Dear Jovana, Among alternative methods suggested by Michael, you can consider regularized versions of the PCA that can handle these types of collinearity-based singularities. Moreover, this often improves the estimation of the explained variance (i.e. the eigenvalues), as classical approaches have a systematic upward bias for the first eigenvalues when the sample size is not very large. You can perform the ‘phylogenetic PCA’, or the ‘regularized phylogenetic PCA’ using the ‘mvgls.pca’ function in mvMORPH.
Best wishes, Julien ________________________________________ De : R-sig-phylo <r-sig-phylo-boun...@r-project.org> de la part de Liam J. Revell <liam.rev...@umb.edu> Envoyé : jeudi 21 novembre 2024 13:39 À : Michael Zyphur <mzyp...@instats.org>; Jovana Malikovic <jovana.maliko...@anatomy.uzh.ch> Cc : r-sig-phylo@r-project.org <r-sig-phylo@r-project.org> Objet : Re: [R-sig-phylo] Handling negative values_phyl.pca function Dear Jovana. The only thing I'd add to Michael's comment is that a common cause of zero (negative because of numerical imprecision) eigenvalues is when one column of your data matrix is a perfect linear function or linear combination of one or more other columns. To inspect the eigenvalues (as opposed to their square-roots, which is what is output by the "phyl.pca" print method) just enter diag(obj$Eval) at the command prompt in R (in which obj is the object created by your phytools::phyl.pca function call). Best wishes, Liam Liam J. Revell Professor of Biology, University of Massachusetts Boston Web: http://faculty.umb.edu/liam.revell/ Book: Phylogenetic Comparative Methods in R <https://press.princeton.edu/books/phylogenetic-comparative-methods-in-r> (/Princeton University Press/, 2022) On 11/21/2024 7:23 AM, Michael Zyphur via R-sig-phylo wrote: > CAUTION: EXTERNAL SENDER > > Hi Jovana > > The warnings you’re encountering (NaNs were generated) are likely > because the phyl.pca function uses eigenvalue decomposition. When > eigenvalues are negative or very close to zero, taking the square root > as part of the process (e.g., sqrt(V[i, i] * obj$Eval[j, j])) results > in NaN values. > > Negative eigenvalues can arise due to numerical imprecision, or they > may indicate a problem with the structure of your dataset or the > phylogenetic covariance matrix. If you've already identified Z-score > transformation as a potential issue, it might be worth investigating > further. Here are a few steps to try: > > Check the Data: > Ensure that the input data (data.mean) and the phylogenetic tree > (tree) are appropriately scaled and formatted. Look for extreme values > or cases where all variables for a particular taxon are identical. > > Inspect the Phylogenetic Tree: > Ensure the tree is ultrametric and well-resolved. If the tree has > branch lengths that are zero or very small, these can lead to > computational issues. Tools like ape::is.ultrametric(tree) can help > verify this. > > Use Alternative Transformations: > If the Z-score transformation introduces negative values that may be > problematic, you could experiment with other normalization techniques > (e.g., range scaling to [0,1]). > > Try Alternative PCA Methods: > If the problem persists, consider using other PCA methods tailored for > phylogenetic data, such as prcomp with pre-transformed data or robust > PCA options in the phytools or vegan packages, depending on your data > structure. > > If none of these approaches resolve the issue, sharing a reproducible > example of your data and code could be useful? > > > Best wishes > > Michael Zyphur > Director > Institute for Statistical and Data Science > instats.org > > > > Best wishes > > Michael Zyphur > Director > Institute for Statistical and Data Science > instats.org > > > On Thu, 21 Nov 2024 at 23:08, Jovana Malikovic > <jovana.maliko...@anatomy.uzh.ch> wrote: >> Dear all, >> Can someone explain the correct way to handle this warning message? >> >> data.ppca <-phyl.pca(tree, data.mean) >> data.ppca >> >> Warnings: >> 1: In sqrt(V[i, i] * obj$Eval[j, j]) : NaNs were generated >> 2: In sqrt(V[i, i] * obj$Eval[j, j]) : NaNs were generated >> 3: In sqrt(V[i, i] * obj$Eval[j, j]) : NaNs were generated >> 4: In sqrt(V[i, i] * obj$Eval[j, j]) : NaNs were generated >> 5: In sqrt(V[i, i] * obj$Eval[j, j]) : NaNs were generated >> >> >> I understand that this happened because of the negative values in the >> dataset after the Z-score transformation, but I don't know if there is a way >> to solve it. >> >> Thank you in advance, >> Jovana >> >> [[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 athttp://www.mail-archive.com/r-sig-phylo@r-project.org/ > _______________________________________________ > R-sig-phylo mailing list -R-sig-phylo@r-project.org > https://stat.ethz.ch/mailman/listinfo/r-sig-phylo > Searchable archive athttp://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/ _______________________________________________ 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/
