Exploring Factors Contributing to Pedestrian Injury Severity in Pedestrian–Vehicle Crashes: An Integrated XGBoost–SHAP, Latent Cluster, and Mixed Logit Approach

Abstract

Everyone faces the risk of road safety, making the precision in predicting factors influencing pedestrian injury severity crucial. To accurately explore factors affecting pedestrian injury severity within crashes data, a three-step analytical framework is proposed. A data set containing 6,593 pedestrian-vehicle crashes in North Carolina between 2013 and 2017 is analyzed. Preliminary feature importance screening is conducted using an extreme gradient boosting–Shapley additive explanations (XGBoost-SHAP) model. Then, the entire data set is categorized into six subclasses based on driver age, road class, road type, and light conditions through latent cluster analysis. Subsequently, the six subclasses undergo further analysis using a mixed logit model, and the marginal effect values of key variables are calculated. The results reveal random coefficients in scenarios involving two-way undivided lanes, weekends, and nighttime without lighting; urban and rural areas exhibited distinct accident characteristics; and double yellow lines exert opposite effects during nighttime with lighting and without lighting. This study provides valuable insights for road safety policymakers, offering a comprehensive understanding of the factors influencing pedestrian injury severity and highlighting key considerations for the development of effective safety measures and policies.