Solutions for loss of information in high‐beta‐diversity community data

Summary Dissimilarity measures, which gauge compositional resemblance between sample units, tend to lose information with increasing distance along ecological gradients. This undesirable property is especially common in high‐beta‐diversity community data, yet analysts seldom acknowledge it when relating multivariate attributes (e.g. species composition) to underlying gradients. With 1000 simulated and 14 real community data sets, I systematically varied beta‐diversity to evaluate the effects of seven dissimilarity adjustment methods (plus one baseline method) on outcomes from multivariate gradient analyses. Performance was determined by plotting dissimilarities vs. environmental distances, by ordination vs. environmental distance correlations and by Procrustean concordance of ordinations with known environmental grid structure. Performance of all methods declined as narrower niches and greater competition asymmetry led to greater beta‐diversity. With simulated data, methods based on probabilities of joint occurrences (Beals smoothing, Swan's method) most effectively resolved the loss‐of‐information problem, followed closely by three step‐across methods (Shortest‐path, Extended, Geodesic); the CY and Diffusion methods did not outperform unmodified Bray–Curtis dissimilarities. With real community data sets, the Beals, Shortest‐path and Extended methods excelled. Electing to adjust dissimilarities is recommended when many pairs of sample units share few species in common, although unmodified dissimilarities may still be appropriate when beta‐diversity is low. Dissimilarity adjustments are appropriate not only for species diversity and ecological purposes, but also for other applications where nonlinear relationships among attributes and gradients are expected in zero‐rich and highly variable multivariate data.