Highway accidents are the result of a complex interaction of a diverse an‎d often hidden set of factors, many of which are not recorded in official data, especially police reports. This lack of data availability limits the power of traditional predictive models an‎d highlights the need to develop frameworks that can capture the effects of unobserved heterogeneities. This study presents an innovative analytical framework for analyzing the duration between traffic incidents an‎d predicting the occurrence of accidents by taking into account unobserved heterogeneities. To this end, accident data were first classified into distinct clusters using a latent class cluster analysis model based on influential variables. The dataset was then transformed into longitudinal data by discretizing time into 15-minute intervals. Finally, by considering the cluster membership variable as a heterogeneity control factor, parametric duration models were estimated for two separate groups of accidents with property damage an‎d accidents with injuries an‎d fatalities to identify the factors affecting the time to the next accident. The experimental results showed that this framework outperformed conventional methods an‎d was able to significantly increase the prediction accuracy. In addition, the analysis of key variables shows that accidents with property damage are more related to the structure an‎d complexity of the road, while accidents with injuries an‎d fatalities are strongly affected by critical environmental an‎d climatic conditions. The findings of this research can help develop intelligent traffic management systems an‎d make efficient preventive decisions to reduce accidents on the highway network.

احمدرضا طالبيان

محمدمهدي بشارتي , سينا صاحبي