Impacts of terraces on phylogenetic inference

TL;DR

This paper investigates the phenomenon of terraces in phylogenetic inference, revealing their effects on confidence measures, model assumptions, and data partitioning, and proposes strategies to mitigate their impact.

Contribution

The study provides five new insights into terraces, including their dependence on partitioning, influence on bootstrap and posterior probabilities, and implications for data heterogeneity and inference accuracy.

Findings

Terraces can cause high bootstrap support for incorrect clades.

Partitioning schemes influence the size of terraces and inference ambiguity.

Patterns of missing data can lead to apparent inconsistency in phylogenetic analysis.

Abstract

Terraces are potentially large sets of trees with precisely the same likelihood or parsimony score, which can be induced by missing sequences in partitioned multi-locus phylogenetic data matrices. The set of trees on a terrace can be characterized by enumeration algorithms or consensus methods that exploit the pattern of partial taxon coverage in the data, independent of the sequence data themselves. Terraces add ambiguity and complexity to phylogenetic inference particularly in settings where inference is already challenging: data sets with many taxa and relatively few loci. In this paper we present five new findings about terraces and their impacts on phylogenetic inference. First we clarify assumptions about model parameters that are necessary for the existence of terraces. Second, we explore the dependence of terrace size on partitioning scheme and indicate how to find the partitioning scheme associated with the largest terrace containing a given tree. Third, we highlight the impact of terraces on bootstrap estimates of confidence limits in clades, and characterize the surprising result that the bootstrap proportion for a clade can be entirely determined by the frequency of bipartitions on a terrace, with some bipartitions receiving high support even when incorrect. Fourth, we dissect some effects of prior distributions of edge lengths on the computed posterior probabilities of clades on terraces, to understand an example in which long edges "attract" each other in Bayesian inference. Fifth, we show that even if data are not partitioned, patterns of missing data studied in the terrace problem can lead to instances of apparent statistical inconsistency when even a small element of heterotachy is introduced to the model generating the sequence data. Finally, we discuss strategies for remediation of some of these problems.