Behavioral risk profiles in urban pedestrian-vehicle crashes: Insights from latent class and SHAP analysis

Traffic crashes remain a major global public health threat, yet most studies focus on crash severity rather than pre-crash factors. This study examines fault attribution as a classification criterion in pedestrian-vehicle accidents, separately analyzing accidents where only the driver or only the pedestrian is at fault, and identifying latent behavioral profiles within each stratum using class membership derived from LCA as the primary analytical outcome.

A hybrid methodology combining Stratified Latent Class Analysis (LCA) and Random Forest (RF)-supported SHapley Additive exPlanations (SHAP) was applied to a dataset of 7,213 vehicle-pedestrian interaction crashes in 2022 and 2023 in Istanbul. Datasets were stratified by fault attribution group (driver fault; pedestrian fault), and LCA was independently applied to each stratum. The optimal number of latent classes was determined using Bayesian Information Criterion (BIC) and Bootstrap Likelihood Ratio Tests (BLRT). RF surrogate models were then trained on LCA-derived class labels using iteratively selected features, and SHAP values were computed to quantify and interpret the contribution of each variable to class membership.

Three distinct latent class profiles were identified for each group. In crashes where only pedestrians were attributed fault, traffic control infrastructure-particularly the presence of traffic signals-emerged as the most dominant variable (SHAP ≈ 0.70). In crashes where only drivers were attributed fault, temporal and behavior-based variables played a more decisive role, with seasonal conditions (SHAP ≈ 0.50) and vehicle movement direction (SHAP ≈ 0.45) identified as the primary discriminating parameters. The RF surrogate model achieved classification accuracies of 99.86% and 98.86% for pedestrian and driver strata, respectively.

Fault attribution in urban pedestrian-vehicle crashes is structurally heterogeneous and cannot be adequately captured by global severity models. The results indicate that urban traffic safety management requires not only infrastructure investment but also behavioral intervention strategies tailored to specific user groups. For pedestrian-fault crashes, active signal control measures are critical. For driver-fault crashes, season-specific speed regulations and targeted awareness programs for young drivers are recommended. Additional references can be found in the bibliography in the Appendix.