Why some clades have low bootstrap frequencies and high Bayesian posterior probabilities

Abstract

Bayesian posterior probabilities are wrongly considered by many systematists as indicative of character support, and equivalent to non-parametric bootstrap frequencies. Here I argue against this view. Non-parametric bootstrap is indicative of the amount of evidence in a data matrix supporting each clade in the tree, while Bayesian posterior probabilities are not intended to represent that property. Clades with high posterior probability may not have a large amount of characters favouring them, and their frequencies are the result of the particular sampling procedure of the Bayesian Markov chain Monte Carlo method, which tends to sample very similar topologies according to their posterior probabilities. Both metrics may relate to the notion of confidence, but depict different properties.