Dear Colleagues,
I uploaded a new version of RRPP to GitHub that corrects a problem, that was
thankfully revealed by Carlo Meloro. The bug was subtle and only affected
linear models with GLS estimation for which the number of variables (p) exceeds
the number of observations (n). This would have impact on procD.pgls in
geomorph, for example. The issue was only in the calculation of SS, not the
estimation of coefficients. For calculating SS with high-high-diemnsional (p >
n) data, it makes sense to project data on the n - 1 possible principal
components (PCs) of the data, which makes computation faster over many
permutations. SS calculations follow coefficient calculations, which are
performed on GLS-transformed data (not PCs). Therefore, the criterion was if p
> n, reduce the transformed data to n - 1 dimensions. However, this produces
slightly different results than first reducing the untransformed data to n - 1
PCs and then transforming the PC projections.
Although this should be an obvious problem, it evaded checks using other
software, which requires univariate data or (n > p). I believe it was
introduced, incidentally, with RRPP 2.0.
If it sounds like I am describing an analytical situation similar to one you
might have, please update RRPP. Unfortunately, this issue does not cause any
errors. Of course, re-installing RRPP or geomorph from Github every now and
then is just a good practice to be most up to date.
To install RRPP from Github:
devtools::install_github("mlcollyer/RRPP", build_vignettes = TRUE)
Regards!
Mike
--
You received this message because you are subscribed to the Google Groups
"Morphmet" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To view this discussion visit
https://groups.google.com/d/msgid/morphmet2/9D221744-A8EA-4CA9-9F62-36DEFDD12365%40gmail.com.
