Abstract
Spatial estimators usually provide lower prediction errors than their aspatial counterparts. However, most of the standard techniques require a large number of operations. Fortunately, for a given observation only a relatively small number of nearby observations typically exhibit correlated errors. This means that most of the elements of the n by n spatial matrices are zero. The use of sparse matrix techniques can dramatically lower storage requirements and reduce execution times. In addition, adopting a first differencing model allows the use of GLS which avoids the necessity of evaluating an n by n determinant. This also greatly reduces computational costs.