You are on page 1of 27

574977

research-article2015
CADXXX10.1177/0011128715574977Crime & DelinquencyBierie

Article
Crime & Delinquency
2017, Vol. 63(8) 899­–925
Assault of Police © The Author(s) 2015
Reprints and permissions:
sagepub.com/journalsPermissions.nav
DOI: 10.1177/0011128715574977
journals.sagepub.com/home/cad

David M. Bierie1

Abstract
The assault of law enforcement officers is an important but understudied topic.
To better understand this form of violence, this study drew on the National
Incident–Based Reporting System—the nation’s largest data set tracking
assaults against police alongside detailed information on situations surrounding
attacks. Risk factors were examined within a fixed-effects logistic regression
framework via a case-control method in which all incidents involving assault
against police (n = 20,140) were compared with a random sample of arrest
encounters that did not involve this form of aggression (n = 20,118). The
data showed that a number of victim, offender, and situational characteristics
predicted violence against officers, and the models were able to explain a
substantial portion of the variance. Implications for research are discussed.

Keywords
policing, violence, criminal justice administration

Policing has long been considered one of the most dangerous careers in the
United States and it remains so today (Brandl & Stroshine, 2003, 2012;
Bureau of Labor Statistics, 2011). National reporting on police victimization
shows approximately 10% of officers are assaulted each year—a rate that has
remained consistent in recent years whereas other violent crime has declined
(Federal Bureau of Investigation [FBI], 2011). Of those officers who are
assaulted, more than one quarter will be injured and a small number killed
(FBI, 2011).

1U.S. Department of Justice, Washington, DC, USA


Corresponding Author:
David M. Bierie, United States Marshals Service, U.S. Department of Justice, 2604 Jefferson
Davis Highway, CM4, 11th floor, Alexandria, VA 22301-1025, USA.
Email: david.bierie2@usdoj.gov
900 Crime & Delinquency 63(8)

Violent attacks against police represent a particularly important form of


violence for social scientists and policymakers to understand. First, this is
because of unique risks they may present to officers, agencies, and the com-
munity. Officers may experience a greater level of injury than other types of
victims (all else held constant). We might expect this because police are prob-
ably less likely than other victim types to flee when physically outmatched.
Their sworn duty to protect others from an assailant or to make an arrest
likely heightens the risk of continued engagement when flight would be the
more rational choice. As an engagement prolongs, the potential escalation of
force by an assailant or officer, including the use of a firearm, poses risk to
those involved as well as bystanders.
Second, attacks against officers also can lead to important costs for agen-
cies and thus tax payers. These include medical treatment for injuries, time
spent completing paperwork and holding hearings on the use of force, and
reduced or disrupted resources when officers are placed on leave. Costs also
include potential losses in terms of community relations. When an officer
responds to an attack with force, the press and the public may be quick to
judge that action as excessive use of force. Regardless of the assertion’s
merit, the cost to police legitimacy may be large.
Finally, this form of violence is important to understand because these acts
reflect, to some degree, an attack on the rule of law and defiance of the justice
system itself. For these reasons, police, policymakers, and the public would
benefit from a deeper understanding of violence directed at police.

Research on Aggression Toward Police


Although there is a long-standing literature on force used by police (Alpert
& Dunham, 1997; Fyfe, 1978), there are few studies of violence against
police. Most studies to date have focused on very minor forms of resistance
(e.g., refusing to obey an officer’s commands) or have used broad defini-
tions that are weighted heavily by more prevalent, minor forms of aggres-
sion (e.g., Kavanaugh, 1997; Mastrofski, Snipes, & Supina, 1996). This is
important because there appears to be some etiological differences between
minor and serious resistance. For example, both Mastrofski et al. (1996)
and McClusky, Mastrofski, and Parks (1999) found that females are more
likely than males to resist officers (e.g., to fail to obey commands) and
Covington, Huff-Corzine, and Corzine (2014) found the same pattern with
respect to battery of police. In contrast, studies of more serious violent
resistance (i.e., lethal or near-lethal encounters) show that male offenders
pose greater risk than women (Bierie, Detar, & Craun, 2013; Craun, Detar,
& Bierie, 2013). Studies of minor forms of resistance generally find younger
Bierie 901

offenders pose the most risk to officers, whereas studies that examine seri-
ous or lethal attacks find risk increases linearly with offender age (Bierie
et al., 2013; Mustard, 2001).
The research has some consistencies regardless of the seriousness of vio-
lence examined. First, the probability of aggression rises when (a) crimes
officers are responding to increase in seriousness or (b) offenders are under
the influence of alcohol (Bierie et al., 2013; Covington et al., 2014; Craun
et al., 2013; Engel, 2003; Johnson, 1999; Kavanaugh, 1997; McClusky et al.,
1999). Second, several variables have shown little or no impact on resistance
across numerous studies. For example, the research to date does not support
the commonly held myth (e.g., see Steffensmeier, Ulmer, & Kramer, 2006)
that minority offenders resist law enforcement officers more often than
Caucasians (Bierie et al., 2013; Covington et al., 2014; Craun et al., 2013).
Despite the importance of the body of literature mentioned, it exhibits
some notable limitations. First, incident-level studies of resistance have been
based on small sample sizes. Belvedere, Worrall, and Tibbetts (2005) drew on
200 events of resistance, Craun et al. (2013) used below 300 events of firearm
violence, Engel (2003) identified just over 400 cases of noncompliance or
verbal resistance, and Kavanaugh (1997) had just fewer than 500 cases of
resistance available for analysis. The largest study to date, Bierie et al. (2013),
drew on 860 firearm-involved attacks against police. Each of these studies
included comparison groups with adequate sample sizes to support their sta-
tistical models of risk (e.g., logistic regression). Yet, the relatively small sam-
ple sizes imply less variation across independent variables than desired and
potential instability in measurement among covariates that lack balance. For
example, a study of a few hundred resistance events may only see 30 or 40
instances of a female offender—we must wonder whether multivariate mod-
els with a gender effect were biased by small cell counts after parsing this
group of women across other covariates in a regression equation.
Second, samples in the literature have been based on a few and potentially
select geographical areas. Most research to date used one jurisdiction
(Belvedere et al., 2005; Engel, 2003; Mastrofski et al., 1996) or a handful of
cities (e.g., McClusky et al., 1999). One problem implied by a lack of geo-
graphic diversity is the potential that findings from one area may not general-
ize well. Localized sampling has inhibited the field from exploring contextual
contamination—to assess what risk factors vary in magnitude across places
or are reflective of aggregate-level processes (Iversen, 1991) or local demo-
graphics. Only two have created incident-level analyses from a national per-
spective (Bierie et al., 2013; Craun et al., 2013). However, both of these
studies restricted consideration to extreme violence (e.g., shooting at offi-
cers). Although important, patterns observed among serious violent attacks
902 Crime & Delinquency 63(8)

against police may not be the same as relationships found in studies of less
serious but more pervasive forms of aggression.
Third, there are a number of variables that could be tested with more preci-
sion. Some have concluded risk increases with the seriousness of the underly-
ing crime (Engel, 2003; Johnson, 2001; Kavanaugh, 1997). However, most of
this research only partitions crimes into very rough categories such as binary
indicators of “violent/not violent.” Yet, there may be important differences
between the perpetrators of different crimes within each of those categories
with respect to their underlying propensity to violence as well as the sanc-
tions an offender faces if apprehended (the two paths by which crime serious-
ness is thought to generate risk). In short, criminals responsible for committing
assault, rape, robbery, homicide, or kidnapping may vary in meaningful ways
with respect to risk of violence against officers. Therefore, more precise mea-
surement may prove beneficial.
Finally, there are also wholly omitted constructs in the extant literature—
variables potentially related to violence yet rarely (or never) tested. For
example, Meyer and Carroll (2011) noted, “It is widely believed that domes-
tic violence calls pose the greatest threat to police officers’ safety and that law
enforcement officers are most likely to be injured or killed responding to this
category of call” (p. 24). And yet, they note, there is little or no empirical
evidence to adequately test whether this is the case. Thus, we do not know
whether responding to crimes involving strangers is more risky than those
involving acquaintances, family, or romantic partners. The literature suggests
other constructs may be important to understanding risk as well: the size and
composition of offender groups (Dodge, Coie, Pettit, & Price, 1990; Warr &
Stafford, 1991), the size and composition of victim groups (Ganpat, Leun, &
Nieuwbeerta, 2013; Levine, Taylor, & Best, 2011; Mullen, 1986), and loca-
tion types (Bierie et al., 2013).

