Date:         Tue, 30 Jul 2002 09:51:34 -0400
Reply-To:     "H. Lawrence Hotchkiss" <larryh@ZEBRA.US.UDEL.EDU>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "H. Lawrence Hotchkiss" <larryh@ZEBRA.US.UDEL.EDU>
Subject:      Re: SAS-L Digest - 29 Jul 2002 to 30 Jul 2002 (#2002-1054)
              (comment)
Comments: cc: larryh@kodiak.us.udel.edu

Hi,

I'm responding to the question posed in the cut-and-paste (from SAS-L
digest) below the line.

I don't think you can use logistic regression to estimate a multinomial
logistic regression -- even if you do run separate binomial regressions.
There are 2 1/2 to 3 reasons:

     1. The predicted probabilities in a multinomial logistic regression
        add up to 1.0.  This would not be forced with multiple runs of
        proc logistic.

     2. The disturbance terms are uncorrelated in a multinomial logistic
        regression; they would not be in multiple runs of proc logistic.

     3. Unless you eliminate all cases except the reference group and the
        current "treatment" group, you would be comparing each
        "treatment" to all other groups instead of to the reference
        group.  This problem can be fixed, of course, by subsetting the
        sample.  But the statistical tests would not be correct.

Larry Hotchkiss

-----------------------------------------------------------------------------
Date:    Tue, 30 Jul 2002 00:51:20 GMT
From:    Hongjiew <hongjiew@AOL.COM>
Subject: use logistic regression for multinomial model

Dear Friends:

I wonder if there are any examples or reference papers to use multiple logistic
regression models to replace a multinomial model?

Suppose I have Y as a nominal variable (say segment membership) with 8 levels
(Y=A,B,C,D,E,F,G,H) where H is my reference category. I could use proc catmod
to estimate a multinomial model. But suing proc catmod is messy. I wonder if
there is a way I can use 7 logistic regression models to do the same job.
Suppose, I build models A vs. H, B vs. H ... (or should I build models A. vs.
NON A, B vs. NON B?) Also, suppose I end up with 7 scores for observation from
7 models. How do I normalize the scores (would the score for H be 1- sum(7
scores)? Would that make sense at all? How do I use these scores to classify
records into segments? Using Proc catmod, the sum of the scores will be
constrainted to 1 and therefore comparable probabilities.  I wonder if this is
still true for 7 logistic regression models?

Thanks for your help.

------------------------------