HIGH-INTERACTION DIAGNOSTICS FOR GEOSTATISTICAL MODELS OF SPATIALLY REFERENCED DATA

Post-modelling diagnostics can be used to perform a model criticism, to explore the data using the model, to decompose the model spatially and to detect anomalies. Development of diagnostic tools for spatially referenced data, a type of data that often arises in the geological sciences, has been largely neglected. This is partly because until recently there has been no effective way of displaying and analysing the plots of such diagnostics. High interaction graphics provide such a method. They have been used to great effect in general Exploratory Data Analysis (Cleveland & McGill, 1988), and recently with spatially referenced data (Haslett et al., 1991). An interactive graphics platform can display and manipulate large quantities of statistical information. This can include pointwise, pairwise and subsetwise measures. This paper uses these facilities to develop some new diagnostics and more effectively use the current ones. This involves examining and decomposing the model fitted to the covariance structure of the data. The diagnostics are demonstrated and compared using the software package REGARD in the context of a data set from the soil sciences.