Statistical methods
The analyses the Lab uses, and what each is designed to show.
- Status
- Current
- Version
- 1.0
- Date
- October 10, 2026
Methods entries document how EWRL research is produced. Findings and conclusions appear in the Reports and Evidence Base collections.
On this page
What published reports report
Reports state whether findings are descriptive (counts and proportions from the data as observed) or inferential (statistical tests and models). Each figure states the number of records it covers and the number excluded. Records with a missing value for a variable are excluded from that variable’s figure; they are not estimated or filled in.
Methods used
| Method | What it is used for |
|---|---|
| Descriptive counts and proportions | Describing how cases distribute across categories — the primary output of most published figures |
| Mann-Whitney U test | Comparing a measured or ranked quantity between two groups without assuming the data are normally distributed; reported as medians |
| Binomial logistic regression (aggregated) | Testing whether a binary indicator (e.g., CIT involvement, mental illness) changed across years; the annual count of positive cases is modelled as a function of year, which is more powerful than an unordered chi-square because it uses the ordering of years |
| Multivariable logistic regression | Estimating the association between a yes/no outcome (e.g., officer injury) and a set of factors simultaneously; standard errors clustered by agency as a robustness check |
| Quasi-Poisson interrupted time-series | Modelling monthly case counts over time with a structural break, allowing the dispersion parameter to be estimated from the data rather than fixed at 1 |
| Pearson chi-square with Yates’ correction | Testing whether two categorical variables are associated where all expected cell counts are at least 5; Fisher’s exact test is used otherwise |
| Spearman rank correlation | Measuring whether two quantities tend to rise or fall together, without assuming a linear relationship |
| Cohen’s kappa | Measuring agreement between two categorical codings; weighted kappa for ordinal variables |
| Intraclass correlation, ICC(2,1) | Measuring agreement between two continuous codings; two-way random-effects, absolute agreement |
| Benjamini-Hochberg false discovery rate | Applied when a family of related tests is conducted simultaneously, to limit the chance that any one finding is a false positive |
Models use complete cases
Statistical models are fitted to the records that have a non-missing value for every variable in the model. Each report states the resulting sample size. See Limitations — missing data for discussion of what records are excluded and why.
What statistical results can and cannot show
The dataset captures fatal encounters only. It has no denominator — no count of all encounters, calls for service, or encounters that resolved without a death. Statistical associations therefore describe patterns within this fatal-case sample; they do not estimate rates, do not establish causation, and do not generalise to encounters that did not end in a death. See Limitations.
Agency clustering
Where a model pools encounters across many agencies, the residuals for encounters from the same agency may be correlated. Published regression models are re-estimated with standard errors clustered by agency (identified by name and state) as a standard robustness check, and results are reported for both the unclustered and clustered estimates.
Period comparisons
Where reports compare an early period to a later one, the boundary is researcher-defined and exploratory. Findings are accompanied by a sensitivity check using an alternate cut point. See Limitations.
How to cite
Edged-Weapon Response Lab. (2026). Statistical methods. In EWRL Methods Documentation (Version 1.0). https://droptheknife.org/methods/statistical-methods