[R-sig-ME] mixed-effect model with multinomial response

Rune Haubo rhbc at imm.dtu.dk
Thu May 5 11:57:11 CEST 2011

On 5 May 2011 01:08, Chris Gast <cmgast at gmail.com> wrote:
> I'm having trouble replying directly to the list, so hopefully this gets
> through as intended.
> Multinomial mixed models can be programmed in ADMB (www.admb-project.org),
> although there is no direct model interface (response ~ fixed effects +
> random effects); rather, it must be programmed manually in a C-like syntax.
> I haven't used the clmm() function in the ordinal package, but I do notice
> that it only allows a single random effect. This limitation does not exist
> in ADMB.

Indeed, ordinal::clmm only allows for a single scalar random effect.
However, this constraint is alleviated in clmm in package ordinal2
currently on R-Forge using an lmer-like syntax for the random effects.
Merging the ordinal and ordinal2 packages is work in progress.

Another restriction is that clmm is designed for *ordered* multinomial
observations, so it is fitting mixed equivalents of ordinal::clm and
MASS::polr rather than the mixed equivalents of nnet::multinom, the
latter being designed for *unordered* or nominal multinomial
observations. The first class of models is known as proportional odds
models, cumulative link models, ordinal regression models, ordered
probit/logit/... models, and several more, while the models fitted by
nnet::multinom are often denoted baseline logit/probit/... models. All
models are concerned with multinomial observations.

(author of ordinal[2])

More information about the R-sig-mixed-models mailing list