The dependent variable in a multivariate explanatory model of accident occurrence ought to be number of accidents and not accident rate. This is because: 1) the number of accidents has known residual distribution: poisson distribution. The accident rate has unkown residual distribution. Normal distribution cannot be assumed unless empirically verified; 2) by using number of accidents as dependent variable, relative contributions of exposure and risk factors can be easily calculated; and 3) improvements at junctions are made as a result of an abnormal number of accidents, but if this is due to random fluctuations, safety measures might be counterproductive. Whether random variation is involved can only be determined if the residual distribution is known. There are two `laws' of accident risk: a) When the number of new items of information per unit time increases, it is more probable that some relevant item will be overlooked; and b) The more rarely road users encounter a risk factor, the more dangerous the encounter will be. The only effective safety measures are those which both reduce information load and make junctions more conspicuous. Channelisation of junctions is of doubtful value, while roundabouts greatly simplify road user tasks; left-turning manoeuvres in front of oncoming vehicles are eliminated. Safety of urban junctions may be most effectively improved by red light cameras. For paper by U. Brüde, see C 6528.
Abstract