Road traffic collisions present a significant challenge in developing nations, largely attributable to factors such as vehicular congestion, ex-potential increases in population and motor vehicle density, inadequate driver training, urbanization, and other infrastructural and traffic-related deficiencies. The resolution of these challenges has been the focus of comprehensive research efforts, with a multitude of studies already concluded and several others currently in progress. In efforts to mitigate traffic accidents, various feature selection models or algorithms have been employed to isolate the critical factors contributing to accident frequency and severity. In this study, several methods including filter-based approaches, tree-based embedded models, and recursive feature elimination (RFE) have been utilized on the US (United States) Accidents: a countrywide traffic accident dataset to identify key determinants of accident severity. Performance comparisons reveal that the tree-based embedded method outperforms other approaches, demonstrating a 4.575% enhancement in accuracy and a 43.695% reduction in model training time when applied in conjunction with random forest algorithms. By contrast, the filter method (ANOVA) results in a 0.784% improvement in accuracy and a 5.77% decrease in training time, while the RFE method yields a marginal 0.075% accuracy gain accompanied by a 3.36% increase in training time. These results indicate that ensemble-based tree models exhibit superior performance over both wrapper and filter-based methodologies.