Wednesday, October 13, 2010

Estimating Diversification Rates

A new study by Wertheim & Sanderson (Evolution; Online Accepted Articles) investigates the sensitivity of existing methods for estimating diversification rates to various types of phylogenetic error. The topic is somewhat related to the recent, provocatively titled Evolution paper by Dan Rabosky clearly showing that when the assumption of constant birth is violated, death (extinction) rates can no longer be reliably estimated from molecular phylogenies.

Especially given Rabosky's (2010) main result - that is, the high sensitivity of extinction rate estimates to certain model assumptions - the new study by Wertheim & Sanderson is particularly intriguing. Although these methods typically assume that the tree and branch lengths are known without error, Werheim & Sanderson demonstrate in their study that diversification rate estimates are not particularly sensitive to phylogenetic errors either in branch length or topology. In fact, they note in the abstract that even a "crude estimate" of the tree provides substantially more power (e.g., 1.6 x more for the conditions of their study) than, for instance, a comparable non-phylogenetic method, the widely used Slowinski-Guyer test (Slowinski & Guyer 1993; Am. Nat.).

Considered together, these two studies remind us that the robustness of a given statistical method cannot be illustrated by a broad brush. Rabosky's study shows that the estimation of extinction rates from phylogenies of extant species is quite sensitive to the underlying assumption that speciation rates are constant throughout the tree. Conversely, Werheim & Sanderson show that the estimation of speciation rate is not sensitive to the underlying assumption that the phylogenetic tree and branch lengths are known without error - and, furthermore, that even a "crude" tree will do.

Note that the figure above is from neither study - but from my 2005 paper (with Dechronization bloggers Harmon & Glor) about the sensitivity of diversification rate estimates to model parameterization. (We found it to be high.)

No comments: