Given that correlation-based evidence is insufficient for designing safety performance countermeasures, identifying causal relationships and treatment effect heterogeneity across traffic conditions remains critical for targeted intervention development. This paper proposes a heterogeneous causal inference framework to investigate the heterogeneous treatment effects (HTE) of left-turn control mode on left-turn conflicts. Left-turn conflicts are identified using post-encroachment time and conflict speed thresholds. A heterogeneous causal graph is constructed to identify causal relationships, followed by Pearson correlation analysis to address redundancy among candidate confounders. Forest Doubly Robust Learning models quantify the HTE of Permissive-only and Protected-Permissive Left-Turn (PPLT) control modes using data from Bellevue, Washington (September 13-19, 2019). Results show PPLT control yielded 0.272 additional conflicts/h relative to Protected-only (compared with 0.510 for Permissive-only), with narrower confidence intervals suggesting more stable effects. Moreover, the main effects and interaction effects of specific safety-related factors are thoroughly analyzed based on Generalized Additive Model (GAM). Main effects analysis reveals both left-turn volume and opposing through volume significantly influence conflict frequency, with nonlinear interaction effects demonstrating that Permissive-only control exhibits rapid conflict escalation at opposing through volumes exceeding 400 (veh/h/lane), regardless of left-turn demand. PPLT effectively manages left-turn volumes below 200 (veh/h) even under high opposing through volumes (>800 veh/h/lane). In practice, comparisons with existing traffic engineering handbooks reveal discrepancies between current warrant criteria and empirically derived safety thresholds. These findings provide preliminary, jurisdiction-specific evidence that agencies may consider when reviewing and refining local left-turn phasing warrants.
Xu et al. (Wed,) studied this question.