Abstract
Parsimony methods have historically been used to reconstruct ancestral character states on a phylogeny in order to test hypotheses about evolution and adaptation. Maximum likelihood methods have recently been considered, with the advantages of quantifying uncertainty and incorporating different models of evolution. However, there have been no conclusive studies as to how much confidence evolutionary biologists should place in maximum likelihood ancestral state reconstruction methods for discrete binary characters. This study compares the accuracy of two maximum likelihood methods used to reconstruct discrete ancestral character states. In the global method, the transition rate parameters for the likelihood model are estimated only once by maximizing over all states. In the local method, for each node, the node is fixed in a certain state and the transition rate parameters are estimated conditional on that state assignment.
In order to examine the accuracy of these two methods, marginal and joint reconstructions were performed on data simulated on a variety of model trees, including those based on phylogenies of real biological organisms, while considering the effects of the model of evolution and the state frequencies at the root of the tree. A custom software program called LASRDISC was written in order to perform all of the necessary simulations and reconstructions for this study. Results show that use of the global and local methods in either marginal or joint reconstructions produce generally the same accuracy rates on average, with both methods having higher accuracy than the parsimony method. There are certain cases when the local method performs better, including when the reconstruction method is too simple, which indicates that the local method is more robust to model violations than the global method. The local method also has higher accuracy when the rates of character evolution are high. The global method has certain advantages also in that for the marginal reconstructions, it is less biased than the local method, leading the local method to be overconfident in the reconstructed probability of a state at a node. Also, the global method is more likely than the local method to produce the same result for marginal and joint reconstructions on the same tree. The major difference in the methods is the time necessary to estimate the reconstructions. There is a significant cost to be considered in terms of speed for a joint reconstruction using the local method, especially for trees with a large number of taxa.
|
