Cases in which bayesian phylogenetic methods will be positively misleading

GOLOBOFF, PABLO.; POL, DIEGO

Simposio; Symposium on "Competing Methods on Phylogenetics," XXII Annual Meeting of the Willi Hennig Society; 2003

Bayesian phylogenetic analysis using Monte CarloMarkov chains has become very popular in certaincircles. It is based on estimating the likelihood surface ofdifferent tree topologies; the prior probabilities ofdifferent trees are generally considered as identical, withwhich one would expect the Bayesian analysis toproduce results identical to those of standard maximumlikelihood. However, there are significant differencesin the results. Supposedly, Bayesian phylogeneticanalysis allows assigning confidence to hypotheses ofmonophyly of groups without ever finding the individual trees of highest likelihood (maximum posteriorprobability), thanks to the use of Monte Carlo simulations. This creates serious problems when the likelihoodsurface has certain characteristics. Thus, if data aregenerated under a given model and tree, the Bayesiananalysis of the data may lead to infer that there is a highposterior (greater than 90 or 95%) of the existence ofcertain groups not present in the model tree. The proofof consistency of the maximum likelihood estimatesdoes not apply in this case; those demonstrations referto the individual trees of maximum likelihood, andBayesian phylogenetic analysis calculates (estimates) theprobability of monophyly of a given group as the sum ofthe likelihoods in different trees with the group. Bayesian phylogenetic analysis would be consistent if theindividual tree of maximum likelihood was used, but inthis case Monte Carlo Markov chains (which producethe ??speed?? that made the method so popular) cannotbe applied and one is forced to go backto standard trialand-error techniques to find optimal trees. In the case oflikelihood, those techniques are extremely slow andinefficient.