The robustness of two phylogenetic methods: four-taxon simulations reveal a slight superiority of maximum likelihood over neighbor joining

The robustness (sensitivity to violation of assumptions) of the maximum-likelihood and neighbor-joining methods was examined using simulation.Maximum likelihood and neighbor joining were implemented with Jukes-Cantor, Kimura, and gamma models of DNA substitution. Simulations wereperformed in which the assumptions of the methods were violated to varyingdegrees on three model four-taxon trees. The performance of the methods wasevaluated with respect to ability to correctly estimate the unrootedfour-taxon tree. Maximum likelihood outperformed neighbor joining in 29 ofthe 36 cases in which the assumptions of both methods were satisfied. In133 of 180 of the simulations in which the assumptions of themaximum-likelihood and neighbor-joining methods were violated, maximumlikelihood outperformed neighbor joining. These results are consistent witha general superiority of maximum likelihood over neighbor joining undercomparable conditions. They extend and clarify an earlier study that foundan advantage for neighbor joining over maximum likelihood forgamma-distributed mutation rates.