Explaining road speeds with spatial lag and spatial error regression models

Abstract

Spatial regression is applied to commonly available planning data to build an economical predictive model of link speeds on a road network.. The method is described and validity is briefly assessed. GPS floating car measurements in the Canton Zurich, matched to a network model, are explained by link type, time of day, and spatial structure (population and employment density) using weighted least squares and spatial autocorrelation and spatial error terms. Weighted least squares corrects for heteroscedastic variance of the speeds by road type. Two different types of spatial neighborhoods were investigated for their suitability in correcting for spatial correlations between links: one based on a nearest neighbor criterion using Euclidean distance, and a second based on network distance, defined by the number of intersections between links. The significant spatial correlation coefficients estimated using either type of neighborhood indicate the presence of both correlated spatial error and autocorrelation of speeds. The best-fit models for the two types of neighborhood have different coefficient estimates, and the neighborhood based on network distance provides higher log-likelihood and adjusted R-square. Speed predictions are made against a holdout sample for validation. The performance and sensitivity indicate that this is a promising approach for monitoring the road system.