Study 1
Miller and Chillar (2021) used a quasi-experimental design with annual panel analysis and propensity score matching to examine the impact of body-worn cameras on fatal police-citizen encounters over a 3-year period.
The analysis focused on subsamples of U.S. police departments drawn from the Bureau of Justice Statistics’ Law Enforcement Management and Administrative Statistics Body-Worn Camera Supplement (LEMAS–BWCS) conducted in May 2016. This survey targeted all primary state law enforcement agencies, local police departments, and sheriff’s offices with 100 or more full-time sworn officers, along with probability samples of smaller agencies. Yearly data points (defined here as May through April) were included for these agencies from before and after the beginning of the LEMAS–BWCS survey fieldwork. The primary independent variable was the presence or absence of a body-worn camera system in law enforcement agencies, by year, and three LEMAS–BWCS survey questions were used to develop this variable. The first question was “Has your agency acquired any of the following tools to record officer–citizen interactions…?” The list of available tools included body-worn cameras and respondents were then able to select whether the agency had acquired or not acquired body-worn cameras. The second question, for those agencies that indicated having acquired body-worn cameras, was when the agency first obtained them. For the core and matched sample analyses discussed below, a group of control agencies that did not acquire body-worn cameras in the postsurvey period to 2018–19 was included, against which body-worn camera-acquiring agencies could be compared. Agencies that indicated they had acquired body-worn cameras in the period after the beginning of survey fieldwork to the point of survey completion were excluded. Further exclusions were made using a third survey question, which asked, “How likely are you to consider acquiring body-worn cameras in the next year?” Control agencies were selected from those that responded that this was “Unlikely” or “Very unlikely.”
This process identified the subset, or core sample, for the analysis, which was a total of 2,376 agencies made up of 1,346 treatment agencies that acquired body-worn cameras only during the period of 2013–14 to 2015–16, and 1,030 control agencies that likely did not acquire cameras at all between 2005–06 and 2018–19. This core sample was then further reduced through a propensity score match. Model covariates for the match included agency state, a variable representing the interaction of agency size and type (based on LEMAS–BWCS), the 2012 Uniform Crime Report population served, county poverty rates from 2011 to 2015 American Community Survey estimates, and earlier counts of agency fatal encounters from 2005–06 to 2007–08. State agencies were excluded from the match because they did not have equivalent demographic measures to the other agencies. The final logistic regression used to estimate the propensity score included a number of interaction terms alongside the main effects to achieve balance (a few binary or “dummy” variables and corresponding cases were dropped by the model where perfect prediction occurred). Matching was conducted with replacement, using two nearest neighbors, and produced reasonable balance. This matched sample included a total of 1,795 agencies (1,178 treatment agencies and 617 control agencies).
In the core sample, treatment agencies (n = 1,346) were mostly local police (75.7 percent) or sheriff’s offices (23.6 percent), and the remainder (0.67 percent) were state agencies. Agency size in the core sample was distributed as follows: 100 to 499 full-time officers (20.4 percent), 10 to 24 fulltime officers (19.9 percent), and fewer than 5 fulltime officers (18.4 percent). In the core sample, control agencies (n = 1,030) were mostly local police (75.6 percent) or sheriff’s offices (22.7 percent), and the remainder (1.7 percent) were state agencies. Agency size among the control agencies was distributed as follows: 10 to 24 full-time officers (24.3 percent), fewer than 5 full-time officers (19.9 percent), and 100 to 499 fulltime officers (16.1 percent).
In the propensity score model, treatment agencies (n = 1,178) were 74.6 percent local police and 25.4 percent sheriff’s offices. Agency size among the treatment agencies was distributed as follows: 10 to 24 full-time officers (22.2 percent), 100 to 499 full-time officers (20.8 percent), and 5 to 9 full-time officers (18.2 percent). Control agencies (n = 617) were 75.0 percent local police and 25.0 percent sheriff’s offices. Agency size among the control agencies was distributed as follows: 10 to 24 full-time officers (22.7 percent), 5 to 9 full-time officers (18.4 percent), and 100 to 499 full-time officers (18.3 percent).
The dependent variable was the count of fatal citizen encounters for each law enforcement agency, by time period. The fatal encounters database used aimed to capture all deaths through police interaction in the United States since Jan. 1, 2000, and relied primarily on media research by the database administrator complemented by public records requests and crowdsourced data. Because of improvements in data capture processes, only fatal encounter data from 2005 through June 24, 2019, were used. Annual measures of fatal encounters were conducted for each agency. These are based on a 12-month (May–April) cycle, to match the beginning of the LEMAS–BWCS survey fieldwork. To create the analytic dataset, LEMAS–BWCS and fatal encounter data sources across common agencies were combined. Both data sources were matched to a 2012 roster of all U.S. police agencies, including 28,065 primary agencies. Fatal encounter data were matched to the 2012 roster, using an iterative matching process involving agency names, counties, and states, which used exact and manually checked fuzzy-matches across records. Vehicle-related deaths, deaths that were credited to multiple agencies, and deaths associated with irrelevant agency types (such as federal agencies or corrections departments) were excluded, and 18,660 fatal encounter records were matched from 2000 to midyear 2019.
Difference-in-difference analyses were used to estimate a body-worn camera treatment effect on treated law enforcement agencies. This approach compared changes in outcomes for agencies that experienced the intervention at particular points in time with those that did not, while “differencing out” group- or time-invariant confounders. An annual panel analysis (referred to as “Model A”) was conducted with the core sample of agencies in which body-worn camera-acquiring agencies (the treatment group) obtained cameras between 2013–14 and 2015–16. The analysis balanced treatment agencies in the pre– and post–body-worn camera acquisition relative time period by including only treatment agencies that had at least 8 years’ pre-intervention time and 3 years’ post-intervention time during the observation period (2005–06 to 2018–19), and by excluding agency years outside this relative timeframe. Control agencies in the core sample were those that did not acquire body-worn cameras at any point in the survey (and did not anticipate getting them in the year following), using years for the entire period (2005–06 to 2018–19). Treatment agencies all had 12 years of data, while control agencies (covering the same years found in the treatment agency sample) all had 14 years of data. Cluster robust standard errors, by agency, were used to help take account of the clustered nature of data points, given that annual or period counts of fatal encounters are correlated within agencies. This model used 819 agencies that reported at least one fatal encounter across the periods analyzed (agencies with consistent zero outcomes are automatically dropped by Poisson fixed effects).
Additionally, two difference-in-difference analyses involving two-group (and two time-period) specifications were conducted to provide a validity check on the findings from the annual panel analysis, provide additional flexibility regarding the assumed time between body-worn camera acquisition and their effects, and make use of additional fatal encounters data in the last 2 years. Both two-group analyses measure outcomes for 3-year periods pre- and post-acquisition: 2010–11 to 2012–13 and 2016–17 to 2018–19, respectively. The first two-group analysis (referred to as “Model C”) focused on the same agencies from the core annual panel analysis (including the zero-fatality agencies). Common trends in the pre-period for this comparison were checked graphically and statistically modeled, and also built credibility from the analysis of common trends in the annual panel analysis conducted on the same group of agencies. The second two-group analysis used a reduced set of local and sheriff’s agencies following the propensity score match described above (“Model D”) for a total of 1,795 agencies (1,178 treatment and 617 control) in the matched sample. No subgroup analysis was conducted.
