Quantification and relative severity of inflated branch-support values generated by alternative methods: An empirical example

Abstract: A supermatrix of 272 terminals from Rubiaceae tribe Spermacoceae that were scored for up to 10 gene regions (two nrDNA, eight plastid) was used as an empirical example to quantify sources of error in heuristic parametric (Bayesian MCMC and maximum likelihood) phylogenetic analyses. The supermatrix includes dramatic disparities in which terminals were sampled for which gene regions. The sources of error examined include poor quality tree searches, requiring a single fully resolved optimal tree, undersampling-within-replicates and frequency-within-replicates bootstrap artifacts, and extrapolation from one character partition to another such that synapomorphies that would only be ambiguously optimized by parsimony are optimized with high probability by parametric methods. Four of our conclusions are as follows. (1) The resolution and support provided by parametric methods for clades that lack unambiguously optimized (by parsimony) synapomorphies are less robust to the addition of terminals and characters than those clades that have unambiguously optimized synapomorphies. (2) Those tree-search methods which can create phylogenetic artifacts (frequency-within-replicates resampling, undersampling-within-replicates resampling, requiring a single fully resolved optimal tree, non-independent resampling among replicates) provide the greatest resolution and support irrespective of whether that resolution or support is corroborated by more conservative and better justified methods. (3) Partitioning data matrices cannot be relied upon to consistently obviate potentially dubious resolution and support caused by missing-data artifacts in likelihood analyses when the models require linked branch lengths among partitions. (4) Undersampling-within-replicates and frequency-within-replicates resampling artifacts are not unique to parsimony and should be accounted for in likelihood analyses by allowing multiple equally likely trees to be saved within each resampling pseudoreplicate and applying the strict-consensus bootstrap rather than the frequency-within-replicates bootstrap. [Copyright &y& Elsevier]