Safety evaluation of protected bike Lane treatments at Intersections: A causal framework

Intersections are a critical focus in bicycle safety research, as approximately one-thirds of bicycle-related crashes occur at these locations. Although protected bike lanes (PBL) at intersections, such as Lateral Shift and Bend-out treatments have been implemented, there is limited crash-based research on their safety performance. Furthermore, the prevailing use of before-after study designs for safety evaluation makes this approach susceptible to selection bias. To address this issue, this study proposes a causal inference framework that combines the advanced generalized causal random forest (GRF) and multimodal large language model (LLM). The LLM is used to extract contextual features from street view images, improving control over unobserved confounding bias. The GRF model is used for effectiveness evaluation by addressing selection bias through residual-based orthogonalization of treatment and outcome. The framework was applied to evaluate the safety impacts of Bend-out and Lateral Shift treatments at intersections. The results indicate that the proposed method outperforms both the baseline and comparative models across all metrics. The average treatment effect (ATE) of Lateral Shift treatments is 1.35 for total crashes and 1.21 for bicycle crashes, suggesting that these treatments tend to increase crashes. For Bend-out treatments, the ATE is -1.61 for total crashes and -0.55 for bicycle crashes, corresponding to a 32.2% reduction in total crashes and a 22.4% reduction in bicycle crashes. Analysis of road user behavior reveals that for Lateral Shift treatments, the low rate of drivers yielding to cyclists is a major issue, with only 30.7% of drivers yielding. To effectively implement Lateral Shift treatments, strengthening enforcement measures should be considered. Furthermore, riding in the wrong direction is a potential risk for both Lateral Shift and Bend-out treatments.