Uncovering nonlinear crash risk patterns among instant delivery riders

Abstract

The rapid growth of instant delivery services has raised increasing concerns about the road safety of delivery riders, many of whom rely on electric two-wheelers in dense urban traffic. Previous studies have linked rider crash risk to sociodemographic characteristics, workload, and psychological factors, but most have relied on linear modeling frameworks. As a result, limited evidence is available on whether latent variables are associated with actual crash involvement through nonlinear patterns. To address this gap, this study conducted a revealed preference survey in Shanghai, China, yielding 276 valid responses from delivery riders. We developed an integrated framework that combines a latent variable model, CatBoost, Shapley additive explanations (SHAP) values, and partial dependence plots to examine crash involvement using sociodemographic, delivery-related, and latent variables. The latent variable model identified two interpretable latent variables, rule-breaking propensity and unhealthy lifestyles, which were then incorporated into the nonlinear CatBoost model. Model comparisons and robustness checks supported the reliability of the framework. The results show that rule-breaking propensity, unhealthy lifestyles, and daily delivery mileage form an operationally stable leading predictor group within the observed sample, with rule-breaking propensity most consistently ranked first. Partial dependence analyses indicate broadly increasing latent-score associations with modest nonlinearities, whose overall shapes are closely reproduced by Logistic RCS. These findings provide preliminary evidence for group-level safety planning, including recurrent safety education, supportive workload management, and rider well-being programs.