Abstract
The police must on occasion use physical force and weapons in order to apprehend and control subjects and fulfil the police function. It is inevitable that some of these interactions will result in injuries to both subjects and officers, with a range of both tangible and intangible harms and costs. It is therefore important to study injuries related to the use of force with an eye toward identifying opportunities to minimize injury and reduce the harms and costs. Injuries to both subjects and officers were examined in a sample of more than 10,000 use of force incidents drawn from 81 agencies located in 8 states. In addition to describing injury rates across a broad spectrum of situational and agency characteristics, we present multilevel logistic regression models predicting subject and officer injury. Among key findings, we report that the likelihood of injury for both subjects and officers is lower when force incidents end quickly and with the minimal necessary superior level of force relative to subject resistance, and higher for both subjects and officers when subjects flee. At the agency level, we find that the likelihood of injury varies by agency size and type. Finally, we explored possible higher-level variation and found that agencies in the sample from Midwestern states (primarily Wisconsin) have substantially lower injury rates that appear to be associated with their less frequent use of weapons and greater reliance on low-level physical force tactics, as compared to agencies in the sample from Western and other states.
The United States has entered an era of heightened attention on police behaviour, particularly with regard to the use of force. This was prompted in large part by several high-profile deaths of minority citizens during 2014 and 2015, growing citizen outrage most notably manifest in the #BlackLivesMatter movement, as well as additional controversial deaths of minority citizens. The President’s Task Force on 21st Century Policing was empanelled in late-2014, and the Task Force conducted meetings and listening sessions, followed with reports identifying six broad areas of needed reform (President’s Task Force, 2015). Since that time, the attention on police behaviour has arguably broadened from the use of force to a more general focus on both racial inequities in policing and a perceived lack of meaningful legal constraints on police behaviour. We are also witnessing renewed focus on an old problem, the lack of information (Fyfe, 2002; Hickman & Poore, 2016; Kane, 2007), and the growth of open-source data collection (The Guardian, 2016; The Washington Post, 2020) as well as new Federal data collection efforts (Federal Bureau of Investigation, 2018). In the scientific realm, epidemiologists and health policy scholars have increasingly focused on police use of force as a public health problem (e.g., Cooper et al., 2004; Obasogie & Newman, 2017). In the popular realm, YouTube videos depicting a broad range of negative police-public interactions are endemic (Brown, 2016).
As disturbing as these interactions may be, they underscore a reality of policing in the United States as it is presently structured and practiced. The core function of the police is to maintain order and enforce the law while preserving individual rights, but it is through the exercise of their coercive authority that they ultimately achieve those democratic goals. While the police perform many complex and important roles within the communities they serve, the single defining characteristic of the police is their capacity to both verbally and physically coerce individuals to do things that they are not otherwise inclined to do, particularly those individuals who are not obeying the rules (Bittner, 1970). Sometimes, the police must use physical force to protect citizens and themselves, or to apprehend criminal subjects, and we expect police officers to rely on their training and good judgment in doing so. Because the police must on occasion engage in these physical acts of coercion, which may include the use of weapons, it is somewhat inevitable that a subset of these interactions will result in physical injuries to subjects and officers.
These injuries have real costs (such as the costs associated with medical treatment, lost work, and municipal liability) but also less tangible collateral costs such as the erosion of police legitimacy and public trust, which is much harder to enumerate but arguably more significant in shaping long-term public perceptions than individual injuries alone. It is important to understand what factors are associated with injury especially if some of these factors may be amenable to training or policy interventions. We note that there is a lack of knowledge about regional variation in the use of force and both officer and subject injury. Understanding that law enforcement training and use of force practices are not uniform, we may suspect agency, state and regional differences that would complicate practical discussions about use of force. The goal of this study is threefold: (1) to describe injuries to subjects and officers in what we believe to be the largest and most diverse database of police administrative records concerning the use of force; (2) to examine key bivariate relationships between injury and subject, incident, and agency characteristics; and (3) to develop multivariate models predicting injury using these subject, incident, and agency characteristics. In the next section we provide a brief literature review relating to injury and police use of force.
Literature Review
Medical scholarship is an important source of research on injuries related to police use of force, particularly when examining detailed injury data that requires the interpretation of medical records. Collaborations between physicians and criminologists may be particularly fruitful when analysing police generated records and/or when merging police records with medical records (e.g., Bozeman et al., 2017; Strote & Hickman, 2018, 2020). The integration of medical and police records can help discern the nature of injuries attributable to police use of force versus other sources of injury. The absence of medical scholars from a project should not be taken to mean that the research is somehow ill-informed or invalid, but injury outcome measures are typically less detailed in these cases and are often derived from police reports rather than medical records, with greater variation in definitions of injury (Kaminski et al., 2015). On the other hand, this research also tends to focus more on characteristics of the events themselves, and is often more detailed with regard to pre-cursors of force than is typical of medical scholarship. Some of this research is focused at the national-level from an epidemiological or public health framework (e.g., Chang et al., 2016; Miller et al., 2017), while other studies conducted at the agency-level focus more on determining the situational factors and other correlates of injury to subjects and officers (e.g., Castillo et al., 2012), as well as specific concerns such as mental health status (Morabito & Socia, 2015; Rossler & Terrill, 2017). A sizeable body of work focuses specifically on Conducted Energy Devices (CEDs) and OC Spray, and their role in injury or injury reduction as well as their overall effectiveness (e.g., Brandl & Stroshine, 2017; Jenkinson et al., 2006; MacDonald et al., 2009; Paoline et al., 2012; Smith et al., 2007; Taylor & Woods, 2010; Terrill & Paoline, 2012).
