Using Geographically Weighted Poisson Regression for county-level crash modeling in California

Development of crash prediction models at the county-level has drawn the interests of state agencies for forecasting the normal level of traffic safety according to a series of countywide characteristics. A common technique for the county-level crash modeling is the generalized linear modeling (GLM) procedure. However, the GLM fails to capture the spatial heterogeneity that exists in the relationship between crash counts and explanatory variables over counties. This study aims to evaluate the use of a Geographically Weighted Poisson Regression (GWPR) to capture these spatially varying relationships in the county-level crash data. The performance of a GWPR was compared to a traditional GLM. Fatal crashes and countywide factors including traffic patterns, road network attributes, and socio-demographic characteristics were collected from the 58 counties in California. Results showed that the GWPR was useful in capturing the spatially non-stationary relationships between crashes and predicting factors at the county level. By capturing the spatial heterogeneity, the GWPR outperformed the GLM in predicting the fatal crashes in individual counties. The GWPR remarkably reduced the spatial correlation in the residuals of predictions of fatal crashes over counties.