ordered multinomial mixed model - R-SIG-mixed-models

Mon, Jun 6, 2011 12:18 PM #

Dear list,

I'm still there, but a bit (!) overloaded at the moment, and still 
interested in building something like polmer(). But not at this time 
(overload and annoying personal impediments).

Could you please keep me posted on my mail address (emm dot charpentier 
at free dot fr) ? I have trouble currently regularly reading this list.

I'd also suggest cross-checking your results against a Bayesian model 
with "somewhat vague" priors ("overly vague" priors will give you 
numerical errors at initialization...), which is usually not too 
difficult to whip out in BUGS.

I am currently thinking of building a framework for such (and analog) 
tasks by using a set of R macros (see the relevant packages in CRAN), 
taking a (po)l(me)r-like statement as input and evaluating to a BUGS 
model that would be fed to BUGS ; alternatively, rewriting the (po)l(me)r 
model in a correct mcmcGLMM call (not as trivial as it seems (been there, 
done that, didn't even got the T-shirt)) might also give another 
milestone.

Sincerely yours,

					Emmanuel Charpentier

On Sun, 29 May 2011 13:25:01 +0200, Thomas Mang wrote?:

Hi,

On 28.05.2011 22:02, Rune Haubo wrote:

On 27 May 2011 17:03, Thomas
Mang<thomasmang.ng at googlemail.com>  wrote:

Hi,

Consider the need for a mixed-effects model with an ordered
multinomial response variable. To the best of my knowledge, lme4 would
not provide such a thing.

You are right; the lme4 package does not provide functions to fit
ordinal mixed models - at least not without modifications. However,
there are other CRAN (e.g., ordinal and MCMCglmm) and R-Forge (e.g.,
ordinal2) packages that does so.

Thanks for these pointers, very helpful indeed !

However, I was wondering if such a model can be built using a
generalized mixed model with Bernoulli response variables. Here is my
thinking, with the question if this is possible as outlined below:

This is a good idea - so good in fact that others have had it before.
Recently Ken Knoblauch posted the polmer function to this list based on
this idea and using glmer
(https://stat.ethz.ch/pipermail/r-sig-mixed-models/2011q1/005069.html).
I am not sure if polmer is exactly the translation of your description
into an R function, but it seems to me that it is fairly close.

I have checked briefly the function by Ken, and the code snippets by
Emmanuel Charpentier. On first sight *I think* these functions are
different, and by the way it is not imminent at all to me that they
would do the same as a cumulative logit model if the model contains
covariates (and quick check on an adaptation of Emmanuel's code,
https://stat.ethz.ch/pipermail/r-sig-mixed-models/2011q1/005060.html,
seems to confirm my point). Also, the likelihood is necessarily
different (including for an intercept-only model).

The version I had proposed is different from the standard cumulative
model too, as can be seen in the likelihood description I had provided.
But on the plus side, it would relax the constraint that the effect of
the covariates at the latent scale is identical across all factor
transitions (i.e. shift of the cutpoints has the same magnitude).

I will investigate the material more in depth and then get back to you,
maybe through private email if it's too technical. So more later.

best and thanks,
Thomas