Current Study
Few studies of resisting arrest have emerged in the past 20 years, and this work
is limited by small sample sizes, select sampling strategies, and measurement
challenges. These three limitations may lead to missed opportunities to identify
important relationships. This study attempts to fill these gaps by drawing on the
National Incident–Based Reporting System (NIBRS) to model assault against
police. This data set contains the most comprehensive perspective available on
violence against police across the United States (i.e., more geographic and juris-
dictional variety, and a far larger sample size than prior studies). It also contains
detailed information on the incidents (offenders, victims, and crimes) that sup-
port testing new or refined constructs when modeling assault against police.
Bierie 903

NIBRS contains information that is operationally available to most or all


police—details police usually receive before arriving on a crime scene or
surmise immediately after arrival. As such, it provides a data source for
research that delivers actionable information to police. With this context in
mind, two research questions are addressed:

Research Question 1: What factors measured in NIBRS are associated


with increased risk that a law enforcement officer will be assaulted in the
incident?
Research Question 2: Does a risk model using NIBRS data increase pre-
diction in a meaningful way relative to chance?

Method
Data Sources
NIBRS is a data collection and repository system designed to capture informa-
tion about offenders, victims, crimes, and situational characteristics of the 59
most serious crime types (Maxfield, 1999). By 2010, the most recent year data
were available when this study was conducted, 37 states and approximately
6,159 police departments were using NIBRS. This represented approximately
34% of the nation’s police departments and overrepresented small- to medium-
sized agencies. In general, data are uploaded directly from police record man-
agement systems to the NIBRS repository held by FBI on an annual or monthly
basis. Despite modest participation, the full data set holds well over 100 million
rows of data regarding victims, offenders, and situations within those incidents,
making it the largest and most comprehensive police data system in the United
States (Bierie, 2014; Dunn & Zelnock, 1999; Maxfield, 1999).
The use of NIBRS data sets for statistical analyses generally involves col-
lapsing each of the primary file segments (administrative, offense, property,
victim, offender, and arrestee) to the desired unit of analysis (e.g., incident
level) and then merging tables based on police agency identifier (i.e.,
Originating Reporting Identifier (ORI)), year, and a unique number assigned
to each case within that agency (Akiyama & Nolan, 1999; Bierie, 2014;
Maxfield, 1999).1 NIBRS is a publicly available data set, and the codebook,
data collection details, and data can be accessed via the Interuniversity
Consortium for Political and Social Research (ICPSR).2

Sample
In 2002, NIBRS began identifying whether the victim of a crime was a law
enforcement officer (on duty at the time of the victimization). The current
904 Crime & Delinquency 63(8)

study drew on all cases from 2002 to 2010 (the most recent year of data avail-
able when this study was conducted) that involved an officer who was
assaulted when responding to a crime (n = 20,104).3 Next, a random sample
of 20,200 incidents that resulted in an arrest and occurred in the same period
was drawn to create a comparison group representing typical police encoun-
ters. The comparison group included only those incidents that resulted in an
arrest to ensure an encounter occurred between police and at least one
offender. The two groups (assault and comparison) were cross-referenced,
and overlapping cases were deleted from the comparison group. A total of
40,258 incidents remained after this step.

Dependent Variable
The dependent variable for this study was binary. Incidents involving assault
against police were coded as “1” (assault group), and incidents without this form
of aggression were coded as “0” (comparison group). Incidents with an officer
assaulted contained several different types of violence against police, including
73 resulting in homicide, 160 involving kidnapping, 4,221 involving aggravated
assault (i.e., egregious violence), and 146 events involving sexual assault of an
officer. The remainder of events may be considered less serious forms of vio-
lence against police (e.g., 8,368 incidents were classified as simple assault).

Covariates
Cleaning and coding of specific items described below occurred at the file
level. Data sets were then collapsed to the incident level and merged into a
final incident-level data set. Law enforcement officer victimization data were
removed before coding any information from offender, victim, or crime files,
a necessary step to ensure features of the incident referred to the underlying
crimes requiring a law enforcement response (i.e., the criminal acts, loca-
tions, and victims that initiated the police response).4

Offender characteristics.  Offender characteristics refer to victim- or police-


reported information about offenders involved in the crime (i.e., it is not lim-
ited to arrestees). The total number of offenders in the crime was coded as a
count, starting at one and truncated at five. Truncation occurred because there
were a handful of cases in either group (less than 1%) with a large number of
offenders that would have led to bias in estimates (e.g., high leverage). Age
was recorded in a year metric and referred to the average age of all offenders.
A second dummy variable was created to indicate whether any offender was
a juvenile. Race referred to the composition of the offender group and was
Bierie 905

coded into the mutually exclusive categories of White (NIBRS does not dis-
tinguish between Caucasian and Hispanic offenders), Black, Other (Native
American and Asian), and multiple races present. Gender likewise was coded
into mutually exclusive categories: male, female, or both genders present.

Criminal acts.  NIBRS allows 59 different crime types to be recorded for each
victim in an incident. These were recoded into 13 categories of attempted or
completed crimes: homicide, kidnapping, sexual assault (rape, sodomy,
incest, fondling, production of child pornography, trafficking), robbery,
assault (including intimidation), weapon (possession, sale, or use), destruc-
tion of property (vandalism or arson), burglary, motor vehicle theft (MVT),
property (e.g., theft, shoplifting, receiving stolen goods), and swindling (e.g.,
extortion, welfare fraud, wire fraud, impersonation, forgery). Two victimless
crime categories were created as well: drug (possession, sales, or distribu-
tion) and petty (e.g., prostitution, gambling). Items were coded as dummy
variables, but not mutually exclusive. This allowed crime types to be mod-
eled as additive characteristics of the incident.
Two summary measures of the criminal acts were created. First, a count of
the unique number of crimes occurring in the incident was coded (range =
1-30). However, due to the skewness of the frequency distribution, the item
was truncated at five. This resulted in a change to less than 1% of cases.
Second, a count of unique crime types was created. For example, an incident
with one robbery was coded as “1,” as was an incident that included five rob-
beries and no other offenses, whereas an incident with one robbery and one
sexual assault would be coded as “2” types of crimes. This metric captured
the diversity of offending in the incident rather than volume.
Collectively, these two measures were intended to capture latent propen-
sity for violence (e.g., offenders who have low self-control or who are highly
agitated may engage in more offending and more types of offending during
an incident). Using these two variables is consistent with observations that
offenders with the greatest prior offending, and also the most diverse criminal
histories, tend to have the greatest risk of aggression (e.g., Gottfredson &
Hirschi, 1990; Sweeten, 2012). Although these theoretical positions are
grounded on empirical work that used criminal histories, it is likely that the
same patterns would apply at the micro level of a single incident (Bierie
et al., 2013). In fact, offending behavior within an incident may well provide
an even more valuable signal of risk than criminal history due to the proxi-
mate and immediate relevance of the incident-level measure.