National-level research has provided evidence on injury trends, primarily derived from Centers for Disease Control (CDC) data. For example, Chang et al. (2016) examined national data for the period 2003–2011, and noted that while the Federal Bureau of Investigation’s (FBI) Law Enforcement Officers Killed and Assaulted (LEOKA) statistics showed a decline in assaults on police officers during that time period, CDC data showed an increase in hospital admissions and nonfatal injuries to citizens related to police action. Miller et al. (2017) examined 2012 CDC data, as well as data from the Healthcare Cost and Utilization Project, and both the Guardian and Washington Post data collections. They estimated that the police killed or injured about 55,400 people in 2012, with the vast majority (51,678, or 93%) being cases where subjects were treated in the Emergency Department (ED) and released. The total injury estimate includes 1,063 fatalities (or 2%) and 2,665 hospital admissions (or 5%). Miller and colleagues (2017) used FBI arrest data in tandem with street- and traffic-stop data from the Police-Public Contact Survey in order to estimate rates of injury (the resulting rates of injury were one in every 291 stops or arrests, or 34 per 10,000 stops and arrests). While rates varied by gender (with substantially higher rates for males), and rates increased with age, rates did not differ significantly by race/ethnicity (the authors clarify that while Blacks clearly have higher rates of both stops and arrests, they are no more likely to be injured or killed during those stops and arrests). Medical costs were estimated at $2,390 for patients treated in the ED and released, $36,550 for hospital admissions, and $9,550 for fatalities. The total costs attributable to injuries or fatalities from police interventions for 2012 were estimated to be $231 million.
Research at the agency-level typically benefits from greater detail on incident-related variables and can greatly benefit from linkages to medical records. For example, Strote et al. (2010) documented injuries in ED records for use of force incidents recorded by the Seattle Police Department during a one-year period. Among 888 subjects, 187 (21%) were seen in the ED within 24 hours of the incident. Of the subjects seen in the ED, 29 (15.5%) were admitted, with more than half of them for psychological evaluations. There were 12 trauma admissions which arguably represent more serious injuries, and there were two fatal injuries (dead on scene), but these more serious and fatal injuries are clearly rare in these data. The 158 subjects (84.5%) who were treated and released were diagnosed predominately with abrasion or contusion (32%), acute toxic encephalopathy (29%), and laceration (28.5%). Descriptive information on use of force events was provided (e.g., number of officers, time of day, incident type, and type of force) but the authors did not analyse their relationship with subject injury.
In related research, Strote and Hickman (2018, 2020) examined citizen complaints about police use of force, focusing on incidents where citizens were seen in the ED within seven days of the incident, and compared cases in which citizens filed complaints about the use of force to a control group where citizens did not file complaints. About one-third of the complaint cases did not involve a reportable level of force by police or had no incident report associated with them. Overall, major injuries were infrequent and Injury Severity Scores were low, with no significant differences across groups on any of the medical variables. Ultimately, the findings suggested that ED records may suffer from poor documentation about the use of force, and that citizen complaints about police use of force may not be a reliable indicator of the extent of injuries resulting from police intervention.
In a similar vein, Bozeman et al. (2017) looked at all use of force incidents within three mid-sized police departments over a two-year period, and reviewed the severity of injuries across force modalities. Injuries were classified by physicians as mild (such as minor contusions, lacerations, and abrasions), moderate (such as long bone fracture, hemo/pneumothorax, liver/spleen laceration), or severe (such as severe head injury, need for life saving surgery, loss of limb or eye, ventricular dysrhythmias) using predetermined criteria. Among 914 subjects, 61 percent sustained no injury while 39 percent sustained some level of injury, and almost all of these injuries (95%) were classified as mild. Just 12 of the 914 subjects sustained moderate injuries, and four sustained severe injuries. Not surprisingly, moderate and severe injuries were more likely when firearms and canines were used by police, which was relatively rare in comparison to unarmed physical force and the use of conducted energy weapons.
In sum, this research demonstrates that the majority of injuries resulting from police use of force can be medically characterized as minor. While this conclusion is not intended in any way to minimize more serious injuries or the loss of life, it is fair to characterize serious physical injury and fatalities resulting from police action as relatively rare phenomena (we also do not intend to minimize the potential mental health effects – see Bor et al., 2018, but also see Nix & Lozada, 2019). There is much to learn from, and about, ED records but they are arguably the most useful source of information about the extent of injuries. Police reports are typically useful for an indication that injury occurred, but less so for the severity or extent of injury, and citizen complaints about police use of force are not a useful proxy for injury occurrence or severity.
There is a sizable body of work at the agency-level examining the predictors of injury. Overall injury rates reported in the literature vary across agencies and may in part be attributable to variation in definitions of injury (such as inclusion/exclusion of Taser probe and chemical irritation) as well as data quality, but can be considered common at about 40 percent for subjects and in the range of 10-50 percent for officers (Stroshine & Brandl, 2019). In general there are higher rates of injury for male subjects and some studies showing higher rates for white subjects (Castillo et al., 2012; MacDonald et al., 2009; Morabito & Socia, 2015; Rossler & Terrill, 2017; Smith et al., 2007, 2019). Higher subject resistance and officer force levels are also associated with higher injury likelihoods, and from the force factor approach, lower force relative to subject resistance is associated with a greater likelihood of officer injury (Hine et al., 2018; Wolf et al., 2008, 2009). 1 Results for impairments are more mixed, including alcohol and drug use, and mental health status (Morabito & Socia, 2015; Rossler & Terrill, 2017). For example, Rossler and Terrill (2017) examined 4,254 incidents drawn from three agencies over a two-year period, and reported an overall 26 percent injury rate, with a higher injury rate for persons with mental illness (33%). However, this relationship did not hold in multivariate models where subject resistance levels, males, whites, age, subject weapon use, and officer use of CEDs and impact weapons were associated with higher likelihoods of injury.
