Limitations of locally sampled characters in phylogenetic analyses of sparse supermatrices

Empirical and simulated examples were used to demonstrate the following four points in the context of sparse supermatrices. First, locally sampled characters, when analyzed with low quality heuristic parsimony, likelihood, or Bayesian searches, can create high resolution and resampling values for clades that are properly unsupported because there is no comparable information among sets of terminals. Second, arbitrary factors that should have no effect on phylogenetic inference can create large fluctuations in congruence of trees inferred by parsimony, likelihood, and Bayesian methods with the simulated topology. Third, phylogenetic signal present in locally sampled characters may be interpreted in radically different ways depending upon the phylogenetic signal present in globally sampled characters. Fourth, application of Bayesian MCMC analyses with unlinked branch lengths among character partitions cannot be expected to universally obviate missing-data artifacts, even when numerous characters are sampled from each partition. The first three points may be addressed by conducting thorough tree searches while allowing numerous equally optimal trees to be saved from each replicate rather than relying entirely upon subtree pruning and regrafting (SPR) while saving a single optimal tree, as is the case in many contemporary empirical sparse-supermatrix analyses. (C) 2014 Elsevier Inc. All rights reserved.