On Sample Reuse Methods for Spatial Data

We describe a resampling method for constructing distribution estimators, and hence for calculating confidence intervals, in the context of statistics computed from nonreplicated spatial data. Our method is related to the spatial block bootstrap, but differs in that full spatial pattern is not actually simulated. Instead, an algorithm is employed, as a first step, to compute distribution estimators in the special case of data from a subset of the observation region. In the second step these estimators are recalibrated, using a device based on mixtures of distributions, to produce distribution estimators for statistics computed from the full data set. An empirical method is suggested for selecting the appropiate subset of the observation region for the first step of the algorithm. A numerical application of the technique is illustrated.