Abstract:
Hierarchical Bayesian regression models, with differing hyper-prior distributions, are considered as accident prediction models to be fitted on data collected over several years on the Portuguese motorway network. A sensitivity analysis is performed by way of simulation to investigate the practical implications of the choice of informative hyper-priors (Gamma, Christiansen and Uniform) and non-informative Gamma, as well as various sample sizes and years of aggregated data, on the results of a road safety analysis, in particular, at detecting high accident risk locations. It was concluded that informative hyper-priors were best at detecting hotspots when small sample sizes were considered. For bigger samples the various hyper-priors produced equivalent outcomes. Furthermore, more accurate results were obtained when more years of data were analyzed.
