Bayesian modelling of the spatial distribution of road accidents

This research aims to develop Hierarchical Bayesian models for road accident counts that take account of the spatial dependency in the neighbouring areas or sites. The Poisson log-linear model is extended by introducing a second level of random variation that includes a conditional autoregressive (CAR) component. Both models for accidents at the area level and models for accidents on a road network are developed. Areal models are fitted using data for counties and districts in England covering two different periods and data for
wards in the West Midlands region in 200l. Network models are fitted to link data for the MI motorway and to junction data for the city of Coventry.
Results show that, in most cases, adding a spatial (CAR) component to conventional models produces better estimates of the expected number of accidents in an area or at a site. Signs of the coefficients for explanatory variables, including level of traffic and road characteristics, are consistent with expectation. Levels of the spatial effects in a CAR model reflect the relative influence of the unknown or unmeasurable explanatory variables on the expected number of accidents. Results from models at the local authority level in the 2000s
show that spatial effects are positive in London boroughs and are negative in most metropolitan districts. For accidents at the ward level in the West Midlands, the performance of the CAR model is similar to that of the non-CAR model which includes log-normal random effects and metropolitan county effects.For models of accidents on the MI, several links are identified to have positive and fairly large spatial effects. For Coventry junction accidents, the CAR model does not perform better than the non-CAR model. Approaches to including temporal effects in spatial models when data cover two or more periods and jointly modelling different types of accidents are also proposed and examined.
Two applications of the CAR models developed in this research are introduced. The first application is about predicting the number of accidents in a local authority in a new year based on previous years' data. One advantage of using the CAR model is that it produces more precise predictions than the non-CAR model. The second application of the CAR model is a new approach for site ranking. The sites selected by such a criterion are those with high risks caused by some unknown or unmeasured factors for instance, curvature or gradient of roads) which are spatially correlated. Further on-site investigation will be needed to identify such factors.