Content uploaded by David C. Pyrooz
Author content
All content in this area was uploaded by David C. Pyrooz on Jan 07, 2018
Content may be subject to copyright.
1
WAS THERE A FERGUSON EFFECT ON CRIME RATES IN LARGE U.S. CITIES?
David C. Pyrooz*
Department of Sociology
University of Colorado Boulder
Scott H. Decker
School of Criminology and Criminal Justice
Arizona State University
Scott E. Wolfe
Department of Criminology and Criminal Justice
University of South Carolina
John A. Shjarback
School of Criminology and Criminal Justice
Arizona State University
* forthcoming in the Journal of Criminal Justice
doi: http://dx.doi.org/10.1016/j.jcrimjus.2016.01.001
This is the authors’ pre-print copy of the article.
Please download and cite the post-print copy published on the JCJ website (link above)
*Correspondence concerning this article should be addressed to David C. Pyrooz, Department of
Sociology and Institute of Behavioral Science, UCB 483, University of Colorado Boulder,
Boulder, CO, 80309-0483, USA. David.Pyrooz@colorado.edu
2
WAS THERE A FERGUSON EFFECT ON CRIME RATES IN LARGE U.S. CITIES?
1
ABSTRACT
Purpose: There has been widespread speculation that the events surrounding the shooting death
of an unarmed young black man by a white police officer in Ferguson, Missouri—and a string of
similar incidents across the country—have led to increases in crime in the United States. This
study tested for the “Ferguson Effect” on crime rates in large U.S. cities.
Methods: Aggregate and disaggregate monthly Part I criminal offense data were gathered 12
months before and after August 2014 from police department data requests and websites in 81
large U.S. cities. The exogenous shock of Ferguson was examined using a discontinuous growth
model to determine if there was a redirection in seasonality-adjusted crime trends in the months
following the Ferguson shooting.
Results: No evidence was found to support a systematic post-Ferguson change in overall,
violent, and property crime trends; however, the disaggregated analyses revealed that robbery
rates, declining before Ferguson, increased in the months after Ferguson. Also, there was much
greater variation in crime trends in the post-Ferguson era, and select cities did experience
increases in homicide. Overall, any Ferguson Effect is constrained largely to cities with
historically high levels of violence, a large composition of black residents, and socioeconomic
disadvantages.
Conclusions: The national discourse surrounding the “Ferguson Effect” is long on anecdotes and
short on data, leaving criminologists largely on the sidelines of a conversation concerning one of
the most prominent contemporary issues in criminal justice. Our findings are largely consistent
with longstanding criminological knowledge that changes in crime trends are slow and rarely a
product of random shocks.
KEYWORDS: crime trends; Ferguson; policing
1
The authors wish to acknowledge the dozens of law enforcement agencies who willingly made their data available
to us. We are indebted to countless crime analysts, police chiefs and command staff for their efforts on our behalf.
We would also like to thank the many colleagues that helped in this data acquisition effort as well.
3
Crime is one of the most important influences on the quality of life in the United States.
Beyond the harm faced by crime victims, crime rates are a key structural feature of communities
that are associated with long-term negative consequences for residents and the well-being of
cities. These consequences include the stigma of being identified as a “bad neighborhood” or
“dangerous city” to outsiders and insiders alike, and the resulting deleterious consequences for
the area’s economic landscape (Besbris, Faber, Rich, & Sharkey, 2015; Sampson, 2012; Xie &
McDowall, 2010). Accordingly, understanding changes in crime rates is important for a wide
range of academic disciplines and has direct policy implications for the criminal justice system
and other social institutions. Since the early 1990s, the United States has enjoyed the longest
sustained decline in crime since the FBI began compiling crime statistics in the early 1930s
(Blumstein & Wallman, 2006; Fagan, Zimring, & Kim, 1998). However, since the shooting of
Michael Brown, an unarmed black man in Ferguson, MO on August 9, 2014, the subsequent
civil unrest, and social media attention to his shooting, there has been speculation that a
“Ferguson Effect” has ended the great crime decline (Bialik, 2015; Davey & Smith, 2015; Mac
Donald, 2015; Rosenfeld, 2015).
Could the events surrounding Ferguson have changed the trajectory of crime trends in the
United States? Such a hypothesis is consistent with three potential explanations. The first of
these is de-policing, where negative publicity and public protest regarding police behavior leads
officers to withdraw from enforcing the law for fear of criticism and lawsuits. From this
perspective, police officers throughout the United States may have become hesitant to be
proactive out of concerns for being subjected to negative media scrutiny for racial profiling or
the use of excessive force (Wolfe & Nix, 2015; see also: Oliver, 2015). Reduced guardianship
and lack of enforcement—if widespread enough—may lead to increases in crime rates (Braga,
Papachristos, & Hureau, 2014; Levitt, 2002; Marvell & Moody, 1996; Rosenfeld, Deckard, &
Blackburn, 2014). The empirical evidence for de-policing is mixed (Shi, 2008; Stone, Foglesong,
& Cole, 2009).
2
For example, Shi (2008) showed that in the wake of a highly publicized incident
involving a white police officer shooting an unarmed African-American teenager in Cincinnati,
OH and subsequent Department of Justice investigation, arrests fell substantially (i.e., evidence
of de-policing). However, research showed that an LAPD consent decree did not result in any
form of de-policing (Stone et al., 2009). In fact, pedestrian and motor vehicle stops doubled after
the consent decree and a higher proportion of such stops resulted in an arrest. Total arrests were
also shown to increase post-consent decree. It must be noted, however, that de-policing is
difficult to measure and requires measurements of police activity which are generally not
publicly available.
Unlike the events spurring the studies of de-policing in Cincinnati and Los Angeles, the
shooting in Ferguson occurred in the era of social media. The massive social media response
following the events in Ferguson may have precipitated de-policing through contagion, the viral
2
It is important to note that Kane (2005) demonstrated that under policing was not associated with crime rates in
New York City in structurally disadvantaged communities. However, aggressive styles of policing—“over
policing”—were shown to be associated with increases in crime over time.
4
spread of information across social media. Indeed, the search term “Ferguson shooting” yields
around 30 million hits on Google, and has led to speculation about such a response by law
enforcement officials including several municipal Chiefs and the Director of the FBI. The
juxtaposition of concerns about de-policing, the resulting unrest following the initial shooting,
and the decision not to pursue criminal charges against the officer coupled with extensive social
media coverage of issues in Ferguson make this a compelling issue to examine empirically.
