The paper seeks to discuss some of the important issues surrounding the development and application of predictive accident models. The discussion has been conducted using as examples the models developed from the long-running series of junction accident studies carried out by the UK Transport Research Laboratory. Typically, these studies have used a representative cross-sectional sample of several hundred sites, at which something in the order of 2000 accidents have occurred over a five year period. Models relating accident frequency to flows and site design have been developed using what has now become accepted as the most appropriate methodology: generalized linear modelling. Models have been developed at a number of levels, ranging from coarse ones which relate total site accidents to some simple flow function, through to detailed ones which are disaggregated by accident type and arm of the junction and which relate accident frequency to relevant turning flows and design/control variables. As well as developing the predictive models themselves, a through analysis of prediction uncertainty has been carried out. The models have now been implemented in the three TRL junction design programmes, so that is now possible for engineers to examine the implications of any proposed design not only in operational terms but also in safety terms. Work has been done to combine the accident models with a traffic assignment model, to enable network-wide effects to be studied. This network model can be run not only to determine the user equilibrium flow pattern, but also in a system-optimising model using an objective function which is a linear combination of network travel time and network accidents. Finally, the paper discussed some of the issues concerning the interpretation and validity of such predictive models and looked forward to future research which might help to resolve some of them. (A)
Abstract