An Extension of Geographically Weighted Regression with Flexible Bandwidths

Geographically weighted regression (GWR) (Brunsdon et al. 1996; Fotheringham et al. 2002) is a useful technique for modelling local spatial relationships between variables. The essential idea of GWR is that observations near to a model calibration point have more influence in the estimation of regression coefficients than observations farther away do. The standard GWR model employs a single bandwidth to control the distance-decay in this influence. In practice however, such a uniform bandwidth may not be sufficient in reflecting complex spatial variations in relationships between dependent and independent variables. In an attempt to produce a more realistic model, this paper develops an extension to GWR, where flexible bandwidths are found providing coefficient surfaces that vary at different spatial scales. Experiments are carried out on simulated datasets to test the model.