Maximum likelihood overcomes long branch attraction bias caused by integrating over parameter uncertainty.
Nucleotide sequences (500 replicates) of increasing length were generated on the topology shown with two long (0.75 substitutions/site) and two short (0.05) terminal branches and a variable internal branch. In each row, the left two panels show the proportion of replicates from which each resolved topology was inferred, plotted against increasing sequence length (left) or internal branch length on the true tree (right). The right two panels plot the mean posterior probability over replicates of each resolved topology. For plots over increasing sequence length, the internal branch length was fixed at zero. The true evolutionary model was used in all analyses. a, Sequences were generated using a simple model with no heterogeneity. b, Half the sites evolved with elevated G+C content (45%), and the other half had reduced G+C (5%); data were analyzed using a correctly partitioned model. c, Sequences were generated under a heterotachous model in which half the sites evolved along a tree with long terminals to B and D, while the other half had taxa A and C with long branches; data were analyzed using a correctly partitioned model. d, Sequences were generated under the same heterotachous model as in c and analyzed with a two-class heterotachous mixture model.