Exploring the modeling and site-ranking performance of Bayesian spatiotemporal crash frequency models with mixture components

Abstract

The current study introduces the flexible approach of mixture components to model the spatiotemporal interaction for ranking of hazardous sites and compares the model performance with the conventional methods. In case of predictive accuracy based on in-sample errors (posterior deviance), the Mixture-5 demonstrated superior performance in majority of the cases, indicating the advantage of mixture approach to accurately predict crash counts. LPML (log pseudo marginal likelihood) was also calculated as a cross-validation measure based on out-of-sample errors and this criterion also established the dominance of Mixture-5, further reinforcing the superiority of the mixture approach from different perspectives.The site ranking evaluation results demonstrated the advantages of adopting the mixture approach. In terms of total rank difference (TRD) results, there were several discrepancies between the two approaches, suggesting that two approaches designate unsafe sites differently. Another site ranking criterion, site consistency test (SCT), was employed to explore the difference in identification of unsafe sites based on two datasets: estimated crash count (traditional) and the spatial variability across time. The advantage of mixture models to act as a complimentary approach for site ranking was revealed by the spatial variability SCT results. The method consistency test (MCT) results also indicate the advantages of mixture models over the Base one. These findings suggested that mixture approach may prove helpful in the network screening step of safety management process to identify sites which may turn unsafe in the future and escape the detection from traditional methods.