Abstract

In this paper, we propose a spatio-temporal Bayesian multinomial mixture model to characterise the distribution of monthly crash severity outcomes (fatal, serious, slight, and no-injury crashes) across police stations in the City of Cape Town between January 2015 and October 2023. The model jointly estimates spatially varying crash type probabilities, captures temporal evolution through structured and unstructured random effects, and enables probabilistic detection of atypical police station level behaviour. In addition, the model enables the identification of police stations exhibiting increasing, decreasing or stable overall trends for certain crash types. We illustrate the utility of the approach through a comprehensive analysis of the nine-year crash dataset, highlighting spatial clusters, emerging temporal patterns, and areas requiring heightened monitoring. The results provide an evidence base for data-driven transport safety strategies and demonstrate the value of multinomial Bayesian modelling for traffic crash monitoring.