Considering spatial heterogeneity in modeling taxi speeding frequency: An advanced geographically weighted negative binomial regression approach

Speeding is a major contributor to traffic crashes, posing significant risks to drivers, passengers, and other road users. This issue is exacerbated among urban taxi drivers due to factors such as time pressure for passenger pickups, potentially leading to higher speeding frequency compared to other drivers. While prior studies have identified factors influencing speeding, they often assume spatial stationarity, overlooking how these effects may vary geographically. This gap limits targeted interventions. To address spatial heterogeneity in count data with potential overdispersion, two geographically weighted negative binomial models are developed: a standard Geographically Weighted Negative Binomial (GWNBR) model to account for overdispersion, and an enhanced Geographically Weighted Negative Binomial with spatially varying dispersion (GWNBRg) that further incorporates spatially varying dispersion parameters. This suite captures both local coefficient variation and unobserved heterogeneity. To optimize the spatial resolution of regression results and capture local heterogeneity in speeding behaviors, five kernel function methods (i.e., Gaussian, Exponential, Bi-square, Tricube, and Boxcar) are adopted for adaptive bandwidth calculation. Further, a sensitivity analysis of Traffic Analysis Zones (TAZs) is conducted to identify the optimal zoning scale. The study uses over 15 million GPS trajectory data points collected from September 1 to 14, 2020, in Chengdu, China, and identifies 298,114 speeding incidents. Results indicate that the GWNBRg outperforms GWNBR in terms of Akaike Information Criterion and residual Moran's I, confirming its robustness in handling overdispersion and spatial dependence. Global estimates show that several factors significantly increase taxi speeding frequency, such as total travel distance, bicycle lanes, on-street parking management, no parking rules, and the density of public service points of interest (POI). Conversely, median dividers, central fences, overpasses and tunnels, and residential POI density decrease taxi speeding frequency. The localized coefficient maps of these factors exhibit spatial heterogeneity across TAZs, which tend to cluster regionally when estimated with the same sign. This spatial clustering suggests that regional context drives most of the spatial heterogeneity. Additionally, the spatially varying overdispersion parameter in GWNBRg further identifies zones with unobserved risk factors, highlighting its utility for precision interventions. This framework enables policymakers to propose spatially targeted strategies for taxi speeding mitigation.