pgilbert at bank-banque-canada.ca
Fri Jan 21 21:48:29 CET 2005
I released a new package called GPArotation in the devel area of CRAN.
This package uses the gradient projection algorithm of Bernaards and
Jennrich <http://www.stat.ucla.edu/research/gpa> to do factor rotation.
The R package is based on code from their web site. Available rotation
objective criteria are "oblimin", "quartimin", "target", "pst",
"oblimax", "entropy", "quartimax", "varimax", "simplimax", "bentler",
"tandemI", "tandemII", "geomin", "cf", "infomax" and "mccammon".
I have done a certain amount of testing of "oblimin" and it appears to
work well. More extensive testing of all criteria and comparisons with
known results would be very much appreciated. Beware that the default is
not to do Kaiser normalization, which often is the default for
commercial implementations of some of these criteria (but it does not
make sense for others).
In superficial testing of some of the criteria I have found that the
loadings matrix appears to stabilize in the sense that it does not
change when the number of iterations is increased, but the gradient
based convergence criterion does not signal convergence. I suspect this
is because the gradient is large even in close proximity to the optimum.
As many of you are much more familiar with this problem than I am, I
would certainly appreciate suggestions for better convergence tests.
More information about the R-devel