The Effects of Trend over Time on Accident Model Predictions
MOUNTAIN L, FAWAZ B, University of Liverpool, and M.AgER M, Napier University, UK
Given that currently some 230 000 people are killed or injured annually on roads in the UK, while the annual cost to the nation of road tratfie accidents exceeds £10 000 million, it is vital that road safety spending is effectively targeted so as to maxim
Given that currently some 230 000 people are killed or injured annually on roads in the UK, while the annual cost to the nation of road tratfie accidents exceeds £10 000 million, it is vital that road safety spending is effectively targeted so as to maximise the potential social and economic benefits to be gained by reducing the annual accident toll. The selection of potential remedial sites and their evaluation is dependent on a knowledge of the accidents which would occur without treatment. However, the rare and random nature of accident occurrence means that estimating the underlying mean .accident frequency at candidate sites prior to treatment is not straightforward.
There is evidence to suggest that an appropriate approach is to use an empirical Bayes fEB) technique in which the Irue underlying mean accident frequency (m) at a site is estimated as a weighted combination of the observed accidents (xb) and a predictive model estimate of the expected accident frequency (/t). The inclusion ofxb allows some account to be taken of specific site characteristics not included in the prediction model while/z smoothes out random variation, controlling for regression-to-mean effects. The quality and usefulness of the EB estimates will depend on the availability of suitable predictive models and the cost of the d_at~ required to apply them.
Over the years numerous models have been developed relating accidents to various measures of flow, site characteristics and geometry. While early models were based on the assumption of a normal error structure using classical least squares regression modelling, it is now accepted that the use of generalised linear models, with a quasi-Poisson or (preferably) negative binomial error structure, is more appropriate (see, for example, Joshua & Garber 1990, l~aou & Lure 1993, Maher & SummersgiU 1996). Thus more rec~ models have been developed using statistical programs specifically designed for fitting generalised linear models, such as GUM (Francis et al 1993) and GENSTAT (Lane et al 1988). In spite of these advances, from the point of view of remedial site selection and monitoring, there remain a number of problems associated with the available predictive models. Two of these form the focus of this paper.
Association for European Transport