Burnin Estimation and Convergence Assessment
(English)Manuscript (preprint) (Other academic)
Estimating the burnin length and assessing convergence purely from the output of an MCMC run is increasingly important in Bayesian phylogenetic inference. Previously, methods for estimating the burnin and assessing convergence have been ad-hoc, such as the minimum number of effective samples or the deviation in split frequencies. In this paper we compare the currently used methods to convergence assessment methods from the mathematical literature, namely the Geweke test and the Heidelberger-Welch test. The latter two show strong advantages in being statistically consistent and unbiased. Statistical consistency and unbiasedness was verified on simulated data with known posterior distributions. Both methods consider convergence as the Null hypothesis. The Null hypothesis is rejected based on standard p-values, which are easier to interpret than a threshold as used by the eeffective sample size. We extend these convergence assessment methods for single and multiple chains. Furthermore, we test the performance of the convergence assessment methods on an empirical dataset and conclude that tests for convergence to the same stationary distribution from independent runs are most adequate. Additionally,we developed an automatic procedure that finds the optimal burnin in the cases we studied. All methods we tested are implemented in the open source software RevBayes (http://www.revbayes.net/).
Research subject Mathematical Statistics
IdentifiersURN: urn:nbn:se:su:diva-64743OAI: oai:DiVA.org:su-64743DiVA: diva2:458403