Research is presented, which initially developed accident prediction models for a range of junction types. The estimates of expected accidents, based on these models, were compared with those using the COBA (cost benefit analysis) models of the Department of Transport (DOT), with accident counts and with EB (empirical Bayes) estimates. For those junction types where COBA models were not available, comparison was made with the simplest forms of Transport Research Laboratory (TRL) models for those types. The models used data for 662 junctions in 12 UK counties: 111 signal-controlled junctions, 161 roundabouts, and 390 major-minor priority junctions. The data included information about junction characteristics, and details of accidents and traffic flows for periods of 5 to 15 years. The models were developed using a generalised linear modelling approach, as implemented in the GLIM statistical software. Tables of the results of the statistical tests and analyses, included a summary of best-fit models and error estimates. Of the three methods used to estimate expected junction accidents, the EB method was better than the accident counts and predictive model estimates, especially where junctions were selected for a high accident frequency in the 'before' period.
Abstract