Examining macro-level traffic crashes considering nonlinear and spatiotemporal spillover effects.

Understanding the impacts of traffic crashes is essential for safety management and proactive safety protection. Current studies often hold the assumption of linearity and spatial dependence, which may lead to underestimated results. To address these gaps, this study considers both nonlinear and spatiotemporal spillover effects to explore the intricate relationships between vehicular crashes and their influencing factors at a macro level. Spatiotemporal spillover effects are captured by creating exogenous variables from neighboring zones and their historical status through a geographically and temporally weighted method. Then, the extracted spillover factors are combined with factors from internal zones to construct independent variables. Their nonlinear characteristics are modeled by the gradient boosting decision trees model and interpreted through accumulated local effect plots. A case study was conducted in New York City spanning four years from 2016 to 2019, considering six categories of influencing factors: street view imagery, exposure, land use, points of interest, traffic network, and socioeconomic attributes. The experimental results demonstrate that model performance is improved by incorporating nonlinear and spatiotemporal spillover effects. Additionally, the proposed model highlights the significant nonlinear effects of factors including mixed land uses, sidewalks, and junction density, and emphasizes the presence of spatiotemporal spillover effects, such as building density, bike parking density, and education attainment. These findings offer insightful implications for transportation practitioners and policymakers to devise safety countermeasures and policies, emphasizing the importance of collaboration across neighboring urban regions.