Research also generally demonstrates that the use of less-lethal weapons such as Tasers and OC spray can result in lower injury rates with some qualifications and exceptions (Jenkinson et al., 2006; MacDonald et al., 2009; Paoline et al., 2012; Smith et al., 2007, 2019; Taylor & Woods, 2010; Terrill & Paoline, 2012), and a fair degree of variability in injury rates is associated with other specific weapons and tactics, such as use of canines (e.g., Smith et al., 2019). Duration of the event has also been associated with increased rates of injury, where duration has been measured in terms of the need for more than one application of force (Castillo et al., 2012) and up to three iterations of force-resistance exchange (Wolf et al., 2008).
Agency-level variation is an important but neglected (and admittedly challenging) area of research on use of force injuries. MacDonald et al. (2009) examined records provided by 12 police departments, including about 24,000 records across varying time spans. In bivariate analyses, the use of OC spray and CEDs were associated with lower rates of subject injury, as were departmental policies restricting their use to situations where subjects exhibited defensive or greater resistance. Multi-level logistic models demonstrated that the individual-level relationships held (such as the effects of gender, race, force, and resistance), but no effect of restrictive OC or CED policies on the likelihood of subject injury.
Building off MacDonald et al.’s (2009) multi-agency analysis, what is needed, and what we intend to present here, is a more comprehensive examination of predictors of subject and officer injury, drawing information from a larger and more diverse sample of agencies. A larger and more diverse sample of agencies will increase our ability to generalize the findings, as well as our ability to examine predictors of both incident and agency-level variation in the likelihood of injury. In the next section, we discuss the research goals and hypotheses.
Research Goals and Hypotheses
A key goal of this study is to identify and model variation in the likelihood of both subject and officer injury at the agency-level. That is, while there are a number of incident-level predictors of injury known from the research literature, we need also to understand variation across agencies. We discuss here our hypotheses at both the incident- and agency-level.
With regard to subject demographics (gender, race, age) we anticipate that when subjects are female, both the subjects and officers will have lower likelihoods of injury. Subject race is less clear, but the available literature suggests that White subjects may be anticipated to have higher likelihoods of injury. We expect opposite effects of age on subject and officer injury, such that as age increases the likelihood of injury to subjects will increase, but the likelihood of injury to officers will decrease.
The role of mental illness in use of force and injury to both subjects and officers is not entirely clear (the research is limited and somewhat mixed), but we would expect injury to be more likely for both subjects and officers where possible mental illness was indicated prior to the use of force. Likewise, the role of drug/alcohol use is not clear, but we anticipate that where there was an indication that the subject was under the influence of drugs and/or alcohol prior to the use of force, the likelihood of injury would be higher for both subjects and officers.
Resistance and force levels are known to be related to subject and officer injury, as is the relative level of force and resistance (i.e., force factor). Specifically, we anticipate that as force factors increase, the likelihood of injury to subjects will increase while the likelihood of injury to officers will decrease. The type of force, whether physical, weapons, or both, is also an important explanatory variable, although the effect may be contingent on injury definitions. We anticipate that the likelihood of subject injury will be higher when weapons are used, but lower for officers. We also anticipate that greater numbers of officers on scene will be related to both subject and officer injury.
The duration of the incident is also known to be related to injury, and we anticipate that as the number of force sequences (iterations of force-resistance exchange) increases, the likelihood of injury to both subjects and officers will also increase. We also include several incident characteristics related to offense severity, the threat posed to officers/self/others, subject resistance, and flight. We anticipate that officer knowledge prior to the use of force (e.g., from dispatchers, witnesses, officer observations, or subject admissions) of more serious underlying offenses, greater threats to officers/self/others, higher levels of subject resistance, warrants, and both attempted and actual flight, will result in higher likelihoods of injury to both subjects and officers.
With regard to the agency-level, administrative data collections such as the Law Enforcement Management and Administrative Statistics (LEMAS) program administered by the Bureau of Justice Statistics, as well as many decades of policing research, have long documented the variability in police data by agency size (both in terms of number of officers as well as the size of jurisdiction) and type (e.g., municipal police, Sheriff’s offices, etc.). While agency size and type are not true organizational measures in themselves (e.g., Langworthy, 1986; Maguire, 1997) they are clearly meaningful stratification variables and it is important at this early stage of research to understand how the likelihood of force related injuries to subjects and officers might vary across these strata. We regard these as purely exploratory hypotheses.
An additional exploratory hypothesis concerns possible higher-level variation. In the course of the ongoing data collection that we report herein, we received anecdotal information from informal discussions and interviews with police trainers in different states indicating that agency variation in use of force and injury may be reflective of how law enforcement training is delivered in those states. For example, the idea that officers in Washington are trained to go to the Taser, while officers in California are trained to use strikes and batons. In some Wisconsin agencies, there have been revisions to academy and in-service training influenced by martial arts techniques that emphasize achieving control without the use of strikes, takedowns, or weapons, and that build greater confidence in the use of physical force (see Torres, 2018), which could potentially result in less injury. We offer these as examples of the variability in training practices generally that could account for observed state and/or regional differences in the use of force. Broader state-level variation in law enforcement training has been well documented (Reaves, 2016) so there is also some empirical grounding to support these more anecdotal hypotheses. To assess this rigorously would require a fairly substantial sampling and data collection strategy, but we believe we are in a reasonable position to begin exploration of this question and return to this in the methodology section.