Second, high-profile incidents such as Ferguson may convey to the public that justice is
being administered unfairly and lead to challenges to the legitimacy of the law. One response to
the belief that the law is not administered fairly is increased participation in crime (Jackson et al.,
2012; Tyler, 2006). It is possible that the killing of unarmed citizens by officers sends the signal
to some citizens that law enforcement’s values and behaviors are inconsistent with their
expectations about how the law should be administered, therefore reducing stakes in conformity,
and leading to crime and disorder. Public trust in the police can be precarious, and such
shootings may upend that balance, particularly for the most disadvantaged members of American
society (Kane, 2005). Public criticism of the police on social media has spread the potential
impact of such an effect well beyond the geographic bounds of the St. Louis metropolitan area,
often in response to other officer-involved shootings. For example, the death-in-custody of
Freddie Gray led to violent rioting in Baltimore and the shooting of John Crawford III in
Beavercreek, OH resulted in a Department of Justice investigation. Public outcry regarding the
shooting of Laquan McDonald in Chicago has led to calls for the resignation of Mayor Rahm
Emanuel. In short, it is clear that citizens in a number of U.S. cities are calling into question the
legitimacy of police use of force.
The third explanation is that crime declines had reached their nadir and any increases in
crime were due to factors unrelated to Ferguson. After all, the large declines in crime observed
over the past two decades likely will eventually level out or even increase at some point. Such a
view is consistent with several threats to internal validity, including history and regression to the
mean (Shadish, Cook, & Campbell, 2002). This argument finds the relationship between a
Ferguson Effect and increased crime rates attributed to extraneous factors and thus spurious. This
is a plausible argument, given the length and magnitude of the crime decline which may not be
sustainable. Nonetheless, whether this nadir occurred coincidentally at the same time as
Ferguson would raise serious questions about turning points in crime trends.
Even so, should the Ferguson Effect be observed systematically throughout the United
States or are claims of de-policing and challenges to the legitimacy of the police idiosyncratic to
particular cities? The heterogeneity among large U.S. cities makes it likely that exogenous
factors will be experienced differently. Social media has played an important role in drawing
attention to Ferguson and related events (Wolfe & Nix, 2015), making it possible for events to be
observed in one city and their impact to be felt in others. In less than two weeks after the
shooting, the Wall Street Journal reported over 7.8 million tweets using the #Ferguson hashtag,
and the New York Times reported considerable misinformation on social media about the
shooting as well as the ensuing social unrest (Bilton, 2014; Zak, 2014). The effects of social
5
media may lead to de-policing or erosion in the legitimacy of the law, which in turn could lead to
increases in crime across the United States.
Importantly, the Ferguson Effect has been blamed for apparent increases in violent crime
in several U.S. cities by government leaders, law enforcement executives, and academics alike.
FBI Director James Comey even recently suggested that the Ferguson Effect has led to increases
in violent crime in some cities by stating, “I don’t know whether that explains it entirely, but I do
have a strong sense that some part of the explanation is a chill wind that has blown through
American law enforcement over the last year” (Schmidt & Apuzzo, 2015). Ironically, Director
Comey heads the agency responsible for producing the Uniform Crime Reports in the United
States and he himself lacked the full data on crime to draw the conclusions he announced.
Indeed, there is little empirical evidence and lots of speculation about this question. A brief paper
examined St. Louis crime rates over time and found very limited support for such an effect
(Rosenfeld, 2015). Another analysis focused on year-by-year differences in homicide across 60
cities and found a 16 percent increase in 2015 from 2014 (Bialik, 2015). A recently released
Brennan Center for Justice report examined homicide changes between 2014 and 2015 in 25 of
the 30 largest U.S. cities, finding that homicide rates increased by 15 percent (Friedman, Fortier,
Cullen, & James, 2015). While these studies reveal important insight into crime in the United
States, they are either narrow due to the number of cities or the number of crime types included
in their analyses. These studies lack the broad theoretical rationale presented here to anticipate
that a Ferguson Effect may extend to a broader cross-section of our largest cities and to forms of
crime other than homicide.
This study examines whether crime trends changed systematically after the Ferguson
shooting and if there were idiosyncratic changes across U.S. cities. We conduct our analysis
among 81 U.S. cities with populations exceeding 200,000 persons using discontinuous growth
modeling that incorporates between-city variability in FBI Part I aggregate and disaggregate
crime trends before and after Ferguson. After all, there is considerable heterogeneity across U.S.
cities. Even among large cities, there is variation in region, crime level, policing style, and the
size of the police force, among other things which could either buffer or bolster any “Ferguson
Effect” on crime. It is important to note that the city-level crime data used in this analysis cannot
establish whether loss of legitimacy or de-policing is at the root of an observed increase in crime,
or whether contagion induced by social media was responsible for transmitting these changes.
Rather, our central goal is to provide the most comprehensive investigation into the Ferguson
Effect on crime trends in the United States. If there is a Ferguson Effect, it could challenge the
decades-long decline in serious crime, particularly homicide. But of course, without data, these
arguments remain speculative. FBI Director Comey echoed this sentiment by suggesting that
“Data is a dry word, but we need better data” concerning the Ferguson Effect. We agree—this
issue is important because considerable public and private resources are committed to responding
to crime and their use should be guided by data and not speculation. Data and analysis must
serve as the foundation of evidence-based decision making and policy development.
6
METHODS
Data This study examined official crime data in U.S. cities. Monthly crime data were collected
from police department data requests and police department websites in 81 of the 105 U.S. cities
with populations exceeding 200,000 persons in 2010. Larger cities were the focus of this study
because crime recording practices are more reliable, and the volume of criminal activity is
greater and less subject to random fluctuation in the numerator, than in smaller cities (Maltz,
2006). Using 200,000 persons as our target population threshold also provided the twofold
benefit of increasing generalizability to a larger group of cities than prior studies (Bialik, 2015;
Friedman et al., 2015; Rosenfeld, 2015), which works to temper concerns about analyses being
underpowered to detect statistical differences.
