Co-occurrence and similarity revisited

Sci. Adv. 8, eabj9204 (2022)

Assessing the degree of similarity or co-occurrence among different sets of observations is a standard and important task in many different fields. To mention a few examples, in biogeography and ecology, one may need to compare the similarity between different sites based on a species of interest; in clinical research, identifying similarities between diseases can yield insights into disease etiology and suggest repurposing of treatments; in population genetics, computing genetic dissimilarity between individuals can help analyze diversity within and among populations. Metrics such as Jaccard, Sørensen–Dice, and Simpson are widely used to evaluate co-occurrence and similarity, but they are very sensitive to the prevalence of the elements that they describe: for instance, when computing the similarity between two species across different, multiple sites, the value will vary depending on the occurrence rate of the species over all available samples. In a recent work, Kumar P. Mainali and colleagues demonstrated this existing sensitivity issue with popular metrics and, in addition, proposed a new metric that is insensitive to prevalence.