Methodology
Data were compiled by Police Strategies LLC, a company that was built by law enforcement professionals, attorneys and academics with the primary goal of helping police departments use their own incident reports to make data-driven decisions and develop evidence-based best practices. The primary data systems include the Police Force Analysis System (PFAS) and Police Force Analysis Network (PFAN). PFAS is a relational database that contains 150 fields of information extracted from law enforcement agencies’ existing incident reports and officer narratives. PFAS uses data visualization software to display the information on dynamic dashboards, which can be used by police management to identify trends and patterns in use of force practices and detect high risk behaviour of individual officers. PFAN allows agencies to compare their use of force data with other agencies in the system. The data can also be used for research purposes, which is the present application. This research was approved by the lead author’s Institutional Review Board.
Trained coders read the police reports and code the content of those reports, including administrative items as well as report narratives, following the structure of a detailed coding manual and using data-entry windows facing a Microsoft Access database. All coders are trained with coding exercises using police reports from a variety of jurisdictions, and training cases increase in complexity as the training progresses. Assigned cases are treated as a series of tests designed to assess coding skill. Coders who successfully pass the training process are permitted to continue working on the PFAS. All coding is reviewed as part of an ongoing Quality Assurance Process in order to ensure complete, accurate and consistent data entry. Informational bulletins are sent to coders when issues are discovered that need improvement.
Agencies vary in terms of their data recording systems, with some using specialized forms, some using commercial solutions such as IAPro, and a few having their own internal systems. It is not necessary for an agency to have a specialized data collection system since incident reports and narratives are the primary source of the data. All agencies included in PFAS require all their officers who use force to complete a narrative detailing what occurred, but there is some variation among agencies in whether lower levels of force must be reported (such as grabbing or holding a subject, with no other force used). This can drive some agency-level variation in injury rates, with the exclusion of lower levels of force tending to lead to the observation of higher injury rates (Smith et al., 2007).
The core of the PFAS is force factor analysis (e.g., Alpert & Dunham, 1997, 2004; Hickman et al., 2015; Terrill, 2001, 2003, 2005), wherein the levels of officer force and subject resistance are recorded in dyadic exchanges throughout the incident, combined with legal criteria established in Graham v. Connor (1989). PFAS data can be used to help determine the risk that a use of force incident would be found to be unnecessary and excessive. Data elements captured in PFAS include: incident details (call type, temporal and geographic information); subject demographics; officer information and dispatch; subject impairments; offense category and severity level; active resistance; subject threat to officers, self, or others; flight; officer weapon use; officer tactics; force factors, both sequential and static; and injury.
The sample used in the present analysis consists of 10,564 subject-incidents, drawn from 81 law enforcement agencies across eight states. The sample is non-probability based but does incorporate a diverse range of agency types and sizes, jurisdiction sizes, and geographic variation. Similar to MacDonald et al. (2009), agency records in this study spanned varying time periods and include from one to six years of data, with most coming from the period 2014-2018.
The majority of the agencies included in this study are municipal police departments (61 agencies), followed by county sheriff’s offices (12), university police departments (5), and state law enforcement special jurisdiction agencies (3). In terms of size, most of the agencies fell into the category of 10 to 49 officers (54 agencies), followed by 50 to 99 officers (14), and those with 100 or more officers (13). Jurisdiction sizes ranged from less than 10,000 persons (22 agencies), 10,000-50,000 persons (40), 50,000-100,000 persons (9), to greater than 100,000 persons (9).
About two-thirds of the agencies are from Washington state (55 agencies), and another one-fifth are from Wisconsin (16). Five agencies are from California, and the remainder are from other states (we do not report those here in order to avoid possible statistical disclosure of individual agency data). Because of this, when exploring state variation we have pooled two Midwestern states with Wisconsin for analyses examining state variability, as will be seen in the descriptive analysis section. About 94 percent of the force records come from agencies in three of the eight states: Washington (45%), California (32%), and Wisconsin (17%).
Our analyses began with bivariate descriptive statistics examining injury rates to subjects and officers across all study variables. We then developed multivariate models predicting subject and officer injury. Analytic findings are reported below.
Results
Descriptive Statistics
Injury
Across all agencies and incidents, subject injuries resulting from the use of force (including physical force and weapons, alone or in combination, and including Taser probe and chemical irritation injuries) were indicated in just over half of all cases (5,471 incidents, or 52%), including both visible injuries (4,763 incidents, or 45%) and complaints only (708 incidents, or 7%). Officers were injured in about 16 percent of cases. These overall rates may appear high, but are consistent with recent studies by Smith et al. (2007), who reported subject and officer injury rates of 56 and 17 percent, respectively, in their Miami-Dade data; Stroshine and Brandl (2019) who reported injury rates of 50 and 20 percent, respectively, in their anonymous agency; and Smith et al. (2019), who reported rates of 54 and 12 percent, respectively, in their Tulsa, OK, data. Smith et al. (2007) point out that studies using police administrative records generally report higher rates than those using citizen or officer surveys, and the higher rates might be due to higher reporting thresholds that eliminate incidents where minor force is applied. The inclusion or exclusion of Taser probe injuries and chemical irritation due to use of OC spray can also account for variation in injury rates, although to varying degrees (Kaminski et al., 2015; Paoline et al., 2012; Stroshine & Brandl, 2019; Terrill & Paoline, 2012).
Detail on the specific type of injury was available for visible injuries recorded in all but one large agency. Consistent with prior research, subject injuries tended to be less serious with the most common injuries including bruises and scrapes (35%), cuts and punctures (25%), and Taser probe (19%). More serious injuries included broken bones or teeth (2.5%), gun/knife wounds (0.7%), and fatalities (0.7%) (it should be noted that one agency did not provide officer involved shooting cases). Injury locations included head (44%), torso (31%), hand/arm (30%), and feet/leg (21%) (multiple injury locations could be involved). Nearly half were treated at the scene and taken to a hospital (24%) or taken directly to a hospital (25%). About a third (36%) were treated at the scene, and 16 percent refused or were not offered treatment (Table 1).
