Assessing species and community functional responses to environmental gradients: which multivariate methods?

Question How do multivariate methods perform in relating species- and community-level trait responses to the environment? Location (1) Field data from grazed semi-natural grasslands, NE Germany; (2) artificial data. Methods Research questions associated with traitenvironment relationships were briefly reviewed and seven available methods evaluated. The main distinction between research questions is whether traitenvironment relationships should be addressed at community or species level. A redundancy analysis (RDA) of mean trait values of species in a plot weighted by their abundances (CWM-RDA) is exclusively suitable for the community level. The other six methods address the species level. A double inertia analysis of two arrays (RLQ) and double canonical correspondence analysis (double CCA) use combinations of ordinations to simultaneously analyse species and trait responses to the environment. A combination of the outlying mean index with generalized additive models (OMI-GAM) predicts the response of species to environmental variables on trait gradients. RDA-RegTree first analyses species responses to the environment with RDA and then uses a regression tree to classify trait expressions according to scores of species responses on the ordination axes. Cluster regression uses cluster analyses and logistic regression to search for trait combinations with the best response to the environmental variables. This method models the distribution of functional groups on environmental gradients. All methods and data are available as R scripts. Results All methods consistently revealed the main trait responses to environment in the field data set, namely that life history was associated with available phosphorus while grazing intensity was related to leaf C:N ratio and canopy height. At community level, CWM-RDA gave a good overview of traitenvironment relationships, as also provided by the species-based methods RLQ and double CCA. OMI-GAM revealed non-linear relationships in the field data set. Field and artificial data gave that the number and stability of functional groups produced by Cluster regression and RDA-RegTree varied more strongly than RLQ, double CCA and OMI-GAM. Conclusions Each method addresses particular ecological concepts and research questions. If a user asks for the response of average trait expressions of communities to environmental gradients, CWM-RDA may be the first choice. However, species-based methods should be applied to address questions regarding co-existence of different life histories or to assess how groups of species respond to environmental changes. The artificial data set revealed that the methods differed in sensitivity to gradient lengths and random data.