ALFARO, M. E.; ZOLLER, S.; LUTZONI, F.: Comparative performance of the bootstrap and Bayesian MCMC sampling in assessing phylogenetic confidence: a simulation study

The use of Monte Carlo-based Bayesian methods to estimate confidence in phylogenetic results is becoming increasingly common. Although the results of such analyses are often thought to be roughly equivalent to traditional maximum likelihood bootstrapping, few studies have attempted to explicitly compare these confidence methods. We present the results of simulations comparing Bayesian Markov chain Monte Carlo (MCMC) sampling and bootstrapping over a range of topologies and relative branch lengths. Both methods attribute high confidence to nodes when branch lengths are all relatively long and character evolution is relatively high. When internodes are short or trees consist of a mixture of short and long internodes, confidence estimates of the two methods sometimes varied substantially depending on the relative branch lengths and position of these internodes within the tree. In general, Bayesian MCMC sampling assigned high confidence values to more correct nodes than did traditional bootstrapping. Bootstrapping rarely assigned high confidence to incorrect nodes, but was also less likely to assign high confidence to correct nodes. In addition, in situations where branch lengths were short and rates of character evolution were low, Bayesian MCMC sampling often assigned high confidence to correct nodes. In contrast, bootstrapping rarely assigned high or moderate confidence to correct nodes under these conditions. Our results suggest that confidence limits estimated from bootstrapping and Bayesian MCMC sampling are not equivalent. Reasons for these differences and their implications will be discussed.