Estimating Unreported and Unobservable Uses of Force With Police Body Camera Data

• In Event: Criminal Justice Policy Reform: Sanctions, Supervision and Public Trust

Abstract

Analyses of police behavior have long relied on self-reported officer accounts, which may contain omissions, errors, or falsehoods. Body-worn camera (BWC) footage promised a more objective record of officer conduct, but resource and technological constraints leave most footage effectively unseen. As a result, police oversight depends not on direct observation of force, but on the partial and imperfect records generated by administrative reporting systems and video archives. Understanding what these systems capture—and what they miss—is central to evaluating police accountability and the administrative data used by researchers, supervisors, and policymakers.

In this paper, we draw on more than 700,000 BWC videos and associated police administrative records from a large police department to assess the accuracy and completeness of traditional use-of-force reports. Using an original computational classifier in concert with human annotators, we first estimate the prevalence of "missing" uses of force: events captured on video but not reported by officers. Reporting rates vary sharply by force type. Weapon uses are relatively well documented in traditional administrative files, while most undocumented force involves tackling and/or manually restraining civilians.

We then use a modified capture-recapture statistical technique, adapted from ecology, to bound for the first time the share of force incidents that are fully "hidden"—events not captured in either paper or video records. Together, these steps allow us to construct a 2x2 table of force incidents by whether they appear in administrative records and whether they are captured on video, distinguishing reported force, unreported force visible on video, and force hidden from both systems.

Finally, we show that BWC footage remains difficult to analyze computationally even with leading vision-analysis techniques. Current models can assist with search and triage, but they remain limited in the fine-grained action recognition required for reliable oversight. To facilitate progress, we discuss policy changes and technological advances that could improve the usefulness of these data in the near term. We also provide the first human-annotated testbed of public BWC footage against which other researchers and industry leaders can objectively evaluate vision-model performance.

Author

• Gregory Lanzalotto, University of Pennsylvania
  Presenting Author