Nonparametric prediction for univariate spatial data: methods and applications

Abstract

We introduce five nonparametric kriging-type predictors for spatial data where only the variable of interest, without covariates, is recorded. The proposed methods seek to fully exploit the information contained in the spatial closeness and also in the similarity between neighbourhoods of the variable of interest. This is managed using different combinations of kernels (one or two kernels), and different combinations of distances (multiplicative and additive). The good performance of the proposed methods is shown via simulation studies and housing price prediction applications.