A chemo-ecologists' practical guide to compositional data analysis

Compositional data are commonly used in chemical ecology to describe the biological role of chemical compounds in communication, defense or other behavioral modifications. Statistical analyses of compositional data, however, are challenging due to several constraints (e.g., constant sum constraint). We use an ontogenetic series of defensive gland secretions from larvae, three nymphal stages and adults of the oribatid model species Archegozetes longisetosus as a typical chemo-ecological data set to prepare a practical guide for compositional data analyses in chemical ecology. We compare various common and less common statistical and ordination methods to depict small quantitative and/or qualitative differences in compositional datasets: principal component analysis (PCA), non-metric multidimensional scaling (NMDS), multivariate statistical tests (Anderson's permutational multivariate analyses of variance = PERMANOVA; permutational analyses of multivariate dispersions = PERMDIPS), linear discriminant analysis (LDA), the data mining algorithm Random Forests, bipartite network analysis and dynamic range boxes (dynRB). We summarize which methods are suitable for different research questions and how data needs to be structured and pre-processed. Network analyses and dynamic range boxes are promising tools for analyzing compositional data beyond the "classical" methods and provide additional information.