Macro-level geospatial determinants of E-bike crash risk: town-level analysis

Electric bicycles have become a dominant mode of transportation in many Chinese cities, including Guangzhou, raising growing concerns about traffic safety resulting in fatalities and injuries among riders. This study performs a macro-level geospatial analysis to explore the spatial variability of factors influencing town-level counts of fatal and injury (FI) and property damage only (PDO) e-bike crashes across 170 towns in Guangzhou, which in this study represent spatial patterns in crash risk at the town level. One year of e-bike crashes and traffic violation data, integrated with environmental and infrastructural variables, such as population density, road network characteristics, land use diversity, and public facility densities, was used to examine spatial associations with FI and PDO crash counts. Spatial autocorrelation and Local Indicators of Spatial Association (LISA) analysis revealed significant clustering, with 64 towns (37.6%) identified as high-risk hotspots. To explore spatial heterogeneity in predictor associations, three models were applied: Ordinary Least Squares (OLS), Geographically Weighted Regression (GWR), and Multiscale Geographically Weighted Regression (MGWR). Within this exploratory Gaussian comparison, MGWR yielded the lowest AICc and highest adjusted R 2 ; however, most predictors exhibited near-global, spatially stationary associations, with localized variation observed only for selected predictors. Count-data diagnostics confirmed substantial overdispersion; therefore, geographically weighted negative binomial regression with a global dispersion parameter (GWNBRg) was applied, yielding deviance pseudo-R 2 values of 0.356 for FI and 0.251 for PDO crashes. Positive Gaussian associations for population density, traffic violations, and intersection density were not consistently corroborated by the negative binomial analyses. By contrast, higher school and healthcare facility densities were negatively associated with FI crashes in the global log-transformed negative binomial analysis (p = 0.014 and 0.027) and received greater nominal local GWNBRg support (29.4% and 17.1% of towns, respectively). These associations are observational and noncausal. These exploratory findings provide a preliminary data-driven basis for evidence-informed e-bike safety planning while highlighting that most predictors exhibit spatially stationary associations across towns.