B0646
Title: Intepretation of outliers in compositional data using the R-package mvoutlier
Authors: Karel Hron - Palacky University (Czech Republic) [presenting]
Peter Filzmoser - Vienna University of Technology (Austria)
Abstract: Multivariate outliers are often the most interesting data points because they show atypical phenomena. Several classical and robust methods have been proposed for their identification also in the context of compositional data - multivariate observations carrying relative information (typically proportions, percentages, mg/kg, etc.). The proposed tools help to better understand the reason, why these samples are defined as atypical by representing them in maps, in a compositional biplot, in univariate scatter plots, and in parallel coordinate plots. In all plots, the same special colors and symbols can be selected, referring to the relative position of the outliers in the multivariate data cloud and thus supporting an interpretation of these observations. Since compositional data are considered, a relation to the single compositional parts can be established only by special isometric logratio coordinates. The developed tools are freely available in the R package mvoutlier. The function mvoutlier.CoDa() computes the multivariate outliers, and it prepares the information for the symbols and colors. The resulting object can then be used for the plot function. All arguments are consistent for the presentations, which makes it possible to see the same symbol and color choices in different views, revealing the structure of the multivariate outliers.