Location.  NIBRS allows the recording of more than 20 categories of location


(e.g., bank, school, bar, field/woods, hotel, church, road/alley, home/residence).
906 Crime & Delinquency 63(8)

These were combined into substantively similar categories: outdoors (field/


woods, waterway, or construction area), buildings (e.g., school, diner, gas sta-
tion, office), residence (homes), roads (road, alley, or highway), parking lots,
and other (those identified only as “other” in NIBRS).5 Finally, a binary item
was created to indicate when locations moved during the incident. When this
movement occurred, the other location types were recoded as “0,” thereby cre-
ating a mutually exclusive location variable.

Substance use. The NIBRS offense segment indicated whether alcohol or


other drug use was associated with each offender in an incident. These were
coded into two dummy variables: drug (1/0) and alcohol (1/0). A final dummy
was also created (substance use) and coded as 1 if either alcohol or drug use
by an offender was indicated in the crime. This item was used to decompose
the substance abuse effect—to test the value of measuring drug versus alco-
hol use separately. The items were then collapsed to the incident level and
coded as 1 if any of the crimes in the incident met the criteria.

Victim characteristics.  Measures for victims followed the same strategy used
for offenders. The count of victims was recoded into an ordinal scale of 0, 1,
2, and 3 or more victims. Age of victims referred to the average of the group
and was measured in a year metric. A juvenile variable was created to indi-
cate whether any victim was younger than age 18. Race was coded into mutu-
ally exclusive dummies of Black, White, Hispanic, Other, and multiple races.
NIBRS does record Hispanic ethnicity with regard to victims (but not offend-
ers, as previously noted). NIBRS allows more than 15 relationship categories
between victims and offenders. These were collapsed into a series of mutu-
ally exclusive variables describing the relationship(s) present in an incident:
acquaintance, romantic partner, stranger, family, or multiple. Finally, a non-
resident variable was created to indicate whether the victim was visiting the
jurisdiction where the incident occurred.

Analytic Strategy
The analysis described and then modeled the difference between incidents
that involved the assault of a law enforcement officer and a representative
sample of arrests without violent encounters. All analyses were conducted at
the incident level, and models were estimated within a logistic regression
framework.6
The analysis initially contained all incidents regardless of crime type.
However, because a large portion of the crimes were victimless, this first
analysis examined all incidents but limited consideration to the impact of
Bierie 907

offender and situational variables alone to avoid listwise deletion of incidents


which did not have victim details (e.g., drug and other public order crimes).
Next, the analysis was repeated but restricted to incidents that contained vic-
tims. In this case, information on offenders, situations, and victims was col-
lectively analyzed.
The large sample size here implies that even small differences were likely
to express as statistically significant. This is appropriate, as statistical signifi-
cance merely indicates whether a coefficient is different from the null hypoth-
esis (usually zero) and how likely that coefficient is to be observed again in
future samples from the same population. It does not indicate whether that
real difference is large enough to matter (Ziliak & McCloskey, 2011).
Coefficients can be statistically significant at the p < .0001 level but have
substantive magnitudes that are meaningless. This is true of small data sets or
large ones as used here. But this is especially important to recall in the analy-
sis of larger data sets because misinterpretation of statistical significance as
an indicator of meaningfulness is likely to lead to widespread error. The anal-
yses that follow discuss items that are not only statistically significant but
also substantively large in magnitude.7
The analysis utilizes a fixed-effects framework (Allison, 2009) in which
cases were grouped on state. This limits the potential of contextual contamina-
tion of individual-level coefficients with state-level forces (see Allison, 2009;
Bierie et al., 2013; Iversen, 1991; Shermer, Bierie, & Stark, 2013). This is par-
ticularly important when using NIBRS because of the large number of states
represented, the fact that the value of independent variables may vary across
states, and that states may themselves vary in important ways that affect officer
assaults, such as policies of police agencies with respect to training and use of
body armor (Fridell, Faggiani, Taylor, Britto, & Kubu, 2009) or other legal
structures (Kent, 2010; Mustard, 2001). For example, scholars have noted that
states with a greater use of “mandatory minimum” and “three strike” laws have
greater levels of resisting arrest and homicide of police officers—presumably
because serious violent offenders in these states have less to lose and more to
gain from resistance (Moody, Marvell, & Kaminski, 2002). The fixed-effects
framework solves this problem by computing incident-level coefficients within
each individual state and therefore eliminates the ability of aggregated levels of
these factors, or state-level differences in average violence against police, to
affect coefficient estimates (Rabe-Hesketh & Skrondal, 2008).
Alternatively, a means-only random-effects specification is plausible with
these data and mimics many of the advantages of the fixed-effects model
here. The random-effects model eliminates aggregation bias (contextual con-
tamination) in individual-level covariates and also corrects for clustering in
standard errors to generate unbiased and efficient estimates of Level I
908 Crime & Delinquency 63(8)

coefficients. Random-effects models are often preferred when addressing


multilevel data because they allow one to accurately estimate individual-
level effects while also modeling differences in intercepts between groups or
variance in individual-level coefficients between groups. However, when
hypothesis tests are only focused on accurate estimation of individual-level
coefficients, the fixed-effects approach is better. This is because the random-
effects approach uses information from within and between clusters and as a
result assumes no correlation between any omitted variable at the group level
and any included item in the model—that there is no endogeneity (Rabe-
Hesketh & Anders, 2008). The fixed-effects approach does not make this
assumption as it is a true within-group estimate (Allison, 2009). As such, the
fixed-effects model is preferred for the context of this study.

Results
As shown in Table 1 (column 7), a typical police encounter involved 1.3
offenders, and the distribution was highly skewed (i.e., 78% had one
offender). Offender groups were just below 30 years old on average, 71%
contained all males, and 69% included only White offenders. A sizable pro-
portion of encounters involved substance use (23%) and most were located
either in a home (37%) or building (such as restaurants, stores, offices; 28%).
Only 1% of typical police encounters involved crimes that moved between
two or more locations. Most incidents involved the assault of a citizen (33%),
property crime (22%), or drug crime (29%) prior to police arrival. On aver-
age, these incidents involved 1.33 criminal acts across 1.12 offense types.
There was an average of 1.21 victims among those cases with victims
(approximately 75%). Victim groups were usually slightly older than offender
groups and just over half of victim groups consisted of all females (53%).
Most victim groups only involved White victims (71%), and more than half
the victims were an acquaintance (26%) or romantic partner (33%) of the
offender. Again, note that these data refer to characteristics of the crime an
officer responded to. The data do not refer to officers themselves.
The descriptive statistics in Table 1 showed several large and substan-
tively interesting bivariate differences between the two groups. In terms of
offender characteristics, the data showed assaultive incidents were more
likely to have male and Black offenders. The largest difference was found in
the use of alcohol—incidents where an officer was assaulted were almost 3
times as likely to involve an intoxicated offender. There were several crime
characteristics that distinguished between groups. Although incidents rarely
spanned two or more locations in either group, they were 10 times as likely
to do so in cases leading to violence against an officer. Assaultive incidents
Table 1.  Descriptive Statistics of Study Sample (N = 40,258).

