Taxon Sampling and the Accuracy of Large Phylogenies

this paper. In our analysis, we found no signi#cant difference between the performance of the MP and NJ methods in estimating phylogeny on the model tree of Hillis (1996) when we used the average proportion of identical bipartitions between the true tree and the inferred tree (the B distance) as our accuracy criterion for both methods. Figure 4 compares the performance of MP and NJ in terms of B distances. In contrast, Hilliss (1996) results appear to suggest that MP outperforms NJ for short sequences (see his Fig. 2). The explanation for this discrepancy lies in the treatment of polytomies on the estimated tree. If one possible resolution of a polytomy on the estimated tree is consistent with a taxon bipartition on the true tree, then we would consider the accuracy to be 1/ b, where b is thenumber of possible resolutions of the polytomy, whereas Hillis (1996) considered the accuracy to be 0.5 (another possible solution is to treat the accuracy as zero for that resolution because the method failed to correctly identify the partition). Hence, for sequences of length zero, MP (or any other method that allows polytomies) would be either 1/ b or 0.5 correct, depending on which measure of accuracy is used. Because the number of unresolved polytomies typically increases with an increase in s, the discrepancy between the two measures of tree similarity matters most when large numbers of taxa are studied. These different ways of resolving polytomies normally would not be a source of bias except that some methods, such as NJ, will either (1) arbitrarily resolve some branches to have a small, but nonzero, length because of stochastic error in the process of substitution or rounding errors in computer memory, or (2) be represented in computer memory as a strictly bifurcati...