The Problem of Missing Values and Rounded Zeros in Compositional Data

• Activity Number: 56
• Type: Topic Contributed
• Date/Time: Sunday, August 3, 2014 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Matthias Templ*+ and Karel Hron and Peter Filzmoser
• Companies: Vienna University of Technology and Palacky University Olomouc and Vienna University of Technology
• Keywords: Compositional Data ; Missing Values ; Rounded Zeros ; Partial Least Squares Regression ; Robust Estimation

Abstract

Methods on compositional data in general need complete data and therefore missing values and rounded zeros have to be imputed before analysis. The challanges are to impute below detection limit for rounded zeros, to deal with outliers and high-dimensional data.

EM-based (robust) imputation methods for the estimation of missing values and rounded zeros are introduced. The algorithms are modified to impute below detection limit in case of rounded zeros. Partial least squares regression methods are used in case of high-dimensional data.

Simulation results show that the algorithms perform very well. The usefulness of the procedures is demonstrated on data from geochemistry and metabolomics.