Hi Graham, Thanks for sharing your insights and for your detailed suggestions for thorough model fit checks!
Yes, I am using exclusively extant taxa; there's no fossil record for these plants, unfortunately. I am aware of the difficulties with OU processes, and of parametric bootstrap checks as in Boetigger et al. 2012; I also recently saw yours and Matt's posterior predicitive sims. paper, but haven't read it yet. I've had these sorts of checks in mind for a post model fitting phase in the analysis, and your prodding is a very good reminder! I was not, however, aware of the phylogenetic signal check you espouse. I definitely see the rationale and phylogenetic signal expectations for the paraphyletic stem and monophyletic subclade. However, in my case phylogenetic signal may be low for the monophyletic subclade for other reasons. In my mind (and there is some ecological evidence to support this), the trait in questionis evolving under negative frequency dependent selectionin the monophyletic subclade. So, often close relatives have very different trait values. A phylogenetic continuous trait model that models negative frequency dependence would be a better choice (at least biologically) than BM, but we don't currently have such a model in our PCM toolbox. Whether phylogenetic signal is lower in the paraphyletic stem than in the monophyletic subclade remains to be seen, though, so it might still be a good check. ...and, of course, all of this is complicated by the tree's topological uncertainty, which I won't get into. As todeveloping a model for negative frequency dependent trait evolution, that's a completely different "can of worms" and perhaps a good topic for a separate R-sig-phylo discussion. However, as it relates to my data, I would likely need better data (more species, better mv trait(s) measurements, and more loci to get a better tree) to show a better fit of such a model over BM. Cheers, Dan. Slater, Graham wrote: > Hi Dan et al, > > Sorry I won't be of much help analytically - my functions were written > specifically for the time slice problem. It wouldn't be too difficult > to write a function to fit the model you describe, but it seems as > though Julien has this all covered in mvMorph already, so that does > seem to be the best route to go. > > A word of caution though. From your last email it sounds as though > you're using a tree of extant taxa. Though it is probably less of a > concern for cases where you have a subclade-based shift in mode, > rather than a temporal, clade-wide shift, these kinds of processes are > still a) very difficult to detect using trees of extant taxa only, and > b) fitting of OU processes exhibit notorious levels of type I error. > It's worth using additional tests on top of the model fitting. For > example, you might use simulations under your best fitting model to > see if it actually does a better job than simpler processes of > replicating your data (see Pennell et al 2014 > http://biorxiv.org/content/early/2014/04/07/004002.short or Slater and > Pennell 2014 http://sysbio.oxfordjournals.org/content/63/3/293 and > Boettiger et al. 2012 > http://onlinelibrary.wiley.com/doi/10.1111/j.1558-5646.2012.01574.x/abstract). > > > > In addition, a simple yet elegant (in my opinion) way of examining > whether an OU process that constrains variance explains the background > macroevolutionary regime and release from this process explains your > subclade pattern would be to use a form of Brian's censored test but > wherein you test for "phylogenetic signal" separately in the > paraphyletic stem and monophyletic subclade. You'd expect > significantly low signal in the background and high (i.e. not > significantly different) signal in the subclade. See Hopkins and Smith > 2015 http://www.pnas.org/content/112/12/3758.abstract for a nice use > of this kind of test. > > These are all nice supplementary tests that would help assuage a lot > of peoples' concerns about OU processes and their tendency to fit well > to any kind of noisy comparative data. > > g > > ------------------------------------------------------------ > Graham Slater > Peter Buck Post-Doctoral Fellow > Department of Paleobiology > National Museum of Natural History > The Smithsonian Institution [NHB, MRC 121] > P.O. Box 37012 > > > (202) 633-1316 > [email protected] <mailto:[email protected]> > www.fourdimensionalbiology.com <http://www.fourdimensionalbiology.com> > > > > > >> On May 29, 2015, at 3:44 PM, Julien Clavel <[email protected] >> <mailto:[email protected]>> wrote: >> >> Hi Dan, >> >> Fitting models with two distinct modes of trait evolution is possible >> with the mvSHIFT function in mvMORPH. You just need to map on the >> tree the two clades (or selective regimes, groups... otherwise) of >> interest rather than two times periods. >> >> The mapping can easily be done using the make.simmap or the >> paintSupTree functions in phytools. The package handle both >> univariate and multivariate data. >> >> Best, >> >> Julien >> >> ------------------------------------------------------------------------ >> Date: Fri, 29 May 2015 08:27:31 -0700 >> From:[email protected] <mailto:[email protected]> >> To:[email protected] <mailto:[email protected]> >> CC:[email protected] >> <mailto:[email protected]>;[email protected] >> <mailto:[email protected]>;[email protected] <mailto:[email protected]> >> Subject: Re: [R-sig-phylo] clade-specific release and radiate model? >> >> Hi Brian, >> >> Thanks for your thoughtful response! Those are very good points >> about identifiability and penalization of the OU mean for the >> "released" clade. >> >> The censored model is an appealing approach, especially since the >> timing of the release along the (stem) branch leading to the released >> clade is unknown, and somewhat beside the point. In the case in >> question several novel morphological-mechanical structures had to >> evolve to enable the release, and whether those evolved gradually or >> not along the stem branch is an open question -- perhaps unknowable >> from analyzing extant taxa. >> >> In thinking about this a bit more, though, I think methods for an >> epoch release may be able to accommodate a clade-specific scenario, >> at least if the timing of the mode shift is specified using SIMMAP >> trees and not a shift-time parameter. Slater 2013 has R code >> associated with it that implements release models, and there's a >> multivariate implementation of release (and other mode shift) models >> in the new mvMORPH package (which is on CRAN, but not yet published >> as a manuscript as far as I can tell). >> >> In skimming Slater's code it doesn't seem to use SIMMAP trees. >> However, mvMORPH's shift specification can be done with SIMMAP trees, >> and I see no reason why its mvSHIFT function would care whether the >> shift is clade-specific or not. I'm cc'ing Graham and Julien to see >> if they have something to add. Regarding the use of mvSHIFT, I don't >> have multivariate data; hopefully that won't be a problem. >> >> Cheers, >> Dan. >> >> Brian O'Meara wrote: >> >> Hi, Daniel. It's a bit arguable whether as alpha -> 0, OU -> BM: >> I think it should, but IIRC in OUCH this doesn't happen, and >> that's a deliberate choice. That said, I think that an OU with >> alpha near zero would be ok for your question, though you might >> want to think about how to penalize parameters (that is, for that >> clade there'd be an OU mean parameter that is unidentifiable >> (alpha of zero, so no pull, so no evidence for it): should you >> count this as a parameter when doing model comparison? I'd argue >> no: you're doing OU with alpha of zero as a kludgy hack to treat >> it as BM). >> >> Another approach would be to resuscitate the "censored" model of >> O'Meara et al. 2006. Slice your tree on the edge leading to the >> released clade (I guess this truly releases the clade to roam >> free of its relatives) so you have the paraphyletic non-released >> set and the released clade as separate trees. Then you can try >> fitting the same or different models to the two trees. The >> downside of this is that you must use an additional ancestral >> state (at the MRCA of the released clade); the upside is that any >> weird changes happening on the edge leading to the released clade >> aren't in the model and so don't affect the fit (you could >> imagine that whatever led the clade to be ecologically released >> happened somewhere on the stem edge, but you don't know where, >> and it could be associated with a major single shift in your >> continuous trait, too). You could try an OU on the non-released >> tree (let's call this A) and an OU on the released clade (B), OU >> on A and BM on B, etc. The only practical difficulty with this is >> constraining the cases where A and B are supposed to have the >> same model: by default, optimization will happen separately in >> different trees, but you can create a wrapper function that calls >> OUwie.fixed() separately on A and B but with the same parameters >> and adds the likelihood and then optimize the parameters in this >> wrapper function. >> >> Hope this helps, >> Brian >> >> >> On Fri, May 29, 2015 at 2:02 AM, Daniel >> Fulop<[email protected] <mailto:[email protected]>>wrote: >> >> Hi All, >> >> Do you know if there are any methods out there for fitting >> ecological release and release and radiate models that are >> clade-specific? That is, in which the change in mode (OU to >> BM) happens at the root of a clade instead of at specific >> time for the whole phylogeny (as in Slater 2013). >> >> As far as I know the closest out there are models in OUwie, >> say with a clade-specific second OU process with a very low >> alpha. However, I don't think that biologically OU is a good >> model for the trait and clade in question (though it is for >> the rest of the tree). I know that at the limit as alpha -> >> 0 OU converts to BM, but I would ideally still like compare >> standard models' fits (including OUwie models) to the fits of >> clade-specific release models. >> >> Any leads or suggestions would be much appreciated, >> especially about how to implement these clade-specific models >> with current tools or about how to "roll my own". >> >> Thanks! >> Dan. >> >> -- >> Daniel Fulop, Ph.D. >> Postdoctoral Scholar >> Dept. Plant Biology, UC Davis >> Maloof Lab, Rm. 2220 >> Life Sciences Addition, One Shields Ave. >> Davis, CA 95616 >> >> _______________________________________________ >> R-sig-phylo mailing list [email protected] >> <mailto:[email protected]> >> https://stat.ethz.ch/mailman/listinfo/r-sig-phylo >> Searchable archive >> athttp://www.mail-archive.com/[email protected]/ >> <http://www.mail-archive.com/r-sig-phylo%40r-project.org/> >> >> >> >> -- >> Daniel Fulop, Ph.D. >> Postdoctoral Scholar >> Dept. Plant Biology, UC Davis >> Maloof Lab, Rm. 2220 >> Life Sciences Addition, One Shields Ave. >> Davis, CA 95616 >> >> 510-253-7462 >> [email protected] <mailto:[email protected]> > -- Daniel Fulop, Ph.D. Postdoctoral Scholar Dept. Plant Biology, UC Davis Maloof Lab, Rm. 2220 Life Sciences Addition, One Shields Ave. Davis, CA 95616 510-253-7462 [email protected] [[alternative HTML version deleted]] _______________________________________________ R-sig-phylo mailing list - [email protected] https://stat.ethz.ch/mailman/listinfo/r-sig-phylo Searchable archive at http://www.mail-archive.com/[email protected]/
