Bias in estimates of the classic and incidence-based Jaccard similarity indices: insights from assemblage simulation

Abstract

Similarity indices are often used for measuring b-diversity and as the starting point of multivariate analysis. In this study, I used simulation to examine the direction and amount of bias in estimates of two similarity indices, Jaccard Coefficient (J) and incidence-based J (J). I design a novel simulation to generate three sets of assemblages that vary in species richness, species-occurrence distributions, and b-diversity. I characterized assemblage differences with the ratio of [proportion of rare species in all shared species / proportion of rare species in all unshared species] (i.e., PRss/PRus) and the Pearson’s correlation in the probabilities of shared species between two assemblages (i.e., share-species correlation). I found that J was subject to strong positive or negative bias, depending on PRss/PRus. J was mainly subject to negative bias, which varied with share-species correlation. In both indices, bias varied substantially from one pair of assemblages to another and among datasets. The high variation in the bias across different comparisons of assemblages may compromise b-diversity estimation established at low sampling efforts based on the two indices or their variants.