A comprehensive methodology for the fitting of predictive accident models

Recent years have seen considerable progress in techniques for establishing relationships between accidents, flows and road or junction geometry. It is becoming increasingly recognized that the technique of generalized linear models (GLMs) offers the most appropriate and soundly-based approach for the analysis of these data. These models have been successfully used in the series of major junction accident studies carried out over the last decade by the U.K. Transport Research Laboratory (TRL). This paper describes the form of the TRL studies and the model-fitting procedures used, and gives examples of the models which have been developed. The paper also describes various technical problems which needed to be addressed in order to ensure that the application of GLMs would produce robust and reliable results. These issues included: the low mean value problem, overdispersion, the disaggregation of data over time, allowing for the presence of a trend over time in accident risk, random errors in the flow estimates, the estimation of prediction uncertainty, correlations between predictions for different accident types, and the combination of model predictions with site observations. Each of these problems has been tackled by extending or modifying the basic GLM methodology. The material described in the paper, then, constitutes a comprehensive methodology for the development of predictive accident models.