繰り返し事故に関与したドライバーにおける事故間隔の分析：代替的な未観測異質性モデリングアプローチ

Repeatedly crash-involved drivers, due to their short inter-crash durations and high crash frequencies, represent a critical risk to public safety. Identifying the significant factors influencing crash recurrence is therefore of fundamental importance for reducing crash frequency. Using traffic accident records from a city in Inner Mongolia, China, from 2014 to 2023, we analyze drivers involved in two or more crashes and examine the factors influencing their inter-crash durations. To address unobserved heterogeneity, two alternate modeling approaches are applied: a correlated random-parameters survival model with mean heterogeneity and a latent-class survival model based on class probability functions. Both models effectively capture multilayered unobserved heterogeneity of the crash data. The correlated random-parameters model identifies heterogeneous effects for spring and novice drivers, with a strong correlation between random parameters (0.746, p < 0.001) significantly influencing crash intervals. Other factors-such as previous PDO crashes, young drivers, truck driving, hilly terrain, median barrier, wet roads, and darkness with lights unlit-exert significant but homogeneous effects. The latent class survival analysis model identifies two distinct latent classes (Latent Class 1 with class membership probability of 0.577 and Latent Class 2 with class membership probability of 0.423) at a 0.001 significance level. And substantial differences are observed in the effects of explanatory variables on accident recurrence intervals across the classes. For instance, previous PDO crashes shorten recurrence durations in Class 1 (-0.789) but lengthen them in Class 2 (0.788); similarly, the wet-road indicator reduces recurrence in Class 1 (-0.815) yet increases it in Class 2 (1.040). Furthermore, estimation results from both heterogeneity-based models indicate that most fixed-effect variables share consistent directional signs, However, the average effect size differs substantially. For example, the effect of the Low visibility indicator is 1.199 in the random-parameters model compared to 0.189 in Class 2 of the latent-class model, while darkness with lights unlit yields -0.526 and -0.250 (Class 2), respectively. By applying two advanced econometric modeling approaches, this study offers a novel theoretical perspective on capturing heterogeneity in inter-crash durations among repeatedly crash-involved drivers. Meanwhile, the key determinants identified through these models provide a robust scientific foundation for developing precision risk-intervention strategies, thereby supporting more targeted behavioral interventions for high-risk driver groups.