Total sample Typical encounter Assault Difference

Variable Minimum Maximum Observations M SD M SD M SD Typical − assault


Offender
  Count of offenders 1 5 40,258 1.32 0.71 1.32 0.70 1.31 0.72 0.01
  Age (years) 10 97 40,045 29.39 11.51 28.78 11.74 30.00 11.25 −1.22***
 Age-squared 0 1 40,045 0.16 0.36 0.18 0.38 0.13 0.34 0.04***
 Age-cubed 0 1 40,168 0.74 0.44 0.71 0.45 0.76 0.43 −0.05***
  Juvenile (any) 0 1 40,168 0.19 0.39 0.22 0.41 0.16 0.37 0.05***
  Male (all) 0 1 39,865 0.74 0.44 0.71 0.45 0.76 0.43 0.02***
  Female (all) 0 1 39,865 0.07 0.26 0.07 0.25 0.08 0.27 −0.01**
  Multiple gender 0 1 39,865 0.37 0.48 0.32 0.47 0.41 0.49 −0.10***
 Race
  Black 0 1 39,865 0.37 0.48 0.31 0.47 0.41 0.49 −0.10***
  White 0 1 39,865 0.64 0.48 0.69 0.46 0.59 0.49 0.10***
  Other 0 1 39,865 0.02 0.13 0.02 0.12 0.02 0.14 0.00***
  Multiple race 0 1 39,865 0.03 0.16 0.02 0.15 0.03 0.16 0.00*
Situation
 Alcohol 0 1 40,258 0.19 0.39 0.10 0.29 0.27 0.45 −0.18***
 Drugs 0 1 40,258 0.15 0.35 0.15 0.36 0.14 0.35 0.01*
  Substance (either) 0 1 40,258 0.30 0.46 0.23 0.42 0.37 0.48 −0.14***
 Locations
  Building 0 1 40,173 0.23 0.42 0.28 0.45 0.17 0.38 0.11***
  Outdoors 0 1 40,173 0.01 0.10 0.01 0.10 0.01 0.10 0.00
  Home 0 1 40,173 0.36 0.48 0.37 0.48 0.36 0.48 0.01*
  Road 0 1 40,173 0.29 0.45 0.23 0.42 0.34 0.47 −0.11***

909
  Other location 0 1 40,173 0.06 0.25 0.07 0.25 0.06 0.24 0.00
(continued)
910
Table 1.  (continued)

Total sample Typical encounter Assault Difference

Variable Minimum Maximum Observations M SD M SD M SD Typical − assault


 Parking 0 1 40,173 0.07 0.25 0.05 0.22 0.08 0.27 −0.02
  Location moves 0 1 40,206 0.06 0.24 0.01 0.11 0.11 0.31 −0.09***
 Offenses
  Homicide 0 1 40,258 0.00 0.05 0.00 0.04 0.00 0.05 0.00
  Kidnapping 0 1 40,258 0.01 0.08 0.01 0.08 0.01 0.08 0.00
  Sexual 0 1 40,258 0.01 0.10 0.02 0.13 0.00 0.06 0.01***
  Robbery 0 1 40,258 0.01 0.12 0.01 0.12 0.01 0.11 0.00
  Assault 0 1 40,258 0.46 0.50 0.33 0.47 0.59 0.49 −0.27***
  Weapon 0 1 40,258 0.06 0.24 0.04 0.20 0.09 0.28 −0.05***
   Destruction of property 0 1 40,258 0.13 0.33 0.08 0.27 0.18 0.38 −0.10***
  Burglary 0 1 40,258 0.04 0.19 0.04 0.20 0.03 0.18 0.01***
   Motor vehicle theft 0 1 40,258 0.02 0.13 0.02 0.13 0.01 0.12 0.00
  Property 0 1 40,258 0.15 0.36 0.22 0.42 0.08 0.27 0.14***
  Swindle 0 1 40,258 0.04 0.20 0.06 0.23 0.03 0.16 0.03***
  Drug 0 1 40,258 0.26 0.44 0.29 0.45 0.24 0.43 0.04***
  Petty 0 1 40,258 0.01 0.08 0.01 0.08 0.01 0.08 0.00
  Count of crimes 1 5 40,258 1.20 0.48 1.33 0.68 1.42 0.80 −0.10***
  Diversity of crimes 1 5 40,258 1.20 0.48 1.12 0.36 1.28 0.56 −0.16***
(continued)
Table 1.  (continued)

Total sample Typical encounter Assault Difference

Variable Minimum Maximum Observations M SD M SD M SD Typical − assault


Victim (if applicable)
  Count of victims 1 3 31,980 1.25 0.68 1.21 0.61 1.29 0.73 −0.08***
 Age 0 97 18,506 34.43 14.56 33.12 14.87 35.91 14.05 −2.79***
  Juvenile (any) 0 1 18,506 0.12 0.32 0.14 0.35 0.09 0.29 0.05***
  Male (all) 0 1 19,043 0.40 0.49 0.39 0.49 0.42 0.49 −0.02***
  Female (all) 0 1 19,043 0.50 0.50 0.53 0.50 0.46 0.50 0.07***
  Multiple genders 0 1 19,043 0.10 0.30 0.08 0.27 0.12 0.33 −0.04***
  Multiple races 0 1 18,574 0.03 0.17 0.02 0.14 0.04 0.19 −0.02***
 Black 0 1 18,574 0.24 0.43 0.23 0.42 0.25 0.43 −0.02***
 White 0 1 18,574 0.71 0.45 0.71 0.45 0.71 0.45 0.00
 Hispanic 0 1 18,681 0.06 0.24 0.07 0.25 0.06 0.23 0.01*
  Other race 0 1 18,574 0.02 0.13 0.02 0.12 0.02 0.13 0.00
 Acquaintance 0 1 14,350 0.23 0.42 0.26 0.44 0.20 0.40 0.06***
 Family 0 1 14,350 0.15 0.36 0.16 0.37 0.15 0.35 0.01*
  Romantic partner 0 1 14,350 0.28 0.45 0.33 0.47 0.24 0.43 0.09***
 Stranger 0 1 14,350 0.22 0.41 0.16 0.37 0.28 0.45 −0.12***
  Multiple relations 0 1 14,350 0.11 0.31 0.09 0.28 0.13 0.34 −0.04***
  Nonresident of area 0 1 17,337 0.16 0.36 0.15 0.36 0.17 0.37 −0.02***

*p < .05. **p < .01. ***p < .001.

911
912 Crime & Delinquency 63(8)

were substantially more likely to have occurred in roadways or alleys (and


less likely to have occurred in a building). In addition, they were more likely
to involve an underlying crime of assault (59% vs. 33%), a weapon offense
(9% vs. 4%), or destruction of property (18% vs. 8%). There were slightly
more crimes occurring in cases with violence against police (1.42 vs. 1.22)
and diverse offenses (1.28 vs. 1.22). Finally, incidents that resulted in assault
of officers were more likely to have all-male or mixed-gender victim groups,
stranger victims, and a slightly greater number of victims. Assaultive inci-
dents were less likely to involve juvenile victims or romantic partners.
Table 2 presents estimates from a fixed-effects logistic regression frame-
work. Model 1 illustrates the impact of incident-level variables on victimless
as well as victim crimes, whereas Model 2 restricts consideration only to
incidents with a victim. Both models of Table 2 generate substantial improve-
ments over chance, as indicated by the classification scores reported at the
bottom of the table. A comparison of the classification scores between Models
1 and 2 indicates these covariates explain an exceptional amount of risk fac-
ing officers—increasing correct classification from 50% with no model (the
mean) to 82% to 84% via the models.
Model 1 shows that risk increases with the count and average age of
offenders. It also suggests the effect of age on risk increased as the offender
group aged (a nonlinear effect). The nonlinearity component of age was sta-
tistically significant but not meaningful in terms of magnitude (even if inter-
preted in terms of large changes in age, such as a jump of 20 years).
Importantly, these multivariate estimates showed that the presence of at least
one juvenile in the offender group actually increased risk to officers after
controlling for age (contrary to Table 1). Comparing Table 1 and Table 2
implies (a) incidents with at least one juvenile tend to have other characteris-
tics that are less risky on average (i.e., other items in Table 1 that signal less
risk), and (b) the juvenile effect is confounded with the measurement of age
(because age and juvenile have opposite effects). Holding age and other vari-
ables constant, incidents with a juvenile were 21% more likely to result in an
officer assaulted than incidents that only included adults.
The data showed that several offender characteristics were important.
Female or all-female groups presented the least risk of assault to officers,
male or all-male groups presented significantly more risk, and multiple-
gender groups of offenders presented the most risk. The data showed that
incidents with Black offenders were more likely to result in violence against
police than incidents with White or Hispanic offenders. There was no differ-
ence between Black and Asian/Native American race groups (“Other”).
Incidents with multiple races present had significantly higher odds of assault-
ing officers than any single race group.
Table 2.  Assault of Police Modeled on Offender and Situational Characteristics: Fixed-Effects Logistic Regression.

