This study investigates the occurrence of road traffic accidents in Great Britain at a national scale. STATS 19 data for road accidents, vehicles involved in road accidents and casualties occurring over several years were analysed and modelled using various statistical techniques. The main aims of this research were to investigate the use of different statistical model formulations and to investigate the numbers of road accidents, casualties, and vehicles involved that occur on each day. Generalized linear model (GLM), generalized estimation equation (GEE), and hierarchical generalized linear model (HGLM) formulations were investigated for this purpose. The variables of weekday 3 (weekday, Saturday, Sunday), seasons (Spring, Summer, Autumn, Winter), month, time, Public holidays, Christmas holidays, new-year holidays, road type and vehicle class, together with certain interactions between them, were found to be important in developing models of risk per unit of distance travel. Additional variables of distance travelled per vehicle, vehicles per head of population, population density, meteorological factors were also investigated, and population, age group and gender were used to develop models of casualty rate per person-year. The GLM model structure with log link function was found to fit data for the occurrence of road accidents reasonably well when the negative binomial distribution was adopted to accommodate over-dispersion beyond Poisson levels. The GEE with negative binomial error together with autoregressive (AR1) structure was preferred over the GLM as it can also accommodate serial correlation that was found to be present in the data due to the natural order of the observations. The coefficients and significance levels of some variables were found to change significantly if the presence of serial correlation is not respected. Finally HGLM with Poisson-gamma errors and log link function was used to estimate the number of casualties involved in road accidents on each day. The advantage of HGLM over GLM and GEE is that it can account for variability within and between clusters using both random effects and dispersion modelling: this was found to be substantial. However, unlike GEE, HGLM cannot accommodate time series structure so that the coefficients and the associated standard errors of some of the variables should be viewed with caution. From the model results, it is found that distance travelled provided a good measure of exposure to risk in most cases, and that each of distance travelled per vehicle, population density and rain is associated with greater risk for road accident per unit of travel whereas risk diminishes with increase in each of numbers of vehicles per person and mean minimum monthly temperature. The risk per unit of travel was also estimated for each of 5 classes of vehicles on each of 5 different kinds of roads. Finally the age and gender specific rate of casualty per person-year was estimated for each combination of age group and gender. The results obtained from this study will lead to the promotion of safe usage of road and vehicle class combinations by raising travellers’ awareness. On the other hand the casualty rates estimated for each of the 8 age groups and two gender groups by vehicle class will help to identify those that need more attention. These results will help various educational, planning, and rescue agencies to identify target groups for education and engineering initiatives to improve road safety. (Author/publisher)
Abstract