Study sample

Data are from The Counted and the Law Enforcement Management and Administrative Statistics (LEMAS) databases [69, 70]. The Counted, compiled by The Guardian newspaper, is a public-use database of the total number of people killed by US law enforcement during 2015–2016 (N = 2,238). The Counted database was selected because it includes all police agencies involved in each civilian death as well as each decedent’s age, race, gender, armed status, and location and cause of death. Reporters validate data through open source reporting, police records, witness accounts, and family/friends. The database is promptly maintained and updated by The Guardian reporters and interactive journalists by collecting tips from the public, crowdsourcing, and data from the Federal Bureau of Investigation. Like many crowdsourced databases of police killings [71, 72], The Guardian’s database has been shown to have greater validity and reliability than government databases due to government agencies underreporting or misclassifying police-involved deaths [71, 73, 74].

LEMAS is a widely-used cross-section of information collected approximately every three years from a nationally representative sample of US law enforcement agencies regarding organizational characteristics (e.g., administrative policy, personnel characteristics, officer requirements) [13, 17, 18, 24, 57, 58, 75]. To ensure policies were implemented prior to each killing, the 2013 LEMAS database was used to match 2,826 agencies to 2015–2016 fatalities in The Counted according to the agency name(s) reportedly involved and the state in which the killing occurred. Because the unit of analysis was each agency involved, multiple agencies could be linked to any given fatality. The matching procedure resulted in a final sample of 481 agencies (17% of LEMAS) involved in 1,085 civilian deaths (49% of The Counted). Bivariate analyses of the matched and unmatched groups were assessed for variation in civilian characteristics (S1 Table) and policy implementation (S2 Table). We exclude the small number of state agencies (N = 30) because (1) state agencies and jurisdictions are very different than municipal and county agencies in demographic factors and their primary role in enforcing the law within their entire jurisdiction (e.g., state agencies are the main source of law enforcement in very small proportions of their jurisdictions) and (2) state agencies have different policies, training, and resources that may bias associations with fatality rates (e.g., community evaluation is less likely to exist within state agencies). This method resulted in 323 municipal agencies and 128 county agencies for analysis (N = 451).

Exposure assessment

Thirteen LEMAS-reported policies expected to reduce lethal intervention were evaluated as exposures. Each policy aims to enhance organizational and officer performance or indirectly promote performance. Therefore, we expected each policy to help reduce avoidable killings and hypothesized lower predicted fatality rates in agencies with versus without these policies, especially initiatives specifically targeting reductions in officer use of deadly force and racial disparities (e.g., less-lethal weapons use, community policing). Data were coded such that a higher value indicates that the policy was reportedly implemented. Based on prior empirical and theoretical evidence, we evaluated policies related to agency climate and accountability, officer requirements, and personnel incentives that may mechanistically influence police killings. Each policy is described in more detail below: the proportion of full-time personnel [17], union representation (offered and active) [17], an established mission statement [8, 65], minimum educational requirements [20, 21, 42], additional training for new hires [11, 22], community policing training [57], less-lethal weapon use [62], video use [59, 76, 77], evaluation of citizen satisfaction with police [78–80], incentive compensation [21, 81, 82], and statistical analysis or research (any and external) [22].

The following variables were assessed dichotomously (No = 0; Yes = 1): Collective bargaining ‘offered’ among sworn personnel; amid the subset of those that offered bargaining, ‘active’ collective bargaining; having an established mission statement; at least 8 hours of training community policing training offered; additional training requirements for new-hires; community evaluation of resident satisfaction with agencies; agency accountability as any statistical analysis or research conducted; among the subsample reporting analysis, external statistical analysis (i.e., employing an organization outside of the agency); and some college required for new-hires.

