Naturalistic Driving Data Analytics: Safety Evaluation With Multi-state Survival Models

Naturalistic driving data analysis offers insights into risky driving behaviors at the trajectory level, which are critical to traffic safety. However, few studies discuss the modeling challenges of vehicle interactions that are multi-state and recurrent. In addition, escalation and de-escalation transitions are two competing events by nature, requiring extra care in statistical modeling. We propose Markov renewal survival models along with cause-specific and cumulative incidence function approaches for such trajectory analysis. This study aims to quantify transition hazards and predict duration to assess the impact of off-ramps on driving behaviors at 2 highway segments in Germany. We use non-parametric, semi-parametric, and parametric estimations and select the best-fitted models based on the corrected Akaike Information Criterion (AICc). The results show that off-ramps significantly increase de-escalation durations by 27% during risky states, while vehicle types show statistically significant impacts on escalation transitions as well. Furthermore, we discuss the limitations of the cause-specific approach and recommend the use of the cumulative incidence function for predicting the marginal survival function in the presence of competing events.