Modelling driver behaviour as a continuum between defensive and neutral states

Improving road safety requires more than identifying collisions or near misses. It calls for understanding how drivers perceive, interpret, and adapt to risk in real time. We introduce a behavioural dimension of conflict severity: the probability a driver is in a defensive state in response to perceived risk, contrasted with a neutral state reflecting routine, goal directed driving. Using a dual state latent discrete choice model, driving is represented as a probabilistic blend of these latent modes rather than a binary safe/unsafe label. The framework is grounded in the affect–behaviour–cognition (ABC) triad in psychology. It assumes drivers interpret a dynamic spatial risk field in which proximity, motion, and relative direction guide adaptive shifts in state. The model is applied to naturalistic road user trajectories and yields interpretable state membership probabilities. The defensive state shows stronger sensitivity to spatial and temporal risk indicators. The neutral state captures context appropriate but less reactive patterns. We also propose a diagnostic to assess the quality of the estimated defensive state probabilities. Due to the duality of states, consistency in one implies reliability in the other. A multi-step validation across five driving contexts, including free flow and diverging scenarios, tests generalizability beyond the estimation sample. Results indicate behaviourally consistent and reasonably stable estimates across contexts. From a theoretical perspective, this work draws on behavioural science to inform traffic safety analysis by modelling a dimension of conflict severity that has largely remained implicit. From a practical standpoint, the continuous defensive state probability offers a richer diagnostic than binary conflict classifications, with potential applications in real time safety monitoring and infrastructure design. These findings frame probabilistic behavioural states as proactive safety indicators and offer a path to integrate perception, attention, and adaptive control into quantitative traffic safety analysis.