Geostatistical modelling for spatial interaction data with application to postal service performance

The geostatistical spatial modelling framework envisions a random spatial process model Y(s) over a region S which is observed at spatial locations, say, s1,s2,...,sn. In the stationary Gaussian case, a mean function and a covariance function fully specify the process. Here, we consider data which, instead, is related to pairs of locations. Such data, denoted by Y(si,sj) for the pair si,sj is customarily referred to as spatial interaction data. We adopt a much broader view of such data but still seek a geostatistical modelling framework. We also wish to accommodate samples of independent but not identically distributed responses at each pair. These responses need not follow a Gaussian model. Our approach is to introduce a hierarchical model capturing spatial interaction at the second stage using a bivariate Gaussian spatial process. Very flexible classes of models result. We illustrate with an application to postal service performance with regard to delivery of priority mail using a designed experiment involving 79,557 letters. © 2000 Elsevier Science B.V.

Duke Scholars

Author Alan E. Gelfand Statistical Science