Parsimony and Bayesian Phylogenetics

GOLOBOFF, P.A.; POL, D.

Parsimony, Pylogeny, and Genomics

Oxford University Press

Lugar: Oxford; Año: 2005; p. 148 - 159

Methods of phylogeny reconstruction are often divided into statistical methods (which require an explicit model of evolution) and non-statistical methods.  Among methods with an explicit satistical justification, the most widely used are the methods of maximum likelihood resulting from Felsenstein´s work, and more recently, bayesian phylogenetic methods based on Monte Carlo Markov chains. Bayesian methods have become very prominent among model-based methods, in part because of computational advantages, and in part because they estimate the probability that a given hypothesis is true, given the observations and model assumptions.Bayesian approaches attempt to integrate the likelihood across the different possible parameter values (i.e., branch lengths and substitution model parameters) through Monte Carlo Markov Chains. Up to now, the frequency of the clades in the chain has been interpreted by most bayesian phylogeneticists as the posterior probability that the clade is true. Here, we will focus on some undesirable properties found on this frequently used option to summarize MCMC results.  Other alternative ways to summarize the results are less frequently used, and they differ from this one in depending much more on whether the chain has succeeded in finding the actual MAP(s).  As the chain is not conceived as a search mechanism, but instead as a sampling mechanism, it is extremely unlikely that it will find the individual trees of maximum a posteriori probability, except in very small data sets.