Topics in constrained and unconstrained ordination

In this paper, we reflect on a number of aspects of ordination methods: how should absences be treated in ordination and how do model-based methods, including Gaussian ordination and methods using generalized linear models, relate to the usual least-squares (eigenvector) methods based on (log−) transformed data. We defend detrended correspondence analysis by theoretical arguments and by reanalyzing data that previously gave bad results. We show by examples that constrained ordination can yield more informative views on effects of interest compared to unconstrained ordination (where such effects can be invisible) and show how constrained axes can be interpreted. Constrained ordination uses an ANOVA/regression approach to enable the user to focus on particular aspects of species community data, in particular the effects of qualitative and quantitative environmental variables. We close with an analysis examining the interaction effects between two factors, and we demonstrate how principal response curves can help in their visualisation. Example data and Canoco 5 projects are provided as Supplementary Material.