Conventional economic models of traffic congestion assume that the relation between traffic flow and speed is a technical one. This paper develops a behavioural model of traffic congestion, in which drivers optimize their speeds by trading off time costs, expected accident costs and fuel costs. Since the presence of other drivers affects the latter two cost components and hence the Nash equilibrium speed, a 'behavioural' speed-flow relationship results for which external congestion costs include expected accident costs and fuel costs, in addition to the time costs considered in the conventional model. It is demonstrated that the latter in fact even cancel in the calculation of optimal congestion tolls. The overall welfare optimum in our model is found to be off the speed-flow function, and off the average and marginal cost functions derived from it in the conventional approach. This full optimum requires tolls to be either accompanied by speed policies, or to be set as a function of speed. Using an empirically calibrated numerical simulation model, we illustrate these qualitative findings, and attempt to assess their potential empirical relevance. © 2004 Elsevier Inc. All rights reserved.