Stable factors associated with traffic conflict risk in expressway weaving segments: A causal modeling framework

This study aims to identify stable factors associated with traffic conflict risk in expressway weaving segments, with a particular focus on addressing the challenge of unobserved data distribution bias between training and test datasets, which can compromise model reliability. To mitigate distribution bias and enhance result robustness, a causally regularized logistic model (CRLM) with a global causal regularizer was employed. To validate the stability of the CRLM, multi-dataset validation and model parameter consistency tests were conducted using five datasets collected from the field and simulation in two weaving types. Meanwhile, classic logistic regression (LR) and eXtreme Gradient Boosting (XGBoost) were developed for comparison. In the multi-dataset validation test, the average area under receiver operating characteristic curve (AUC) of the CRLMs is close to that of the XGBoosts, but with a lower standard deviation, suggesting that the CRLM provides more stable predictive performance across different combinations of training and testing datasets. In the model parameter consistency test, the CRLM can identify more stable factors across heterogeneous traffic environments. Furthermore, the causal mechanisms underlying traffic conflict risk in Type A and Type B weaving segments are distinct. The hazardous traffic flow characteristics for each weaving type were discussed in detail. These findings provide a novel and robust methodological framework for traffic conflict risk analysis. In addition, the model results have practical implications for developing proactive traffic control strategies and enhancing automated driving systems (ADS) to improve traffic safety in expressway weaving segments.