3
We focused on the seven Part I Uniform Crime Report (UCR) offenses that include
measures of violent crime (criminal homicide, forcible rape, robbery, and aggravated assault)
and measures of property crime (burglary, larceny-theft, and motor vehicle theft). Included in the
UCR are crimes reported to the police and the hierarchy rule applies to UCR offenses. Only the
most serious of multiple-offense incidents is recorded. Detailed information about the UCR
reporting can be found in the summary reporting system manual (Federal Bureau of
Investigation, 2013).
Crime was standardized by city population as monthly crimes per 100,000 citizens.
Twelve months of pre- and post-Ferguson data were included. This provides us with a full year
of crime data before and after Ferguson, allowing us to determine if crime trends were indeed
redirected in the months following Ferguson. There were two reasons for inducing symmetry in
our research design. First, our data were temporally censored in the period after Ferguson. We
collected and analyzed the data between October and December 2015, which made September
the last available month, permitting a 12-month post-Ferguson period of observation. Second, a
12-month pre-Ferguson period was long enough to assess crime trends, but short enough to
protect the findings from threats to internal validity. Indeed, a longer pre-Ferguson window of
study would have made our findings more susceptible to historical, maturational, and statistical
regression factors. Future research might consider alternative research designs, namely, longer
periods of observation, especially as additional data become available.
Complete or partial crime data were included for 81 of the 105 largest U.S. cities.
Jurisdictions were excluded if they did not maintain at least four months of post-Ferguson crime
data. Homicide was the modal crime type included (N=81 cities), while data on rape was
3
FBI thresholds are populations of 250,000 or more for Group I cities (76 total), while populations between
100,000 and 249,999 for Group II cities (209 total). There is no bright line that guides population thresholds in
macro-level criminological research. In addition to the reasoning described above (measurement, generalizability,
power), our selection strategy also aimed to maximize crime, demographic, economic, political, and social variation
in the cities represented in this study and operated under the practical restraint that the authors collected data city-
by-city, negotiating the unique policies and practices of data collection in 81 cities. While we are grateful for the
support of the dozens of police chiefs, crime analysts, and command staff in this effort, it would have been
unmanageable to extend this data collection to FBI Group II cities, or even smaller cities like Ferguson, MO
(population 21,111).
7
available for 76 cities. Jurisdictions without full data on total crime (N=5), violent crime (N=5),
and property crime (N=2) were excluded from aggregate analyses. An average of 11 months of
post-Ferguson data were available across the crime types for the 81 cities. Overall, our data
include 87 percent—18,321 of the 20,250 possible data points—of the data available across the
10 crime types analyzed in 81 cities over a 25-month period. Appendix A includes a list of the
cities that were included in the study.
All 2013 monthly data were obtained from the UCR, while we retrieved 2014 and 2015
data either from police department websites (N=34 cities) or agency contacts (N=47 cities).
4
There were no statistically significant differences in pooled pre- or post-Ferguson crime rates
between the sample cities based on the approach used to obtain the data, nor did any of the
empirical Bayes predictions differ statistically at the 0.05 level. However, across the seven crime
types, data obtained from department websites produced fewer post-Ferguson valid points
(mean=71 months) compared to requested data (mean=77 months). This difference is largely
attributed to the fact that more recent information was available from data received in response
to requests to departments. On average, we had about an additional month of data across all
seven crime types; however, the total number of valid data points did not differ statistically.
Overall, this means that the potential bias introduced by our data collection strategy is
exceptionally low.
There were no statistically significant differences in the pooled pre-Ferguson aggregate
crime rates (total, property, and violent) between the excluded and included cities. However,
when disaggregated by crime type, homicide and aggravated assault in excluded cities were
about 65 percent the size of included cities. All of the structural characteristics of cities were
statistically equivalent between included and excluded cities. The cities included in the study,
nonetheless, provide a good representation of large U.S. cities. Moreover, our analyses focus on
changes in crime trends within-cities rather than between-cities, which helps minimize the
significance of the difference between excluded and included cities.
As we note below, there are many challenges when gathering data to address
contemporary issues related to crime, such as the Ferguson Effect. This article represents a
systematic attempt to bring comprehensive evidence to bear on a topic that has commanded the
attention of public officials at the highest levels, mainstream national news media, and the
general public on social media outlets. This article addresses a topic of prominent national
discourse that is long on anecdotes and speculation and short on data.
Analytic Strategy
A discontinuous growth model was used to assess the Ferguson Effect on within-city
changes in crime trends (Singer & Willett, 2003). A multilevel model was constructed to analyze
the research questions and our primary level 1 model takes the following form:
𝑌
𝑡𝑖 = 𝜋0𝑖 + 𝜋1𝑖𝑇𝑅𝐸𝑁𝐷𝑡+ 𝜋2𝑖𝑃𝑂𝑆𝑇𝑇𝑅𝐸𝑁𝐷𝑡+ 𝜖𝑡𝑖
4
Even at the time of this writing, January 2016, the 2014 UCR data were still not deposited on ICPSR, therefore it
remained necessary for us to rely on our data collection strategy.
8
where Y represents the crime rate at a given month t for city i, 𝜋0𝑖 is the crime rate for city i in
August 2014, 𝜋1𝑖 is a linear trend for city i where TREND is incremented by 1 for each month
and is centered at August 2014, and 𝜋2𝑖is a linear trend for city i where POSTTREND is
incremented by 1 for each month succeeding August 2014. Errors, 𝜖𝑡𝑖, are assumed to be
independent and normally distributed; residual diagnostics confirmed this. Some outliers (±4
standard deviations) led us to use robust standards to relax the assumption of correct
specification.
This study treats the events surrounding Ferguson as an exogenous shock that cities
experienced. Consistent with discontinuous growth models, the two linear trends, 𝜋1𝑖 and 𝜋2𝑖 , are
interpreted additively. In the months leading up to Ferguson, the crime trend for city i is
represented by 𝜋1𝑖 because POSTTREND is fixed to 0. In the months following Ferguson, the
crime trend is represented by 𝜋1𝑖i + 𝜋2𝑖 , where coefficients for 𝜋2𝑖 indistinguishable from 0
indicate no change in the crime trend post-Ferguson, while coefficients above or below 0
represent a redirection in crime trends occurring in the months succeeding Ferguson. All of the
models are adjusted for seasonality, a well-known correlate of crime rates relevant to the study of
crime trends (Baumer & Wright, 1996; Rosenfeld, 2015), by introducing fixed effects for the
month using August as the reference category.
