Investigating ?Optimal? Kriging Variance Estimation Using an Analytic and a Bootstrap Approach

Kriging is an interpolation technique for predicting unobserved responses at target locations from observed responses at specified locations. Kriging predictors are best linear unbiased predictors (BLUPs) and the precision of the BLUP is assessed by the mean square prediction error (MSPE), commonly known as the kriging variance. Both the BLUP and the MSPE depend on the covariance function describing the spatial correlation between locations and on specific parameters. The parameters are usually treated as known, whereas in practice they invariably have to be estimated and the empirical BLUP (that is, the EBLUP) so obtained. The empirical or estimated mean square prediction error (EMSPE), or the so called ?plug-in? kriging variance estimator, underestimates the true kriging variance of the EBLUP, at least in general. In this paper five estimators for the kriging variance of the EBLUP are considered and compared by means of a simulation study in which a Gaussian distribution for the responses, an exponential structure for the covariance function, and three levels of spatial correlation ? weak, moderate, and strong ? are adopted. The Prasad-Rao estimator obtained using restricted or residual maximum likelihood (REML) is recommended for moderate and strong spatial correlation and the Kacker-Harville estimator for weak correlation in the random fields.