Rotations for PCA?

[r1-q]

Rotations for PCA?

[sdray]

You can do the same with $li instead of $l1. The second contains the same information than $li but we a variance = 1 (variance equals the eigenvalue for $li). If you aims to draw biplot (superposition of row and columns), you should use $co and $l1 (correlation biplot) or $c1 and $li (distance biplot). If you use separate plots for row and columns, you can draw $li and $co.

[sdray]

Hi,
there is no function to do that in ade4. Varimax is a procedure of rotation allowing to perform such task. It is mainly used in psychometry and available in package psych or function varimax in package stats. If you want, you can apply on the outputs of ade4:

library(ade4)
irisX <- iris[,1:4] 
pca1 <- dudi.pca(irisX, scannf = FALSE)
vv < - varimax(as.matrix(pca1$co))
vv$loadings # rotated variables
as.matrix(pca1$l1)%*%vv$rotmat  # rotated individuals

This is equivalent to

psych::principal(irisX, rotate="varimax", nfactors=2, scores=TRUE)

Cheers

[r1-q]

Thanks a lot Stephane! So, that is pretty cool - I wasn't aware of that stats::varimax function!

One last thing though: I usually only use the $li components of the dudi objects, which I believe are the coordinates of the individuals on the principal axes (I always use that with s.label... I hope I've got that right!) - what does the $l1 component refer to and why do we use that one specifically in combination with the rotation matrix?

Thanks a lot
Erwan

[r1-q]

Thanks Stephane! That shed some light on the subject indeed