Divergent maximum-likelihood-branch-support values for polytomies

We applied simple 4-taxon simulations with 3-way character conflict or a hard polytomy to check for false positive branch support, with a focus on the bootstrap and recently introduced likelihood-based phylogenetic-inference programs. Given that there are only three possible bifurcating topologies, discrepancies among methods identified in this study should generally be restricted to factors other than topological search heuristics. Our four major conclusions are as follows. First, Bayesian MCMCMC posterior probabilities are not the only means of quantifying support that can produce dramatically inflated values when applied to cases of strong character conflict. Rapid bootstrapping with the GTRCAT model in RAxML can provide still greater support values for polytomies and we suggest that it generally be avoided. Second, the SH-like approximate likelihood-ratio test outperforms the bootstrap when applied to polytomies. We suggest that the SH-like aLRT be widely applied to likelihood-based empirical studies to complement the bootstrap by collapsing those branches with an SH-like aLRT percentage of ≤ 10, irrespective of how high the likelihood bootstrap support is. Third, the 70% bootstrap cutoff does not equate to a 5% error rate and we suggest that the idea that ≥ 70% bootstrap generally equates to 95% probability of accuracy in empirical analyses finally be abandoned. Fourth, rapid bootstrapping with the GTRCAT model in RAxML can generate values with very low precision, which reinforces our assertion that this method should be avoided, let alone be entirely relied upon for phylogenetic inference.