Geographically-Weighted Regression Models for Improved Predictability of Urban Intersection Vehicle Crashes

Most of current intersection vehicle crash models are calibrated using global regression analysis methods that are often inaccurate in crash predictions as some localized crash contributing effects are not explicitly addressed. This paper employs the Geographically-Weighted Regression (GWR) technique to calibrate statistical models for predicting intersection injury, property damage only (PDO), and total crashes using data on 245 intersections in City of Chicago for period 2001–2008. In the calibrated GWR models, factors contributing to intersection vehicle crashes identified include major and minor road daily traffic, number of major and minor road through and left-turn lanes, and household income level. The analysis of variance (ANOVA) test reveals that improved model predictability is achieved from all crash models developed using the GWR technique compared with those models calibrated based on the Ordinary Least Squares (OLS) technique. The Monte Carlo test identifies significance of spatial variability of explanatory variables in the GWR models.