Date:         Tue, 6 Apr 2004 13:54:39 -0400
Reply-To:     Peter Flom <flom@NDRI.ORG>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Peter Flom <flom@NDRI.ORG>
Subject:      Re: Analysis of ordinal data
Content-Type: text/plain; charset=US-ASCII

I agree that Agresti is a great reference, in general.  But not for
this, in particular, as he does not cover the RC model.  I think a very
good reference for this particular model (and related ones, is Clogg and
Shihadeh, Statistical Models for Ordinal Variables pub by Sage.  I found
it clearer than Goodman.  However, they do not provide code in any
computer language.  They say it should be relatively easy to program
them in GAUSS or S.

I asked a similar question to Boris's a while back on SAS-L, and found
that no one had programed these models in SAS

I have a problem with the traditional loglinear models for data that is
really ordinal, in that the loglinear model is not, strictly speaking,
an ordinal one. When you assign 'an equally spaced score' you are
implicitly stating a distance between the different levels, and this is
precisely what ordinal models do NOT do.

For bivariate analysis where one variable is ordinal, I like PROC
FREQ's Jonckheere Terpstra test, but AFAIK there's no equivalent in SAS
for muliple predictors.

Peter

Peter L. Flom, PhD
Assistant Director, Statistics and Data Analysis Core
Center for Drug Use and HIV Research
National Development and Research Institutes
71 W. 23rd St
www.peterflom.com
New York, NY 10010
(212) 845-4485 (voice)
(917) 438-0894 (fax)



>>> Agustin Calatroni <acalatr@UMICH.EDU> 4/6/2004 1:33:15 PM >>>
Boris,
You can perform loglinear modeling with proc genmod by fitting a
Poisson
regression model (the ML estimates for the Poisson parameters are
identical to the corresponding ML estimates in the loglinar model). An
excellent reference is "Categorical Data Analysis" second edition by
Alan Agresti that describe in an appendix section the SAS code.

Using Agresti table 9.5 of political ideology data you can fit the row
effect model as follows

data party;
 input party $ ideology $ score count;
cards;
democ 1liberal  1 143
democ 2moderate 2 156
democ 3conserv  3 100
indep 1liberal  1 119
indep 2moderate 2 210
indep 3conserv  3 141
repub 1liberal  1 15
repub 2moderate 2 72
repub 3conserv  3 127
run;

proc genmod data=party;
 class party ideology;
 model count = party ideology score*party /dist=poisson link=log;
run;

In the above example party is treated as nominal and ideology as
ordinal. For the ordinal ideology I created a equally space score and
since party is nominal no score is needed. The SAS output uses dummy
variables for the first two categories of each classification, and the
interaction term is the product of the score for ideology and a
parameter for party.

HTH
Agustin Calatroni



-----Original Message-----
From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On Behalf Of
Boris Tawakoli
Sent: Tuesday, April 06, 2004 9:05 AM
To: SAS-L@LISTSERV.UGA.EDU
Subject: Analysis of ordinal data


Hello,

I have to analyze categorical data with an ordinal scale. I want to
use
loglinear models like the row effects model or the RC Model.

Using Proc Catmod by directly entering the design matrix allows me
estimate the row effects model, but the programming of the design
matrix
is quite time consuming and error prone. Has anybody an idea of how to
proceede more rapidly/ automatically? Does anybody know of any macro
that helps on that issue?

Another thing is the RC Model (Row Column Model by Goodman (?)). I
have
no clue how to estimate that one. It seems there is no easy way
because
the model is not linear. Unfortunately I can't find any example in
SAS.
Can anybody help?

Best regards and many thanks in advance

Boris Tawakoli