Partitioning variation in ecological communities: do the numbers add up?

Abstract

Summary 1. Statistical tests partitioning community variation into environmental and spatial components have been widely used to test ecological theories and explore the determinants of community structure for applied conservation questions. Despite the wide use of these tests, there is considerable debate about their relative effectiveness. 2. We used simulated communities to evaluate the most commonly employed tests that partition community variation: regression on distance matrices and canonical ordination using a third‐order polynomial, principal components of neighbour matrices (PCNM) or Moran’s eigenvector maps (MEM) to model spatial components. Each test was evaluated under a variety of realistic sampling scenarios. 3. All tests failed to correctly model spatial and environmental components of variation, and in some cases produced biased estimates of the relative importance of components. Regression on distance matrices under‐fit the spatial component, and ordination models consistently under‐fit the environmental component. The PCNM and MEM approaches often produced inflated R2 statistics, apparently as a result of statistical artefacts involving selection of superfluous axes. This problem occurred regardless of the forward‐selection technique used. 4. Both sample configuration and the underlying linear model used to analyse species–environment relationships also revealed strong potential to bias results. 5. Synthesis and applications. Several common applications of variation partitioning in ecology now appear inappropriate. These potentially include decisions for community conservation based on inferred relative strengths of niche and dispersal processes, inferred community responses to climate change, and numerous additional analyses that depend on precise results from multivariate variation‐partitioning techniques. We clarify the appropriate uses of these analyses in research programmes, and outline potential steps to improve them.