A POISSON REGRESSION APPROACH TO MODELING TRAFFIC ACCIDENT FREQUENCY IN URBAN AREAS

Abstract

Urban traffic safety research frequently models crash frequency using Poisson-family regression because it is interpretable, extensible, and policy relevant. This systematic review synthesizes how these models are specified, diagnosed, validated, and translated into practice for urban contexts. Following PRISMA protocols, we searched multidisciplinary and transportation databases from inception through 2022, applied dual-stage screening with explicit inclusion criteria, extracted standardized methodological and results fields, and appraised reporting quality; 110 studies met all criteria and were included. The modeling landscape shows a clear center of gravity around the Poisson family, with negative binomial as the most common primary specification at 40.0 percent, followed by canonical Poisson at 29.1 percent, zero-inflated or hurdle variants at 12.7 percent, Poisson-lognormal or multivariate forms at 7.3 percent, and mixed or spatial CAR/ICAR primaries at 5.5 percent each, while Poisson appears as a baseline in most studies. Practice quality is uneven. Offsets are specified in 83.6 percent of papers and are associated with stronger validation and more stable inference, yet only 35.5 percent report any out-of-sample validation and calibration plots appear infrequently. Across covariates, higher speeds, turning shares, and access density typically increase risk, whereas medians and coordinated signals are protective; pedestrian and cyclist volumes often exhibit safety-in-numbers curvature. Translation to policy commonly occurs through Safety Performance Functions paired with empirical-Bayes adjustment; among studies reporting re-ranking, median turnover in top-site lists is about 30 percent, underscoring the operational impact of correct EB use and calibration. Overall, the evidence supports a disciplined workflow: construct mechanism-matched offsets, diagnose dispersion, zeros, and dependence, escalate model complexity only when warranted, and pair likelihood-based selection with explicit predictive checks to ensure credible, auditable safety decisions in cities.