The remaining policies were assessed continuously. The proportion of full-time sworn personnel (i.e., 0–100%) was measured among those full-time and part-time sworn and non-sworn. Summary scores were generated for less-lethal weapon use, video use, and incentive compensation. Offering personnel less-lethal weapon authorization was coded dichotomously across 9 types of weapons or actions (e.g., soft projectiles, neck restraints) with a summary score (range 0–9) representing more types of less-lethal weapons offered. Four video camera types (e.g., body, vehicle) were coded dichotomously: due to small cell sizes, the top two levels were combined resulting in a summary score (range 0 to 3) reflecting increasing video use types. Incentives included nine domains related to increased salary or additional paid leave (e.g., bi/multi-lingual, merit/performance), were coded dichotomously, and then summed (range 0–9) to represent increasing incentives offered.

Outcome assessment

Although measures of police–civilian encounters and civilian arrests have been used to evaluate racial disparities in police use of force, population-level measures are likely the most robust [10, 16, 23, 38, 83–86]. This is because police encounters are systemically biased against minorities [23, 38, 83–85]. A countervailing opinion is that not every member of the population is at risk of being a police fatality victim [38, 84, 85]. In fact, all Black civilians are at risk of a police encounter even while doing acts that would not subject Whites to the same risk (e.g. jogging in a predominantly White neighborhood). Therefore, conditioning estimates on police–civilian encounters or subsequent arrests using administrative records can produce biased results. This type of statistical bias can lead to underestimation of racial inequities: non-White rates have inflated denominators (more encounters) while White rates have deflated denominators (fewer encounters) thereby making racial groups more statistically similar [84]. Hence, incidence (i.e., fatality) rates were calculated using the total number of killings within an agency (numerator) divided by the total jurisdiction population size (denominator). Because racial inequities in police use of lethal force point to Black civilians being at the greatest risk [5, 12, 35, 39, 87, 88], agency policies found to be protective for Black civilians may benefit all minoritized racial groups. Therefore, we conducted separate analyses for agencies involved in all, White, Black, and non-White civilian killings. Race-specific rates were estimated using race-specific population size to avoid bias from conditioning on denominators that may over- or under-represent racial groups within a given jurisdiction. Apart from “Hispanics”, racial categories were non-Hispanic. The search tool Social Explorer was used to collect demographic data from the 2016 American Community Survey to measure the total and race-specific population distributions within a given jurisdiction [89].

Covariates

Empirically established and theoretical confounders were identified using a directed acyclic graph (DAG) (S1 Fig); civilian race and agency and jurisdiction-specific characteristics [17, 20, 21, 28, 57, 58, 68, 90]. Race categories included White, Black, and non-White (Black, Asian/Pacific Islander, Hispanic, Native American, Arab-American, Other/Unknown). Agency type included municipal and county agencies. Using LEMAS-established cut-points, number of officers was measured dichotomously (<100, ≥100). Consistent with the US Census Bureau [91], U.S. region was categorized as Midwest, Northeast, South, and West. All remaining covariates were evaluated continuously. Although the link between agency diversity and police killings is not well-established [68, 90], use of force has been shown to vary by officer race and gender [20, 92–94] suggesting that agency demographic composition may have implications for policy implementation; hence, we controlled for both racial and gender diversity. Gender diversity measured the number of female sworn personnel among all sworn personnel. Following previous work, racial diversity was measured as the proportion of full-time “White” sworn personnel [57]. Data from Social Explorer [89] measured percent native, educational attainment (≥HS & ≥BA/BS), poverty rate, owner occupied housing units, and married, as well as Gini index (perfect income equality to perfect income inequality [range 0–1]), whereas the median was used for age and income. The Uniform Crime Report measured violent crime rate—the average crimes reported per 100,000 people [95].

Statistical analysis

Analysis was restricted to agencies involved in one or more civilian deaths during 2015–2016. Missingness was negligible (3%), nonetheless we performed multiple imputation (MI) (m = 10) [96]: the MI assumption of data missing at random, although ultimately unverifiable, was supported by the similarity of MI and complete case results. Eighteen civilian deaths were dropped due to mismatches between agency name and state. One tribal agency with one Native American fatality was omitted due to inaccessible information on jurisdiction characteristics. One municipal agency outlier (i.e., exceptionally large jurisdiction size) was excluded. The final sample was comprised of 449 municipal and county agencies involved in at least one civilian killing during 2015–2016. Descriptive statistics reported for all agencies include all races. Because multiple agencies can be involved in one civilian killing, race-specific sample sizes will not sum to equal that of all agencies. Bivariate analyses (e.g., χ²) evaluated associations between covariates and each agency policy and each fatality group.