Descriptive Statistics for Injury Outcomes and Detail.
Subject Demographics and Basic Incident Characteristics
Use of force subjects were predominately male (83%), and males have substantially higher injury rates as compared to females (55% versus 36%). In terms of race/ethnicity, 44 percent of subjects were White, 27 percent Black, 22 percent Hispanic, five percent Asian or Pacific Islander, two percent Native American, and the remainder were in other or undetermined race/ethnicity categories. In comparison to the overall injury rate of 52%, injury rates varied from a low of 47.5 percent for Black subjects to a high of 75 percent among subjects in the other category, with rates for other groups including Hispanic (57%), Asian/Pacific Islander (62%), Native American (64%), and White (50%) subjects. The average age was 32 years, ranging from seven to 98 years (standard deviation = 12 years) (Table 2).
Descriptive Statistics and Injury Rates for Subject Demographics.
Subject Impairments, Behaviours, and Incident Characteristics, Prior to the Use of Force
This represents information that the officer had prior to the use of force, from any source including dispatchers, witnesses, the officer’s observations, or an admission by a subject. Prior to the use of force, officers had some indication that the subject was under the influence of drugs and/or alcohol in about half of cases, with slightly higher injury rates for both subjects and officers. Possible mental illness was indicated in 22 percent of cases, with slightly higher injury rates for both subjects and officers. Injury rates to subjects were higher when officers had information that a warrant had been issued, although officer injury rates did not vary substantially in these cases (Table 3).
The underlying offense (meaning the crime the officer was investigating immediately before force was used) as indicated in the report was categorized on an ordinal scale based upon approximate relative severity. This ranged in terms of severity from “none” (14% of cases) up to “violent offense with weapon” (5%). In general, subject injury rates increased with the ordinal offense severity categories from 35 to 66 percent. Officer injury rates were mixed with the lowest rates for either “none” or “violent offense with weapon” (13%) and the highest for “drug/trespass/disorderly” (18%).
The threat to officers, self, or others, as indicated in the report was categorized on an ordinal scale based upon approximate relative degree of threat. This ranged from “none” (55% of cases) up to “deadly weapon” (4%). Excepting the “none” category, subject injury rates increased with the ordinal threat categories from 39 percent (“verbal threat”) to 68 percent (“deadly weapon”). Officer injury rates were again mixed and were lowest for “deadly weapon” (11%) but highest for “assault or self-harm” and “less lethal weapon” (22% each).
Descriptive Statistics and Injury Rates for Officer Knowledge (Prior to the Use of Force) Regarding Subject Impairments, Behaviours, and Incident Characteristics.
Subject resistance as indicated in the report was categorized on an ordinal scale based upon approximate relative degree of resistance. This ranged from “none” (0.5% of cases) up to “deadly weapon” (2% of cases). In general, subject injury rates increased with the ordinal resistance categories from 37 percent (“none”) to 75 percent (“deadly weapon”); an exception to this general increase was the “physical non-compliance” category. Officer injury rates were relatively low in all but the top three categories of resistance, with the highest in the “active physical” (30.5%) and “less lethal weapon” (34%) categories.
Subject flight from officers as indicated in the report resulted in higher injury rates, whether attempted (49%) or actual (59%) (attempted flight means the subject tries to leave the scene and is caught before they run or drive away). Officer injury rates were also elevated in these cases (19% and 17%, respectively).
Force Sequences and Force Factor
Up to six dyadic “back-and-forth” force exchanges were recorded for each incident, which may be taken as a proxy indicator of the duration of the force event. The median number of force sequences was three (mean = 3.4, standard deviation = 1.4). Both subject and officer injury rates increased with the number of force sequences, from 44 percent (1 sequence) to 69.5 percent (6 sequences) for subjects, and from five percent (1 sequence) to 35 percent (6 sequences) for officers. The greater the number of force sequences, the more likely injury is to both subjects and officers (Table 4; Figure 1).
Injury Rates by Number of Force Sequences.
Force factors were computed by subtracting subject resistance levels from officer force levels using parallel ordinal scales, both overall and sequentially (i.e., at each force sequence). Looking at the maximum sequential force factor recorded across all incidents, subject injury rates were lowest (36%) when officer force was directly proportional to subject resistance (i.e., a force factor equal to zero). Subject injury rates increased with positive force factors, up to a 77 percent injury rate for force factors of four or greater; conversely, officer injury rates were lowest when force factors were high. Interestingly, both subject and officer injury rates were elevated when force factors were negative (36% and 19%, respectively), as compared to when force factors were proportional. In sum, as the force factor increases from zero, injury to subjects is more likely and injury to officers is less likely, while injury to both is slightly elevated when force factors are negative (i.e., where the officer is using less force relative to subject resistance) (Figure 2).
Injury Rates by Maximum Sequential Force Factor.
Descriptive Statistics and Injury Rates for Number of Force Sequences, Maximum Force Factor, and Specific Types of Force.
Officer Tactics and Weapons Use
Two-thirds of the force incidents involved physical force only, while 22 percent involved both physical force and the use of a weapon, and the remaining 12 percent involved a weapon only. Physical only force events had the lowest injury rate for subjects (34%), while incidents involving a weapon had overall rates of 74 percent (both physical and weapon) and 78 percent (weapon only). Officer injury rates were lowest (3%) when using a weapon only, followed by physical force only (15%) and incidents involving both physical force and a weapon (25%) (Table 4).
Specific tactics used by officers were recorded, including grabbing, takedowns, use of weight to hold subjects down, pushing, pain compliance holds, physical strikes, wrestling with subjects, and lateral neck restraints (LNR). The median number of officers was equal to two, ranging from one to 18, and the total number of tactics used in force incidents was three, ranging from one to 33. The median number of different types of tactics used was three, ranging from one to 10. Subject injury rates were highest for physical strikes (63%) and LNR (62%), followed by wrestling (56.5%). These three categories of tactics were also associated with the highest injury rates for officers (29% to 33%).
