Understanding traffic collision severity's contributing factors: A mixed effect multinomial logistic regression and machine learning approaches

This study aims to understand the influence of various contributing factors on traffic collision severity. With a focus on variables such as pedestrian involvement, cyclist presence, motor vehicle roles, weather conditions, road characteristics, geographical contexts, and among others. The objective of this study is to shed light on the in-depth behavioral dynamics that underlie the severity of accidents. The dataset utilized in this study is retrieved for the Virginia Road department and contains over 500,00 data points with 18 different variables. This study utilized two statistical models and one machine learning model—Multinomial Logistic Regression, Multi-level (Mixed Effect) Multinomial Logistic Regression, which captures the group level heterogeneity, and Random Forest model—to analyze and understand the relationship between various factors and collision severity outcomes. The results show that the Multi-Level Multinomial Logistic Regression model overcomes the Multinomial Logistic Regression model. Moreover, the results show that the existence of vulnerable road users, including pedestrians and bikes would likely increase the odds of fatalities. The odds ratios for fatality and major injuries of collisions involving unbelted drivers are higher than 10, raveling the higher likelihood of sever outcomes compared to belted drivers. Collision occurs are traffic controls (e.g., signalized intersections) are likely to be more severe compared to collision occurs at regular road. These results were in alignment with what were reveled from the Random Forest model. Overall, these findings can help policymakers to design strategies that can reduce severity outcomes in different regions.