Modeling road-segment-level speeding risk of new energy vehicle taxis using a multistage framework with spatial spillover, endogeneity, and nonlinear effects.

Speeding poses a critical risk to urban traffic safety, especially for new energy vehicle taxis, whose capacity of high torque and rapid acceleration may exacerbate this risk. To investigate the speeding risk of new energy vehicle taxis, this study proposes a multistage analytical framework, which combines endogeneity-adjustment methods and machine learning to account for spatial spillovers, endogeneity, and nonlinear effects. The framework employs geographically weighted generalized propensity score matching to capture nearby road segments' spillover effects and balance observable endogeneity, and combines spatial instrumental variables with two-stage residual inclusion method to account for potential unobservable endogeneity. Besides, the study develops a generalized speeding risk index that incorporates speeding frequency, severity, and exposure to comprehensively measure the speeding risk at the road segment level. Further, three Extreme Gradient Boosting models with Shapley Additive Explanations values and Accumulated Local Effects plots are developed to capture the nonlinear relationships between speeding risk and four categories of influencing factors (i.e., traffic conditions, road characteristics, traffic management strategies, and points-of-interest density) and enhance model interpretability. The proposed framework is applied using GPS trajectory data of new energy vehicle taxis in Chengdu, China. Results show that the proposed framework significantly outperforms conventional geographically weighted regression and linear regression models. Medium traffic flow, moderate intersection spacing, 4-5-lane segments, low commercial points-of-interest density, and roads near charging stations are associated with higher speeding risk, whereas no-parking rules, higher residential points-of-interest density, physical separators, and higher speed-limit categories are associated with lower risk. The findings further indicate that traffic management strategies become more influential after accounting for spatial spillover effects and correcting endogeneity, highlighting the importance of structurally aware modeling for urban traffic safety analysis. The study provides theoretical insights and practical implications for urban traffic safety management in the era of vehicle electrification.