Coding rules and reliability

Who codes, and from what

Variables are coded by the Lab’s researcher from official investigative sources, body-worn camera footage where available, and contemporaneous news coverage. The researcher has substantial professional experience in law enforcement and use-of-force training. That expertise informed the coding framework but may also introduce professional-perspective bias in interpretive coding. Variables are operationalised through a structured codebook, and inferential claims are restricted to observable case characteristics.

A sample of records is independently coded by a second reviewer so that agreement can be measured. See Reliability below.

Key variable definitions

Engagement distance

The straight-line distance, in whole feet, between the decedent and the person toward whom the precipitating hostile conduct was directed — an officer, bystander, or hostage — measured as the minimum separation reached during the hostile approach sequence up to the first police firearm discharge.

The variable captures how close the threat came before firearm intervention, not necessarily the separation at the precise instant of discharge.

Contact range. If the decedent’s weapon or body made physical contact with any person before the first shot, the encounter is classified as contact range. These events are coded as 1 ft by convention — the value marks a category, not a measured separation.

Civilian contact. If a civilian was injured and moved away before officers arrived or before the shooting sequence, the encounter is not coded as contact range. The closest non-contact distance to that civilian is coded instead.

Qualitative descriptions. When a source gives only a verbal description, “a couple of feet” is coded as 2 and “a few feet” as 3.

Conflicting accounts. The shortest distance perceived by law enforcement up to the first shot is recorded, and the choice is noted in case records.

Whose distance is coded. For encounters in which no civilian was threatened (TC = 0), the coded distance represents the officer’s own minimum pre-discharge exposure distance. For encounters in which a civilian was threatened (TC = 1), it represents that civilian’s minimum exposure distance. The dataset covers both types.

Unknown. If distance cannot be determined or reasonably inferred from video or source text, the value is left blank. Blanks are treated as missing and are never converted to zero.

Encounter duration

The elapsed time from first police contact with the decedent to the first police firearm discharge, categorised into three bands:

Where an exact timestamp is available, the band is assigned from that value. Where it is not, the band is assigned from explicit wording in media reports or official summaries (“immediately upon exiting the vehicle”; “a prolonged standoff”). If neither a timestamp nor a timeline description is available, the value is left blank.

These band boundaries are a pre-existing coding convention. They are not derived from PERF guidance or any externally validated threshold.

Documented mental illness

Coded when the case history references the decedent as a known user of mental health services or as having documented mental illness at the time of the encounter. This is a documentary indicator, not a clinical diagnosis. Undocumented mental illness is not captured.

Officer injury

Coded when any physical injury to a responding officer is reported in connection with the encounter. Severity is coded separately (minor, moderate, major), as is the timing relative to the first shot (pre-shot, post-shot, indeterminate) and the injury mechanism.

Third-party threat status

Coded true (TC = 1) when, at any point in the encounter, the decedent’s conduct threatened, endangered, or injured a person other than a responding officer. Encounters coded TC = 0 are those in which no third party was threatened at any point.

Documented CIT involvement

Coded 1 if a dedicated Crisis Intervention Team unit, mental health co-responder, or officer whose regular daily assignment is mental health crisis response was physically on-scene before lethal force was used. Standard patrol officers who have completed mandatory 40-hour CIT training but work regular patrol beats are coded 0. See also the CIT glossary entry.

Missing information

Distance could not be determined for 569 of 1,970 records (28.9%). The primary reasons are no body-worn camera footage and no sufficiently detailed witness or investigative account of the encounter’s final geometry.

Precise timing (to the second) was available for only 980 records (49.7%); ordinal duration for 330 (16.8%) records could not be determined from available sources.

Charts state how many records were excluded for each variable. Missing values are never estimated or filled in.

Reliability

Coding quality is assessed through a stratified random audit of 197 cases (10% of the analytic dataset), independently recoded by a second reviewer blind to the original coding.

The sample was stratified to ensure adequate representation of rare variables: CIT-positive cases were oversampled (25 of the 44 CIT-positive cases in the full dataset), and cases with recorded duration but no body-worn camera involvement were oversampled to ensure independent verifiability from written records alone.

Agreement statistics

Distance reliability was calculated only for the 61 audited cases in which the second reviewer could independently reconstruct distance from written source material alone, without body-worn camera access.

Adjudicated variables

Third-party threat status (TC) showed an initial κ of .20, with disagreements running in one direction only: the second reviewer never coded TC as true where the original coded it false. Case-by-case adjudication against source material found the original coding correct in 79% of disputes; the discrepancy traced to the audit instruction being read too narrowly rather than to unreliable primary coding. Four of 82 disputed cases were revised following adjudication. Post-adjudication agreement with the adjudicated reference: 96.3% (κ = .93 versus adjudicated reference).

Documented mental illness (D_Illness) showed an initial κ of .52, with disagreements in both directions across 48 cases. Adjudication found the original coding correct in 60% of disputes; the second reviewer’s advantage in some cases traced to closer review of lengthy investigative records. Post-adjudication agreement with the adjudicated reference: 90.4% (κ = .81 versus adjudicated reference). Nineteen of 197 audited cases were revised following adjudication.

Adjudicated values are used in all analyses.

How to cite

Edged-Weapon Response Lab. (2026). Coding rules and reliability. In EWRL Methods Documentation (Version 1.0). https://droptheknife.org/methods/coding-and-reliability