Affordable Access

deepdyve-link
Publisher Website

First-order correct bootstrap support adjustments for splits that allow hypothesis testing when using maximum likelihood estimation.

Authors
  • Susko, Edward
Type
Published Article
Journal
Molecular Biology and Evolution
Publisher
Oxford University Press
Publication Date
Jul 01, 2010
Volume
27
Issue
7
Pages
1621–1629
Identifiers
DOI: 10.1093/molbev/msq048
PMID: 20154180
Source
Medline
License
Unknown

Abstract

The most frequent measure of phylogenetic uncertainty for splits is bootstrap support. Although large bootstrap support intuitively suggests that a split in a tree is well supported, it has not been clear how large bootstrap support needs to be to conclude that there is significant evidence that a hypothesized split is present. Indeed, recent work has shown that bootstrap support is not first-order correct and thus cannot be directly used for hypothesis testing. We present methods that adjust bootstrap support values in a maximum likelihood (ML) setting so that they have an interpretation corresponding to P values in conventional hypothesis testing; for instance, adjusted bootstrap support larger than 95% occurs only 5% of the time if the split is not present. Through examples and simulation settings, it is found that adjustments always increase the level of support. We also find that the nature of the adjustment is fairly constant across parameter settings. Finally, we consider adjustments that take into account the data-dependent nature of many hypotheses about splits: the hypothesis that they are present is being tested because they are in the tree estimated through ML. Here, in contrast, we find that bootstrap probability often needs to be adjusted downwards.

Report this publication

Statistics

Seen <100 times