Hi, Others might have a better idea for the root cause of the differences, but one thing you could do to convince yourself which method works best is to simulate some similar data, and see how well each method performs.
E.g. to simulate a 2-state 26 taxon tree under an ER model, you could use Phytools, like this (slightly modified from Liam Revell's blog, here: http://blog.phytools.org/2013/03/marginal-ancestral-state-reconstruction.html require(phytools) tree <- pbtree(n=26,scale=1) Q <- matrix(c(-1,1,1,-1),2,2) rownames(Q)<-colnames(Q)<-c(0,1) x <- sim.history(tree,Q)$states Then just use those states to run your different methods, and see which methods come closer to the true values. You'd expect some variance given that you have a relatively small dataset, especially if your 1's and 0's are clumped together on your tree, in which case there won't be many changes with which to infer rates. A better test of the methods themselves would be to use a really big tree, ensuring that there's plenty of information to reconstruct rates, e.g. tree <- pbtree(n=250,scale=1) Cheers, Rob On 5 April 2013 01:15, Jingchun Li <[email protected]> wrote: > Hi all, > > I am exploring different methods for ancestral state reconstruction for a > small phylogeny I have (26 OTUs). I have one binary trait coded as 0s and > 1s for all the taxa. And I am more interested in the transition rates > between the two states. > > If I understand correctly, a maximum likelihood ancestral state > reconstruction using a MK1 model or an asymmetrical rates model should give > me the optimized estimated transition rates between the two states. The > formal will assume the two rates are the same, the later will > assume independent rates. > > So I firstly did this using the ape function in ace. > > ERreconstruction <- ace(traits_data, tree, type="discrete", model="ER") > ERreconstruction$rates > ERreconstruction$loglik > > ARDreconstruction <- ace(traits_data, tree, type="discrete", model="ARD") > ARDreconstruction$rates > ARDreconstruction$loglik > > Then I tried using Mesquite, using MK1 model and the "Asymmetrical 2-param. > Markov-k Model " > > Then I tried using BayesTraits, with the MultiState -> Maximum Likelihood > option, with no restrictions on q10 and q01. > > What puzzled me is that I'm getting different results from the three > softwares. > -------------------------- > aceER: > transition rate: 0.0059, likelihood: -7.00 > aceARD: > transition rate: 0.019 and 0.00, likelihood: -6.11 > > Mesquite MK1: > transition rate: 0.0059, likelihood: -7.70 > Mesquite Asymmetrical: > transition rate: 0.015 and 0.005, likelihood: -7.30 > > BayesTraits no restriction: > transition rate: 0.0057 and 0.0057, likelihood: -14.09 > -------------------------------- > > I can see that in both ace and Mesquite, > it slightly favors the asymmetrical model with two different rates, and the > rates and likelihoods are more of less comparable. But BayesTraits seems to > think the two rates should be equal, and it has a different likelihood. > > Does this has something to do with different searching algorithms? Or am I > missing something here? I am aware that the scaled likelihoods in ace are > the scaled conditional likelihoods, not the joint or marginal > reconstructions. But this should not affect the estimated transition rates? > > > Thank you very much! > > [[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]/ > -- Rob Lanfear Research Fellow, Ecology, Evolution, and Genetics, Research School of Biology, Australian National University phone: +61 (0)2 6125 3611 www.robertlanfear.com [[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]/
