Spatial Analysis of Inter-city Crashes, Using Geographically Weighted Regression

In the recent years a considerable attention has been carried out to explore the relationship between crashes and related factors such as land use, demographic characteristics and network parameters in urban areas. Trip production and attraction as the first-hand results of trip generation models along with total road kilometers might be proper surrogate predictors to estimate the crashes in aggregate spatial units. Generalized Linear Models are known as the common techniques to investigate the relationship between crashes and predictors based on estimating the global fixed coefficients. Geographically Weighted Poisson Regression is suggested to investigate the non-stationary effects in relationship between crash frequencies and contributing factors. The model was run using the trip production and attraction parameters and total road kilometers over 253 Traffic Analysis Zones in Mashhad and the existence of spatial autocorrelation was examined among residuals. The goodness-of-fit tests indicate a considerable improvement in the model performance comparing the global models. The results also show that the significant non-stationary state in relationship between crash frequency and selected predictors. Based on the results, there is a direct and positive relationship between the independent variables and dependent variable; although the coefficient magnitudes vary over different spaces. The examination of residuals for global and local models reveals that the spatial autocorrelation has successfully been removed from the local residuals.