Testing methods of homeland detection using synthetic data

There are two families of quantitative methods for inferring geographical homelands of language families: Bayesian phylogeography and the 9diversity method9. Bayesian methods model how populations may have moved using the backbone of a phylogenetic tree, while the diversity method, which does not need a tree as input, is based on the idea that the geographical area where linguistic diversity is highest likely corresponds to the homeland. No systematic tests of the performances of the different methods in a linguistic context are available, however. Here we carry out performance testing by simulating language families, including branching structures and word lists, along with speaker populations moving in areas drawn from real-world geography. We test five different methods: two versions of BayesTraits; the random walk model of BEAST; our own RevBayes implementations of a fixed rates and a variable rates random walk model; and the diversity method. Each method is tested on the same synthetic family of 20 languages, evolving in 1000 different random geographical locations. The results indicate superiority in the performance of BayesTraits and different levels of performance for the other methods, but overall no radical differences in performance.