Estimate traffic cyclist crashes using Poisson-Tweedie models

Abstract

Cyclist safety remains a critical issue in urban transportation, where infrastructure configuration and spatial dynamics play a key role in crash occurrence. This study estimates cyclist crash frequencies in Lisbon between 2015 and 2019 using Poisson-Tweedie models, which are well-suited for overdispersed count data. A total of 541 cyclist crashes were analyzed, spatially structured into 250 × 250 meter grid cells and supplemented with covariates such as road length, intersection types, and various cycling infrastructure elements. Two models were developed: a base model with aggregated variables and a disaggregated model distinguishing road types, intersection forms, and cycleway categories. Both models incorporated spatial autocorrelation to account for neighboring effects. The key findings indicate that intersection density and road length are strongly associated with crash frequency, while cycleway length has a more modest yet significant effect. The disaggregated model offers greater interpretability but does not outperform the base model in predictive accuracy or goodness-of-fit, suggesting that a simpler specification may be more effective for policy applications. Elasticity analysis revealed that intersections have the greatest influence on crash risk, followed by road length and cycleways. Spatial predictions aligned with observed crash clusters and highlighted latent high-risk zones, reinforcing the model’s utility for proactive safety planning. The study concludes that improving intersection design is likely to yield greater safety benefits than merely increasing cycling infrastructure length. These results provide actionable insights for data-driven urban mobility planning and emphasize the value of predictive modeling tools for cyclist safety management.