Quantifying the responses of macroinvertebrate assemblages to simulated stress: are more accurate similarity indices less useful?

1. Similarity indices are commonly used for multivariate analyses, estimations of beta-diversity and assessments of ecological impacts of disturbances. Both statistical and ecological criteria have guided index selections. The former criteria includes accuracy (i.e. how close an estimate is to the true value) and precision (i.e. how repeatable different estimates are) while the latter is focused on detection of known assemblage changes or environmental gradients. 2. These two types of criteria, however, often lead to conflicting recommendations of similarity indices. We hypothesized that a key statistical criterion, accuracy, is either irrelevant or incompatible with the ecological criteria. 3. To test this hypothesis, we compared eight commonly used similarity indices for quantifying the responses of stream macroinvertebrate assemblages to a simulated stress gradient based on both types of criteria. 4. Chao's abundance-based Jaccard and Sorenson coefficients, and Morisita index were most accurate but strongly nonlinear in response to linear community changes and the least or less capable of discriminating stress levels. By comparison, the Bray-Curtis index, CY similarity index and Canberra Metric were less or least accurate, but outperformed the first three indices in all ecological criteria used. Multiple linear regressions confirmed that accuracy is irrelevant to or incompatible with the ecological criteria. 5. Synthesis and applications. By examining the relationships between the statistical and ecological performances of similarity indices, our study provides a critical synthesis of index evaluations and clarified much existing confusion about index selection.