Model 1 Model 2

  Odds ratio Coefficient SE p Odds ratio Coefficient SE p


Offender
  Count of offenders 1.59 0.46 0.03 .000 1.50 0.41 0.07 .000
  Age (years) 1.20 0.18 0.02 .000 1.15 0.14 0.04 .001
 Age-squared 1.00 0.00 0.00 .000 1.00 0.00 0.00 .003
 Age-cubed 1.00 0.00 0.00 .000 1.00 0.00 0.00 .007
  Juvenile (any) 1.21 0.19 0.07 .006 1.19 0.17 0.14 .217
  Female (all) 0.85 −0.17 0.04 .000 1.08 0.08 0.08 .318
  Multiple gender 1.60 0.47 0.07 .000 0.96 −0.05 0.15 .765
 Race
  White 0.84 −0.18 0.04 .000 1.03 0.03 0.10 .798
  Other 0.92 −0.08 0.14 .565 1.53 0.43 0.30 .150
  Multiple race 1.48 0.39 0.11 .001 1.00 0.00 0.26 .991
Situation
 Locations
  Building 1.90 0.64 0.16 .000 0.79 −0.23 0.29 .415
  Outdoors 2.42 0.88 0.22 .000 1.43 0.36 0.48 .459
  Home 1.37 0.32 0.16 .045 0.82 −0.20 0.28 .470
  Road 3.10 1.13 0.16 .000 1.20 0.18 0.28 .516
  Other location 1.66 0.51 0.16 .002 0.81 −0.22 0.30 .473
  Parking 2.51 0.92 0.16 .000 0.91 −0.10 0.29 .740

(continued)

913
914
Table 2.  (continued)

Model 1 Model 2

  Odds ratio Coefficient SE p Odds ratio Coefficient SE p


  Substance use
  Alcohol 2.82 1.04 0.05 .000 2.26 0.81 0.07 .000
  Drugs 1.27 0.24 0.05 .000 1.99 0.69 0.16 .000
 Offenses
  Homicide 2.13 0.76 0.35 .033 0.32 −1.14 0.62 .065
  Kidnapping 2.80 1.03 0.22 .000 5.89 1.77 0.36 .000
  Sexual 1.00 0.00 0.19 .988 0.61 −0.50 0.45 .265
  Robbery 3.36 1.21 0.14 .000 0.77 −0.26 0.41 .526
  Assault 7.41 2.00 0.06 .000 1.55 0.44 0.41 .279
  Weapon 3.18 1.16 0.08 .000 0.40 −0.91 0.32 .004
   Destruction of property 4.61 1.53 0.07 .000 3.56 1.27 0.29 .000
  Burglary 1.62 0.48 0.10 .000 3.60 1.28 0.34 .000
   Motor vehicle theft 1.77 0.57 0.14 .000 1.09 0.08 0.56 .880
  Swindle 1.72 0.54 0.10 .000 2.35 0.86 0.59 .148
  Victimless 2.06 0.72 0.06 .000 12.62 2.54 0.33 .000
  Count of crimes 0.17 −1.78 0.03 .000 0.05 −2.96 0.08 .000
  Diversity of crimes 5.70 1.74 0.07 .000 5.52 1.71 0.26 .000

(continued)
Table 2.  (continued)

Model 1 Model 2

  Odds ratio Coefficient SE p Odds ratio Coefficient SE p


Victims
  Count of victims 28.31 3.34 0.11 .000
 Age 1.02 0.02 0.00 .000
  Juvenile (any) 0.99 −0.01 0.11 .943
  Female (all) 1.09 0.09 0.07 .206
  Multiple genders 1.09 0.09 0.14 .539
 Race
  Multiple races 1.05 0.04 0.23 .844
  White 1.14 0.13 0.10 .209
  Hispanic 1.06 0.06 0.16 .722
  Other race 0.59 −0.53 0.29 .073
 Relationships
  Acquaintance 0.45 −0.80 0.10 .000
  Family 0.52 −0.66 0.13 .000
  Romantic partner 0.36 −1.01 0.11 .000
  Multiple relations 0.68 −0.38 0.14 .005
  Nonresident of area 1.16 0.15 0.09 .096
 Constant 0.19 −1.68 0.34 −4.890 0.19 −1.68 0.69 .016
n 32,873 10,236
Correctly classified 83.61% 81.80%

Note. Reference category includes offender race = White; substance use = none; location = location moves; offender gender = male (all), and

915
relationship = stranger.
916 Crime & Delinquency 63(8)

Several situational features were important. Offenders under the influence


of alcohol had almost 3 times the odds of assaulting police. Offenders under
the influence of other substances were slightly more likely to assault police
(relative to offenders not under the influence of any substance). The data also
showed several important differences in risk between locations. Incidents
that “moved” areas had the lowest risk of assault for officers (the reference
category). All other location types were more risky, and roadways presented
the highest risk of assault to officers.
Finally, crime types signaled some of the largest risk for police. The low-
est risk scenarios for officers were found in property crime (the reference
category) and sex crime. The most risk was indicated by the presence of
assault (odds ratio [OR] = 7.4), destruction of property (OR = 4.6), and rob-
bery (OR = 3.4). Incidents with more crimes present signaled less risk, but
only when controlling for the occurrence of each crime type during an inci-
dent. Consistent with expectations, a greater diversity of offenses in the inci-
dent corresponded to substantially more risk to officers.8
Model 2 presents estimates after restricting the sample to incidents with
victim data. This not only allows the model to estimate the effect of victim
characteristics but also reduces the sample size substantially from Model 1
(i.e., the usable sample decreases from over 32,000 cases in Model 1 to just
over 10,000 cases in Model 2). Both groups (assault and comparison) showed
nearly identical proportion of cases dropped when moving to Model 2. Some
of the difference is due to missing data on victim characteristics. The major-
ity of the difference, however, reflects a drop in the number of qualifying
incidents. If an incident only involved victimless crimes, or a “victim” crime
against an entity such as a business, then the incident was discarded due to
listwise deletion. Victimless crimes can still be present in the analysis if the
incident included another offense in the incident with victim data.
The estimates from Model 2 show some important similarities to the esti-
mates from Model 1. For example, the average age of offenders maintains its
significant relationship and a similar magnitude of effect. Likewise, both
alcohol and substance use signal risk of assault to officers at approximately
the same levels as Model 1, as does the count of offenders.
There are also several important differences as compared with Model 1.
For example, offender race is no longer significant and location types no
longer matter. Having a juvenile present no longer signaled significant
increases in risk. The coefficients for offender gender also changed—The
data showed this item no longer predicted assault. This suggests females are
less likely to engage in violence overall (Model 1), but once we examine only
the subpopulation of females who victimize others (Model 2), gender is no
longer a factor.
Bierie 917