Random effects were introduced to allow the intercept and slopes to vary across cities,
with our level 2 models represented as follows:
𝜋0𝑖 = 𝛾00 + 𝜁0𝑖
𝜋1𝑖 = 𝛾10 + 𝜁1𝑖
𝜋2𝑖 = 𝛾10 + 𝜁2𝑖
where 𝛾00 represents the mean crime rate the month of the Michael Brown shooting (August
2014), 𝛾10 represents the mean crime trend prior to Ferguson, and 𝛾20 represents the mean
change in post-Ferguson crime trends. These coefficients are represented as the fixed effects in
Tables 1 and 2. By introducing random effects, it permits us to not only investigate city-level
variation in crime rates at the time of Ferguson, but also variation in the trends before and after
Ferguson. This allows us to detect if there were truly changes in crimes trends within-cities, as
each city maintains its own individual intercept and trajectories. All analyses were conducted
using the me suite in Stata 14.0 (StataCorp, College Station, TX).
Results
Table 1 shows that the total crime rate was decreasing in the 12 months prior to Ferguson
[b=-1.43, P<0.05], consistent with a long-term trend of declining crime rates in the U.S. (Xie,
2014; Zimring, 2006). The overall pre-Ferguson crime decline was driven by trends in property
crime [b=-1.52, P<.01] but not violent crime [b=0.03, P=0.84]. Whereas the violent crime rate
trend was essentially flat in the 12 months before Ferguson, property crime rates were decreasing
by over 1.6 crimes per capita each month.
9
Table 1. Unstandardized Coefficients from Discontinuous Growth Models of Trends in
Total, Violent, and Property Crime Rates.
Fixed effects
Total Crime
Rate
Violent Crime
Rate
Property Crime
Rate
Constant
452.06
(18.10)**
73.43
(5.13)**
376.27
(14.64)**
Crime trend
-1.43
(0.65)*
0.03
(0.17)ns
-1.52
(0.59)**
Post-Ferguson trend
1.00
(1.17)ns
0.34
(0.22)ns
0.72
(1.05)ns
Random effects
Constant 𝜁0𝑖
19582.20
1445.55
14099.80
Crime trend 𝜁1𝑖
25.98
1.77
23.00
Post-Ferguson trend 𝜁2𝑖
79.114
2.17
69.60
Sample size
N of cities
76
76
79
N of cities*months
1764
1769
1845
NOTES: Robust standard errors are given in parentheses. Random effects report between-city variance. All results
are seasonality-adjusted.
ns, P>0.10; +, P<.10; *, P<0.05; **, P<.01
After the shooting of Michael Brown, and the subsequent social unrest and social media
responses, was there a systematic change in crime trends in large U.S. cities? We find no
evidence in support of this contention. Our results reveal that the post-Ferguson trends in total
[b=1.00, P=0.39], violent [b=0.34, P=0.12], and property [b=0.72, P=0.50] crime were
statistically insignificant. While the post-Ferguson trend coefficients for all three aggregate crime
were positive, they were not large enough to be statistically distinguishable from the pre-
Ferguson crime trend, which was flat for violent crime and declining for property crime.
Altogether, we can conclude that there is no systematic evidence of a Ferguson Effect on
aggregate crime rates throughout the large U.S. cities represented in this study.
Table 2 disaggregates the total crime rate by Part I offenses, as aggregate crime rates
might mask important trends within specific offense types and there is a theoretical rationale
motivating such analyses. In the 12 months leading up to Ferguson, monthly crime trends were
flat for homicide, aggravated assault, larceny, and motor vehicle theft, increasing for rape
[b=0.06, P<.01], and decreasing for robbery [b=-0.13, P<.10] and burglary [b=-0.50, P<.01]. For
the post-Ferguson trends, the only crime type that resembles the expected changes associated
with a Ferguson Effect was robbery, which increased at a monthly rate of 0.26 incidents per
capita [P<.05]. Given that our post-trends are estimated additively with the pre-trend, this means
that robbery rates were falling by 0.13 per capita monthly, only to shift in the opposite direction
resulting in an increase of 0.12 robberies per capita monthly [-0.132+0.255]. We provide a
graphical illustration of this redirection in Figure 1. All of the remaining outcomes revealed post-
Ferguson crime trends that were statistically indistinguishable from pre-Ferguson crime trends.
While there was a clear absence of evidence for a Ferguson Effect among aggregated crime
10
Table 2. Unstandardized Coefficients from Discontinuous Growth Models of Crime Trends Disaggregated by Crime Type.
Fixed effects
Homicide
Rate
Rape
Rate
Robbery
Rate
Aggravated
Assault Rate
Burglary
Rate
Larceny
Rate
Motor Vehicle
Theft Rate
Constant
0.944
(0.109)**
4.85
(0.30)**
25.17
(2.01)**
41.86
(3.35)**
82.94
(4.36)**
248.04
(10.70)**
45.29
(3.43)**
Crime trend
-0.003
(0.005)ns
0.06
(0.02)**
-0.13
(0.07)+
0.09
(0.13)ns
-0.50
(0.18)**
-0.88
(0.56)ns
-0.12
(0.11)ns
Post-Ferguson trend
0.015
(0.009)ns
-0.04
(0.03) ns
0.26
(0.12)*
0.13
(0.14)ns
0.19
(0.23)ns
0.26
(0.89)ns
0.23
(0.23)ns
Random effects
Constant 𝜁0𝑖
0.6200
5.358
216.77
696.69
1275.14
7746.04
685.50
Crime trend 𝜁1𝑖
0.0005
0.004
0.35
1.15
2.07
23.29
0.83
Post-Ferguson trend 𝜁2𝑖
0.0030
0.012
0.77
0.96
2.14
52.45
3.05
Sample size
N of cities
81
76
79
80
79
79
79
N of cities*months
1902
1781
1850
1865
1850
1845
1850
NOTES: Robust standard errors are given in parentheses. Random effects report between-city variance. All results are seasonality-adjusted.
ns, P>0.10; +, P<.10; *, P<0.05; **, P<.01
11
Figure 1. Predicted Values of the Pre- and Post- Ferguson Trends in Monthly Robbery
Rates Per 100,000 Persons (N=79).
types, the conclusion from this disaggregated analysis is that changes in robbery rates constitute
the lone exception to a spurious Ferguson Effect.