Model specification involved theory-based (e.g., DAGs) and data-driven approaches to assess confounding, collinearity, and best model fit. To address over-dispersion of the count outcome variable, negative binomial regression using robust standard errors estimated incidence rate ratios, adjusted for covariates either significant with two-sided P<0.05 or theoretical (i.e., “forced-in” regardless of statistical significance). The a priori minimum sample size required was calculated as 388 agencies (15 potential predictors, 1-β = 0.80, α = 0.05, and f² = 0.05) [97]; p-values were reported at P<0.10 to capture marginally significant estimates within the smaller race-specific sample sizes. Each policy was examined in a separate adjusted model. Sensitivity analyses involved adjusted models restricted to agencies with jurisdiction sizes of 35,000 people or more because this was about the minimum-sized county population whereas the maximum-size of municipal and county populations were similar. STATA 16 was used for all statistical analyses.

Sample characteristics

Table 1 shows the sample distribution of agency policies by those involved in all, White, Black, and non-White civilian killings. Table 2 shows the sample distribution of covariates. Overall, agencies involved in civilian killings were largely municipal (72%), had 100 or more officers (72%), were located in the South (46%), and had high Black population densities compared to the general population (46% vs. 13%). Full-time sworn personnel were mostly White (75%) and male (90%). Bivariate analyses are presented in S3 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Distribution of 2013 LEMAS policies among agencies involved in all and race-specific police killings in The Counted during 2015–2016.

https://doi.org/10.1371/journal.pone.0252749.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Distribution of organization and jurisdiction characteristics among agencies involved in all and race-specific police killings in The Counted during 2015–2016.

https://doi.org/10.1371/journal.pone.0252749.t002

Regression model specification

Number of officers, percent female personnel, percent with a HS diploma or bachelor’s degree, median income, percent poverty status, violent crime rate, percent Black population, and total population size contributed to 99.3% of the principal components (eigenvalues>1). Among these, the highly correlated socioeconomic and demographic variables (P<0.05) contributing the least to the variance were excluded; percent with a HS diploma, poverty, and total population size. Bivariable analysis showed that each remaining covariate confounded the exposure–outcome association independently (P<0.05). Nested modeling and Akaike and Bayesian information criteria indicated that violent crime rate did not improve model fit; however, it was considered an important theoretical confounder and thereby was included in the analysis.

Multivariable negative binomial regression

Table 3 shows adjusted incidence rate ratios (IRRs) comparing rates of police killings between US law enforcement agencies with and without selected policies. All models were adjusted for agency type, number of officers, percent female personnel, median income, percent with a bachelor’s degree, and violent crime rate. To account for variation in jurisdiction size, race-specific models were further adjusted for total population size. Given that total population size was used to calculate the incidence rate for all civilian killings, models including all agencies were further adjusted for Black population density to account for size differences in the subpopulation at highest risk of being killed by police [2, 3, 7, 12].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Multivariable negative binomial regression of municipal and county-level police killing rates on 2013 LEMAS policies by agencies in The Counted during 2015–2016, counts by all and race-specific killings.

