How do human factors affect the explanatory power of driver models? A methodology for comparative model assessment and validation

The evaluation of theories and models is key to the progress of science. This paper shows that current methodologies for assessing and comparing driver models – typically based on error distributions – are intrinsically flawed. A comprehensive methodology is proposed to evaluate driver models under both nominal and safety-critical driving conditions. Rooted in Popper's epistemology, the methodology relies on the execution of “risky tests” and on model “verisimilitude”, i.e., the idea of a degree of better or worse correspondence to truth. The approach involves extensive testing of models via pair-wise calibration against individual trajectories and variance-based sensitivity analysis, enabling a systematic examination of how the trade-off between model accuracy and uncertainty evolves with increasing model complexity. The methodology is applied to 800 variants of the Intelligent Driver Model (IDM) and its improved formulations, augmented with human factors (HF) layers – namely perception errors and delays, temporal and spatial anticipation, and adaptive driving behaviours. A full factorial design isolates the explanatory contribution of each HF addon and its interaction effects. A novel formulation of the IDM, the M-IDM, is introduced and consistently outperforms both the original and all tested improved formulations across three diverse naturalistic trajectory datasets. The study, therefore, provides a falsifiability-oriented foundation for rigorous, transparent comparison and validation of HF-augmented driver models.