Treatment Effectiveness Evaluation of Mandatory Pre-Right-Turn Stops for Large Vehicles at Signalized Intersections: Combining Causal Inference with the Wasserstein Generative Adversarial Network with Gradient Penalty Method

Abstract

In China, more than 30% of intersection traffic crashes are related to right-hand turns, with large vehicle crashes consistently being the most prevalent. Novel countermeasures termed mandatory pre-right-turn stops for large vehicles at signalized intersections were widely implemented. Understanding their effects is critical for improving road safety. Traditional statistical approaches are affected by confounding bias, and because of the problem of overfitting, causal machine learning methods often fail to produce accurate results in small sample size . To overcome these challenges, a novel treatment effect evaluation system that combines a causal inference method with a data generation method is proposed in this paper. The Wasserstein generative adversarial network with gradient penalty (WGAN-GP) is used to generate additional control group data, thereby increasing the sample size. Findings demonstrate that the WGAN-GP-generated data closely match real data. The effects of countermeasures are evaluated using dou-bly robust estimation with neural networks (DR-NNs). This method effectively models nonlinear relationships between variables and addresses confounding bias. Both the combination of a right-turn stop sign for large vehicles and right-turn danger zones, and the right-turn stop sign alone, significantly reduce right-turn crashes, with crash modification factors (CMFs) of 0.53 and 0.56, respectively. The analysis of heterogeneous treatment effects indicates that the combination is less effective at skewed intersections, and intersections without bicycle lanes show higher CMFs for the right-turn stop sign alone. These results underscore the importance of bicycle lanes and highlight the need for improved implementation of right-turn danger zones at skewed intersections.