Similar predicted fatal-crash risk, different contribution architectures: A temporally validated analysis of vehicle–pedestrian crashes in Japan

Vehicle-pedestrian crashes with similar predicted fatal-crash risk may arise from different combinations of contributing conditions, while the distinction between risk magnitude and risk composition has received limited attention. This study investigated whether reproducible multivariable contribution architectures exist among vehicle-pedestrian crashes with comparable model-predicted fatal-crash risk. Nationwide police-reported crash data from Japan were analyzed using temporally separated 2023 and 2024 cohorts, comprising 74,002 vehicle-pedestrian crashes, including 1,838 fatal crashes. Four classification models were evaluated using the 2023 cohort, with the 2024 cohort reserved for independent temporal validation. XGBoost was selected as the nonlinear base model and achieved a PR-AUC of 0.203 and ROC-AUC of 0.877 in the 2024 test cohort. SHapley Additive exPlanations (SHAP) were then used to decompose model predictions, and signed SHAP contributions were conditioned on overall predicted risk, aggregated across seven domains, normalized, and clustered. Two contribution profiles were identified. One showed relatively greater contributions from road-user characteristics, whereas the other showed greater contributions from environmental conditions, roadway geometry, and traffic control and speed regulation. The profiles were highly reproducible in 2024, with high centroid cosine similarities and an adjusted Rand index of 0.975 relative to an independently fitted 2024 solution. Both profiles occurred across the full predicted-risk distribution and were not well reproduced by simple partitions based on predicted risk, pedestrian age, or day/night conditions. Profile-specific SHAP interaction analysis further revealed systematic differences in nonlinear predictor combinations, particularly the interaction between pedestrian age and day/night conditions. These findings demonstrate that similar predicted fatal-crash risk can arise through distinct and temporally reproducible contribution architectures, suggesting that predicted risk magnitude alone does not fully characterize the multivariable circumstances underlying fatal vehicle-pedestrian crashes.