Uncovering Probabilistic Crash Risk Among Coach Drivers Using a Bayesian Network Based on Violation and Crash Records

Abstract

Passenger transport is frequently associated with high-fatality crashes, highlighting the need for enhanced risk management. However, most risk-assessment approaches rely on cross-sectional data and fail to capture the cumulative nature of unsafe driving patterns. This study developed a static Bayesian network and a dynamic Bayesian network using violation and crash records from 9,894 coach drivers in Shanghai (2021–2024) to examine causal factors and cumulative crash risk. Overall, 14.75% of drivers were involved in one crash and 6.64% were involved in multiple crashes, while 24.83% committed one violation and 53.95% committed repeated violations. Violations were classified into five categories covering unsafe conditions and behavioral faults. A static Bayesian network identified a key causal pathway: License type/Company size → Passenger-transport violations → Roadway maneuver violations → At-fault crash. Sensitivity analysis showed large variability in crash probability (0.063–0.152), with no violations and consistently compliant operations as the strongest protective factors. Granger causality analysis identified habitual violation patterns (p<0.05), including improper lighting use and poor speed control, which were associated with increased subsequent crash risk. These findings informed an empirical dynamic Bayesian network, which achieved strong performance (Mean Absolute Error = 1.56%, Mean Relative Error = 19.02%) and over 80% predictive consistency. Overall, the study provides a behaviorally interpretable and temporally grounded approach for identifying high-risk drivers, enabling dynamic risk classification and supporting proactive safety-management strategies in coach-transport operations.