Quantitative causality assessment between traffic states and crash risk in freeway segments with closely spaced entrance and exit ramps

Segments with closely spaced entrance and exit ramps are recognized as crash black-pots on freeways. Accurately identifying hazardous traffic states before crash occurrence can help implement targeted control measures, thereby reducing crash risk in such segments. As such, this study developed a structural causal model (SCM) to explore the causality between traffic states and crash risk in segments with closely spaced entrance and exit ramps. Firstly, high-resolution traffic flow data related to three types of lane configurations and three crash types were collected. A deep clustering based on generative adversarial networks (clusterGAN) was developed to classify traffic flow into seven states. Secondly, the causal graph connecting traffic states and crash risk was generated. Backdoor adjustment from the SCM was used to identify seven confounding variables (e.g. traffic volume, ramp volume ratio, speed difference on the inside lanes between the beginning and end of the segment, the shoulder and median width, etc.), which influence both traffic states and crash risk. The inverse-probability-weighted regression adjustment estimator was adopted to estimate the average causal effects of traffic states on crash risk conditional on different lane configurations. The results suggest that hazardous traffic states in each lane configuration are different, even for the same crash types. In addition, the comparison results with the standard logistic models suggest that ignoring confounding effects may cause unreasonable conclusions and biased estimation. These results have the potential to help advanced traffic management systems develop proactive traffic control strategies to improve traffic safety in freeway segments with closely spaced entrance and exit ramps.