Limitations

Fatal outcomes only

Every record is an encounter that ended in a death. The dataset describes how those encounters unfolded; it cannot show how often similar encounters ended without a death. A tactic that appears absent from the fatal caseload may be absent because it was never used, or because it worked — the dataset cannot distinguish between those explanations. Rates, effectiveness estimates, and comparisons to non-fatal encounters require a different data source.

Counts, not rates

The figures are counts. Annual case totals run from 171 to 213 for 2015 to 2023; the count for 2024 is 261, reflecting a full calendar year now in the dataset. The dataset has no information on the total volume of police-public contact, calls for service, or the population at risk. A count going up or down does not establish that the underlying rate changed.

Publicly reported encounters only

Records are identified from public databases and public reporting. An encounter that was never reported or documented publicly is not in the dataset.

Missing data is not random

Values are missing for 569 records and duration for 330 records in the ordinal category (short / medium / extended), representing 28.9% and 16.8% of the analytic sample respectively. Precise duration (to the second) was missing for 990 records (50.3%).

Missing values are not random. A comparison of distance-recorded and distance-missing cases found:

The higher officer-injury rate among distance-recorded cases has two likely explanations. First, an injury produces more detailed investigative and medical documentation of the encounter’s final moments, which is the same documentation from which distance is typically coded. Second, under the dataset’s coding rules, an encounter in which the decedent’s weapon reached an officer is coded at contact range (1 ft), so injury-producing encounters are inherently easier to assign a distance.

The practical implication is that figures involving engagement distance — particularly the contact-range injury relationship — describe the distance-recorded subset, in which officer injury was considerably more common than in the dataset as a whole (12.3% versus 9.3% overall). The magnitude of any distance-based finding cannot be assumed to hold unchanged among cases where distance could not be documented.

A bounding analysis found the direction of the distance-injury relationship is unlikely to reverse under any plausible allocation of the missing cases, but the magnitude cannot be independently verified from missing cases alone.

The contact-range coding convention

Contact-range events are coded as 1 ft by convention. That value is a category, not a measured separation — it marks that the minimum distance reached before firearm intervention was zero. Because an officer who is injured in an edged-weapon encounter is by definition at contact or near-contact range, there is a partly definitional element to the contact-range injury finding: it identifies where officer injuries occur within these encounters more than it independently estimates the causal effect of distance. This is acknowledged in the first report and is part of why the finding is described as a within-fatal-case association.

Engagement distance is not always the officer’s own distance

The coded distance represents the minimum separation between the decedent and whoever the precipitating hostile conduct was directed toward — which may be a civilian rather than the officer who fired. For encounters with a third-party threat (53% of distance-coded cases), the coded distance is the civilian’s exposure distance, not the officer’s. Findings are replicated within the officer-directed subset (TC = 0) as a standard validity check. See Coding rules and reliability.

Documentary indicators undercount true involvement

Documented CIT involvement and documented mental illness are coded from media reports, official statements, and investigative records. Cases in which these facts were not reported or documented are coded 0. The true prevalence of mental illness and crisis-resource involvement in this caseload is certainly higher than the coded rates suggest. Trends in these indicators may reflect changing documentation practice as well as, or instead of, real change.

Single researcher

Narrative coding was conducted by a single researcher with professional law enforcement experience. That expertise informed the framework but may also introduce professional-perspective bias. The inter-rater reliability audit (see Coding rules and reliability) found substantial to excellent agreement on the variables most central to published findings, with greater uncertainty for third-party-threat status and documented mental illness.

Researcher-defined period comparisons

Where reports compare an “early” period to a “later” period (e.g., 2016–2018 versus 2019–2024), those boundaries are researcher-defined and exploratory. No agency-level training-adoption dates were available to anchor them empirically. Findings that depend on a period boundary are reported alongside a sensitivity check using an alternate cut point.

Aggregates only

The Lab publishes aggregate counts and suppresses cells with fewer than five records, which limits how finely any chart can be subdivided. See Suppression and privacy.

Figures can change

Records are corrected and added as new information becomes available. Each report names the snapshot it draws on, and changes across snapshots are recorded in the version history.

How to cite

Edged-Weapon Response Lab. (2026). Limitations. In EWRL Methods Documentation (Version 1.0). https://droptheknife.org/methods/limitations