Comparing several treatments with a control

Assessing statistical reliability of phylogenetic trees via a speedy double bootstrap method

Evaluating the reliability of estimated phylogenetic trees is of critical importance in the field of molecular phylogenetics, and for other endeavors that depend on accurate phylogenetic reconstruction. The bootstrap method is a well-known computational approach to phylogenetic tree assessment, and more generally for assessing the reliability of statistical models. However, it is known to be biased under certain circumstances, calling into question the accuracy of the method. Several advanced bootstrap methods have been developed to achieve higher accuracy, one of which is the double bootstrap approach, but the computational burden of this method has precluded its application to practical problems of phylogenetic tree selection. We address this issue by proposing a simple method called the speedy double bootstrap, which circumvents the second-tier resampling step in the regular double bootstrap approach. We also develop an implementation of the regular double bootstrap for comparison with our speedy method. The speedy double bootstrap suffers no significant loss of accuracy compared with the regular double bootstrap, while performing calculations significantly more rapidly (at minimum around 371 times faster, based on analysis of mammalian mitochondrial amino acid sequences and 12S and 16S rRNA genes). Our method thus enables, for the first time, the practical application of the double bootstrap technique in the context of molecular phylogenetics. The approach can also be used more generally for model selection problems wherever the maximum likelihood criterion is used.