As might be expected, the use of weapons by officers is clearly connected to higher subject injury rates: any use of a weapon saw subject injury rates of 76 percent, compared to 34 percent when weapons were not used. Subject injury rates ranged from 71 percent for impact weapons such as batons, to 97.5 percent when canines are employed and bite (i.e., not used for tracking and barking only). Officer injury rates were highest in incidents where impact weapons were used (29%) and lowest where canines were used (3%).
Agency Characteristics
Injury rates for both subjects and officers varied by agency size (number of officers), jurisdiction size (population served), type of agency, and, interestingly, geographical region. Subject injury rates ranged from 28 percent among agencies having 10 to 49 officers, to 61 percent among those having 100 or more officers. Officer injury rates were highest in agencies with 100 or more officers (19%) (Table 5).
Jurisdiction size (population served) and number of officers tend to be correlated, so it is perhaps not surprising that a similar pattern emerges for injury rates by jurisdiction size. Injury rates are higher for both subjects and officers in larger jurisdictions as compared with smaller jurisdictions.
By type of agency, county sheriff offices had the highest rate of both subject and officer injury (67% and 20%, respectively), followed by municipal police departments (52% and 16%). University police departments and special jurisdiction agencies had relatively lower injury rates.
As previously noted in the methodology section, the majority of the agencies in the sample (and the majority of the use of force records) come from three states: WA, WI, and CA. While we do not have the data structure to be able to model state-level variation, nor to include a true regional control in the sense of being able to examine all regions of the Country, we can at least examine variation in this limited set of agencies. Injury rates are substantially lower in the Midwestern agencies as compared to the other five states represented in the sample. Specifically, among 18 Midwestern agencies in 3 states, injury rates for subjects and officers are 20.5 and 10.5 percent, respectively, compared to 59 and 17 percent among 63 non-Midwestern agencies.
Descriptive Statistics and Injury Rates by Agency Characteristics and Region.
Multivariate Models
Given the binary and nested nature of our dependent variables (incidents within agencies) and our research questions, we decided to use multilevel logistic regression. This method allows variation in the intercept (or average levels of the dependent variable, across agencies) to be properly modelled. We used HLM 7.01 (Scientific Software International, Inc., 2013) and specified a two-level Bernoulli distributed outcome variable, which applies a logit link function for a binary outcome.
We began with the subject injury model. In terms of process, an initial unconditional model confirmed statistically significant level-2 (agency) variation in the intercept (variance component = .950; χ2(80) = 2448.762, p < .001). The Intraclass Correlation Coefficient (ICC) was equal to .224, indicating that 22% of the variance is at the agency level, and 78% at the incident level. We then built the level-1 model, with all level-1 variables centered around their group-means, and confirmed that significant level-2 variation remained. Finally, we built the full model including level-2 covariates, all of which were centered around their grand-means.
2
List-wise deletion of cases with missing data reduced the available analytic sample to 8,866 incidents (or 84%) within 76 agencies (or 94%). The level-1 and level-2 models were specified as follows, where Prob(SUBJECTINJURYij = 1|βj) = ϕij, and log[ϕij/(1–ϕij)] = η
ij
:
With regard to demographic variables, the results indicate that when the subject is female there is a 26 percent decrease in the odds of injury; when the subject is black, there is a 17 percent decrease in the odds of injury; and each additional year of age increases the odds of injury by 0.8 percent (Table 6).
Certain offense categories had higher odds of subject injury relative to those cases where officers had no knowledge of the offense prior to using force: Violent with weapon, 63 percent greater odds; property/warrant, 35 percent greater odds; and traffic/liquor/infraction, 24 percent greater odds.
While none of the threat variables were significantly related to subject injury, resistance with a deadly weapon increased the odds of subject injury by 188 percent and active physical resistance increased the odds of injury by 132 percent. Actual flight from officers increased the odds of subject injury by 22 percent. Force incidents in which officers used a weapon only had 342 percent greater odds of subject injury as compared to incidents involving only physical force; the use of both physical force and a weapon increased the odds of subject injury by 181 percent.
Longer force incidents, as indicated by the number of dyadic force sequences, resulted in greater odds of injury to subjects. Each additional sequence increases the odds of subject injury by 14 percent. This suggests that ending force incidents as quickly as possible may minimize injury to subjects. However, this must be balanced with the degree of force used in order to gain control of subjects. The higher the maximum sequential force factor (i.e., the greater the level of force used, relative to the level of subject resistance), the greater the odds of subject injury. Each additional force factor step increases the odds of subject injury by 21 percent. This would suggest that ending force incidents as quickly as possible, but with the minimum superior level of force necessary, would minimize the likelihood of injury to subjects.
Subject injury was more likely on average in larger agencies (those with 100 or more officers) as compared to small agencies; when the agency is large, the odds of subject injury are increased 100 percent. Finally, Midwestern agencies (driven primarily by Wisconsin agencies) as compared to agencies from other regions in the data, have a substantially lower average likelihood of subject injury. When the agency is Midwestern, the odds of subject injury are decreased 74 percent.
The variance component for the full model is equal to .688 (χ2[69] = 784.621, p < .001), indicating that the remaining unexplained level-2 (agency) variance is significant. The model explains 26% of the total variance in subject injury; of the remaining unexplained variance, 13% is at level-2 and 61% is at level-1.
Multilevel Logistic Regression Models Predicting Injury to Subjects and Officers.
p ≤ .05, **p ≤ .01, ***p ≤ .001 (two-tailed tests).