Perhaps the largest change between the two models occurred with respect
to crime types. These effects were substantially smaller than in Model 1. In
Model 2, most offenses showed no statistically significant differences in risk
relative to the reference category (i.e., 6 out of 11 crime types were not sig-
nificant). The most risky offenses were kidnapping, burglary, and destruction
of property. In addition, the riskiest crime type was “victimless”—a drug or
prostitution crime. However, recall these incidents required another victim-
based crime to be used in Model 2. This is different from the victimless
crimes in Model 1, which were primarily composed of victimless crime inci-
dents that did not have any other crimes against victims.
Model 2 also indicated that several victim characteristics mattered. Risk
rose significantly as the number of victims grew and as the average age of
victims increased. The data showed that victim race and gender combinations
did not differentiate risk, although relationship type did. Incidents with
stranger victims presented the greatest risk to officers: approximately twice
the risk as acquaintance or family relation categories and approximately
thrice the risk as incidents involving romantic partners. Contrary to popular
belief regarding domestic violence, these data showed the groups “romantic
partner” and “family” posed the least risk to officers. Residential status of
victims (i.e., whether that victim resides in the incident jurisdiction) had no
impact on risk to officers.
It is not clear whether differences between Models 1 and 2 were due to the
inclusion of victim-level covariates in Model 2 (e.g., victim measures account
for the variance otherwise attributed to coefficients for crime types) or a dif-
ference in samples (i.e., Model 2 only includes incidents with victims).
Regenerating a fixed-effects model using the items from Model 1 but with the
sample used in Model 2 indicates consistency between coefficients in Model
1 and the new model (data not shown). This suggests the differences between
Model 1 and Model 2 are most likely due to the omitted-variable bias of vic-
tim characteristics rather than differences in cases available for Model 2.

Discussion
Little is known about risk factors for assaults on police officers. Only a hand-
ful of studies exist on this topic, and these tend to focus on minor forms of
resistance (e.g., refusing officer commands) or very serious but rare attacks
(e.g., gun violence directed at officers). Few studies examine assault specifi-
cally. This is an important limitation in the field because this form of aggres-
sion embodies a great deal of risk to officers, carries the potential to strain or
damage public relations, and generates other costs to an agency or commu-
nity. It is also noteworthy because risk factors for assaulting police may well
918 Crime & Delinquency 63(8)

be different from the rare lethal attacks or the more common noncompliance
scenarios examined in prior studies.
The current study drew on NIBRS, the nation’s largest data set for track-
ing information on offenders, victims, and situations surrounding criminal
incidents. The study utilized a case-control methodology in which all events
involving assault of police were compared with a random sample of incidents
that did not generate this form of violence. The models were computed to
answer two key questions: (a) What incident-level factors predict violence
against officers? and (b) What portion of violence can be explained or pre-
dicted by data available in NIBRS?
The data showed a number of factors predicted risk to officers. Consistent
with prior literature, these data showed risk increasing with the count of offend-
ers present (Bierie et al., 2013; Covington et al., 2014). This may support argu-
ments from researchers who anticipate increased resistance when the chance for
success is higher. For example, perhaps the reason resistance is more likely on
roadways or outdoors is that these locations offer the most hope of escape.
Likewise, perhaps situations having more offenders present may imply a greater
chance of overpowering an officer to facilitate flight. Or, perhaps these situa-
tions have other explanations. For example, these findings may support other
researchers who suggest a higher likelihood of resistance when a larger audience
is present to witness the display of aggression (Mastrofski et al., 1996).
These data also highlight an interesting and important conflict in the resis-
tance literature. Most studies find male and female offenders resist arrest at
similar rates (Engel, 2003; Johnson, 1999; Kavanaugh, 1997; McClusky
et al., 1999) or that females resist more than males (Mastrofski et al., 1996).
However, these studies tend to be saturated with less serious crime encoun-
ters and measure minor forms of resistance (e.g., disobeying an officer’s
command). Research examining serious or lethal violence against officers
finds males are riskier (Bierie et al., 2013; Craun et al., 2013). This study
showed a middle ground.
In general, all-female offender incidents were significantly less likely to
assault officers (Table 2, Model 1) than all-male offender incidents. However,
the difference between male and female offenders largely dissipated (and
even reversed) when only predatory crimes were examined (Model 2). Thus,
gender differences in assault may emerge, in part, by selection into offense
types: Men are more likely than women to engage in crimes with direct vic-
tims, and it is this kind of crime that signals risk to officers. Within incidents
containing victims, gender no longer signals risk.
A similar pattern is found when examining mixed-gender groups across
Model 1 and 2. In general, offender groups with more than one gender pres-
ent were the riskiest (all else held constant). These findings remain after
Bierie 919

controlling for the number of offenders present, and they also hold when
restricting the analysis only to incidents with more than one offender (data
not shown). These data do not indicate whether males, females, or both com-
mit a given act of assault against officers in the incident. For this reason, it is
particularly difficult to interpret gender effects. It may be that men act more
aggressively in front of female accomplices. Or, perhaps female offenders
behave more aggressively in an effort toward identity management in a male-
dominated field (crime). Or, perhaps the pattern observed here is a function
of omitted variables—spuriousness due to the lack of important covariates
(e.g., perhaps crimes committed by mixed-gender groups tend to be gang-
related, and gang member arrests pose risk). Importantly, the effects of
offender gender dissipate after controlling for victim characteristics and
examining only those cases in which a person was victimized (Model 2).
Investigating gender may be an important avenue for future research.
As noted, the findings change in some important ways when consideration
is limited to victim-involved crimes. The most dramatic change is found in
the magnitude of ORs for criminal acts. Fewer crimes matter, and their
impacts diminish or reverse when moving from Model 1 to Model 2. This is
due in part to case differences in the two models, but more so from the inclu-
sion of victim characteristics in the latter model. The reason crimes such as
robbery and assault have such a dramatic impact on risk to officers (Model 1)
is likely that they have different kinds of victims—and these attributes of
victims signal varying amounts of risk to officers. Once victim attributes are
added to the equation (Model 2), the crime coefficients diminish.9
Finally, the data showed an interesting pattern with respect to domestic
violence. The literature has generally suggested that these incidents pose
elevated risk to officers. Theories of that risk revolve around the intense emo-
tionality of parties involved, or perhaps diminished perceived legitimacy of
police to interfere with such a personal matter (McClusky et al., 1999). These
data suggested that there are many other types of incidents which were just as
dangerous for police officers or even more so—that domestic violence was
not unique and extraordinary relative to other types of calls for service.
Caution is warranted when interpreting this finding until more research or
replications can be done; there is relatively little empirical work on this topic
to date (Meyer & Carroll, 2011). Future research should also investigate
whether there are subgroups of domestic violence that do pose special risk
but were masked within the large pool of cases defined here.
This study is not without limitations. First, the NIBRS data are not com-
prised of a random or systematic sample. The data oversample small- and
medium-size departments, and this was even more true of earlier years of the
data collection. Although more than 30% of the nation is represented by
920 Crime & Delinquency 63(8)

