Database Provenance and Crash-Scenario Contributions to Fatal-Risk Prediction in Multi-Source Chinese Crash Records

Pooling multi-source administrative crash data may introduce predictive information from differences in inclusion rules, injury coding, and field coverage. This study quantifies database-source and crash-scenario contributions to fatal-outcome prediction across eight Chinese sources with reconstructable outcomes. Nine datasets contained 50,139 records; Hunan was retained for data-quality and descriptive analyses but could not support fatal-outcome reconstruction. Of 19,776 casualty crashes with reconstructable outcomes, 17,866 earlier-year records formed the development set, and 1821 records in total from the most recent eligible year of each source formed the pooled test set. A baseline assigning each source’s development-period fatality proportion had no within-source ranking ability, yet achieved a pooled ROC-AUC of 0.720 and an average precision of 0.494, indicating substantial between-source variation. The two-seed CatBoost ensemble increased ROC-AUC by 0.059 to 0.779 (95% CI 0.756–0.800) and average precision by 0.124 to 0.619 (0.578–0.661); the Brier score was 0.178 and scaled Brier skill score was 0.201. Removing the crash-cause field produced an ROC-AUC of 0.781 and an average precision of 0.610, with no clear paired performance difference. Source-specific performance differences were marked: the equal-weight macro-average ROC-AUC was only 0.628, and Jiangxi showed no clear evidence of within-source discrimination (0.479, 95% CI 0.359–0.599). SHAP and adjusted regression supported stable associations of pedestrian collisions and rollover/run-off-road crashes with fatality, while pooled calibration varied by source. Pooled metrics should be interpreted as performance comparisons within this assembled collection, not as a literal decomposition or proof of a shared crash-severity mechanism; reporting should include source-proportion baselines, within-source metrics, source-level calibration, and sensitivity analyses.