A hybrid machine learning and random parameters approach to truck collision severity: temporal variations and heterogeneous effects

Truck crashes can have severe consequences, yet the temporal stability and contextual heterogeneity of their severity determinants remain unclear. Using 5,385 two-vehicle truck collisions from the UK STATS19 database (2021–2023), this study integrates LightGBM–SHAP analysis, common-specification temporal-stability tests, and a random parameters binary logit model with heterogeneity in means and variances (RPBL-HMV). SHAP highlights counterpart-vehicle characteristics, collision configuration, speed environment, driver attributes, lighting, and roadway context as key predictors. Although SHAP importance rankings differ across years, formal tests find no robust evidence of temporal parameter instability, supporting pooled estimation with year effects. The RPBL-HMV model identifies substantial crash-level heterogeneity, particularly for lower speed limits, whose association with severe outcomes varies in mean and dispersion by collision configuration and counterpart-vehicle type. These findings distinguish predictive importance from parameter stability and contextual heterogeneity, informing targeted truck-safety strategies.