5
Nonetheless, the post-Ferguson crime trends were not universally equivalent to pre-
Ferguson crime trends across large U.S. cities. Tables 1 and 2 report the variance in the pre- and
post-Ferguson trends for all 10 outcomes examined in this article. Overall, the variance in the
post-Ferguson trend was much greater than the pre-Ferguson trend. Indeed, the variance in total
crime rates was 3 times as large after Ferguson than before; for homicide rates, the variance was
5.7 times greater. This means that crime trends in the post-Ferguson era varied considerably
5
For those concerned with multiple testing we used the Benjamini-Hochberg (1995) procedure to compute critical
values for rejecting the null hypothesis, as follows: pk<(i/m)*Q, where p is the p value for variable k, i is the rank of
the p value in ascending order for k variables, m is the total number of tests, and Q is the false discovery rate. For
robbery rates, p=0.027, i=1 (smallest p value), m=4 (m of violent crimes), and Q=0.10 (false discovery rate of 10%),
the Benjamini-Hochberg critical value of 0.025 exceeded that of the p value from the mixed effect models.
Adjustments for multiple testing, particularly the Bonferroni correction which is somewhat commonly seen in
criminological outlets, are highly criticized because there are few established rules for correction. The criterion for
deciding the family wide error rate (FWER) is vague and what constitutes simultaneous inference is unclear (Cabin
& Mitchell, 2000). And, as Cabin and Mitchell pointed out, is the correction table- or manuscript-wide, or even the
career of a dataset? It is also subject to misuse. Indeed, using Nakagawa’s (2004) hypotheticals, varying the number
of outcomes could push the p value above or below the critical value. While the purpose of these corrections is to
reduce Type I error, it could ultimately introduce Type II error. We view this as a net zero gain, and thus urge future
research examining issues comparable to ours to pay close attention to robbery rates.
0
5
10
15
20
25
30
-12 -10 -8 -6 -4 -2 0246810 12
Monthly Robbery Rates (Per 100,000 Persons)
Months Pre- and Post- Ferguson
Pre-Trend
Post-Trend
Counterfactual
12
across cities, with some cities experiencing large changes in their crime trends, while others
experienced little to no change. It is important to emphasize that such variation across cities does
not alter the overarching conclusion that evidence of an overall Ferguson Effect on crime rates in
large U.S. cities is negligible. But it does mean that there is less stability in crime trends in the
post-Ferguson era, which could point to heterogeneous responses to the exogenous shock of
Ferguson.
Figure 2 plots the empirical Bayes predictions for the post-Ferguson monthly rate of
change in homicide per capita for 81 cities with valid data on homicide. We examined homicide
because it is the most reliably recorded of the seven Part I crimes and a barometer of the overall
crime trend in cities. Notably, St. Louis, the metropolitan area that includes Ferguson, scored
among cities with the largest increases in homicide rate trends. To the extent that any Ferguson
Effect exists, it appears to be constrained to a small number of cities, particularly cities with
historically high homicide rates, as we show below.
Table 3 examines the characteristics of cities based on the magnitude of changes to
homicide trends in the post-Ferguson era. Cities were divided into three categories based on
empirical Bayes predictions of post-Ferguson homicide trends: negative rate of change (N=27), a
low, yet positive, rate of change (N=27), and a high and positive rate of change (N=27). In the
latter group, assuming the rate of change remains stable, it would take about two years for cities
to witness an entire one-unit increase in the homicide rate. We compare these three groupings of
cities in order to determine if there were factors associated with heterogeneity in post-Ferguson
homicide trends, and our findings reveal that there were indeed important differences.
As the rate of change in homicide trends becomes positive and larger in magnitude, cities
tended to have higher pre- and post-Ferguson pooled violent crime rates, a higher rate of police
officers per 1,000 citizens, a greater composition of black residents and a smaller composition of
white residents, and greater socioeconomic disadvantages that are typically associated with high
homicide rates (Pratt & Cullen, 2005). The differences are especially pronounced between the
high/positive change cities compared to the negative and low/positive change cities. Indeed, the
cities with a flat or negative homicide rate trend share a great deal of similarities in their
demographic, social, economic, and criminal profiles. Moreover, heterogeneity in homicide rate
increases is not related to population size. Overall, these increases are confined to a smaller
group of cities; however, these cities are very different from those which experienced little or no
changes in homicide rate trends after the events that propelled Ferguson to the center of national
and international scrutiny surrounding criminal justice system policies and practices in the
United States.
Discussion
Declining crime rates since the 1990s have marked one of the most significant
“civilizing” trends in U.S. history (Pinker, 2011). Large, sustained declines in crime have made
the United States much safer. The sources of the crime decline throughout the 1990s and 2000s
are complex, ranging from improvements in the economy and the decline of the crime prone
13
Figure 2. Empirical Bayes Predictions of the Post-Ferguson Rate of Monthly Change in
Homicide Per Capita by City (N=81)
Note: CI=confidence interval; Mean=post-Ferguson trend unstandardized coefficient
Mean0CI CI-.05 .05 .1 .15 .20
Buffalo
San Bernardino
Corpus Christi
Lubbock
Pittsburgh
Glendale
Jersey City
Boise
Greensboro
Lincoln
Boston
St. Paul
Irving
Las Vegas
Wichita
Scottsdale
San Diego
Austin
Madison
Portland
San Antonio
Seattle
Tucson
San Francisco
Riverside
Raleigh
Arlington
Gilbert
Henderson
Chandler
New York
Toledo
Chula Vista
Louisville
Colorado Springs
Spokane
Stockton
Fresno
Mesa
Phoenix
Los Angeles
Tampa
Atlanta
Virginia Beach
Durham
Tulsa
Lexington
Sacramento
Winston-Salem
Miami
Jacksonville
Houston
Albuquerque
Omaha
Denver
Philadelphia
Garland
St. Petersburg
Dallas
Long Beach
Chicago
Oklahoma City
Indianapolis
Kansas City
Baton Rouge
Fort Wayne
Cincinnati
Minneapolis
Cleveland
Richmond
Oakland
Detroit
Rochester
Milwaukee
Newark
Washington DC
New Orleans
St. Louis
Baltimore
14
Table 3. Differences Across Cities by Monthly Changes in Post-Ferguson Homicide Rate
Trends.
