Date: Tue, 6 Jun 2006 16:36:16 -0700
Reply-To: David L Cassell <davidlcassell@MSN.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: David L Cassell <davidlcassell@MSN.COM>
Subject: Re: Question/rant about AIC and PROC REG and so on
In-Reply-To: <200606061015.k567XsfS002720@mailgw.cc.uga.edu>
Content-Type: text/plain; format=flowed
flom@NDRI.ORG replied:
><<<
> >1) Why does SAS not have AIC in GLM?
> >>>
>David Cassell replied
>
> I don't think PROC GLM is the right place, personally. Most of the
> time, that is not what we want PROC GLM for.
>
>But SAS output is not known for its terseness. :-)
>In PROC MIXED, one gets the AIC automatically. Why not in GLM?
Real reason? Probably because PROC MIXED is a lot newer.
>Is AIC more valuable in a PROC MIXED than in a GLM model that is
>really just a special case of MIXED?
I prefer to think of it as one more reason to switch to PROC MIXED.
:-)
>Also, while I agree that many people don't want AIC and related
>statistics
>when they do a GLM, I think more people should want it, because it helps
>us do model comparisons in a more sensible way. I also think that we
>should
>be looking at multiple models in many cases where we do not.
And it is certainly better than having people sitting around saying
"Well, my R squared is higher, so this must be the better model.
That's what my instructor said last year..."
>Of course, that reflects my background and work, where I mostly work
>with
>social science data. It might be different in other fields.
It varies. In some fields, it is easier to bulid models based on available
research findings. In some fields, there are standard approaches
which may (or may not) help avoid these kinds of issues.
>David:
>
> The SELECTION= bit is annoying. But OTOH, why would people
> be after the AIC or Mallows' Cp unless they were fitting more than
> one model? Under normal circumstances, AIC seems like non-useful
> clutter when fitting a single model in PROC REG. If you have K
> variables and you insist on START=k and STOP=k, then you only
> get one model fit. But I think that's a kludge.
>
>But this gets into one of your own favorite rants, David - that is, the
>use of automatic
>selection methods. If one has come up with several models that make
>substatnive sense,
>then one could run them as several PROC GLMs and compare the AIC (if it
>were output);
>one could do it in PROC REG and get more diagnostics more easily (but
>then one has to do the
>dummy coding on one's own, which can be a nuisance)
Yeah. I opt for other approaches.
Instead of doing the dummy coding by hand, let PROC TRANSREG do
it all for you, and also build a macro variable &TRG_IND that holds the
complete string of newly-created dummy variables for easy insertion
into other procs. If you were doing something like creating all the
dummy variables for A1, A2, and also A1*A2, then this can be a *big*
help.
Or do it all in PROC MIXED.
David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330
_________________________________________________________________
Express yourself instantly with MSN Messenger! Download today - it's FREE!
http://messenger.msn.click-url.com/go/onm00200471ave/direct/01/
