As global measures of spatial association have been supplemented by local indicators, Fotheringham, Charlton, and Brunsdon (1996, 1997) and Brunsdon, Fotheringham, and Charlton (1996) have been developing weighting schemes to allow possible differences in local parameter estimates for regression models to be revealed. Moving from the global to local settings, one would perhaps expect the local parameter estimates to vary, but within the bounds of their global standard error based confidence intervals, that is with divergences of more than 
less than five times in a hundred. The weighting scheme used so far is distance based, weighting zone i with unity, and with weights declining with increasing distance from i. There are similarities with kernel regression techniques, although these use weighting in attribute space, rather than across the observations. Currently, cross-validation is used to select an appropriate global bandwidth parameter, which then determines the form of the distance decay function used to define the weights for each observation. There are clearly substantial difficulties involved in making statistical inferences from results of this kind of procedure, although it has proved very useful in showing up missing variables.