Generalised linear models have proved increasingly useful in traffic accident modelling since their first appearance in this context in the 1980s. Such models are used, for example, in the development of standards and in the isolation of locations where improvements should be made. The aim of this paper to describe traffic accident modelling in tandem with scientific method and thus to set it in its natural context. Four steps in one cycle of the modelling process can be identified: variable selection and measurement, model fitting, goodness of fit testing and finally, development of confidence intervals and prediction intervals. Appropriate tools for the latter two steps have been developed recently and are described. The paper illustrates the cycle with a real example.