Harmonizing police crash registers for cross-border fatal-crash modelling: Poland and Queensland

Road safety researchers and agencies may seek to compare crash patterns or evaluate the applicability of severity models developed in other jurisdictions. However, differences in police crash data definitions, coding, and reporting coverage can undermine such comparisons. This study examines whether independently maintained police crash datasets can be harmonized into a common set of indicators that preserves sufficient information for cross-jurisdiction fatal-crash model evaluation, and identifies the limitations of transferring models between these datasets. Event-level casualty-crash records from Poland (SEWIK) and Queensland, Australia (TMR), covering 2015–2024, were harmonized using documented deterministic mapping rules. A structural-support audit excluded seven indicators without counterparts in both datasets, leaving 20 binary indicators describing weather, collision type, road surface, and location (401,805 crashes). Logistic regression and CatBoost models were trained on 2015–2022 data and evaluated on 2023–2024 records, both within and across jurisdictions. Logistic-regression AUC was 0.71 for Poland-to-Queensland transfer and 0.67 for Queensland-to-Poland transfer, compared with 0.76 and 0.73 for locally trained models evaluated on the corresponding test records. Raw predictions from class-weighted models were poorly calibrated; target-jurisdiction Platt recalibration reduced expected calibration error from 0.43 to 0.005 in Queensland and from 0.36 to 0.019 in Poland. Sensitivity analyses indicated that the principal discrimination findings persisted under alternative modelling and sample definitions. The results demonstrate that, for this dataset pair, harmonization preserves partial fatal-crash ranking information but does not establish equivalence of crash characteristics, reporting practices, or predicted probabilities. The study provides a documented procedure for assessing cross-jurisdiction comparability and model transfer; local validation and recalibration remain necessary before operational use.