example, in Table 5, the first number in the first column is 0.484. This number is the odds
ratio for a Black driver relative to a White driver. Or in other words, it is the odds for a
Black driver divided by the odds for a White driver, thus 0.484 is equal to 2.125/ 4.32 (slight
difference due to rounding). Thus, for a Black driver, the odds of a search yielding
contraband is less than half the odds for a White driver.
Table B1. Odds-Ratio Example with Contraband Found
Variables
y = Pr(ContrabandFound) (for
benchmark group= White)

Race only

Marginal effects
0.8113

	
  

Black

0.484***

-0.132***

	
  	
  
Asian

-0.0576

-0.025

0.993

-0.001

	
  	
  
Native American

-0.449

-0.069

0.319*

-0.233

-0.206

-0.157

0.492***

-0.131**

-0.102

-0.044

	
  	
  
Hispanic
	
  	
  
Note: Data are from Table 5, Model 1.

Similarly, in the case of the search regressions, the baseline probability of being searched
when the driver is White is a little over 1% (0.01058 in the table below). The odds of being
searched are 0.01058/0.98942 = 0.01 to 1 which, in this case, is not very different from the

	
  

27	
  

search rate itself. This will be true when the probability is very low. The search rate for a
Black driver is almost 3 percentage points higher (0.029 from table below) than that White
search rate of 0.01058. Consequently, the odds of a black driver being searched are
(0.029+0.01058)/0.96042 = 0.04 to 1. The ratio of these odds, with a little rounding, is
estimated to be 0.04/0.01 while the precise odds ratio is 3.994. A Black driver is almost 4
times as likely to be searched as a White driver in Vermont.
Table B2. Odds-Ratio Example with Searches
Variables
y = Pr(Searched) for benchmark

	
  	
  

group = White)

	
  	
  

	
  

Black

	
  

Asian

	
  

Native American

	
  

Hispanic

	
  

Note: Data are from Table 2, Model 1.

	
  

Marginal
effects

Race only

28	
  

0.01058

	
  	
  
3.994***

0.029***

-0.208

-0.0019

0.559***

-0.005***

-0.0939

-0.001

2.829***

0.019*

-0.909

-0.009

3.498***

0.025***

-0.322

-0.003

REFERENCES
Alpert, G., M. Smith, and R. Dunham. 2004. “Toward a Better Benchmark: Assessing the Utility of Notat-Fault Traffic Crash Data in Racial Profiling Research.” Justice Research and Policy 6(1): 43-69.
Baumgartner, F., L. Christiani, D. Epp, K. Roach, and K. Shoub. 2017. “Racial Disparities in Traffic Stop
Outcomes.” Duke Forum for Law and Social Change 9: 21-53.
Farrell, A., L. Rumminger, and J. McDevitt. 2005. New Challenges in Confronting Racial Profiling in the 21st
Century: Learning from Research and Practice. Northeastern University Institute on Race and Justice.
Fisher, D., M. Fargen, and V. Morris. 2017. “2016 Traffic Stops in Nebraska: A Report to the Governor
and the Legislature on Data Submitted by Law Enforcement.” Nebraska Crime Commission.
https://ncc.nebraska.gov/sites/ncc.nebraska.gov/files/doc/Traffic_Stops_in_Nebraska_COMPLETE_F
INAL_0.pdf
Fridell, L. 2004. “A Guide for A.01nalyzing Data from Traffic Stops.” Washington, DC. Police Executive
Research Forum.
https://cops.usdoj.gov/html/cd_rom/mayors72nd/pubs/ExecutiveSummaryBytheNumber.pdf
McClean, K. and J. Rojek. 2016. “Traffic Stops, Race, and Measurement.” In B. Huebner and T. Bynum
(eds), The Handbook of Measurement Issues in Criminology and Criminal Justice, pp. 452-472. Wiley and Blackwell.
Missouri Attorney General. Vehicle Stops Report, 2000-2016. https://ago.mo.gov/home/vehicle-stopsreport.
Persico, N. and P. Todd. 2008. “The Hit Rates Test for Racial Bias in Motor Vehicle Searches.” Justice
Quarterly 25: 37-53.
Pierson, E., C. Simiou, J. Overgoor, S. Corbett-Davies, V. Ramachandran, C. Phillips, and S. Goel. 2017.
“A Large-scale Analysis of Racial Disparities in Police Stops Across the United States.” Stanford
University.
Seguino, S., N. Brooks, and K. Mitofsky. 2012. “Racial Disparities in Policing? An Assessment of 2009-10
Traffic Stop Data in Chittenden County, Vermont.” University of Vermont and Cornell University.
Seguino, S. and N. Brooks. 2017. “Driving While Black and Brown in Vermont.” University of Vermont
and Cornell University. https://www.uvm.edu/giee/pdfs/SeguinoBrooks_PoliceRace_2017.pdf

	
  

29