Tuesday, June 3, 2008
Stop Using the Mantel Test
Lots of folks have been using the Mantel test to ask questions about phylogenetic signal and character correlation. Luke Harmon (U. of Idaho) and I just ran a bunch of simulations to see how these tests performed relative to alternatives (Blomberg et al.'s K statistic for tests of phylogenetic signal and independent contrasts for tests of character correlation). The results were not encouraging and strongly suggest that the Mantel test should only be used with data that can only be expressed as pairwise distances (e.g., geographic distances among poplutions). The Mantel test was designed specifically to deal with pairwise distances because this type of data cannot be analyzed using standard statistical methods due to a lack of independence. If individual measures (which can be analyzed using standard statistical methods) are converted to pairwise distances the resulting Mantel tests suffer from low power and, in the case of partial Mantel tests applied to tests of character correlation, elevated type-I error. I'm not really sure why people have been converting their data to pairwise distances and conducting Mantel tests, but this practice should end. I'm reluctant to write more here because we've submitted our results as a note to Evolution and don't want to give than an excuse not to publish. Write to me or Luke if you want more details.