Analysis of word-order universals using Bayesian phylogenetic inference

Abstract

This thesis examines the novel approach by Dunn et al. (2011) that employs the Bayesian phylogenetic inference to compute the Bayes Factors that determine whether the evolutions of a set of word-order traits in four language families are correlated or independent. In the first part of the thesis, the phylogenetic trees of the Indo-European and Bantu language families are reconstructed using several methods and the differences among the resulting trees are analyzed. In the second part of the thesis, the trees are used to conduct various modifications to the original experiments by Dunn et al. in order to evaluate the accuracy and the utility of the method. We discovered that the Bayes Factors computation using the harmonic mean estimator is very unstable, and that many of the results reported by Dunn et al. are irreproducible. We also found that the computation is very sensitive to the accuracy of the data because a one-digit error can alter the Bayes Factors significantly. Furthermore, through an examination of the source code of BayesTraits, the software package that were used compute the Bayes Factors, we discovered that Dunn et al. supplied invalid inputs to the software, which renders their whole calculations erroneous. We show how the results of the computations would change if the inputs were corrected.