Analysis of community-level mesocosm data based on ecologically meaningful dissimilarity measures and data transformations
Résumé
The Principal Response Curves (PRC) method is a constrained ordination method developed specifically for the analysis of community data collected in mesocosm experiments which provides user-friendly summaries and graphical representations of the data. The PRC method is based on redundancy analysis (RDA) usually performed on log-transformed abundance data. The log-transformation is used to lower the weight of the most abundant species in the analysis and implies that the variations in abundance are scaled to the total abundance of each species. Many different measures to calculate dissimilarity between samples on the basis of the abundance of each species have been developed, in particular in the field of ecology. The measure of dissimilarity between samples and the data transformations have a very large impact on the results of the multivariate analysis. The Euclidean distance between samples is implicitly used in PRC, but in the field of ecology, many more dissimilarity measures are currently used. Distance-based redundancy analysis provides a basis for integrating more ecological meaningful distance measures into the PRC than Euclidean distance alone. In this paper, we investigated the ordinations produced with a small selection of ecological meaningful dissimilarity measures, namely the Euclidean distance on log-transformed data, the Hellinger distance, and the Bray-Curtis dissimilarity on raw and log-transformed data. We compared unconstrainted ordination as well as PRC results obtained with these dissimilarity measures for two different macro-invertebrate community datasets, resulting from experiments performed with anti-inflammatory drug diclofenac and the insecticide chlorpyrifos. In both cases, the ordinations obtained with unconstrained ordinations and RDA were similar, indicating that, once time effects had been partialled out, the treatment effects were the major source of variability in the data. In the diclofenac dataset, the Hellinger distance produced interesting significant results that completed those obtained with the Euclidean distance on log-transformed data. The chlorpyrifos dataset showed similar results for the different dissimilarity measures, with slight differences in the number of significant dates. The fraction of the total variance belonging to between-replicate and time-dependent variation appeared to be an important criterion in the selection of a dissimilarity measure.