Crash Severity Risk Modeling Strategies under Data Imbalance

This study investigates crash severity risk modeling strategies for work zones involving large vehicles (i.e., trucks, buses, and vans) when there are crash data imbalance between low-severity (LS) and high-severity (HS) crashes. We utilized crash data, involving large vehicles in South Carolina work zones for the period between 2014 and 2018, which included 4 times more LS crashes compared to HS crashes. The objective of this study is to explore crash severity prediction performance of various models under different feature selection and data balancing techniques. The findings of this study highlight a disparity between LS and HS predictions, with less-accurate prediction of HS crashes compared to LS crashes due to class imbalance and feature overlaps between LS and HS crashes. Combining features from multiple feature selection techniques: statistical correlation, feature importance, recursive elimination, statistical tests, and mutual information, slightly improves HS crash prediction performance. Data balancing techniques such as NearMiss-1 and RandomUnderSampler, maximize HS recall when paired with certain prediction models, such as Bayesian Mixed Logit (BML), NeuralNet, and RandomForest, making them suitable for HS crash prediction. Conversely, RandomOverSampler, HS Class Weighting, and Kernel-based Synthetic Minority Oversampling (K-SMOTE), used with certain prediction models such as BML, CatBoost, and LightGBM, achieve a balanced performance, defined as achieving an equitable trade-off between LS and HS prediction performance metrics. These insights provide safety analysts with guidance to select models, feature selection techniques, and data balancing techniques that align with their specific safety objectives, offering a robust foundation for enhancing work-zone crash severity prediction.