Segment-Level Spatial Heterogeneity of Arterial Crash Frequency Using Locally Weighted Generalized Linear Models

Decades of literature in traffic crash modeling show the popularity of generalized linear models (GLMs). However, because of the failure to accommodate spatial heterogeneity, parameters estimated with these models are inconsistent and inefficient. In light of that, this study aims to investigate the spatial heterogeneity of crashes aggregated at roadway segment levels using geographically weighted Poisson regression (GWPR) and two variants of the geographically weighted negative binomial regression (GWNBR) model. The results indicate: (i) the GWNBR model with global dispersion parameter outperforms conventional Poisson, GWPR, and negative binomial (NB) models; (ii) the performance of the GWNBR model further enhances as the dispersion parameter becomes spatially non-stationary; (iii) tests of spatial heterogeneity and autocorrelation reveal the existence of non-stationarity and less than 1% likelihood of randomness; and (iv) median of parameter estimates reveal a positive association between crashes and posted speed limit, number of lanes, number of three-leg intersections, number of access points, and vehicle miles traveled (VMT). The study concludes that, when spatial heterogeneity is evident, conventional GLMs should be avoided to circumvent the futile estimation of parameters on the arterial segments. The findings are expected to contribute to the small pool of literature on spatial non-stationarity of parameters in segment-based aggregation, identification and selection of segments with similar influence factors in corridor-level studies, and spatial analysis of arterial crashes.