Prediction of the spatial structure of censored data by robust statistical methods.

Abstract

Censored data are a well-known problem when dealing with regional geochemical data. Often many analytical results for some of the most interesting variables, for example gold (Au), are reported as below detection limits. With a high proportion of censored data, distribution estimators and many statistical tests will not perform. The regional structure of the data as displayed in a geochemical map may also get lost or be a poor approximation of reality. However, very many other variable are often available from the same sample sites. A method to recover censored data, based on robust principal component analysis (PCA) and robust multiple regression is introduces. All other available information from each sample site is used to predict the censored data. It is first used to recover the regional data structure for a test data set where the regional distribution is known. Subsequently it is used to predict the regional data structure for a number of elements in topsoil and C-horizon samples from the Kola Project where a large proportion of the predicted values, the regional structure of the data can be recovered, thus permitting more effective use of the geochemical maps in mineral exploration campaign.