Compositional Data Analysis of Some Alkaline Glasses

Abstract

The approach to the analysis of compositional data involving log-ratio transformation of the data has not been generally adopted by researchers wishing to analyse such data. In the context of exploratory methods of multivariate analysis, such as principal components analysis, where the hope is to identify (cluster) structure in the data, this may be because traditional methods can produce more interpretable results than the log-ratio approach. After illustrating this with an example, circumstances under which the log-ratio approach performs poorly when traditional approaches work well are identified. Log-ratio analysis can be dominated by variables having low absolute presence and high relative variation that do not contribute to, and can obscure, structure in the data. Traditional methods can detect certain kinds of structure in the data that correspond to structure on a ratio scale, after a suitable redefinition of the composition. Since traditional methods often detect such structure more directly than log-ratio analysis it can be concluded that claims that the traditional analysis is “inappropriate” or “meaningless” are exaggerated. This conclusion is based on empirical experience rather than theoretical concerns. The arguments are illustrated using compositional data for alkaline glasses, but have more general application.