Modelling two-vehicle crash severity under interwoven heterogeneity and interaction effects: A reliability and system safety perspective

Understanding driver injury outcomes in two-vehicle crashes remains challenging because interaction effects and unobserved heterogeneity jointly shape injury risk in complex and scenario-dependent ways. Although recent studies have combined machine learning based SHAP and random parameters logit structure models, most hybrid frameworks rely on a single learning algorithm and interpret interaction effects qualitatively or treat them as fixed effects, limiting their ability to capture heterogeneous interaction mechanisms. To address these limitations, this study develops a stability-oriented hybrid framework that integrates multi-tree SHAP analysis based on XGBoost, CatBoost, Random Forest, and Extra Trees with a random parameters logit model with heterogeneity in means, using two-vehicle crash data from Shenzhen, China. Cross-model and cross-fold consistency is emphasized to identify robust main effects and interaction candidates, reducing model-specific bias. Class weighting, AUC-PR-oriented evaluation, and probability threshold optimization are jointly adopted to address severe class imbalance. Results show that rear-end crashes are among the most injury-prone crash types and injury severity is highly context-dependent. The Struck × Rear-end interaction exhibits significant random variation, with its mean effect systematically moderated by learner driver status and new energy vehicles, providing direct evidence that interaction effects and heterogeneity are interwoven. Scenario-based probability simulations further reveal substantial variation in predicted injury risk across driver–vehicle–collision role combinations. By embedding stable interaction effects as heterogeneous components within an econometric framework, this study improves the robustness and interpretability of hybrid crash severity models and provides a quantitative basis for targeted risk assessment and interventions.