Exploring the Determinants of Pedestrian Crash Severity: A Comparative Study Using Ordered Probit, Stacking Ensemble, and TabNet

Abstract

Abstract In this study, we aimed to explore the determinants influencing the severity of pedestrian crashes, focusing particularly on the novel application of the TabNet model in pedestrian crash severity analysis. Utilizing pedestrian crash data from Utah for the years 2010 to 2022, our methodology incorporated the ordered probit model, a stacking ensemble approach including a multinomial logistic model, XGBoost, and extremely randomized trees, as well as the innovative use of the TabNet model. A significant advancement in our approach was the application of SHapley Additive exPlanations (SHAP) for the first time to interpret the results of the TabNet model in this specific research area, providing new insights into the interpretability of deep learning models in pedestrian crash severity analysis. Our findings indicated that the TabNet model outperformed other models in predictive accuracy, effectively identifying key factors such as pedestrian age, involvement in left and right turns, lighting conditions, and alcohol consumption as significant in influencing crash severity. These results highlight the effectiveness of using advanced analytical models like TabNet to gain a deeper understanding of the factors impacting pedestrian crash severity, offering valuable insights for transportation safety engineers and policymakers in developing targeted safety measures to improve pedestrian safety in urban environments.