Spatiotemporal path analysis of collision frequency using restricted spatial regression with speed difference as a mediator.

Developing reliable collision analysis models is a fundamental task in road safety research. Understanding the spatiotemporal effects of highway collisions is essential for effective prevention and control strategies. This study examines multiple spatiotemporal random effect models using three key approaches: [i] incorporating path analysis, [ii] introducing speed difference as a mediating variable, and [iii] accounting for both structured and unstructured unobserved heterogeneity. Spatial confounding between fixed and random effects was addressed by constructing a restriction matrix to correct spatial random terms. Data were collected from freeways in the Guangxi Zhuang Autonomous Region, China. The results showed that model combining structured spatial and unstructured temporal interaction terms achieved the best performance across all goodness-of-fit measures. The spatial terms explained most of the random variation. Among fixed effects, slope, speed difference, rainfall, interchange, route length, and average daily traffic volume were all positively and significantly associated with collision frequency. Further, the slope factor exhibited indirect effect on collision frequency through its influence on the mediating variable, speed difference. However, given that the freeways are located in relatively flat terrain with large radii (exceeding 750 m), the curvature radius does not exhibit a statistically significant effect on either collision frequency or speed difference. The findings provide deeper insight into the influence of spatiotemporal factors on highway collisions and offer guidance for future highway design and traffic management.