Bayesian priors for tree calibration: Evaluating two new approaches based on fossil intervals
Ryan W Norris, Cory L Strope, David M McCandlish, Arlin Stoltzfus
Background: Studies of diversification and trait evolution increasingly rely on combining molecular sequences and fossil dates to infer time-calibrated phylogenetic trees. Available calibration software provides many options for the shape of the prior probability distribution of ages at a node to be calibrated, but the question of how to assign a Bayesian prior from limited fossil data remains open. Results: We introduce two new methods for generating priors based upon (1) the interval between the two oldest fossils in a clade, i.e., the penultimate gap (PenG), and (2) the ghost lineage length (GLin), defined as the difference between the oldest fossils for each of two sister lineages. We show that PenG and GLin/2 are point estimates of the interval between the oldest fossil and the true age for the node. Furthermore, given either of these quantities, we derive a principled prior distribution for the true age. This prior is log-logistic, and can be implemented approximately in existing software. Using simulated data, we test these new methods against some other approaches. Conclusions: When implemented as approaches for assigning Bayesian priors, the PenG and GLin methods increase the accuracy of inferred divergence times, showing considerably more precision than the other methods tested, without significantly greater bias. When implemented as approaches to post-hoc scaling of a tree by linear regression, the PenG and GLin methods exhibit less bias than other methods tested. The new methods are simple to use and can be applied to a variety of studies that call for calibrated trees.