Open matsen opened 6 years ago
I wonder if we can compute the BF by taking the ratio of a BEAST marginal likelihood to e.g. a MrBayes marginal likelihood. This way we can use the unrooted, non-clock tree in MrBayes without fussing with relaxed clocks.
Assuming that the likelihoods calculations are equivalent/comparable (which is reasonable if you run both BEAST and MrBayes with Beagle). It looks like MrBayes can use stepping-stone to compute the marginal likelihood. http://hydrodictyon.eeb.uconn.edu/eebedia/index.php/Phylogenetics:_Morphology_and_Partitioning_in_MrBayes#Estimating_the_marginal_likelihood_using_the_stepping-stone_method
You may also need to adjust the marginal likelihoods by a constant ratio, to account for the different number of rooted and unrooted trees on the same number of taxa.
Thanks, @armanbilge ! It's not clear to me that MB does whole-sequence ancestral state reconstruction in the same way that BEAST does. It does ASR, but it seems designed slightly differently. http://mrbayes.sourceforge.net/wiki/index.php?title=Evolutionary_Models_Implemented_in_MrBayes_3&oldid=5249#Ancestral_State_Reconstruction
But if we're just comparing (non-)clock models, we don't care about ASR. The ancestral states get marginalized out in the marginal likelihood.
I agree with you about the comparison, but in the end for this project we're interested in doing ASR. Perhaps I don't follow where you are going with this?
Sorry, maybe I am confused? I thought the point is to test whether using a clock-like tree is OK. Doing a BF test against MrBayes will show if we have evidence to reject a clock model in favor of no clock. So long as all is good, we can proceed using BEAST for ASR while still sleeping at night.
OK, thanks!
It'll be heaps easier to do a test of BEAST relaxed clocks vs strict clock, but the test you propose seems more stringent.
Just my 2 cents, but I'm not 100% sure how easy it'll be to do a relaxed clock in BEAST. It could open up a whole can of mixing issues. Or not. I wouldn't know.
We should run a Bayes Factor test for strict vs uncorrelated lognormal clocks. I don't think that we have enough data to support a more complex model, but worth checking out.