Context-dependent effects of built environment factors on pedestrian-injury severities with imbalanced and high dimensional crash data.

Built environment is an important component that influences pedestrian injury severities in pedestrian-vehicle crashes. Previous studies indicated that the effects of various built environment factors on pedestrian injury severities are heterogeneous, meaning the direction of their effects varies under different conditions. This heterogeneity poses a challenge in developing effective countermeasures. Therefore, this study aims to explore the context-dependent effects of built environment factors and uncover the potential sources of this heterogeneity. Using North Carolina crash data (2017-2022) as a case study, an enhanced machine learning framework that incorporates a conditional tabular generative adversarial network and wrapper-based feature selection techniques is designed to investigate the context-dependent effects. The enhancements seek to mitigate the adverse impacts associated with imbalanced and high dimensional crash data. Model results show that 13 out of the top 19 important factors belong to built environment factors, with 11 of these factors (i.e., restaurants present and road class) being heterogeneous. The context-dependent effects analysis can help identify conditions for the positive effects of heterogeneous factors. For example, restaurants near the local routes or in urban areas, higher-grade roads (i.e., US, Interstate, and NC routes) in better economically developed counties, and signs or signals control roads under dark (with or without) roadway light conditions have clear positive effects on pedestrian injury severities. These findings can provide guidance for urban planning and policy development, thereby promoting the sustainable development of cities, regions, and transportation systems.