Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Re: question about mixed-effects ordinal regression in STATA 13


From   Darcy Hannibal <dlhannibal@ucdavis.edu>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Re: question about mixed-effects ordinal regression in STATA 13
Date   Fri, 14 Feb 2014 11:58:38 -0800

As far as I could determine from trying it, gllamm only allows for all variables in the model to be fit to a proportional odds model or a non-proportional odds model and does not allow a partial-proportional odds model. Because I knew SuperMix would allow this if one or few variables did not meet the proportional odds assumption, I did all comparisons of proportional, non-proportional, and partial-proportional odds models in SuperMix. I only considered programs that allowed a mixed-effect ordinal logitistic model.

On 2/14/2014 10:25 AM, Stas Kolenikov wrote:
I am pretty sure I am missing something (proportional odds sounds like
coming from biometric literature, and I am used to look at the ordinal
models as econometric/social science models, and see the outcome as a
categorized continuous variable, rather than a bunch of different
outcomes "Low", "Medium", "High", "Very high"), but if the alternative
model is that with each category freely estimated, I think there are
several potential ways to fit a multinomial model with (correlated)
random effects, or a set of binary models for each category, using
user-written -gllamm- and/or -cmp- (the latter with random effects, on
top of what's in the award winning Stata Journal paper).

P.S. Stata is not an acronym, so you don't have to shout STATA.


-- Stas Kolenikov, PhD, PStat (ASA, SSC)
-- Principal Survey Scientist, Abt SRBI
-- Opinions stated in this email are mine only, and do not reflect the
position of my employer
-- http://stas.kolenikov.name



On Fri, Feb 14, 2014 at 12:58 PM, Darcy Hannibal <dlhannibal@ucdavis.edu> wrote:
Although I did not receive response to this form the statlist, I have had
several people contact me who have the same question and want to know if I
received a response or found out anything more.  I am writing this email to
post for the record my response to the latest inquiry so that it might be
helpful or expanded on by anyone familiar with mixed ordinal regression.
Please see below:

"No one responded, but after doing quite a bit more reading it is clear that
meologit and meglm use an algorithm that does assume proportional odds.
Unfortunately STATA does not provide a way to test the proportional model
with a mixed model.  Non-proportional odds separate out the coefficients of
the predictors by each category of the ordinal outcome and you can force
STATA to do this by recoding your outcome variable into multiple separate
binary outcomes for your cutpoints.

There is a program called SuperMix that will do it and allows you to do
proportional odds models, non-proportional odds models, or
partial-proportional odds models. It is not very user friendly and you will
have to read an extensive amount of documentation to use it. I ended up
using that to assess the variables in my models for the proportional-odds
assumption and then ran the final model in STATA to produce predicted
probabilities and make graphs. Fortunately, all of the variables in my final
model met the assumption. You can get estimates out of SuperMix, but it is
pretty time consuming.

There is a free version of SuperMix you can download from their website at:
http://www.ssicentral.com/supermix/downloads.html. [additional note: the
non-free version is super expensive at $425 but so super useful I would buy
it at a more affordable price of say $100-$200; unless, of course, STATA 14
makes it possible to do everything SuperMix can do].

Although the website states it limits the size of the model you can analyze,
I did not have any problems and had a data set much larger than what is
supposed to be allowed in the student version. Although they state they do
not provide support for the student version, they were willing to respond to
the few questions I asked, but that was after I read much of the
documentation and had figured most of it out. So, I think they are willing
to help those who've already done as much as possible on their own and are
just a little bit stuck."

I hope that helps,
Darcy


On 12/5/2013 1:15 PM, Darcy Hannibal wrote:
Hello,

I have a question about the assumptions for the models using either the
meologit or meglm (ordinal family) commands. None of the documentation I
have found for these new commands available in version 13 mention anything
about testing for whether the data meet the proportional odds assumption.
Since there are different varieties of ordinal models and not all of them
are constrained by proportional odds I am wondering if the models used in
these two commands do not assume proportional odds. The brant command will
test proportional odds, but it only works after the ologit command.

Can anyone tell me if the proportional odds assumption applies to meologit
and meglm (ordinal family). If so, is there a simple way to test for this in
STATA or does it have to be done by hand?

Thank you in advance for any advice you can give.
--Darcy

--
Darcy L. Hannibal, PhD
Staff Research Associate III Supervisor

McCowan Animal Behavior Laboratory for Welfare and Conservation
Department of Population Health and Reproduction

Behavior Management
Brain, Mind, and Behavior Unit
California National Primate Research Center

University of California at Davis

Office: 3029-B CNPRC
Phone: 530-752-1586

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/

--
Darcy L. Hannibal, PhD
Staff Research Associate III Supervisor

McCowan Animal Behavior Laboratory for Welfare and Conservation
Department of Population Health and Reproduction

Behavior Management
Brain, Mind, and Behavior Unit
California National Primate Research Center

University of California at Davis

Office: 3029-B CNPRC
Phone: 530-752-1586

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index