Date:         Mon, 29 Oct 2001 10:45:01 +0100
Reply-To:     Gijs Dekkers <gd@PLAN.BE>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Gijs Dekkers <gd@PLAN.BE>
Organization: planbureau / bureau de plan
Subject:      Re: question on FACTOR
Content-Type: text/plain; charset=us-ascii

Dear David,

This is the source code:

proc factor data=datin
  method=principal
  scree
  mineigen=0
  priors=asmc
  rotate=varimax
  nfactors=4
  corr
  outstat=factout
  out=datuitF ;
  var spreken sarm cowoko garbeid is_wcomfort
  ziekte is_opl1 is_tevr is_socint1 is_socint2 is_socint3 ;
run ;

As you can see, there is nothing special about it. But the results are
stunning.

1. In the case of an orthogonal rotation, some of the factors all of a
sudden are correlated.
2. In the case of an oblique rotation, it sometimes happens that the
"inter-factor correlations" are not equal (even of the opposite sign) from
semipartial correlations which I derived afterwards, using PROC CORR.

I serially omit all variables (i.e. I let them out of the analysis one after
another) to find out if it was just one variable which caused this, but no.
It is a complete mistery to me. Any help would be greatly appreciated.

Gijs

"David L. Cassell" wrote:

> Gijs Dekkers replied [in part]:
> > That, unfortunately, is not the issue. I agree with you on the scree,
> but my problem
> > arises way after the number of factors have been chosen. It is very
> simple: I choose 3
> > factors and did not rotate the solution. I then used PROC CORR to find
> 3 factors that
> > are uncorrelated.
> >
> > Next, I run the same program, but this time, I rotated orthogonally.
> And, all of a
> > sudden, my factors (_not_ pseudo-factor scores) were highly
> correlated.  And that's
> > when I was completely flabbergasted...
>
> I would be too.  PROC FACTOR by default will generate uncorrelated
> factors.  But not
> all rotations are created equally.  Which rotation did you use?  Many
> optimize other
> features of the data by making a non-orthogonal rotation of the data,
> after which the
> new factors would indeed be correlated.
>
> David
> --
> David Cassell, CSC
> Cassell.David@epa.gov
> Senior computing specialist
> mathematical statistician

--
dr. Gijs Dekkers
Federaal Planbureau
Algemene Directie
Kunstlaan 47-49
B 1000 Brussel
++32/(0)2/5077413
fax  7373