Date:         Fri, 24 Jun 2005 16:03:19 -0500
Reply-To:     "Swank, Paul R" <Paul.R.Swank@UTH.TMC.EDU>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "Swank, Paul R" <Paul.R.Swank@UTH.TMC.EDU>
Subject:      Re: proc GLM output
Content-Type: text/plain; charset="us-ascii"

I don't think the B's are the problem since they just indicate that the
parameters are not unique, which is the case with class variables since
many different parameterizations can be done. The real issue is the
missing statistics for those categories with data but not the "last"
category. This would seem to indicate multicollinearity, that is some
categories are completely confounded with others, making the parameter
redundant. But with some many categorical variables, it will be a
nightmare to figure out which ones.

Paul R. Swank, Ph.D.
Professor, Developmental Pediatrics
Medical School
UT Health Science Center at Houston

-----Original Message-----
From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On Behalf Of
David L. Cassell
Sent: Friday, June 24, 2005 11:21 AM
To: SAS-L@LISTSERV.UGA.EDU
Subject: Re: proc GLM output

Baris Sagiroglu <zawalazingo@yahoo.com> replied:
> I need to know which variables make sense.

Which is a logical thing to want to do.  But you're trying to do too
much with too little.

>                                            I'd love to have stepwise
glm but
> as far as I know there's no such thing.

As Peter pointed out, stepwise selection is a BAD thing.  One thing
which doesn't get discussed in stepwise selection literature is that
using it on categorical variables is EVEN WORSE!  The whole process is
predicated on that continuous-variables-with-normal-errors
underpinnings.  You can't use stepwise selection on categorical
variables and get meaningful results.  So just don't even consider that.

>                                          I also thought about using
proc reg
> to get the collinearity diagnostics but I have to turn those 40
variables
> into dummies and I don't think proc reg can handle that.

It can.  I wouldn't do it, but it can.  And there are SAS tools to do
the work for you, like PROC GLMMOD.

>                                                           I ran proc
glm with
> few variables too (like 4 or 5 variables) and I still have the 'B'
next to
> the estimates so I don't think I can get rid of it. Thanks.

If you're running PROC GLM with 4 or 5 IVs and still getting those 'B'
notes, then there's something wrong with your data or your variables.
Do you have multi-collinearity that is this bad?  Do you have missing
cases or missing data?

Try stuffing those exact same 4 or 5 IV's into PROC REG without
bothering to do any kind of dummy variable creation (you don't care
about the estimates, just the VIFs), and see what kind of VIFs you get.

HTH,
David
--
David Cassell, CSC
Cassell.David@epa.gov
Senior computing specialist
mathematical statistician