Date: Wed, 28 Apr 2010 19:41:09 -0400
Reply-To: Peter Flom <peterflomconsulting@mindspring.com>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Peter Flom <peterflomconsulting@MINDSPRING.COM>
Subject: Re: ordinal logistic vs multinomial logistic models
Content-Type: text/plain; charset=UTF-8
Hi Dale
How odd that everyone says that PROC GENMOD will do this. Of course, I am guilty of the same thing - I read it somewhere and assumed the statement was right.
But you are the wizard of NLMIXED
Peter
-----Original Message-----
>From: Dale McLerran <stringplayer_2@YAHOO.COM>
>Sent: Apr 28, 2010 2:59 PM
>To: SAS-L@LISTSERV.UGA.EDU
>Subject: Re: ordinal logistic vs multinomial logistic models
>
>Peter,
>
>The partial proportional odds model that you mention in your
>NESUG presentation is intriguing. I have not yet found a good
>cogent explanation of how the PPO model can be fit using the
>GENMOD procedure (although I have seen a number of references to
>the use of GENMOD to fit such a model). However, if one searches
>support.sas.com for the exact phrase "partial proportional odds",
>one can find an example demonstrating the use of (what else but)
>the NLMIXED procedure. See: http://support.sas.com/kb/22/954.html.
>
>The example which is produced actually fits a fully non-proportional
>ordinal logistic regression model. Contrasts of the parameters
>for a given variable across levels of the response are employed
>to test the proportional odds assumption for each predictor
>and also for the entire set of predictor variables. It is of
>some interest to observe that the PO assumption is not rejected
>for any of the predictors examined separately, But when the
>entire set of predictors is taken as a whole, the PO assumption
>is rejected. That is something of a conundrum.
>
>I would note that one could constrain the PPO model such that
>a proportional odds model is assumed for some SUBSET of the
>predictors and not assumed for a different subset of the predictors.
>That really is the "Partial Proportional Odds" model. Employing
>the data which were used for the example given with the above
>link and fitting PPO models in which we impose the assumption
>of proportionality for each variable in turn (fit a model in
>which PO is assumed for center only, fit a model in which PO
>is assumed for baseline only, fit a model in which PO is assumed
>for all of the treatment indicators but not for center or
>baseline), we can construct likelihood ratio tests of the PO
>assumption by comparing the value of -2LL for the partially
>constrained model against the value of -2LL for the unconstrained
>model. (We can also construct a likelihood ratio test for the
>entire set of predictors by comparing the value of -2LL for the
>PO model with the value of -2LL for the totally unconstrained
>ordinal model.)
>
>Now, the likelihood ratio tests for center and baseline are both
>significant at p<0.05 (p=0.023 for center, p=0.015 for baseline).
>The contrast statements had produced p-values of approximately
>0.18 for both. The LRT for the treatment effect produces p=0.157
>whereas the contrast statement p-value for the treatment effect
>was p=0.076. The LRT of the PO assumption for all variables
>simultaneously had p-value 0.020 (compared to the CONTRAST
>p-value of 0.006).
>
>The difference in p-values for the LRT vs CONTRAST statements
>is disconcerting, to say the least. However, it would seem that
>the LRT p-value computations are reasonably consistent in that
>there were two variables (center and baseline) for which the PO
>assumption was rejected and the PO assumption was also rejected
>for the full model. Compare these results against what was
>observed when using CONTRAST statements in the fully
>non-proportional odds model. Based on an N of 1 trial, it would
>appear that the use of CONTRAST statements to test PO assumptions
>for variable subsets in a fully non-proportional odds ordinal
>model may not be warranted.
>
>The little bit that I have seen about how to fit a PPO model
>employing the GENMOD procedure indicates that the PPO model is
>fit employing generalized estimating equations (GEE). As such
>
> 1) a different result may be expected when using GENMOD
> than is obtained using NLMIXED
>
> 2) a likelihood ratio test cannot be employed to test PPO
> assumptions when one has fit a GEE
>
> 3) ergo, CONTRAST statements must be relied on for testing
> the PO assumption if using GENMOD
>
>
>While one would expect results for a PPO model fitted employing
>the GENMOD procedure to differ in explicit values from a PPO
>model fitted employing NLMIXED, we might expect that hypothesis
>tests based on CONTRAST statements would be similar. Based on
>the previous observations of the inconsistencies of the hypothesis
>tests in the fully non-proportional odds model, I would have
>concern that results from GENMOD could be equally inconsistent.
>
>Dale
>
>---------------------------------------
>Dale McLerran
>Fred Hutchinson Cancer Research Center
>mailto: dmclerra@NO_SPAMfhcrc.org
>Ph: (206) 667-2926
>Fax: (206) 667-5977
>---------------------------------------
>
>
>--- On Wed, 4/28/10, Peter Flom <peterflomconsulting@MINDSPRING.COM> wrote:
>
>> From: Peter Flom <peterflomconsulting@MINDSPRING.COM>
>> Subject: Re: ordinal logistic vs multinomial logistic models
>> To: SAS-L@LISTSERV.UGA.EDU
>> Date: Wednesday, April 28, 2010, 5:46 AM
>> DMAK posted the following:
>>
>> >
>> > how can one compare the two models in terms of
>> goodness of fit.
>> >
>> >
>>
>> I actually discussed this in a paper for NESUG:
>>
>>
>> www.nesug.org/Proceedings/nesug05/an/an2.pdf
>>
>> I'd be interested in the reactions of other stats type
>> people, and I hope it's useful
>>
>>
>> Peter
>>
Peter L. Flom, PhD
Statistical Consultant
Website: http://www DOT statisticalanalysisconsulting DOT com/
Writing; http://www.associatedcontent.com/user/582880/peter_flom.html
Twitter: @peterflom
| 