https://doi.org/10.1371/journal.pone.0252749.t003

Higher proportions of full-time sworn personnel were associated with lower fatality rates for agencies involved in all (IRR = 0.85; 95% confidence interval [CI] = 0.77–0.93) and non-White civilian killings (IRR = 0.85; CI = 0.73–0.99). Having a mission statement predicted lower rates of all (IRR = 0.70; CI = 0.58–0.84) and White killings (IRR = 0.60; CI = 0.40–0.90). Offering personnel incentives predicted a lower non-White rate (IRR = 0.94; CI = 0.89–1.00). Agencies participating in community evaluation had a lower White rate (IRR = 0.82; CI = 0.68–0.99). Greater video use predicted a higher White rate (IRR = 1.13; CI = 1.01–1.28). Additional training for new-hires predicted a marginally higher rate of all killings (IRR = 1.27; CI = 0.97–1.67). Greater use of less-lethal weapons predicted a marginally higher Black rate (IRR = 1.09; CI = 1.00–1.19). Utilizing an external organization to conduct internal statistical analysis or research had a marginally lower Black rate (IRR = 0.80; CI = 0.62–1.03). Our sensitivity analyses showed that restricting the sample to jurisdiction sizes of 35,000+ people did not meaningfully change the magnitude or direction of any estimates, though, unsurprisingly, precision was reduced or remained unchanged.

Full-time sworn personnel

Our finding that agencies with more full-time personnel had lower rates of all and non-White civilian killings is novel. One study reported that agencies with more full-time sworn officers had lower citizen complaints while those having internal affairs units with full-time personnel had higher complaint rates, which the researchers suggest could represent the presence of more structured internal policies and procedures [17]. We found null associations for White and Black civilian subgroups, suggesting a greater influence of full-time personnel on more racially diverse communities. Our findings support the commonly cited hypothesis that a more structured police organization may lead to greater officer accountability and improved officer behavior [17, 49, 58]. The mechanism linking full-time personnel and police killings through organizational structure is unclear and warrants further consideration.

Mission statement

Having an established mission statement predicted lower rates of all and White civilian killings, which has not yet been reported. Police agency mission/value statements communicate the fundamental beliefs that govern organizational policy, priorities, and commitment to the community served. The null findings for agencies involving racially minoritized groups (vs. White) was surprising given they were slightly more likely to have a mission statement (98% vs. 96% respectively). Future qualitative content analysis of language used in established mission statements (e.g., diversity or equal protection component) as predictors of officer use of deadly force might help elucidate its role.

Community evaluation

Inconsistent with related work, community evaluation was protective in this study, but only for White civilian killings. Agencies utilizing citizen/community review boards have reported increased excessive use of force complaints yet study findings are not disaggregated by civilian race [17, 99]. In the current study, agencies involved in White fatalities had the lowest proportion of community evaluation. Our findings suggest that White civilians may benefit from community evaluation initiatives, more so than other racial groups. Further, jurisdictional differences in the sociodemographic composition may influence whether community evaluation is implemented and to what extent it benefits minoritized populations.

Incentives

We are seemingly the first to report increased personnel incentives associated with lower non-White civilian fatalities. Seminal works by Uelmen (1973) and Sherman (1983) posited that changes to organizational incentives may mitigate avoidable police killings [50, 54]. Similarly, McElvain and Kposowa (2008) recommended officer college education incentives after finding that college-educated (vs. no college) officers were less likely to be involved in a shooting [20]. However, current research on organizational regulations aimed to reduce police killings largely focus on restrictive use of deadly force policies and supervisor oversight [27, 45, 51, 53, 57]. Our finding bolsters the notion that personnel incentives may be one agency policy that can mitigate deadly force, particularly against people of color.

Video use

Opposite our hypothesis, our finding that video use was associated with higher White civilian rates was likewise novel. Most studies, including randomized control trials, solely examine the influence of body cameras on officer force, and findings are mixed [59, 67, 76, 77, 100–103]. To compare findings to this literature, we evaluated body camera use independently (vs. 4 video types) and the association was also positive. These combined results underscore our lack of understanding about how video use truly impacts officer behavior, meriting continued study.

Marginal findings

A marginally significant association suggests that additional training for new-hires trends toward higher rates of all civilian killings, which aligns with previous work showing a positive association between officer field training and police killings [58]. The type of additional new officer training (e.g., racial bias, de-escalation vs. physical restraint) may be reflective of internal agency dynamics, which has a strong influence on officer use of lethal force [48, 104]. Therefore, more officer training is not likely to change officer behavior unless the agency culture and type of training demands it. In LEMAS, the additional training was not characterized, so there is no way to know what type of training(s) may have unintended fatal consequences, further indicating a cautious interpretation of this finding.

