Geographically weighted regression bandwidth selection and spatial autocorrelation: an empirical example using Chinese agriculture data

Seong-Hoon Cho, The University of Tennessee, Knoxville
Dayton M. Lambert, The University of Tennessee, Knoxville
Zhuo Chen, Centers for Disease Control and Prevention

Abstract

This research note examined the performance of Geographically Weighted Regression (GWR) using two calibration methods. The first method, Cross Validation (CV), has been commonly used in the applied literature using GWR. A second criterion selected an optimal bandwidth that corresponded with the smallest spatial error Lagrange Multiplier (LM) test statistic. We find that there is a tradeoff between addressing spatial autocorrelation and reducing degree of extreme coefficients in GWR. Although spatial autocorrelation can be controlled for by using the LM criterion, a substantial degree of extreme coefficients may remain. However, while the CV approach appears to be less prone to producing extreme coefficients, it may not always attend to the problems that arise in the presence of spatial error autocorrelation