Lunch with Wainwright Lab
Friday lunch with Wainwright lab continues to be a very useful chance to bounce ideas off of lab members with an excellent grasp of both the models involved and the real biology, morphology and evolution in question. Today's discussion hit several topics.
- Should regime transitions be restricted to nodes? The general consensus was No. On one hand, the position of nodes may contain information about the timing of an actual environmental change responsible for a regime shift. On the other hand: (1) the exact temporal position of the node is uncertain anyway, as it comes from inference of the tree, (2) the evolutionary response may occur immediately after or before the actual position of the node, and indeed it may be interesting to say which it is, (3) this ignores information about the branch lengths, and that transitions are more likely to occur on longer branches (4) this ignores the fact that many more nodes exist than are inferred, but are absent from the phylogeny because the lineage subsequently goes extinct. That said, restricting transitions to nodes seems like a reasonable place to start.
- log transforming data. This is a minor comment, which we might submit as a note to Evolution. While nondimensionalizing data is good, this makes interpreting the Brownian motion model more challenging, as the average size continues to increase. The problem is slightly remedied by the OU process, in which the average size increases only until the stationary distribution is reached. Discussion seemed encouraging and now I need to think of how to usefully illustrate the concern in four or fewer figures...
- Model choice and AIC scores This is another note I'm preparing to submit. The correct way to choose between BM and OU models may not be as simple as AIC, as any simulation can explore. Accurately characterizing false-positive rates may be more informative and is easily done using existing methods and software. The degree of error will be particular to the tree topology, but this kind of analysis could be performed for each data set automatically.