External statistical analysis and less-lethal weapons use were marginally associated with Black civilian killings. Having an external organization conduct internal research trended toward lower rates. Analysis type was not reported (e.g., crime, geospatial), thus the specific research that may be protective is unknown. Plausibly, agencies that allow external organizations to examine internal performance may promote wider transparency and greater agency liability, which in turn, contributes to less use of officer force, particularly against Black civilians, thereby reducing police killings and perhaps racial inequities.

Conversely, less-lethal weapons use was positively associated with a marginally higher Black civilian rate. However, most publications report on the impact of conductive energy devices (CED) (e.g., Tasers) on police use of deadly force. The overall CED literature points to (1) greater officer use of CED than firearms when confronting potentially lethal suspect resistance [105], (2) reduced civilian and officer injuries when used [62], and (3) more restrictive policies reducing both CED use [51] and use of force overall [27]. Therefore, we assessed Taser use alone (vs. our composite measure) and our results were null, which aligns with other literature [98, 106]. Our finding supports public concern that alternate less-lethal weapon use may leave civilians of color at higher risk (e.g., neck restraints). Future studies should investigate the independent impacts of various non-lethal tactics on police killing rates among racially minoritized groups.

The absence of any statistically significant policy associations among agencies involved in Black killings is an important finding. One interpretation is that our models lacked sufficient statistical power to detect an association due to a modest number of agencies involved in Black killings (N = 165). Another interpretation is that the policies we examined are ineffective at reducing the killings of Black people. This includes policies that specifically purport to protect Black lives, such as the use of body cameras and less-lethal weapons training. This finding has urgent policy implications. Though support for defunding or reforming police departments is increasing, past reform efforts have not reduced police violence or improved community health [107]. For example, officer body camera use has increased since the 2014 killing of Michael Brown yet police shootings largely remain unchanged [108]. Two recent policy statements passed by the American Public Health Association identify law enforcement and the carceral system more broadly as a threat to the public’s health, calling for a disinvestment in and abolition of these systems and investment in alternative strategies to promote equitable community health and wellbeing [109, 110]. Critical to this recommendation is the stance that additional reforms of an inherently violent and racist social institution will fail. Our finding that a host of different policy reforms are ineffective at protecting Black lives supports this conclusion.

Limitations and strengths

Our exploratory study findings should be interpreted with caution. Causal inference cannot be inferred using this cross-sectional design. Though temporality was established as each policy was implemented prior to each fatality, LEMAS does not track or report the ongoing active status of agency policies, hence it is unclear whether policies remained implemented at time of death. Several factors limit generalizability. US law enforcement agencies were not required to participate; however, LEMAS surveyed a random sample and the response rate was high (85%) thus the agencies included should be fairly representative. External validity is further compromised by half of cases being excluded. Despite the high response rate, there may be some self-selection bias introduced in terms of organizational or cultural differences between agencies that opted in and out of participation. Apart from race, we found no differences in matched and unmatched cases (S1 Table). However, differences between the matched and unmatched agencies emerged, suggesting bias resulting from unobserved variation in agency characteristics (S2 Table). Because law enforcement agencies are not required to collect or report information associated with agency policies, police–civilian encounters, or use of force events to a national registry [70, 111–113], there is no way to test this assumption. Moreover, we are unable to predict the direction of the bias without knowing the distribution of the outcome in unmatched agencies during other years. Though not formally validated in association with deadly use of force events, several LEMAS measures appear consistent with other findings [17, 58]. Internal validity also largely depends on accurate data collection and proper validation by The Guardian. Although The Counted database is considered reliable [73, 74], it is improbable that all known cases are accurately reported to any system, conceivably leading to an underestimate of the true prevalence of police killings. Last, due to a small sample size, the likelihood of a committing a Type II error was higher for agency estimates involving race-specific civilian killings.