Re: [R-sig-phylo] weirdness in confidence intervals returned by ace, pic option
On Wed, Mar 23, 2011 at 10:24 PM, Emmanuel Paradis emmanuel.para...@ird.fr wrote: Hi Nick, With method = pic, the CIs are computed using the expected variances under the model, so they depend only on the tree. I've added a paragraph in the man page to explain this. Cheers, Emmanuel Obviously, though, if one set of tip data ranges from 100-1000, and another set of tip data ranges from 0.001-1, and both are mapped on the same tree, the variances confidence intervals will be different at the estimated internal nodes for each trait. But I was getting identical confidence intervals across ~20 traits of widely different magnitudes -- and then when I rescaled the trait data, the problem went away, and each trait had different-sized CIs like you would expect. It was as if there some bizarre bug where there was a memory error or some such and the same CI values were getting copied from one trait to the next in some situations. It's probably pointless to discuss, though, unless someone else can replicate the problem I saw with the code I posted. If the problem disappears on other computers, then it's a local problem. R.app sometimes throws messages about memory errors and the like for no apparent reason (not specifically associated with this, though), so it might not even be an APE issue. Cheers! Nick Nick Matzke wrote on 22/03/2011 12:30: Hi all, This isn't crucial to my work at the moment since I am not using the PIC option of ace to do ancestral character estimation. But while trying it out I noticed a very weird result that I can't explain...basically when I run ace on my raw trait values, I get the same sized confidence interval (97.5% CI minus 2.5% confidence interval) for all of my (drastically different) traits. ___ R-sig-phylo mailing list R-sig-phylo@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-phylo
Re: [R-sig-phylo] weirdness in confidence intervals returned by ace, pic option
Ah -- I get it now. Even my rescaled trait data have the same width 95% CI, e.g. my CI for a particular node could be mean +/- 2.345 whether my input trait data ranges from 0.3-0.5, or 400-800. My fix with scaled data occurred because I was back-transforming, which scaled the size of the 95% CI with the size of the mean. So I guess neither of these options is a real estimate of the CI, unlike when one runs ace, method=ML. I noticed the same behavior using ace, method=gls, so that should be noted as well. I have been using method=ml for ancestral character estimation, the width of its CIs vary as you might expect, so I was just surprised when PIC GLS didn't exhibit the same behavior. Cheers, Nick On 3/24/11 12:21 AM, Nick Matzke wrote: On Wed, Mar 23, 2011 at 10:24 PM, Emmanuel Paradis emmanuel.para...@ird.fr wrote: Hi Nick, With method = pic, the CIs are computed using the expected variances under the model, so they depend only on the tree. I've added a paragraph in the man page to explain this. Cheers, Emmanuel ___ R-sig-phylo mailing list R-sig-phylo@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-phylo
Re: [R-sig-phylo] weirdness in confidence intervals returned by ace, pic option
Hi Nick, With method = pic, the CIs are computed using the expected variances under the model, so they depend only on the tree. I've added a paragraph in the man page to explain this. Cheers, Emmanuel Nick Matzke wrote on 22/03/2011 12:30: Hi all, This isn't crucial to my work at the moment since I am not using the PIC option of ace to do ancestral character estimation. But while trying it out I noticed a very weird result that I can't explain...basically when I run ace on my raw trait values, I get the same sized confidence interval (97.5% CI minus 2.5% confidence interval) for all of my (drastically different) traits. ___ R-sig-phylo mailing list R-sig-phylo@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-phylo