Too good to be true: pitfalls of using mean Ellenberg indicator values in vegetation analyses

Question: Mean Ellenberg indicator values (EIVs) inherit information about compositional similarity, because during their calculation species abundances (or presence-absences) are used as weights. Can this similarity issue actually be demonstrated, does it bias results of vegetation analyses correlating mean EIVs with other aspects of species composition and how often are biased studies published? Methods: In order to separate information on compositional similarity possibly present in mean EIVs, a new variable was introduced, calculated as a weighted average of randomized species EIVs. The performance of these mean randomized EIVs was compared with that of the mean real EIVs on the one hand and random values (randomized mean EIVs) on the other. To demonstrate the similarity issue, differences between samples were correlated with dissimilarity matrices based on various indices. Next, the three mean EIV variables were tested in canonical correspondence analysis (CCA), detrended correspondence analysis (DCA), analysis of variance (ANOVA) between vegetation clusters, and in regression on species richness. Subsequently, a modified permutation test of significance was proposed, taking the similarity issue into account. In addition, an inventory was made of studies published in the Journal of Vegetation Science and Applied Vegetation Science between 2000 and 2010 likely reporting biased results due to the similarity issue. Results: Using mean randomized EIVs, it is shown that compositional similarity is inherited into mean EIVs and most resembles the inter-sample distances in correspondence analysis, which itself is based on iterative weighted averaging. The use of mean EIVs produced biased results in all four analysis types examined: unrealistic (too high) explained variances in CCA, too many significant correlations with ordination axes in DCA, too many significant differences between cluster analysis groups and too high coefficients of determination in regressions on species richness. Modified permutation tests provided ecologically better interpretable results. From 95 studies using Ellenberg indicator values, 36 reported potentially biased results. Conclusions: No statistical inferences should bemade in analyses relatingmean EIVs with other variables derived from the species composition as this can produce highly biased results, leading to misinterpretation. Alternatively, a modified permutation test using mean randomized EIVs can sometimes be used.