Comparative Analysis of Parametric and Non-Parametric Data-Driven Models to Predict Road Crash Severity among Elderly Drivers Using Synthetic Resampling Techniques

As the global elderly population continues to rise, risk of severe crashes among drivers has become a pressing concern. This study presents comprehensive examination crash severity this demographic, employing machine learning models and data gathered from Virginia, United States America, between 2014 2021. The analysis integrates parametric models, namely logistic regression linear discriminant (LDA), as well non-parametric like random forest (RF) extreme gradient boosting (XGBoost). Central is application resampling techniques, specifically, over-sampling examples (ROSE) synthetic minority technique (SMOTE), address dataset’s inherent imbalance enhance models’ predictive performance. Our findings reveal that inclusion these techniques significantly improves power notably increasing true positive rate for prediction 6% 60% geometric mean 25% 69% in regression. Likewise, SMOTE resulted notable improvement performance, leading increase 8% 36% XGBoost. Moreover, established superiority over counterparts when balanced are utilized. Beyond modeling, delves into effects various contributing factors on severity, enhancing understanding how influence road safety. Ultimately, underscore immense potential analyzing complex data, pinpointing heighten informing targeted interventions mitigate risks driving.