Hi, Rafa. One thing to look at is whether the models are effectively identical; if the 3 state is 2 AICc worse than the 2 state it could be that it’s just adding a single extra parameter with no improvement (i.e, the likelihoods might be the same). In some of our other software we now have code to detect functionally redundant models and remove this redundancy, but I don’t remember whether it’s in corHMM yet.
Assuming the models are somewhat different, much as I like model averaging I’d use the one with lower AICc for ancestral state reconstruction (and yep, glad you brought up all the uncertainties with anc recon). With hisse, with model averaging one can basically model the turnover, net diversification, and other diversification-flavored rates: this edge is 0A in this model, and 0C in that model, but one can still just take the turnover rate from each. It’s partly a benefit from hisse’s relative inattention to the variation in transition rates. But corHMM is all about the transition rates: in one model, we have rates 0A->1A, 0B->1B, while in the other we have 0A->1A, 0B->1B, and 0C->1C: it’s not that the overall rate of 0->1 is the unweighted mean of the 0*->1* rates in each model, as state B might be really rare in one model and common in another. One can think about what happens at equilibrium with the estimated rate matrix, or use the simmap functionality in corHMM to look at empirical frequencies of going from 0->1, but it’s an area where I think we need to think a bit more about how to summarize and present results (for one thing, I think summarizing by wait times (reciprocal of rates) rather than rates can be a better way to summarize, so an ephemeral state with a high outgoing rate that a species is expected to occupy for only a short time does not have a huge impact on the average parameter). I hope this helps; I’m CCing Jeremy Beaulieu and James Boyko in case they’re not seeing R-sig-phylo posts. Best, Brian _________________________________________ Brian O’Meara He/Him Professor, Dept. of Ecology & Evolutionary Biology University of Tennessee, Knoxville From: R-sig-phylo <r-sig-phylo-boun...@r-project.org> on behalf of Rafael S Marcondes <raf.marcon...@gmail.com> Date: Thursday, November 14, 2024 at 12:11 PM To: r-sig-phylo <r-sig-phylo@r-project.org> Subject: [R-sig-phylo] Model-averaging corHMM models? Hi all, I have two corHMM models within 2 AICc units of each other. They differ in the number of hidden states (2 versus 3). I'm trying to decide how to proceed from here and wondering if model-averaging would be the way to go. I'm interested in drawing biological conclusions based on the parameter estimates, and in doing ancestral character estimation (strictly for visualization purposes only--I'm well aware of the caveats of over-interpreting ACE states). corHMM doesn't seem to have an in-built model averaging utility. But by analogy with the one in hisse (modelAveRates), it seems as if it's just a matter of averaging the parameter estimates based on model AIC weights. Is it really that simple? How would I deal with the param present in one model and not the other (related to the additional hidden state)? And then would it be appropriate to run ACE using the averaged rate parameters? Or would it be more appropriate to perform ACE separately for each model and then average the node likelihoods instead? Thanks, -Rafa *--* *Rafael S. Marcondes, Ph.D. (he/him)* (The R in Rafael is pronounced like the h in "hat") *https://www.rafaelmarcondes.com/ <https://www.rafaelmarcondes.com/>* Faculty Fellow in EEB Department of BioSciences Rice University Houston TX 77005 *"Eu quase que nada não sei. Mas desconfio de muita coisa"* *"I almost don't know nothing. But I suspect many things"* -João Guimarães Rosa, Brazilian novelist (Portuguese original and free English translation by me) [[alternative HTML version deleted]] _______________________________________________ R-sig-phylo mailing list - R-sig-phylo@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-phylo Searchable archive at http://www.mail-archive.com/r-sig-phylo@r-project.org/ [[alternative HTML version deleted]] _______________________________________________ R-sig-phylo mailing list - R-sig-phylo@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-phylo Searchable archive at http://www.mail-archive.com/r-sig-phylo@r-project.org/
