A spatially adaptive empirical Bayes framework with dynamic dispersion parameters for enhanced crash frequency prediction across rural highway networks

Abstract

Traffic crashes often exhibit strong spatial dependence that is insufficiently captured by the Empirical Bayes (EB) method recommended in the Highway Safety Manual (HSM). This study proposes a Spatially Adaptive Empirical Bayes (SA-EB) framework that integrates advanced spatial models, including Geographically Weighted Poisson Regression (GWPR) and Multiscale Geographically Weighted Regression (MGWR), with Crash Modification Factors (CMFs) to enhance the prediction accuracy of expected crash frequencies across rural divided multilane highways (RDMHs), and generates CMF-adjusted, spatially weighted forecasts for both past and future conditions. The framework was calibrated and validated using crash and roadway data from 1071 km of highways in Hamadan Province, Iran, encompassing 2995 crashes recorded between 2017 and 2019. A dynamic overdispersion parameter (ranging from 0.3 to 0.6) was incorporated to capture spatial variability in crash dispersion. Results revealed substantial spatial heterogeneity in crash predictors: a 1 % increase in roadway slope corresponded to a 3.5-unit rise in crash frequency, while a 1 km/h increase in speed deviation led to approximately 4.5 additional crashes per segment. Implementation of SA-EB–guided geometric improvements reduced predicted crash frequencies by about 20 %, outperforming the conventional EB model. Overall, the SA-EB framework advances both the theoretical understanding and practical application of spatial safety modeling by providing transportation agencies with a data-driven and location-sensitive tool to identify high-risk segments and optimize Highway Safety Improvement Program (HSIP) investments across rural highway networks worldwide.