Date:         Wed, 8 Jan 1997 22:56:01 GMT
Reply-To:     Steve Gregorich <gregorich@PSG.UCSF.EDU>
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From:         Steve Gregorich <gregorich@PSG.UCSF.EDU>
Organization: University of California, San Francisco
Subject:      Re: Nested Logit Model?!?!?
Content-Type: Text/Plain; charset=ISO-8859-1

Well, no proc but SAS DOES have a macro
for nested logit models--otherwise known
as generalized linear mixed models.
You may be interested in GLIMMIX.SAS (at
the SAS web site).
 
Its a sas macro that iteratively
calls PROC MIXED to solve generalized
linear mixed models.  Some of it uses
IML, but mostly it uses PROC MIXED
and macro language--so you can see
what transformations are performed
to generalize the capabilities of
PROC MIXED.  GLIMMIX was written
by Russ Wolfinger.
 
I should say that the Marginal
Quasi-LIkelihood and Pseudo
Quasi-Likelihood approaches
available in GLIMMIX.SAS (and
most any other software
performing generalized
linear mixed models) has
be shown to produce downwardly
biased parameter estimates in
many situations when the logit
link function is used.  See for
instance the Multilevel Projects
Homepage for a short article on the
iterative boostrap bias correction
procedure:
http://www.medent.umontreal.ca/multilevel/workpap.html#biascorr
 
If your model is not too complex (i.e., two
"levels", up to 8? random effects) you may
be interested instead in Don Hedeker's
MIXOR program (DOS & Windows).  His program
uses a Gaussian quadrature approach to
yield Maximum Marginal Likelihood estimates.
This approach apprears to be supperior (when
it can accomodate the dimensions of the problem).
 
Good Luck!
 
Steve Gregorich
University of California, San Francisco
Prevention Sciences Group
74 New Montgomery Street
San Francisco, CA  94105
gregorich@psg.ucsf.edu
 
In article <32D25DE1.5A23@cma-mkting.com>, scot@cma-mkting.com says...
>So, I just got off the phone with SAS/Stat tech support.  'Said that
SAS doesn't have a "proc" for running a nested logit model.  Well, fine,
I'm not sure that I really need a proc, IML code is fine with me.
___________________________________________________________
>       |                                                           |
>       | Scot A. Richardson        Project Mgr Statistical Services|
>       | scot@cma-mkting.com       UNIX System Admin               |
        |                           Conway/Milliken & Associates    |