We next moved to the officer injury model. An initial unconditional model confirmed statistically significant level-2 (agency) variation in the intercept (variance component = .216; χ2(80) = 474.845, p < .001). The ICC was equal to .062, indicating that 6% of the variance is at the agency level, and 94% at the incident level. We followed the same process as for the subject injury model in terms of variable centering and model building. The level-1 and level-2 models were specified identical to the subject injury model, with the officer injury variable as the outcome.
With regard to demographic variables, the results indicate that when the subject is female there is a 23 percent decrease in the odds of officer injury; when the subject is in the other race category (which includes Asian/Pacific Islander and Native American) there is a 24 percent decrease in the odds of officer injury; and the odds of officer injury decrease by 0.9 percent with each additional year of subject age.
Drug/Trespass/Disorderly offenses had an increased likelihood of officer injury, with a 30 percent increase in the odds. None of the variables related to threat or resistance were significantly related to the likelihood of officer injury. However, both attempted and actual subject flight from officers increased the odds of officer injury by 26 and 47 percent, respectively. Force incidents in which officers used a weapon only had 78 percent lower odds of officer injury as compared to incidents involving only physical force. When officers had an indication that subjects were under the influence of drugs or alcohol, the odds of officer injury decreased 19 percent.
Similar to the finding for subject injury, longer force incidents (as indicated by the number of dyadic force sequences) resulted in greater odds of injury to officers. Each additional sequence increases the odds of officer injury by 35 percent, suggesting that ending force incidents as quickly as possible may minimize injury to officers as well as subjects. As mentioned previously, this must be balanced with the degree of force used in order to gain control of subjects, so it is perhaps unsurprising that the higher the maximum sequential force factor, the lower the odds of officer injury while increasing the odds of subject injury. Each additional force factor step decreases the odds of officer injury by 11 percent.
Officer injury was less likely on average in mid-sized agencies (those with 50-99 officers) as compared to small agencies; when the agency is mid-sized, the odds of officer injury are decreased 42 percent. Sheriff’s offices had higher average likelihoods of officer injury relative to “other” jurisdiction types (a nearly 500% increase in odds). Finally, Midwestern agencies (driven primarily by Wisconsin agencies) as compared to agencies from other states in the sample, have a substantially lower average likelihood of officer injury. When the agency is Midwestern, the odds of officer injury are decreased 36 percent.
The variance component for the full model is equal to .193 (χ2[69] = 180.853, p < .001), indicating that the remaining unexplained level-2 (agency) variance is significant. The model explains 29% of the total variance in officer injury; of the remaining unexplained variance, 4% is at level-2 and 67% is at level-1.
Discussion
This study examined physical injuries to subjects and officers resulting from use of force incidents in a large sample of incidents and agencies and included a broad range of relevant explanatory and control variables. The findings demonstrate that there are many similarities across the predictors. Female subjects are less likely to be injured, as are the officers in these incidents, as compared to incidents involving male subjects. Black subjects were less likely to be injured as compared to whites, and officers were less likely to be injured when subjects were in “other” race categories as compared to white. Older subjects are more likely to be injured, while presenting decreased likelihood of injury to officers. Subject flight from officers increases the likelihood of injury to both subjects and officers. Longer incidents, meaning a greater number of back-and-forth force/resistance iterations, increase the likelihood of injury to both subjects and officers. Higher levels of force relative to subject resistance increase the likelihood of injury to subjects but decrease the likelihood of injury to officers. Resistance with a deadly weapon greatly increases the likelihood of subject injury, as does officer use of any weapon type. Finally, there is some evidence of a possible state and/or regional effect in these data, net of other agency-level controls and individual-level predictors.
The finding that longer force interactions lead to higher injury rates for both parties is consistent with and extends prior research (Castillo et al., 2012; Wolf et al., 2008). This provides empirical evidence that supports what police trainers already know from their experience, and reflects the training that is presently delivered in many law enforcement academies: In order to minimize injury resulting from use of force, end it quickly with the minimum necessary superior level of force. This often manifests in guidance to rely on takedowns, where appropriate. The finding that increasingly disproportional force factors (where officers are using increasingly higher levels of force relative to subject resistance) lead to higher injury rates for subjects but lower rates for officers is consistent with prior research (Hine et al., 2018; Wolf et al., 2008, 2009) and likely reflects the utility of less-lethal weapons, a similar conclusion drawn in prior research (Castillo et al., 2012). Determining whether and at what point the trade-off between subject injury and officer injury can be “optimized” would likely require additional data and some consideration of societal goals. In the case of negative force factors, it may be the case that the higher injury rates reflect the officer being overwhelmed by the subject and struggling to gain control. Since the officer is not in control, they may not use force very effectively and will likely take longer, leading to subject injury.
The lower injury rates at the state or possibly regional level is an interesting finding. We caution that this is purely exploratory and should be regarded as speculative. Our goal here is to prompt further investigation. We believe our study provides sufficient evidence to merit further investigation using a more rigorous methodological design. Exploring this a little further, we focus on Wisconsin agencies since they are the majority of the Midwestern agencies in the sample, and we find injury rates for subjects and officers are 21 and 11 percent respectively, much lower than the overall base rates of 52 and 16 percent for the entire sample. Part of this is potentially attributable to less reliance on weapons, and greater reliance on low-level physical tactics, in these Wisconsin agencies. To illustrate this possibility, Figure 3 shows injury rates (panel ‘a’) as well as variation in weapons (panel ‘b’) and physical tactics (panels ‘c’ and ‘d’) for agencies in Wisconsin, California, and Washington states. As can be seen, the California agencies do appear to have greater reliance on physical strikes and batons; Washington agencies do appear to rely heavily on the Taser; and Wisconsin agencies rarely use weapons.