Homicide Rate Trend
Negative Rate
of Change
Mean/%
Low/Positive
Rate of Change
Mean/%
High/Positive
Rate of Change
Mean/%
P
Empirical Bayes Predictions
Pre-Ferguson
0.0058
-0.0015
-0.0130
Post-Ferguson
-0.0150
0.0046
0.0557
Pre-Ferguson Pooled Crime Rates
Homicide Rate
0.59
0.65
1.54
0.000
Total Crime Rate
400.71
376.92
470.10
0.043
Violent Crime Rate
51.97
49.50
89.95
0.000
Property Crime Rate
348.74
327.47
382.34
0.234
Post-Ferguson Pooled Crime Rates
Homicide Rate
0.56
0.66
1.72
0.000
Total Crime Rate
388.11
380.32
458.42
0.082
Violent Crime Rate
56.64
52.45
91.08
0.001
Property Crime Rate
331.47
322.51
364.86
0.367
Structural Characteristics
Population
461,979
888,547
577,967
0.283
Officers Per Capita
2.08
2.06
3.08
0.000
% Consent Decree
7.4%
3.7%
14.8%
0.335
% Minority Mayor
11.1%
22.2%
37.0%
0.079
% White
64.8%
61.4%
48.6%
0.000
% Black
11.6%
18.0%
35.5%
0.000
% Hispanic
25.4%
23.9%
17.0%
0.153
% Female HH w/Children
7.8%
8.7%
10.4%
0.000
% Unemployed
11.5%
12.3%
13.9%
0.047
NOTES: Means and prevalences are reported. One-way ANOVA and chi-square tests are used to determine
statistical significance. Cities were grouped into three post-Ferguson homicide trends: cities with a negative rate of
change N=27, cities with a low/positive rate of change N=27, and cities with a high/positive rate of change N=27.
segment of the population (ages 15-29 years) to the decline of crack cocaine markets and
increases in imprisonment (Blumstein & Wallman, 2006; Levitt, 2004; National Research
Council, 2008). Of course, policing is another important explanation. Any redirection in such a
sustained decline in crime is invariably a cause for concern.
In recent years, the public, the police and elected officials are increasingly subject to the
social contagion effects of social media. The killing of Michael Brown in Ferguson, MO appears
to have sparked one such social process, reinforced by repetitive mentions of the Ferguson Effect
in social and other media by political and law enforcement leaders. Despite widespread coverage
of protests and the responses to officer-involved shootings by social and other media and
speculation concerning the Ferguson Effect, the incident was not a sentinel event that changed
15
the direction of the crime decline. Our analysis was well positioned to identify the existence of a
Ferguson Effect on crime rates among the largest sample of U.S. cities examined to date. We
found that the crime decline has not changed substantially in large American cities in the 12
months after Ferguson. Simply put, we observed no systematic and widespread change in crime
trends among the cities in our study beyond robbery rates. Our analysis thus confirms the long-
held understanding that the causes of crime reflect slow processes and are not amenable to
sudden shocks (LaFree, 1999; National Research Council, 2008). With this in mind, several
issues require further discussion.
First, although the overall null Ferguson Effect was robust for both total and
disaggregated crime scales, it is important to note that our analysis showed that robbery rates
significantly increased in the study cities post-Ferguson. This finding suggests that a Ferguson
Effect may have occurred but its influence is limited to robbery rates. Although our data cannot
offer insight into the causal mechanism behind this observation, it appears that robbery rate
increases began about the same time as Michael Brown’s death. We urge future research
examining issues comparable to ours to pay close attention to robbery rates.
Next, while it would be a rush to judgement to suggest that the Ferguson Effect will end
the decades-long crime decline, we observed substantial variation in crime trends in the post-
Ferguson era. Several cities (Baltimore, St. Louis, Newark, New Orleans, Washington, D.C.,
Milwaukee, and Rochester among others) experienced large increases in homicide rates
following the events in Ferguson. Accordingly, the data offer preliminary support for a Ferguson
Effect on homicide rates in a few select cities in the United States. What is important about these
cities is that they had much higher crime rates before Ferguson, which in turn may have primed
them for increases in crime. Cities with post-Ferguson increases in crime tended to have a higher
proportion of black residents, lower socioeconomic status, and more police per capita—
important macro-level correlates of crime rates (Pratt & Cullen, 2005; Sampson, 2012). Simply
put, these other predictors of crime rates lead to questions that may inhibit any ability to attribute
crime increases specifically to the Ferguson Effect in these cities, and require further—and more
formal—moderator analyses to isolate the constellation of factors that such cities share in
common.
What our analysis cannot speak to is the extent to which de-policing or a crisis in police
legitimacy have occurred post-Ferguson, and if so, the impact it may have had on crime rates.
Indeed, it is not possible to use these data to discriminate between such hypotheses or to
establish the role, if any, that protests may have played in contributing to de-policing by some
officers or de-legitimizing the police in the eyes of many citizens. Anecdotal evidence
concerning de-policing abounds in social media, a medium that has provided a front-row seat to
the civil unrest after a number of police killings of citizens (Wolfe & Nix, 2015). What we do
know, however, is that if de-policing or a legitimacy crisis are occurring, neither is impacting
crime rates systematically across large U.S. cities. Subsequent analyses should examine these
propositions as they are critical to the effective administration of justice in the United States.
Given the complexity of such analyses, it may be beneficial to focus efforts on the apparent
16
systematic robbery rate increase and the homicide rate increases observed in the handful of
violent, racially diverse cities discussed earlier.
This study is not the final word on the Ferguson Effect. There are alternative ways to test
for a Ferguson Effect—we chose one method that addresses a key question in this debate: did the
events in Ferguson lead to a redirection in crime trends? While our study provides important
answers to this question there are several issues that represent opportunities for future research.
The first is the nature of our sample. We chose to use 200,000 as the population cutoff for
inclusion in the sample. While the 81 cities in our sample account for both a large portion of the
U.S. population (17 percent) and violent crime (29 percent) in 2014, it omits smaller cities such
as Ferguson. There may be important differences between the large cities in our sample and
medium and small cities that our analysis does not capture. As such, our results are only
generalizable to those cities included in our sample. This analysis cannot speak to whether a
Ferguson Effect has or has not occurred in smaller towns throughout the U.S. Future research is
needed to answer this question.