NIBRS, it is unclear whether patterns observed in these agencies generalize


to patterns in nonreporting departments. This is a similar limitation as found
with prior studies of violence against police (e.g., a study of a single city may
only generalize to that city). Here, the patterns may only generalize to the
6,000 departments reporting to NIBRS.
Second, it is important to note that NIBRS is a complex system, requiring
thousands of unique individuals to enter data. Criminal Justice Information
Services (CJIS) provides uniform training and consistent guidebooks for all
jurisdictions participating in NIBRS. In addition, CJIS audits the system and
requires an error rate below 3%. However, it seems likely there are still
opportunities for error given that jurisdictions may differ in crime classifica-
tion or other influencers of data entry.
Third, the study did not include information on police strategy in the
encounter. Perhaps police behavior or tactics change across encounters such
that police tend to approach some crimes in a way that decreases their chance
of being assaulted (Pate, Fridell, & Hamilton, 1993). For example, perhaps
the reason incidents with homicides or sexual assault are less risky to officers
(Table 2) is because officers tend to approach suspects in these cases more
carefully than cases they deem less serious.
Finally, and most importantly, there are other variables not included in
NIBRS, which would doubtlessly contribute to the explanation of violence.
NIBRS contains no information on police officers unless they are victims of a
crime. This means police officer characteristics cannot be used in models
comparing incidents with police assaulted to those without—the data would
be uniformly missing in the comparison group. Likewise, there is no historic
information on offenders, such as their criminal history or other personal
background factors that may signal propensity to violence. For example, per-
haps the reason “offender age” signals a positive association with risk is
because older offenders have longer criminal histories and thus a greater pro-
pensity toward violence. The issue of omitted variables is a significant limita-
tion to NIBRS, as in most social science research. It means that some of the
patterns observed here may be nothing more than a proxy for these omitted
variables—a result of spuriousness.
Omitted-variable bias is a pervasive and important issue in social science
research with a goal of identifying the precise relevance of individual con-
structs (i.e., theory testing). However, spuriousness is sometimes a desired
property of risk assessment “in the field.” This occurs when the items included
in a risk analysis reflect measurable information in practice and omitted vari-
ables are difficult or impossible to measure. In this case, the omitted items are
allowed to act on risk signaling through included and measurable items—the
spuriousness is exploited to improve predictions (Burke, 2012). For this
Bierie 921

reason, the potential problem implied by using NIBRS is reduced somewhat


when the study is considered in the context of practical application by police
officers. One may be less concerned, for example, about the omission of
criminal history of offenders because police rarely know this information
when responding to a crime scene—at least not until after they make the
arrest and their risk for assault has mostly passed.
Notwithstanding these limitations, the study makes important contribu-
tions to the study of policing and violence. These data offer a step forward in
the pursuit of tangible risk assessment tools for officers. The items tested here
were derived from operational police records—they reflect information police
can, and do, measure. Furthermore, most items can be ascertained before
arrival at a crime scene or immediately after arrival. Thus, these findings pro-
vide reasonable and relevant information for officers—information they can
access and use in individual cases or planning responses to crimes. However,
it is critically important to note that assault on police is a rare event. National
statistical reporting shows that 10.7 million arrests were made in 2009 (Bureau
of Justice Statistics, 2011, Table 4.3) and there were just over 58,000 officers
assaulted that year (Law Enforcement Officers Killed and Assaulted [LEOKA],
2011, Table 77)—a risk of 5 assaults for every 1,000 arrests. Thus, knowing
that an incident poses 10 times more risk than usual would only indicate a 5%
chance of assault—still an unlikely outcome. It is important that police offi-
cers use this base-rate information in interpreting the risk they face.
The study also offers new ideas to academia. It suggests academics inves-
tigate more thoroughly those items identified previously as significant such
as gender. Scholars may benefit from pursuing a deeper understanding of
when and why the presence of female offenders can lead to greater risk to
officers in this and other studies of noncompliance with police. The study
also identifies items not yet considered (or rarely considered) in this arena of
scholarship, including the importance of the size of offender and victim
group, the impact of locations and specific crime types, and the interplay of
these measures with victim characteristics. Collectively, these findings pres-
ent a new and exciting challenge to academia—new puzzles for policing
scholars to consider and explain.

Declaration of Conflicting Interests


The author(s) declared no potential conflicts of interest with respect to the research,
authorship, and/or publication of this article.

Funding
The author(s) received no financial support for the research, authorship, and/or publi-
cation of this article.
922 Crime & Delinquency 63(8)

Notes
1. National Incident–Based Reporting System (NIBRS) data are complex and con-
sist of interlinked relational tables—each maintained at a different unit of analy-
sis. Readers can contact the lead author for additional written explanation of
computational steps and syntax used to clean, create, and analyze these data.
2. http://www.icpsr.umich.edu/icpsrweb/NACJD/NIBRS/
3. Note that research suggests that approximately 58,000 officers are assaulted each
year during this time period (National Law Enforcement Officers Memorial Fund
[NLEOMF], 2014). The count of assaults in NIBRS is substantially smaller. This is
expected for three reasons. First, the assaults recorded in NIBRS were reported as
crimes, whereas the NLEOMF data also includes minor incidents of victimization
for which officers did not seek formal charges and prosecution. Second, several
states representing a large portion of the nation’s crime do not report to NIBRS
(e.g., New York). Third, the number of agencies reporting to NIBRS has grown
over time to the current 34% representation. In the earliest year of this study, only
22% of the nation’s police departments were represented here. Issues of generaliz-
ability implied by these factors are reiterated in the “Discussion” section.
4. Note that the analysis does not consider police officer attributes. NIBRS does not
record any characteristics of police officers unless they are victims of a crime.
Therefore, the data include information about officers in the assault group alone
but are universally missing in the comparison group.
5. Groupings were made based on theoretical similarity, such as the liklihood of
a witness being present at the scene (Mastrofski, Snipes, & Supina, 1996). The
component locations within a given group (when possible) were also examined
individually and found consistent with this aggregation structure in terms of their
relationship to membership in the assault group.
6. The goal of this research was to understand which incidents present the most
danger to officers. The decision to analyze the data in this way emphasizes an
interest in creating risk prediction tools to assist officers and dispatchers when
responding to incidents and to enhance training and risk mitigation tools in the
field. Thus, all analyses were conducted at the incident level after fixing aggre-
gate (state) level effects.
7. As a sensitivity analysis, the models described below were reestimated on two
separate random subsamples (10% of the data) to identify any changes in coef-
ficients or statistical significance as a function of sample size. Few differences
were observed and conclusions remained the same. For example, in Model 1,
only one independent variable exhibited change between the main and subset
sample (p < .04; Paternoster, Brame, Mazerolle, & Piquero, 1998). In addition,
only a handful of items changed from statistically significant to marginal or non-
significance. These differences were found in rare event items (e.g., homicide,
kidnapping), which indicates the subsamples had too small an n to generate a
reliable analysis for these items rather than a problem with the model itself. I
would like to thank the editor and reviewers for suggesting this specific series of
sensitivity analyses. Tables are available upon request.
Bierie 923

8. Count of crimes and diversity of crimes are collinear (r = .55), which drive down
the value for count of crimes. Using both measures in the same model allows the
most robust approach to controlling for crime-type information when estimating
other independent variables in the model. However, the correlated nature of the
two measures of criminal acts means that interpretation of an individual crime
summary measure alone warrants caution.
9. The analysis does not test which victim attributes wiped out crime-type effects.
Rather, it simply notes that the coefficients for crimes such as robbery are signifi-
cant prior to including the victim controls (Model 1) and are no longer significant
when the various victim characteristics are entered into the model (Model 2).

