This paper reviews models relating 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 new generalised linear models for predicting accidents at intersections in New Zealand is then described. It is shown that the new models fit the empirical data better than existing, simpler models. The use of the models for predicting intersection accidents in three networks is described. The paper concludes that models based on summing intersections in-flows, and estimated using ordinary (Normal) regression, are less appropriate than those based on products of conflicting flows, and estimated using flows, and estimated using Poisson or Negative Binomial regression. The latter models take better account of the mechanisms of accident occurrence, enable prediction of accident types, and provide accurate estimates of accidents in a network and the safety benefits of network flow changes. (A)
Abstract