An easy way to do this is to use PROC GLMMOD, which builds your regression
matrix the way PROC GLM would, building dummy variables, interactions, etc.
Glad I saw this post. Right now I'm working on a model with high overall
fit, and using some selection model seemed like a good way to cut out some
fat. I'll look into your oft-repeated advice and see if PLS is appropriate.
BTW - on a complete tangent - I find it hard to follow threads in the
archives, because when people change the subject line when they reply it
breaks the thread. Is there a FAQ or welcome message that tells people not
to do this?
On Wed, 2 Jun 2004 15:40:51 -0700, David L. Cassell
> "Andrel, Jocelyn" <jandrel@WVMI.ORG> wrote:
> > Is there a Proc where you can use both interaction terms and a
> selection option (like SELECTION = STEPWISE) for linear regression?
> Proc Reg has the selection option, and Proc GLM can handle interaction
> terms, but I haven't found a Proc that does both (similar to what Proc
> Logistic can do).
> Your best option is: do not do stepwise regression.
> Seriously. I have written pages on this issue in SAS-L before.
> (You can bore yourself to tears by looking up my rants in the
> SAS-L archives at http://www.listserv.uga.edu/archives/sas-l.html
> if you want to. Just search for the keyword 'stepwise'.)
> Particularly when you are working with interaction terms, which
> by definition will be highly correlated with other variables
> in your model, stepwise regression can do bad things. You have
> no guarantee that you will get the right term, instead of some
> higher-order term which happens to be correlated.
> Second option: Build your own interaction terms. They're just
> your first-order terms, multiplied together. But be warned of
> the incipient risks. And decide up front how you are going to
> interpret an 'interaction' term which is the product of two
> continuous variables. What exactly is that supposed to mean in
> the real world? That is why you see this feature in PROC GLM but
> not in PROC REG.
> [Side note: I don't like the fact that stepwise selection on
> big heaping wads of interaction terms is even theoretically
> possible in PROC LOGISTIC. It opens up several big steaming
> cans of slimy worms that I would just as soon avoid.]
> David Cassell, CSC
> Senior computing specialist
> mathematical statistician