Modeling pedestrian-injury severities in pedestrian-vehicle crashes considering spatiotemporal patterns: Insights from different hierarchical Bayesian random-effects models

• Published in: Analytic methods in accident research Vol. 28; p. 100137
• Main Authors: Song, Li, Li, Yang, Fan, Wei (David), Wu, Peijie
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.12.2020
• Subjects: Hierarchical Bayesian model; Injury severity; Pedestrian; Random effects only; Spatiotemporal analysis; Pedestrian; Injury severity; Spatiotemporal analysis; Hierarchical Bayesian model; Random effects only
• ISSN: 2213-6657; 2213-6657
• DOI: 10.1016/j.amar.2020.100137

•Spatiotemporal features of the pedestrian-vehicle crashes were studied.•Developed a hierarchical model with random intercepts across spatiotemporal subsets.•Developed a hierarchical model with random-effects-only across observations.•The random-effects-only model significantly improved the fitness and prediction accuracy.•Investigated the effects of contributing factors on pedestrian-injury severities. To systematically account for the spatiotemporal features and unobserved heterogeneity within pedestrian-vehicle crashes, this paper employs the spatiotemporal analysis and hierarchical Bayesian random-effects models to explore the factors contributing to pedestrian-injury severities of pedestrian-vehicle crashes involving single vehicle in North Carolina from 2007 to 2018. Ten spatiotemporal patterns of the crashes are identified by applying an improved spatiotemporal analysis. Significant temporal instability and the spatiotemporal instability of the factors to the pedestrian-injury crashes are identified by the likelihood ratio tests. A hierarchical Bayesian random intercept logit model with random-effects across the spatiotemporal groups is firstly employed for the whole dataset. The comparison between different hierarchical models indicates that addressing random-effects across observations and increasing the number of random parameters could both improve the model performance. Then a hierarchical Bayesian random-effects-only logit model, which allows all parameters to be randomly distributed across observations, is developed to further investigate the unobserved heterogeneity in spatiotemporal segmented datasets. The significant improvements in terms of model fit and the hit accuracy underscore the superiority of the random-effects-only model. The marginal effects of the human, vehicle, crash, locality, roadway, environment, time, and traffic control factors for each spatiotemporal dataset also provide insights into possible inherent reasons for the spatiotemporal instability/tendency of the crash and correlated factors. Meanwhile, specific countermeasures are given to locations especially in which the spatially aggregated patterns of the crashes have new, consecutive, and intensifying temporal tendencies. This study provides a framework for engineers and researchers to identify spatiotemporal patterns of the crashes and explore the factors affecting pedestrian-injury severities especially in those existing crash-prone areas.