A NEW METHOD OF PREDICTION FOR SPATIAL REGRESSION-MODELS WITH CORRELATED ERRORS

This paper deals with minimum mean-squared error, unbiased linear interpolation of a continuous domain spatial process based on a sparse set of irregularly spaced observations. The process is assumed to be governed by a linear regression model with errors that follow a second-order stationary Gaussian random field. A new method of prediction is developed that is compatible with the parameter estimation procedures of Vecchia. The result is a new likelihood-based method for joint parameter estimation and prediction that can be applied to large or small data sets with irregularly spaced data. Simulated and observed data sets are analysed to illustrate the methods.