tag:blogger.com,1999:blog-8499895524521663926.post8069017263513707672..comments2024-11-06T04:50:42.713-05:00Comments on Phylogenetic Tools for Comparative Biology: Stochastic mapping with a bad informative prior results in parsimony reconstructionLiam Revellhttp://www.blogger.com/profile/04314686830842384151noreply@blogger.comBlogger1125tag:blogger.com,1999:blog-8499895524521663926.post-73180221641155613252013-11-16T11:40:20.734-05:002013-11-16T11:40:20.734-05:00hi Liam!
So I've been playing around with thi...hi Liam!<br /><br />So I've been playing around with this again, and I'm comparing the results obtained using the empirical bayesian approach and the full MCMC approach. I'm kinda surprised that the results are so different, and wondering if this is expected.<br /><br />My tree has 47 tips, 7 have state 0 and 40 state 1. The seven tips with state 0 are quite sparsely distributed throughout the phylogeny.<br /><br />here's an example of what I get with the two approaches:<br /><br />> describe.simmap(make.simmap(tree, setNames(x[,1], row.names(x)), model='ARD', pi='estimated'))<br />make.simmap is sampling character histories conditioned on the transition matrix<br />Q =<br /> 0 1<br />0 -22.794297 22.794297<br />1 4.089519 -4.089519<br />(estimated using likelihood);<br />and (mean) root node prior probabilities<br />pi =<br /> 0 1 <br />0.1521182 0.8478818 <br />Done.<br />1 tree with a mapped discrete character with states:<br /> 0, 1 <br /><br />tree has 153 changes between states<br /><br />changes are of the following types:<br /> 0 1<br />0 0 78<br />1 75 0<br /><br />mean total time spent in each state is:<br /> 0 1 total<br />raw 3.9761632 17.6508515 21.62701<br />prop 0.1838517 0.8161483 1.00000<br /><br /><br />> describe.simmap(make.simmap(tree, setNames(x[,1], row.names(x)), model='ARD', Q='mcmc', pi='estimated'))<br />Running MCMC burn-in. Please wait....<br />Running 100 generations of MCMC, sampling every 100 generations. Please wait....<br />make.simmap is simulating with a sample of Q from the posterior distribution<br />Mean Q from the posterior is<br />Q =<br /> 0 1<br />0 -3.221033 3.221033<br />1 0.596259 -0.596259<br />and (mean) root node prior probabilities<br />pi =<br /> 0 1 <br />0.1561995 0.8438005 <br />Done.<br />1 tree with a mapped discrete character with states:<br /> 1 <br /><br />tree has 15 changes between states<br /><br />changes are of the following types:<br /> 0 1<br />0 0 4<br />1 11 0<br /><br />mean total time spent in each state is:<br /> 0 1 total<br />raw 1.70813038 19.9188843 21.62701<br />prop 0.07898133 0.9210187 1.00000<br /><br />I'm kinda surprised there is an about tenfold difference in the estimated transitions between the characters. Of course intuitively it feels that less transitions make more sense, but as you explore in this post, that needn't be the case. But the difference between the transition matrices is quite huge too. Is it because the default priors might not be reasonable for this case? <br /><br />Thanks for any help, cheers!Rafael Maiahttps://www.blogger.com/profile/02395303011366399867noreply@blogger.com