Hello all, I have been using ace() to reconstruct ancestral states for a number of continuous morphological traits. I need confidence intervals for my analysis. However, for some sets of traits, I get the following error:
Warning message: In sqrt(diag(solve(out$hessian))) : NaNs produced The only tangible result of this error is that the second sigma2 estimates is reported as NaN. However, if I get this error, I find several things very suspect when I explore my data: confidence intervals tend to be very large and ancestral states are reconstructed which are outside the range of variation shown by my tips (values below zero for a length trait, for example). I have been unable to repeat this error with random trees or a random trait dataset. I don't get this error when I use getAncCharacters() or when I use ace(,method="pic"). Both of these also estimate character states within the original range of variation. The "pic" method gives confidence intervals, but they are very different from the ones given by ML estimation. They tend to be either overly large for some traits (many times larger than the range of variation for my entire clade) or unreasonably small for other sets of traits. I'm not familiar enough with the math behind CI estimates for ML and least squares to understand why these should be so different for the same trait data and tree. I don't think this matters, but none of the tips are coeval because they are various extinct lineages. The tree also has a large number of short branches (however, all branches have been modified so that no branch is less than 1 MY in length; total length of the tree is about 40 MY). Looking at the mail archives, it appears my office mate, Annat Haber, has reported the same error message. She isn't in today, however, or I would turn around in my office chair and ask her how she dealt with it. -Dave Bapst, UChicago _______________________________________________ R-sig-phylo mailing list R-sig-phylo@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-phylo
