Improving Transferability of Safety Performance Functions by Bayesian Model Averaging

Abstract

A jurisdiction can import a safety performance function from another jurisdiction or another time period through model calibration. For the transfer to be achieved successfully, the calibrated model must sufficiently capture local road and traffic features. As proposed in the AASHTO Highway Safety Manual, the model calibration factor is estimated to be the ratio of the sum of observations in a local sample to the sum of predictions for the sample from the uncalibrated model. Although this approach may be adequate for overall goodness-of-fit measures, achievement of a satisfactory fit over all ranges of the covariates is not guaranteed. This paper seeks to address this limitation by investigating a new methodology for the transfer of models with four groups of sample data from Canada and Italy. First, the calibration factor approach was evaluated by the use of goodness-of-fit tests. Then, local models were developed and evaluated. For these models, a variety of random structures for frequentist and Bayesian approaches was explored with generalized linear regression, nonlinear mixed fitting, or Markov chain Monte Carlo simulation procedures. Finally, a Bayesian model averaging approach that integrated all considered models was investigated as an alternative to traditional model selection. This methodology did improve model transferability over all ranges of covariates, suggesting that Bayesian model averaging can be a sound alternative to conventional model calibration, especially when the flexibility and estimation ease of this technique are considered. Moreover, this approach is conceptually superior to selection of a single best model because it explicitly addresses model uncertainty.