Cryptic failure of partitioned Bayesian phylogenetic analyses: lost in the land of long trees.

Partitioned Bayesian phylogenetic analyses of routine genetic data sets, constructed using MrBayes (Ronquist and Huelsenbeck 2003), can become trapped in regions of parameter space characterized by unrealistically long trees and distorted partition rate multipliers. Such analyses commonly fail to reach stationarity during hundreds of millions of generations of sampling-many times longer than most published analyses. Some data sets are so prone to this problem that paired MrBayes runs begun from different starting trees repeatedly find the same incorrect long-tree solutions and consequently pass the most commonly employed tests of stationarity, including the average standard deviation of split frequencies (ASDSF) and the potential scale reduction factor (PSRF) statistics offered by MrBayes (Gelman and Rubin 1992). In these situations, failure to reach stationarity is recognizable only in light of prior knowledge of model parameters, such as the expectation that third-codon-position sites usually evolve fastest in protein-coding genes. The conditions that lead to the long-tree problem are frequently encountered in phylogenetic studies today, and I present 6 demonstration examples from the literature. Although the effects on tree length (TL) are often dramatic, effects on topology appear to be subtle. Susceptibility to the problem is sometimes predicted by the difference between the true TL and the starting TL. In some cases, the problems described here can be avoided or reduced by manipulation of the starting TL and/or by adjustments to the prior on branch lengths. In more difficult situations, accurate branch length estimation may not be possible with Bayesian methods because of dependence of the solution on the branch length prior.