To transform or not to transform: that is the dilemma in the statistical analysis of plant volatiles

Chemical ecology, be it the study of plant volatiles or insect cuticular hydrocarbons, largely involves the analysis of compositions or "blends" of a mixture of compounds. Compositional data have intrinsic properties such as a "constant-sum constraint" which should be taken into account when statistically analysing these data. The field of compositional data analysis has greatly improved our understanding of the nature of such compositions and has provided us with insights on statistically rigorous ways of analysing such constrained data. Employment of standard multivariate statistical procedures on compositional data necessitates the use of appropriate transformation procedures, which removes the non-independence of data points, thus rendering the data suitable for such analysis. Here we present the current situation of the analysis of compositional data in chemical ecology; the awareness of this constraint of compositional data; and alternative ways of analysing such constrained data using Random Forests, a data-mining algorithm which has many features that facilitate the analysis of such data. Two such features of particular relevance to compositional data are that Random Forests does not incorporate implicit assumptions about the distribution of the data and can deal with auto-correlations between data points.