A Likelihood-Ratio Test of Monophyly

?Many phylogenetic analyses are inspired by or depend upon the monophyly of a group specified a priori. Also, many evolutionary problems for which phylogenies are useful do not require every detail of the phylogeny to be estimated correctly but depend upon the monophyly (or lack thereof) of a particular group. We propose a likelihood-ratio test that compares whether the best trees estimated with and without the constraint of monophyly are significantly different. Simulation suggests that the test is conservative when the null hypothesis (a particular specified constraint) is correct. We applied the likelihood-ratio test of monophyly to the question of the relationship of the presumed-extinct marsupial wolf (Thylacinus). Specifically, we examined the null hypotheses that (1) the marsupial wolf does not form a monophyletic group with dasyurids and (2) the marsupial wolf is a basal member of the Australian radiation of mammals. Both hypotheses were rejected using the likelihood-ratio test of monophyly. [Likelihood-ratio test; max? imum likelihood; monophyly; phylogenetic methods; Thylacinus. ] Systematists are increasingly concerned not only with the accurate estimation of phylogenetic relationships but also with the testing of phylogenetic hypotheses. The method of maximum likelihood pro? vides a powerful framework to address both aspects of a phylogenetic analysis. In terms of the estimation of relationship, maximum likelihood appears to be both efficient and robust (Kuhner and Felsen? stein, 1994; Tateno et al., 1994; Huelsen? beck, 1995a, 1995b). However, likelihood also provides a natural way of testing hy? potheses through the likelihood-ratio test (Edwards, 1972). The likelihood-ratio test has been used to examine (1) the adequacy of models of DNA substitution (Goldman, 1993), (2) the effect of additional parame? ters on the fit of a substitution model (Goldman, 1993; Yang et al., 1994; Yang, 1996), (3) the existence of a molecular clock (Felsenstein, 1981), and (4) whether trees estimated from different data partitions are heterogeneous (Huelsenbeck and Bull,