tag:blogger.com,1999:blog-8499895524521663926.post6927044524049437248..comments2021-09-27T14:55:41.504-04:00Comments on Phylogenetic Tools for Comparative Biology: MCMC method for rate variationLiam Revellhttp://www.blogger.com/profile/04314686830842384151noreply@blogger.comBlogger9125tag:blogger.com,1999:blog-8499895524521663926.post-39077154358005533942011-02-27T11:48:42.521-05:002011-02-27T11:48:42.521-05:00Dave,
This is a really good question - and not so...Dave,<br /><br />This is a really good question - and not so easy to answer. Part of the posterior sample in this case is a position on an edge of the tree - the location of our Brownian rate shift.<br /><br />If you have multiple trees (say, from a Bayesian posterior sample), then you can imagine having posterior density for the rate shift on some edges of one tree that are absent from a second. I think in this case you might want to compute the posterior probability of the change on each edge as the average probability across trees - setting this probability to zero if the edge is absent from the tree.<br /><br />This is just an idea, though.Liam Revellhttps://www.blogger.com/profile/04314686830842384151noreply@blogger.comtag:blogger.com,1999:blog-8499895524521663926.post-39621084721650150452011-02-27T10:41:50.698-05:002011-02-27T10:41:50.698-05:00Liam,
How would you apply this method to datasets ...Liam,<br />How would you apply this method to datasets with multiple valid phylogenetic hypotheses? Average the parameter estimates across MCMC analyses on each tree?<br />-Davedwbapsthttps://www.blogger.com/profile/17606476387441191531noreply@blogger.comtag:blogger.com,1999:blog-8499895524521663926.post-13642107554805846252011-02-26T21:02:03.810-05:002011-02-26T21:02:03.810-05:00Hi Marcio.
I think this is a great idea. I alrea...Hi Marcio.<br /><br />I think this is a great idea. I already have a likelihood method (described in Revell & Collar 2008; Evol.) and R function [evol.vcv()] for fitting multiple evolutionary correlations, conditioned on a discrete character map. A natural next step using MCMC would be to find the shift-points based on only the data for our continuously valued traits.Liam Revellhttps://www.blogger.com/profile/04314686830842384151noreply@blogger.comtag:blogger.com,1999:blog-8499895524521663926.post-30174471842171111342011-02-26T20:42:25.805-05:002011-02-26T20:42:25.805-05:00Hey Liam,
Cool stuff. One possible extension woul...Hey Liam,<br /><br />Cool stuff. One possible extension would be to have, say, two continuous traits and to look for shifts in trait correlation (i.e. to investigate the degree of correlation itself as the trait of interest). That could be a great tool to study phenotypic integration/dissociation. <br /><br />Does that make sense to you?Marcionoreply@blogger.comtag:blogger.com,1999:blog-8499895524521663926.post-25239651814037001212011-02-26T19:16:24.959-05:002011-02-26T19:16:24.959-05:00By the way, if you try to run this with your data,...By the way, if you try to run this with your data, please let me know if it works and if I can be of any help in tuning the control parameters of the function.<br /><br />The result matrix, $mcmc, can be printed to a tab delimited text file and then read into Tracer for typical MCMC diagnostics. [These will be more informative for the likelihoods and quantitative parameters than for the phylogenetic shift-position.] If you want to stick with R, most of the same diagnostics can be computed using functions from the package "coda." I will cover some of this in a subsequent post.Liam Revellhttps://www.blogger.com/profile/04314686830842384151noreply@blogger.comtag:blogger.com,1999:blog-8499895524521663926.post-9839789726664852152011-02-26T19:11:56.149-05:002011-02-26T19:11:56.149-05:00Hi Dave.
Yes, the method is completely appropriat...Hi Dave.<br /><br />Yes, the method is completely appropriate to a phylogeny with non-contemporaneous tips.<br /><br />I don't know of the Venditti method and can't find it on his website. I would be happy to learn more.<br /><br />You're probably right that the time has come for this kind of approach. I starting working on this just as soon as I knew how. . . .Liam Revellhttps://www.blogger.com/profile/04314686830842384151noreply@blogger.comtag:blogger.com,1999:blog-8499895524521663926.post-41108264258805867432011-02-26T14:53:28.956-05:002011-02-26T14:53:28.956-05:00Is this method valid to apply to datasets containi...Is this method valid to apply to datasets containing non-coeval tips? (A stereotypical question for a paleontologist, I know.)<br /><br />Chris Venditti presented an abstract on a MCMC method for identifying shifts in the rate of BM evolution on a tree, at the International Paleontological Congress in London last summer. Meade and Pagel were his co-authors. He applied the method to body-size evolution on the dinosaur supertree. I guess it's just an idea whose time had come?<br />-Davedwbapsthttps://www.blogger.com/profile/17606476387441191531noreply@blogger.comtag:blogger.com,1999:blog-8499895524521663926.post-17611442501063465192011-02-25T18:21:28.029-05:002011-02-25T18:21:28.029-05:00Hi Carl. Sorry this was unclear. "rand.shif...Hi Carl. Sorry this was unclear. "rand.shift" is just for the MCMC chain. [I.e., the chain samples shift points between rates as a random walk, but then jumps to a random position in the tree with probability "rand.shift." We did this mainly to help improve mixing.]<br /><br />Our model is one in which the Brownian rate changes once at some point somewhere in the tree. It is meant as a starting point to model shifts in the OU process and multiple shifts in a similar way. It differs from related methods, such as O'Meara et al. (2006), in that we bring in no extrinsic information about where the rate changes in the tree. This is inferred only from our data for the continuous character.<br /><br />I saw an early version of the Eastman et al. manuscript (which, I will note, seemed very impressive) shortly after submitting our paper. Since their article is unpublished, I won't discuss its details except to comment that indeed the two methods are related.Liam Revellhttps://www.blogger.com/profile/04314686830842384151noreply@blogger.comtag:blogger.com,1999:blog-8499895524521663926.post-80009601455341419902011-02-25T18:03:54.294-05:002011-02-25T18:03:54.294-05:00Liam, very nice post example. I'm unclear how...Liam, very nice post example. I'm unclear how the shift is being modeled -- there is constant probability rand.shift that at any moment in time a shift in rate occurs, with transition from sd1->sd2 occurring at the same rate as sd2->sd1? <br /><br />This assumes the trait diversification rate (sd1 or 2) is determined by a discrete character with constant known transition rate rand.shift? Would it be possible to MCMC over possible values of rand.shift as well, given some prior for it?<br /><br />Exciting developments. I was working over the difference between this and Harmon/Eastman's RJMCMC of diversification rates -- I guess they can relax the assumption of two rates and don't infer a transition rate between the faster and slower diversification, but rather just infer transition points?Carlhttps://www.blogger.com/profile/17998217796660657289noreply@blogger.com