Multivariate space-time modeling of crash frequencies by injury severity levels

Road traffic crashes threaten thousands of drivers every day and significant efforts have been put forth to reduce the number and mitigate the impacts of traffic crashes. Although the last decade has witnessed substantial methodological improvements in crash prediction modelling, several methodological challenges still remain in terms of predicting crash frequencies of different injury severity levels. These challenges include spatial correlation and/or heterogeneity, temporal correlation and/or heterogeneity, and correlations between crash frequencies of different injury severity level. A framework of Bayesian multivariate space-time model is developed to address these challenges. A series of multivariate space-time models are proposed under the Full Bayesian framework with different assumptions on the spatial and temporal random effects. In addition to the ability to consider both temporal and spatial trends, the proposed framework is also capable of addressing complex correlations between crash types. It allows the underlying unobserved heterogeneity to be better captured and enables borrowing strength across spatial units and time points, as well as over crash types. The proposed methodology is illustrated using one-year daily traffic crash data from the mountainous interstate highway I70 in Colorado, which is categorized into no injury crash and injury crash. The results show that multivariate space-time model outperforms other alternatives, including multivariate random effects model and multivariate spatial models. The model comparison results highlight the importance to properly account for spatial effects, temporal effects and correlations between crash types. (C) 2017 Elsevier Ltd. All rights reserved.