Abstract

Developing reliable safety performance functions (SPFs) capable of estimating expected accident frequencies and identifying hazardous sites is a major concern of departments of transportation. In Bayesian accident data analysis, sites are commonly ranked based on their posterior expected accident frequency in order to be selected for safety countermeasures. The primary objective of this research was therefore to propose an alternative method to evaluate the level of accuracy of an SPF and identify potential hazardous sites; both directly through a single step or measurement. A case study of the Trans-Canada highway in New Brunswick was used applying Bayesian statistics with three different likelihoods: Poisson, hierarchical Poisson-gamma, and hierarchical Poisson-lognormal. As a secondary and validating objective the above mentioned models were investigated and compared. At the same time, the effect of environmental exposure on the occurrence of accidents was studied showing minor effects. The posterior means of the model parameters indicated that, for the case study, various likelihoods provided roughly similar estimates. However, there were significant differences in the way in which these likelihoods captured the uncertainty around the posterior mean through the standard deviation, 95% credible interval, and model-fitting. Moreover, a series of computational and graphical goodness-of-fit measures were examined. In particular, the hierarchical Poisson-gamma likelihood presented the best model-fitting. Furthermore, a measure of relative risk was computed for each site based on the error term presented in Poisson mixture models. The rankings of sites using this measure and the posterior expected accident frequency were generated and compared. A positive covariance between the adopted relative risk factor and the expected accident frequency per segment length was observed. The results and discussions suggested that such a factor can be employed (1) to verify the dependability of SPFs and (2) as an alternative to identify and prioritize potential hazardous sites.