This paper reviews models used in practice to relate accidents to traffic flows, with particular emphasis on the appropriateness of the model form and the statistical analysis technique employed for parameter estimation. The development of generalised linear models for predicting individual accident types at intersections in New Zealand is then described. The use of covariate analysis to identify the effect of intersection location, an investigation of the effect of non-collision flows, and the use of the models for predicting intersection accidents in three networks, are also described. It is concluded that generalised linear models for estimating different accident types (based on the conflicting flows) are better than models for estimating total accidents (based on the approach flows), especially when the cost of different accident types is known. It is also found that intersection location affects the number of different accident types, that it is important to consider the interactions between turning flows (to take better account of the mechanisms of accident occurrence), and that non-collision flows are sometimes important. Comparison of the predicted and observed numbers of accidents has shown that there is poor agreement for individual intersections, but fairly good agreement for networks. (Author/publisher).
Abstract