Injury Rates, Weapon Use, and Physical Tactics in California, Washington, and Wisconsin.
Officers used a weapon in just 14 percent of use of force incidents in the Wisconsin agencies, as compared to 36 percent in California and 42 percent in Washington. California agencies in the sample used impact weapons (such as a baton) to a much greater extent (15% of incidents) while Washington agencies relied heavily on the Taser (29% of incidents). In terms of physical tactics, California agencies used strikes (e.g., closed-fisted punch) to a greater extent (28% of incidents), as well as takedowns (59%) and use of weight to hold subjects down (43%). Pushing subjects was more common in Wisconsin (33% of incidents).
The findings reported above could reflect law enforcement training in those states. While we cannot make a direct empirical connection between training and the lower injury rates observed for these agencies, and we again caution that these findings are exploratory and speculative, we offer this as some initial support for examining the variability in training practices generally that could account for observed state or regional differences in the use of force.
More broadly, it is important to ask the question of why use of force training should vary across the states. The most recent available national data on law enforcement training academies (Reaves, 2016) shows great variation in how basic law enforcement training (BLET) is delivered across the states. In some states a centralized academy delivers BLET, while in other states training may occur at the county or regional level, through community colleges or technical schools, or at the individual agency level. While state Peace Officer Standards and Training (POST) commissions generally set the minimum hiring and training standards within states, and an organization of POST Directors (the International Association of Directors of Law Enforcement Standards and Training) seeks to advance the development of professional standards, the decentralization of training delivery can lead to substantial variation both within and across states. Across all training academies in the United States, the average number of hours of training in the areas of weapons use, defensive tactics, and use of force included 71 hours on firearms skills, 60 hours on defensive tactics, 21 hours on use of force, and 16 hours on non-lethal weapons (among academies that provide such training) (Reaves, 2016). About three-quarters (77%) of recruits trained in 2013 received instruction using a force continuum of some type (Reaves, 2016). These national-level statistics mask any detail in terms of the specific training content and delivery methods, and of course the hiring agencies may supplement BLET with additional training (particularly field training, and in-service training) conducted at the agency level.
Limitations
This study has a number of limitations that are important to consider in evaluating the analytic results. A primary limitation is the reliance on police reports as the only source of information about subject and officer physical injuries. While it would be nice to have medical records available to assist in the documentation and measurement of injury, acquiring those records for such a large sample of agencies across multiple states would clearly not be feasible. This study and its findings are contingent on the indication of injury in official police records. Most often this comes from officer narratives or from a supervisor’s review and interview of the subject. Officers are trained to document their actions and observations in police reports and will typically describe the type and location of injury, complaint of injury, and who treated the subject. Injury is sometimes reduced to a check-box indicating whether any injury was visible to the officer, whether the subject complained of any injury, or both, as well as whether treatment was delivered at the scene or by transport to a hospital. We must also acknowledge the possibility that officers may not accurately recall events, deliberately mislead, or fail to document all incidents.
A related concern is possible variation in the interpretation and documentation of injury both within and across police reports. There is often no policy or guidance on the threshold of what may or may not constitute an injury, leaving the determination entirely to officer discretion. While we believe that the determination of an injury based upon officer observations and subject statements is a reasonable general indicator of injury, in the absence of specific criteria one could reasonably expect to see some variation in how officers make the determination for a particular case.
Our use of “any injury” may be misleading in that it could hide associations with severe injury and highlight associations with minor injury. It may be excessively broad, including Taser probe and chemical irritation injuries. While consistent with much prior research, such a measure does not address the severity of injury and as such our outcomes include the full range of injury (the majority of which are minor). We also cannot link specific tactics to specific injuries. As previously noted, we believe that our police records do not provide sufficient detail to accurately measure injury severity or causes and while medical records would be an ideal supplement, it was simply not feasible to obtain these data for the large number of agencies and multiple states included in this study.
It should be noted that the force factor method has a number of limitations, including the reliance on police reports mentioned above as well as concerns associated with the process of content analysis and systematic coding. There is also reasonable concern about the underlying level of measurement. Relevant research examining the reliability of the force factor method has demonstrated an overall acceptable degree of reliability, although there are some tendencies for interrater agreement on force and resistance levels to decline slightly as the number of iterations increases beyond three (see generally Hickman et al., 2015; Stewart, 2013).
Finally, as noted in the Methodology section the data were collected and processed by a private company, and the sample is non-probability based with data covering varying time periods. The agency data were accumulated variously through public records requests, grant-supported projects, and/or contractual engagements with individual agencies. While it is reasonable to be concerned about bias, the sample does include a wide variety of agency types (municipal police departments, sheriff’s offices, university police, state law enforcement agencies), agency sizes (both in terms of number of personnel and size of jurisdiction), and geographic variation. At present, we believe it to be the largest and most diverse database of police administrative records concerning the use of force, and view the potential utility as outweighing the bias concerns. More broadly, we believe it also emphasizes the need for more representative national data collection on the use of force.
Conclusion
Ultimately, use of force will to some degree reflect the “local standards” of the agencies in which the officers work. But the goals are always the same regardless of the state or locality: to apprehend or control a subject, and/or to protect life. Presuming we can know a “best way” or “better way” to do it (with outcomes such as injury potential being part of the evaluation), why would we not want every officer in the United States to do it that way? While we cannot answer this question here, we encourage researchers interested in the use of force generally, or injury specifically, to work on the systematic evaluation of training methods (e.g., Wood et al., 2020) that may lead to better use of force outcomes. We also encourage continued work on multilevel models, particularly broadening the range of agency-level variables as well as incorporating additional spatial covariates that might be captured at incident locations.
Footnotes
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This project was funded in part by a grant from The Joyce Foundation (Grant #18-38996).
