Mistreating birth-death models as priors in phylogenetic analysis compromises our ability to compare models
Time-calibrated phylogenetic trees are fundamental to a wide range of evolutionary studies. Typically, these trees are inferred in a Bayesian framework, with the phylogeny itself treated as a parameter with a prior distribution (a "tree prior"). This prior distribution is often a variant of the stochastic birth-death process, which models speciation events, extinction events, and sampling events (of extinct and/or extant lineages). However, the samples produced by this process are observations, so their probability should be viewed as a likelihood rather than a prior probability. We show that treating the samples as part of the prior results in incorrect marginal likelihood estimates and can result in model-comparison approaches disfavoring the best model within a set of candidate models. The ability to correctly compare the fit of competing tree models is critical to accurate phylogenetic estimates, especially of divergence times, and also to studying the processes that govern lineage diversification. We outline potential remedies, and provide guidance for researchers interested in comparing the fit of competing tree models.