Tree-based approaches to understanding factors influencing crash severity across roadway classes: A Thailand case study

Abstract

Existing studies often overlook the nuanced differences between various road classifications and their respective crash dynamics, hindering the development of targeted interventions to mitigate crash severity. To address this gap, this study investigates factors influencing the likelihood of fatality in road crashes across highways, collector roads, and local roads in Thailand using crash data from 2015 to 2021. Highways connect regions with high-speed traffic and large volumes, collector roads link smaller communities with lower traffic density but allow higher speeds, and local roads primarily pass through villages, with narrow pathways, two traffic lanes, and frequent motorcycle use. The study employs machine learning methodologies utilizing tree-based algorithms, including Decision Trees, Random Forest, Gradient Boosting, AdaBoost, Extra Trees, XGBoost, LightGBM, and CatBoost. The XGBoost model delivered superior performance for highways, while Gradient Boosting slightly outperformed XGBoost for local and collector roads. Both models consistently achieved a test accuracy of 0.70, with precision between 0.66 and 0.67, recall ranging from 0.59 to 0.61, and F1-scores from 0.58 to 0.61. The AUC values also consistently ranged from 0.59 to 0.61. SHAP values reveal key factors influencing fatality risk across road types, including speeding, gender disparities, driving under the influence of alcohol, inadequate lighting, and elderly drivers. Specific concerns include reversing on highways, collisions in poorly lit areas on collector roads, and helmet non-use on local roads. The findings support policy recommendations to address speeding, target male and older drivers, prevent reversing incidents, enhance lighting, and promote helmet use. This research deepens our understanding of factors affecting road crash severity and offers valuable insights for improving road safety across various environments.