Bayesian Dating of Shallow Phylogenies with a Relaxed Clock

Bayesian methods are increasingly being used to estimate divergence times without the restrictive assumption of a global clock. Little is known about their reliability for shallow phylogenies where DNA sequence divergence is low. We analyzed both simulated and real sequences to evaluate dating methods in phylogenies with mid-late Miocene roots. A large number of data sets (5000) with 10 taxa each were simulated under a rate-drift model for trees with 2 topologies (balanced or unbalanced) and with different sets of divergence times (characterized by long or short external branches). Data were analyzed using Bayesian Markov chain Monte Carlo methods in which the prior on divergence times was specified from a birth-death process with species sampling (BDS) or a Dirichlet distribution using the programs MCMCTREE and MULTIDIVTIME. The programs generally performed well on shallow phylogenies, but posterior mean node ages were biased and 95% posterior intervals included true ages in fewer than 95% of trees in some analyses. This typically occurred when the 95% prior interval did not include the true age and/or sequence lengths were < 1 kbp. Widths of posterior intervals were also very dependent on the position of the calibrated node within the tree, particularly when sequences were short. Different divergence times priors within MCMCTREE, MULTIDIVTIME, and BEAST were used to analyze mitochondrial DNA data sets from a Bovid subfamily (the Caprinae) from Asian Laudakia and North African Chalcides lizards. Posterior divergence times were quite sensitive to different BDS priors but less sensitive to different Dirichlet priors. Our study demonstrates the impact of the prior on divergence times in shallow phylogenies and shows that 1) prior intervals on nodes should be assessed as a prerequisite to a dating analysis, 2) >= 1 kbp of quite rapidly evolving sequence may be required to obtain accurate posterior means and usefully narrow posterior intervals.