In this study, we investigate the performance of a non-Euclidean distance metric in calibrating a Geographically Weighted Regression (GWR) model with a simulated data set. Random predictor variable and spatially varying coefficients are generated on a square grid of size 2020. We respectively apply Manhattan and Euclidean distance metrics for the GWR calibrations. The preliminary findings show that Manhattan distance performs significantly better than the traditional choice for GWR Euclidean distance. In particular, it outperforms in the accuracy of coefficient estimates. © 2012 IEEE.