Heavy-weighting rare species in dissimilarity indices improve recovery of multivariate groups

Abstract

Dissimilarity indices differ in the relative weight given to rare species. Heavy-weighting of rare species may be justified in terms of sampling. An index may erroneously estimate high dissimilarity between two identical communities if they are composed of many rare species and the sampling effort is insufficient to observe most of them in both samples. Heavy-weighting of rare species is thought to compensate for this negative bias. I evaluated two quantitative indices that heavy-weight rare species, NNESS (New Normalized Expected Species Shared) and Goodall, and two probability versions of the Sørensen index, one that takes into account shared unseen rare species and the other that does not. They were compared against the widely used Bray-Curtis (or Sørensen quantitative) and the Morisita-Horn. Indices were computed using raw abundance data or coded data that heavy-weight rare species (frequency in sample units, log-transformation and standardization by the maximum abundance within species). Indices were evaluated for their ability to distinguish, using distance-based MANOVA, season-defined (summer, winter) groups of samples of stream macroinvertebrates and groups of samples obtained by simulation. Sørensen corrected for unseen shared species performed poorly in the empirical study and intermediate in the simulations. NNESS was good in the empirical study and intermediate in the simulations. Goodall scored inversely as NNESS, being intermediate in the empirical assessment and very good in the simulations. The Sørensen uncorrected for unseen shared species, Bray-Curtis and the Morisita-Horn presented poor or intermediate results using raw abundance data. Their performance, however, improved consistently using coded data that heavy-weight rare species and made them good or very good. I conclude that heavy-weighting rare species improves the ability to detect multivariate groups. Heavy-weighting of rare species may be achieved either by using specific formulae (NNESS, Goodall) or using coded data.