Integrating generalized linear mixed models and XGBoost for safety performance function development on urban arterials.

Urban arterial corridors present elevated crash risk due to high traffic volumes, complex geometries, and interactions among diverse road users. This study develops Safety Performance Functions (SPFs) for urban arterial segments in Southeast Michigan (USA) and evaluates traditional count models alongside ensemble machine learning approaches for crash frequency prediction. Poisson and Negative Binomial regression models were first estimated using traffic exposure and roadway characteristics, including Annual Average Daily Traffic (AADT), segment length, posted speed limit, and pavement condition. To address overdispersion and unobserved heterogeneity, a Generalized Linear Mixed Model (GLMM) was then applied to capture both fixed effects and segment-level variability while producing an interpretable, length-adjusted SPF formulation. In parallel, Random Forest and Extreme Gradient Boosting (XGBoost) models were implemented to capture nonlinear relationships and complex interactions without restrictive parametric assumptions. Model performance was assessed using 5-fold cross-validation. Among the evaluated approaches, XGBoost achieved the best predictive accuracy (R² = 0.835; RMSE = 28.65). Across modeling frameworks, AADT, segment length, posted speed limit, and pavement condition emerged as the most influential predictors. The resulting models provide transportation agencies with robust, data-driven tools for network screening, identifying high-risk segments, and prioritizing safety investments; application to other Michigan regions is feasible subject to local calibration and data availability.