The state-dependent causal effect of a V2X-based beyond-line-of-sight warning system on young driver response: a...

Due to the perceptual challenges created by obstructed sight lines, drivers in beyond-line-of-sight (BLOS) scenarios often fail to detect hazards until insufficient stopping distance remains, making these situations particularly hazardous for young drivers with limited hazard anticipation capabilities. While vehicle-to-everything (V2X)-based warning systems offer a promising countermeasure, precisely quantifying the system effectiveness remains a challenge, as previous studies often rely on associational methods with limited ability to isolate causal effects or characterize treatment effect heterogeneity. This study aimed to apply a causal machine learning framework to estimate the effect of a V2X-based BLOS warning system on young driver response and explore potential sources of treatment effect heterogeneity in a data-driven manner. Based on data from a randomized within-subject driving simulator experiment for causal identification, double/debiased machine learning (DML) was used to estimate the average treatment effect, with the estimates further examined through robustness analyses. The analysis revealed that the BLOS warning had a positive and statistically significant effect on overall response quality. Among trials with identifiable response initiation, the warning was also associated with shorter reaction times. To explore treatment effect heterogeneity, a causal forest implemented within the DML framework was applied to estimate the conditional average treatment effects. An exploratory analysis utilizing Shapley additive explanations (SHAP) suggested that the estimated treatment effects were highly state-dependent. The heterogeneity patterns were more strongly associated with drivers' instantaneous vehicle control states than with other variables, such as static driver characteristics. A subgroup analysis further revealed that the warning benefits were statistically reliable primarily for observations characterized by stable longitudinal control and more active steering correction behavior. These findings highlight the value of incorporating real-time vehicle control states into the design of adaptive advanced driver-assistance systems to improve warning system effectiveness.