Second, outside of seasonality-adjustments, our models excluded both time-varying
confounders and mediators. As we have mentioned above, securing crime data alone was a time-
intensive task; that burden would have increased considerably or impossibly if we attempted to
secure data that would allow us to simultaneously model our mediating mechanisms (e.g.,
depolicing, legitimacy, social media). Therefore, while our results point to a null hypothesis,
there could be factors suppressing the Ferguson Effect.
Third, we use a 12-month pre- and post-Ferguson series of crime data. More months may
have made the estimates more stable, perhaps increasing the chance of finding statistically
significant changes in crime trends. However, a shorter window of observation does provide
advantages that limit threats to internal validity, namely, history, maturation, and regression to
the mean. Fourth, despite repeated efforts we were unable to obtain all of the crime data for cities
that met our sampling criteria (over 200,000 population). While three-fourths of all large U.S.
cities are represented in our study, including 10 of the largest U.S. cities, some bias in our
findings may remain because of the cities for which data were unavailable.
Before concluding, there is another important issue regarding data worthy of note. In the
wake of concerns about an abrupt change in the decades-long crime decline, the data to assess
this question were not readily available to policy makers, the public, or researchers. Crime
counts were collected from police department websites and through direct requests for crime data
made to police departments. Indeed, this was an arduous and time-consuming process. As others
have noted, important issues of public policy such as crime should be addressed with current and
publicly accessible data (Rosenfeld, 2007). Data should be readily available for addressing
questions such as what causes crime rates to go up or down given the potentially far-reaching
public health and public safety policy concerns associated with such investigations. Reliable and
publicly available crime data is necessary to ground policy decisions on evidence-based,
scientific knowledge.
17
In conclusion, tragic events such as those in Ferguson, Staten Island, Baltimore, and
North Charleston, to name a few, have sparked debate over important issues such as police-
community relations and police legitimacy. Such debates, however, should be informed by solid
data and careful analysis. Policy decisions that are not evidence-based can negatively impact
public safety, curtail debate and action on important issues such as mass incarceration, or, at the
very least, result in ill-advised expenditure of tax dollars. We sought to bring empirical evidence
to the Ferguson Effect debate. On the whole, there is no nationwide Ferguson Effect on crime
rates. At the same time, we did observe an increase in robbery rates in the United States and
homicide rates in several cities that began at the same time as Ferguson. Our hope is that these
results will help provide evidence-based discussions of the Ferguson Effect, specifically, and
changes in crime trends more generally.
18
References
Baumer, E., & Wright, R. (1996). Crime seasonality and serious scholarship: a comment on
Farrell and Pease. Brit. J. Criminology, 36, 579.
Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: A practical and
powerful approach to multiple testing. Journal of the Royal Statistical Society, 57(1),
289–300.
Besbris, M., Faber, J. W., Rich, P., & Sharkey, P. (2015). Effect of neighborhood stigma on
economic transactions. Proceedings of the National Academy of Sciences, 112(16), 4994–
4998. http://doi.org/10.1073/pnas.1414139112
Bialik, C. (2015). Scare headlines exaggerated the U.S. crime wave. Retrieved from
http://fivethirtyeight.com/features/scare-headlines-exaggerated-the-u-s-crime-wave/
Bilton, N. (2014, August 27). Ferguson reveals a Twitter loop. The New York Times. Retrieved
from http://www.nytimes.com/2014/08/28/fashion/ferguson-reveals-a-twitter-loop.html
Blumstein, A., & Wallman, J. (2006). The crime drop in America. Cambridge University Press.
Braga, A. A., Papachristos, A. V., & Hureau, D. M. (2014). The effects of hot spots policing on
crime: An updated systematic review and meta-analysis. Justice Quarterly, 31(4), 633–
663.
Cabin, R. J., & Mitchell, R. J. (2000). To Bonferroni or not to Bonferroni: When and how are the
questions. Bulletin of the Ecological Society of America, 81(3), 246–248.
Davey, M., & Smith, M. (2015, August 31). Murder rates rising sharply in many u.s. cities. The
New York Times. Retrieved from http://www.nytimes.com/2015/09/01/us/murder-rates-
rising-sharply-in-many-us-cities.html
Fagan, J., Zimring, F. E., & Kim, J. (1998). Declining homicide in New York City: A tale of two
trends. Journal of Criminal Law and Criminology, 1277–1324.
19
Federal Bureau of Investigation. (2013). Criminal Justice Information Services (CJIS) Division
Uniform Crime Reporting (UCR) Program: Summary Reporting System (SRS) User
Manual. Version 1.0. Washington, DC: U.S. Department of Justice, Federal Bureau of
Investigation, Criminal Justice Information Services Division. Retrieved from
https://www.fbi.gov/about-us/cjis/ucr/nibrs/summary-reporting-system-srs-user-manual
Friedman, M., Fortier, N., Cullen, & James. (2015). Crime in 2015: A Preliminary Analysis.
New York: Brennan Center for Justice. Retrieved from
https://www.brennancenter.org/publication/crime-2015-preliminary-analysis
Jackson, J., Bradford, B., Hough, M., Myhill, A., Quinton, P., & Tyler, T. R. (2012). Why do
people comply with the law? Legitimacy and the influence of legal institutions. British
Journal of Criminology, 52(6), 1051–1071.
Kane, R. J. (2005). Compromised police legitimacy as a predictor of violent crime in structurally
disadvantaged communities. Criminology, 43(2), 469–498. http://doi.org/10.1111/j.0011-
1348.2005.00014.x
LaFree, G. (1999). Declining Violent Crime Rates in the 1990s: Predicting Crime Booms and
Busts. Annual Review of Sociology, 25, 145–168.
Levitt, S. D. (2002). Using electoral cycles in police hiring to estimate the effects of police on
crime: Reply. The American Economic Review, 92(4), 1244–1250.
Levitt, S. D. (2004). Understanding why crime fell in the 1990s: Four factors that explain the
decline and six that do not. Journal of Economic Perspectives, 163–190.
Mac Donald, H. M. (2015, May 29). The new nationwide crime wave. Wall Street Journal.
Retrieved from http://www.wsj.com/articles/the-new-nationwide-crime-wave-
1432938425
20
Maltz, M. D. (2006). Analysis of missingness in UCR crime data. Columbus, OH: Criminal
Justice Research Center, Ohio State University.
Marvell, T. B., & Moody, C. E. (1996). Specification problems, police levels, and crime rates.
