Generalized relative data and robustness in Bayes spaces

“Relative data” refer to statistical data which are part of a whole, which is typically reflected
in the reported data unit (mg/kg, ppm, %, etc.). The analysis of such data should not depend on the data scale, and this is made possible by the toolbox of “compositional data analysis”. However, nowadays there are more complex data structures: single observations are collected as continuous functions, and even as continuous functions of higher order, or they can be tables because of underlying factors. An example for the latter are the number of Covid-19 infections in a certain time period for different countries, grouped according to gender and/or age groups with the aim to analyze relative relationships between these two factors. As with any other data, data quality issues appear in practice, because some variables or data ranges can only be measured with higher uncertainty, and because of data outliers affecting traditional data analyses. In this project we will use a framework where such generalized forms of relative data can be treated jointly, and where variable and observation weighting, and particularly weighting the combination of variables and observations, can be performed. The development of statistical methods under this framework will not only make it possible to analyse complex data structures, but such methods will also lead to an improved understanding of the analysis results.

Coordinator: TU Wien

TU Wien team: N.N. (PreDoc), N.N. (PostDoc), Peter Filzmoser


Program / Call: FWF International Project

Proposal: I 5799-N

Funding: Austrian Science Fund (FWF)

Start: ?? 2022, duration: 36 months