Date:         Mon, 28 Sep 1998 15:26:34 -0500
Reply-To:     "Gary D. & Franchesca D. Zenitsky" <gdzenitsky@MOCHA.MEMPHIS.EDU>
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From:         "Gary D. & Franchesca D. Zenitsky" <gdzenitsky@MOCHA.MEMPHIS.EDU>
Subject:      AIC in stepwise logistic?
Content-type: text/plain; charset="us-ascii"

I'm new to the list, so please excuse if the following question's topic is
too simplistic or not appropriate for this forum:
 
I'm running PROC LOGISTIC using stepwise model selection with a binary
response.  I'm comparing 4-5 models, with any where from 3 to 5 explanatory
variables selected from 7, with 140 observations.  The procedure  produced
these models from runs on the original data as well as on bootstrapped
data, all using SLE/SLR of 0.25.  To compare the models, I've been looking
at the c-index and the H-L lack-of-fit test.
 
Here's my question:  if low values of AIC indicate a more fit model, then
why are the intercept- only values for AIC much smaller than those with
covariates added in?  Also, the model with the least lack-of-fit (H-L
statistic) has the largest AIC value.  I'm left wondering what it all means
because the model I presumed to be "best" is the one with the largest AIC.
I understand that the number of obs and variables has a role here, and that
there may be a cut-off point that's above the minimum AIC.  The data is of
the field ecology kind, and therefore, whether or not the "true" model
(?-var) is contained within the global model (7-var) is unknown.
 
Oh yes, did I mention that I'm desperate for some expert insight on this
matter.  I do thank everyone in advance for any responses.
 
Cheers!
Gary D Zenitsky
Department of Biology
University of Memphis
Memphis, TN  38152
Phone wk/hm: 901-678-3322/386-6565
Fax: 901-678-4746
E-mail: gdzenitsky@cc.memphis.edu