Criminology, 34(4), 609–646. http://doi.org/10.1111/j.1745-9125.1996.tb01221.x
Nakagawa, S. (2004). A farewell to Bonferroni: the problems of low statistical power and
publication bias. Behavioral Ecology, 15(6), 1044–1045.
http://doi.org/10.1093/beheco/arh107
National Research Council. (2008). Understanding crime trends: Workshop report. (A. S.
Goldberger & R. Rosenfeld, Eds.) (Vol. 10). Washington, DC: The National Academies
Press. Retrieved from http://www.nap.edu/catalog/12472
Oliver, W. M. (2015). Depolicing: Rhetoric or reality? Criminal Justice Policy Review,
0887403415586790. http://doi.org/10.1177/0887403415586790
Pinker, S. (2011). The better angels of our nature: Why violence has declined. New York:
Viking.
Pratt, T. C., & Cullen, F. T. (2005). Assessing macro-level predictors and theories of crime: A
meta-analysis. Crime and Justice, 32, 373–450.
Rosenfeld, R. (2007). Transfer the uniform crime reporting program from the FBI to the Bureau
of Justice Statistics. Criminology & Public Policy, 6(4), 825–833.
Rosenfeld, R. (2015). Was there a “Ferguson Effect” on crime in St. Louis? Washington, DC:
The Sentencing Project. Retrieved from
http://sentencingproject.org/doc/publications/inc_Ferguson_Effect.pdf
21
Rosenfeld, R., Deckard, M. J., & Blackburn, E. (2014). The effects of directed patrol and self-
initiated enforcement on firearm violence: A randomized controlled study of hot spot
policing. Criminology, 52(3), 428–449.
Sampson, R. J. (2012). Great American city: Chicago and the enduring neighborhood effect.
University of Chicago Press.
Schmidt, M. S., & Apuzzo, M. (2015, October 23). F.B.I. Chief Links Scrutiny of Police With
Rise in Violent Crime. The New York Times. Retrieved from
http://www.nytimes.com/2015/10/24/us/politics/fbi-chief-links-scrutiny-of-police-with-
rise-in-violent-crime.html
Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002). Experimental and quasi-experimental
designs for generalized causal inference. Boston, MA: Houghton-Mifflin. Retrieved from
http://impact.cgiar.org/pdf/147.pdf
Shi, L. (2008). Does oversight reduce policing? Evidence from the Cincinnati police department
after the April 2001 riot. Journal of Public Economics. Retrieved from
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=647606
Singer, J. D., & Willett, J. B. (2003). Applied longitudinal data analysis: Modeling change and
event occurrence. Oxford University Press.
Stone, C., Foglesong, T. S., & Cole, C. M. (2009). Policing Los Angeles Under a Consent
Degree: The Dynamics of Change at the LAPD. Program in Criminal Justice Policy and
Management, Harvard Kennedy School. Retrieved from http://watch-command-
1142464325.us-east-1.elb.amazonaws.com/assets/pdf/Harvard-LAPD%20Study.pdf
Tyler, T. R. (2006). Why people obey the law. Princeton University Press.
22
Wolfe, S. E., & Nix, J. (2015). The alleged “Ferguson Effect” and police willingness to engage
in community partnership. Law and Human Behavior.
Xie, M. (2014). Area differences and time trends in crime reporting: Comparing New York with
other metropolitan areas. Justice Quarterly, 31(1), 43–73.
Xie, M., & McDowall, D. (2010). The reproduction of racial inequality: How crime affects
housing turnover. Criminology, 48(3), 865–896.
Zak, E. (2014, August 18). How #Ferguson has unfolded on Twitter. Retrieved from
http://blogs.wsj.com/dispatch/2014/08/18/how-ferguson-has-unfolded-on-twitter/
Zimring, F. E. (2006). The great American crime decline. Oxford University Press.
23
Appendix A. Source of 2014 and 2015 Data
Agency Websites/Open Data
(N=48)
Agency Analysts/Contacts
(N=33)
Missing Cities
(N=24)
Arlington, TX
Atlanta, GA
Austin, TX
Baton Rouge, LA
Chula Vista, CA
Dallas TX
Fort Wayne, IN
Garland, TX
Houston, TX
Irving, TX
Jacksonville, FL
Jersey City, NJ
Lincoln, NE
Long Beach, CA
Mesa, AZ
Minneapolis, MN
Oakland, CA
Oklahoma City, OK
Omaha, NE
Philadelphia, PA
Portland, OR
Richmond, VA
Riverside, CA
San Antonio, TX
San Bernardino, CA
San Diego, CA
San Jose, CA
St. Louis, MO
St. Paul, MN
St. Petersburg, FL
Tulsa, OK
Virginia Beach, VA
Washington, D.C.
Albuquerque, NM
Baltimore, MD
Boise, ID
Boston, MA
Buffalo, NY
Chandler, AZ
Chicago, IL
Cincinnati, OH
Cleveland, OH
Colorado Springs, CO
Corpus Christi, TX
Denver, CO
Detroit, MI
Durham, NC
Fresno, CA
Gilbert, AZ
Glendale, AZ
Greensboro, NC
Henderson, NV
Indianapolis, IN
Kansas City, MO
Las Vegas, NV
Lexington, KY
Los Angeles, CA
Louisville, KY
Lubbock, TX
Madison, WI
Miami, FL
Milwaukee, WI
New Orleans, LA
New York, NY
Newark, NJ
Phoenix, AZ
Pittsburgh, PA
Raleigh, NC
Reno, NV
Rochester, NY
Sacramento, CA
San Francisco, CA
Scottsdale, AZ
Seattle, WA
Spokane, WA
Stockton, CA
Tampa, FL
Toledo, OH
Tucson, AZ
Wichita, KS
Winston-Salem, NC
Anaheim, CA
Anchorage, AK
Aurora, CO
Bakersfield, CA
Birmingham, AL
Charlotte, NC
Chesapeake, VA
Columbus, OH
Des Moines, IA
El Paso, TX
Fort Worth, TX
Fremont, CA
Hialeah, FL
Honolulu, HI
Irvine, CA
Laredo, TX
Memphis, TN
Mesa, AZ
Modesto, CA
Nashville, TN
Norfolk, VA
North Las Vegas, NV
Orlando, FL
Plano, TX
Santa Ana, CA