Classical and robust orthogonal regression between parts of compositional data

• Published in: Statistics (Berlin, DDR) Vol. 50; no. 6; pp. 1261 - 1275
• Main Authors: Hrůzová, K., Todorov, V., Hron, K., Filzmoser, P.
• Format: Journal Article
• Language: English
• Published: Abingdon Taylor & Francis 01.11.2016; Taylor & Francis Ltd
• Subjects: bootstrap inference; Compositional data; Economic analysis; isometric log-ratio coordinates; MM-estimates; orthogonal regression; Parameter estimation; Parameter robustness; Principal components analysis; Regression analysis; Robustness (mathematics); Statistical analysis; Statistical inference
• ISSN: 0233-1888; 1029-4910
• DOI: 10.1080/02331888.2016.1162164

The different parts (variables) of a compositional data set cannot be considered independent from each other, since only the ratios between the parts constitute the relevant information to be analysed. Practically, this information can be included in a system of orthonormal coordinates. For the task of regression of one part on other parts, a specific choice of orthonormal coordinates is proposed which allows for an interpretation of the regression parameters in terms of the original parts. In this context, orthogonal regression is appropriate since all compositional parts - also the explanatory variables - are measured with errors. Besides classical (least-squares based) parameter estimation, also robust estimation based on robust principal component analysis is employed. Statistical inference for the regression parameters is obtained by bootstrap; in the robust version the fast and robust bootstrap procedure is used. The methodology is illustrated with a data set from macroeconomics.