Investigating the risk factors of motorcycle crash injury severity in Pakistan: Insights and policy recommendations

A large number of fatalities and severe injuries are caused by motorcycle crashes worldwide, particularly in developing countries. More than 50% of crashes in Pakistan involve motorcycles. To analyze motorcycle crash severity, various models, including both statistical and machine learning methods, have been applied. Researchers have widely acknowledged that machine learning methods provide superior prediction performance but have weaker interpretability power. However, no study has investigated the consistency of risk factors identified by the two streams of models. The consistency of the findings between these two kinds of methods is vital to improve the interpretability power of machine learning methods for policymaking in an era with more and more applications in the area of traffic safety. This study aims to narrow this research gap by comparing the consistency of crash severity risk factors identified by statistical models and machine learning methods. The study analyzes motorcycle crashes in Rawalpindi city of Pakistan. Multinomial logit model (MNL) and three machine learning models, i.e., the random forest (RF), naive Bayes, and gradient-boosted trees methods, are used to analyze the prediction performance and identify risk factors. The results show that the RF model, with an overall accuracy of 86.7%, outperformed other models. The SHapley Additive exPlanations (SHAP) method was adopted to explore the interpretability of machine learning methods. It was found that the contributing factors to crash injury severity identified by the RF method, such as distracted driving, collisions involving pedestrians, collisions involving a truck, and female riders, are consistent with those determined by the MNL model. These results have clear implications for developing cost-effective safety countermeasures to improve motorcycle safety in Pakistan.