Robustness for Compositional Data

Abstract

Compositional data, data containing relative rather than absolute information, need to be transformed to the usual Euclidean geometry before the standard statistical tools can be applied. Different possible transformations and their properties are discussed. For robust multivariate methods based on a robust covariance estimation, it is crucial to use a transformation that avoids singularity issues. Moreover, the robust location and covariance estimators need to be affine equivariant in order to obtain invariance of the results from the transformation used. Here, different robust multivariate methods are discussed for compositional data analysis, like principal component and discriminant analysis, and applied to a data set from geochemistry.