References
Akiyama, Y., & Nolan, J. (1999). Methods for understanding and analyzing NIBRS
data. Journal of Quantitative Criminology, 15, 225-238.
Allison, P. (2009). Fixed effects regression models. Newbury Park, CA: Sage.
Alpert, G., & Dunham, R. (1997). Understanding police use of force: Officers, sus-
pects, and reciprocity. Cambridge, MA: Cambridge University Press.
Belvedere, K., Worrall, J., & Tibbetts, S. (2005). Explaining suspect resistance in
police citizen encounters. Criminal Justice Review, 30, 30-44.
Bierie, D. (2014). Enhancing the National Incident-Based Reporting System: A
policy proposal. International Journal of Offender Therapy and Comparative
Criminology. Advance online publication. doi:10.1177/0306624X14525926
Bierie, D., Detar, P., & Craun, S. (2013). Firearm violence against police officers. Crime
& Delinquency. Advance online publication. doi:10.1177/0011128713498330
Brandl, S., & Stroshine, M. (2003). Toward an understanding of the physical hazards
of police work. Police Quarterly, 6, 172-191.
Brandl, S., & Stroshine, M. (2012). The physical hazards of police work revisited.
Justice Quarterly, 15, 262-282.
Bureau of Justice Statistics. (2011). Federal justice statistics 2009—Statistical tables
(NCJ 233464). Washington, DC: U.S. Department of Justice.
Bureau of Labor Statistics. (2011). National census of fatal occupational injuries in
2010 (USDL-11-1247). Washington, DC: U.S. Department of Labor.
Burke, R. (2012). Criminal justice forecasts: A machine learning approach. New
York, NY: Springer.
Covington, M., Huff-Corzine, L., & Corzine, J. (2014). Battered police: Risk fac-
tors for violence against law enforcement officers. Violence and Victims, 29(1),
34-52.
Craun, S., Detar, P., & Bierie, D. (2013). Shots fired: Examining lethal force used
against Deputy U.S. Marshals. Victims & Offenders, 8, 56-69.
Dodge, K., Coie, J., Pettit, G., & Price, J. (1990). Peer status and aggression in
boy’s groups: Developmental and contextual analysis. Child Development, 61,
1289-1309.
Dunn, C., & Zelnock, T. (1999). NIBRS data available for secondary analysis. Journal
of Quantitative Criminology, 15, 239-248.
924 Crime & Delinquency 63(8)

Engel, R. S. (2003). Explaining suspects’ resistance and disrespect toward police.


Journal of Criminal Justice, 31, 475-492.
Federal Bureau of Investigation. (2011). Table 70: Law enforcement officers
assaulted, type of weapon and injury 2011. Retrieved from http://www.fbi.gov/
about-us/cjis/ucr/leoka/2011/tables/table-70
Fridell, L., Faggiani, D., Talor, B., Brito, C., & Kubu, B. (2009). The impact of agency
context, policies and practices on violence against police. Journal of Criminal
Justice, 37, 542-552.
Fyfe, J. (1978). Shots fired: An examination of New York City police firearms dis-
charges. Albany: State University of New York.
Ganpat, S., Leun, J., & Nieuwbeerta, P. (2013). The influence of event characteristics
and actors’ behavior on the outcome of violence events: Comparing lethal with
non-lethal events. British Journal of Criminology, 53, 685-704.
Gottfredson, M., & Hirschi, T. (1990). A general theory of crime. Stanford, CA:
Stanford University Press.
Iversen, G. (1991). Contextual analysis. Newbury Park, CA: Sage.
Johnson, R. (1999). Phoenix project: Predictors of suspect use of force (NCJRS:
187776). Washington, DC: U.S. Department of Justice.
Kavanaugh, J. (1997). The occurrence of resisting arrest in arrest encounters: A study
of police-citizen violence. Criminal Justice Review, 22, 16-33.
Kent, S. (2010). Killings of police in U.S. cities since 1980: An examination of envi-
ronmental and political explanations. Homicide Studies, 14, 3-23.
Law Enforcement Officers Killed and Assaulted. (2011). Extent of injury of victim
officer by type of weapon, 2007-2011. Washington, DC: U.S. Department of
Justice. Retrieved from http://www.fbi.gov/about-us/cjis/ucr/leoka/2011/tables/
table-77
Levine, M., Taylor, P., & Best, R. (2011). Third parties, violence, and conflict reso-
lution: The role of group size and collective action in the microregulation of
violence. Psychological Science, 22, 406-412.
Mastrofski, S., Snipes, J., & Supina, A. (1996). Compliance on demand: The pub-
lic’s response to specific police requests. Journal of Research in Crime &
Delinquency, 33, 269-305.
Maxfield, M. (1999). The National Incident-Based Reporting System: Research and
policy applications. Journal of Quantitative Criminology, 15, 119-149.
McClusky, J., Mastrofski, S., & Parks, R. (1999). To acquiesce or rebel: Predicting
citizen compliance with police requests. Police Quarterly, 2, 389-416.
Meyer, S., & Carroll, R. (2011). When officers die: Understanding deadly domestic
violence calls for service. The Police Chief, 78(5), 24-27.
Moody, C., Marvell, T., & Kaminski, R. (2002). Unintended consequences: Three-
strikes laws and the murders of police officers (NCJRS: 203649). Washington,
DC: U.S. Department of Justice.
Mullen, B. (1986). Atrocity as a function of lynch mob composition: A self-attention
perspective. Personality and Social Psychology Bulletin, 12, 187-197.
Mustard, D. (2001). The impact of gun laws on police deaths. Journal of Law &
Economics, 44, 635-657.
Bierie 925

National Law Enforcement Officers Memorial Fund. (2014). Deaths assaults & inju-
ries. Author. Retrieved from http://www.nleomf.org/facts/officer-fatalities-data/
daifacts.html
Pate, A., Fridell, L., & Hamilton, E. (1993). Police use of force: Official reports, citizen
complaints, and legal consequences. Washington, DC: The Police Foundation.
Paternoster, R., Brame, R., Mazerolle, P., & Piquero, A. (1998). Using the correct sta-
tistical test for the equality of regression coefficients. Criminology, 36, 859-866.
Rabe-Hesketh, S., & Skrondal, A. (2008). Multilevel and longitudinal modeling using
STATA. College Station, TX: Stata.
Shermer, L., Bierie, D., & Stark, A. (2013). Endogeneity in prison risk classification.
International Journal of Offender Therapy and Comparative Criminology, 57,
1248-1274.
Steffensmeier, D., Ulmer, J., & Kramer, J. (2006). The interaction of race, gender,
and age in criminal sentencing: The punishment of being young, Black, and male.
Criminology, 36, 763-798.
Sweeten, G. (2012). Scaling criminal offending. Journal of Quantitative Criminology,
28, 533-557.
Warr, M., & Stafford, M. (1991). The influence of delinquent peers: What they think
or what they do? Criminology, 29, 851-866.
Ziliak, S., & McCloskey, D. (2011). The cult of statistical significance: How the stan-
dard error costs us jobs, justice, and lives. Ann Arbor: University of Michigan
Press.

Author Biography
David M. Bierie received his Ph.D. in criminology from the University of Maryland,
College Park, and has published widely on corrections, policing, courts, and quantita-
tive methodology. He has worked as a research scientist within the Federal Bureau of
Prisons, the U.S. Marshals Service, and is currently the Research Coordinator for the
Division of Innovation and Research, U.S. Department of Justice (DOJ)/Office of
Juvenile Justice and Delinquency Prevention (OJJDP). He is a past member of the
Research Advisory Committee of the International Association of Chiefs of Police
and is currently a member of the editorial board of Criminology: An Interdisciplinary
Journal.

You might also like