A double machine learning causal inference framework for estimating the impact of speed limits on traffic operational risks

Accurately assessing the causal effects of speed limit policies on traffic operational risks is critical for enhancing freeway safety. While existing studies have relied on theoretical models or driving simulations, their associational analyses often suffer from model bias and fail to provide robust causal interpretations. This study introduces a dual machine learning (DML) causal inference framework to address this gap. First, integrate aggressive driving behavior and speed fluctuations into a Traffic Order Index (TOI). Second, use the conditional Shapley value index for confounding factor screening. Finally, construct a DML causal forest model to precisely quantify the conditional average treatment effects of speed limit signs on traffic operational risks. Empirical analysis of multi-source traffic data in Shandong Province shows superior estimation accuracy of the proposed method compared to baseline models. On average, the Lane-specific and vehicle type-specific Speed Limit Signs(La_SLS) (ATE = 0.190, p < 0.01) are more effective in improving traffic order than the Traditional Speed Limit sign (Tra_SLS) (ATE = 0.168, p < 0.01). This effect, however, is highly heterogeneous. The perceived safety benefit of La_SLS is most pronounced at interchanges and under adverse weather conditions. Counterfactual analysis reveals La_SLS in Section 34 boosts TOI by 3.09%, while Tra_SLS in Section 33 improves it by 2.37%.