A two-stage joint modeling framework for evaluating the effectiveness of Midblock Pedestrian Signals (MPS)

This study evaluates the safety effectiveness of newly implemented Midblock Pedestrian Signals (MPS) at 14 locations across Florida. Data collection and processing: A total of 2,645 pedestrian-vehicle interactions were extracted from CCTV footage and processed using advanced computer vision techniques. Pedestrian-vehicle conflicts were categorized as serious, moderate, and non-conflicts based on the Relative Time to Collision (RTTC) measure.

To estimate the safety impact, a two-stage joint modeling framework was developed, addressing two key methodological challenges: potential selection bias due to non-random MPS assignment and temporal and baseline differences across sites. The selection model estimated the probability of MPS treatment using a Probit model, while the outcome model predicted conflict severity through a penalized multinomial logistic regression with integrated Difference-in-Differences (DiD)-style variables. Joint likelihood estimation corrected for selection bias by linking treatment assignment to outcome patterns within a unified likelihood structure.

The Average Treatment Effect (ATE) results demonstrated that MPS installations significantly improved pedestrian safety outcomes by reducing both moderate and serious conflicts across all control group comparisons, including locations with Pedestrian Hybrid Beacons (PHBs), Rectangular Rapid Flashing Beacons (RRFBs), and Flashing Beacons. The DiD analysis confirmed that these improvements were not merely driven by general time trends but were directly attributable to the MPS treatment itself.

The findings provide strong empirical support for transportation safety policies that prioritize MPS deployment at midblock crossings, suggesting that MPS can serve as an effective and practical alternative to traditional pedestrian crossing treatments under appropriate conditions. The proposed framework also offers a methodological foundation for evaluating non-randomized interventions and can inform future safety research, policy development, and data-driven signal implementation strategies.