View in EDS

Can surrogate safety measures explain crash patterns at signalized intersections? evidence from network-wide connected-vehicle data using negative binomial, random forest, and diffusion GNN models.

Saved in:
Bibliographic Details
Title: Can surrogate safety measures explain crash patterns at signalized intersections? evidence from network-wide connected-vehicle data using negative binomial, random forest, and diffusion GNN models.
Authors: Nasri, Mehrdad1 (AUTHOR) mehnasri@uw.edu, He, Jingyi1 (AUTHOR) jhe0718@uw.edu, Karim, Muhammad Monjurul1 (AUTHOR) mmkarim@uw.edu, Zhang, Shucheng1 (AUTHOR) sz224@uw.edu, Wang, Yinhai1 (AUTHOR) yinhai@uw.edu
Source: Journal of Safety Research. Jun2026, Vol. 97, p553-564. 12p.
Subjects: Signalized intersections, Traffic safety, Safety, Prediction models, Random forest algorithms, Regression analysis, Graph neural networks
Geographic Terms: Tucson (Ariz.)
Abstract: • DRAC was the most robust surrogate for predicting intersection-level crash frequency. • DRAC and TTC patterns aligned with crash types involving braking and crossing conflicts. • XGBoost and diffusion GNN models captured nonlinear and spatial crash risk patterns. • Connected vehicle data supported scalable, real-time screening of intersection risk. • Findings inform countermeasures targeting abrupt deceleration and turning movements. Introduction: Connected-vehicle trajectories provide large-scale data for proactive, network-wide safety management, yet guidance on transforming these dense data streams into actionable indicators remains limited. This study presents a reproducible framework that extracts three surrogate safety measures (SSMs): critical deceleration rate to avoid a crash (DRAC), time-to-collision (TTC), and post-encroachment time (PET), from 54.8 million second-by-second trajectory points, mapped to 139 signalized intersections in Tucson, Arizona. Critical SSM counts were linked to five years of police-reported crashes at the same sites to quantify crash–SSM relationships. Method: Negative Binomial regression, Random Forest, XGBoost, and a diffusion graph neural network assessed how distributional assumptions, nonlinearities, and spatial context influence model performance. SSMs show the strongest association with low-severity crashes (no-injury and non-incapacitating), whereas severe outcomes require additional contextual and behavioral variables. In terms of collision type, rear-end and left-turn crashes are best explained by these trajectory-based surrogates, consistent with their underlying conflict mechanisms. Results: Results show that across every model family, DRAC is the most influential predictor. In count regressions, an additional one thousand DRAC events raises expected rear-end, left-turn, and angle crashes by 5 to 10%. DRAC ranks first in XGBoost models and pushes the coefficient of determination above 0.60 for rear-end and no-injury crashes. Permuting DRAC in the diffusion network model increases mean absolute error three- to sevenfold, whereas TTC has smaller effects and PET is negligible. Sideswipe and single-vehicle collisions are insensitive to all three surrogates, indicating the need for alternative indicators. Conclusions: These findings indicate that DRAC is a reliable network-wide screen for crash risk. TTC offers supplemental insight for turning and low-injury events, and PET adds little information at a one-second sampling rate. Practical Applications: The proposed framework demonstrates how large-scale connected-vehicle data, coupled with modern modeling techniques, can inform data-driven crash-prevention programs. [ABSTRACT FROM AUTHOR]
Copyright of Journal of Safety Research is the property of Pergamon Press - An Imprint of Elsevier Science and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:• DRAC was the most robust surrogate for predicting intersection-level crash frequency. • DRAC and TTC patterns aligned with crash types involving braking and crossing conflicts. • XGBoost and diffusion GNN models captured nonlinear and spatial crash risk patterns. • Connected vehicle data supported scalable, real-time screening of intersection risk. • Findings inform countermeasures targeting abrupt deceleration and turning movements. Introduction: Connected-vehicle trajectories provide large-scale data for proactive, network-wide safety management, yet guidance on transforming these dense data streams into actionable indicators remains limited. This study presents a reproducible framework that extracts three surrogate safety measures (SSMs): critical deceleration rate to avoid a crash (DRAC), time-to-collision (TTC), and post-encroachment time (PET), from 54.8 million second-by-second trajectory points, mapped to 139 signalized intersections in Tucson, Arizona. Critical SSM counts were linked to five years of police-reported crashes at the same sites to quantify crash–SSM relationships. Method: Negative Binomial regression, Random Forest, XGBoost, and a diffusion graph neural network assessed how distributional assumptions, nonlinearities, and spatial context influence model performance. SSMs show the strongest association with low-severity crashes (no-injury and non-incapacitating), whereas severe outcomes require additional contextual and behavioral variables. In terms of collision type, rear-end and left-turn crashes are best explained by these trajectory-based surrogates, consistent with their underlying conflict mechanisms. Results: Results show that across every model family, DRAC is the most influential predictor. In count regressions, an additional one thousand DRAC events raises expected rear-end, left-turn, and angle crashes by 5 to 10%. DRAC ranks first in XGBoost models and pushes the coefficient of determination above 0.60 for rear-end and no-injury crashes. Permuting DRAC in the diffusion network model increases mean absolute error three- to sevenfold, whereas TTC has smaller effects and PET is negligible. Sideswipe and single-vehicle collisions are insensitive to all three surrogates, indicating the need for alternative indicators. Conclusions: These findings indicate that DRAC is a reliable network-wide screen for crash risk. TTC offers supplemental insight for turning and low-injury events, and PET adds little information at a one-second sampling rate. Practical Applications: The proposed framework demonstrates how large-scale connected-vehicle data, coupled with modern modeling techniques, can inform data-driven crash-prevention programs. [ABSTRACT FROM AUTHOR]
ISSN:00224375
DOI:10.1016/j.jsr.2026.04.008