[R-lang] mixed effect multinomial regression in R?

[rlevy]

Marco Baroni wrote:
> Dear R-Langers,
>
> Does anybody know if R supports multinomial mixed effect regression?
> (By "multinomial regression" I mean that the dependent variable is
> categorical (not ordinal) and it has more than 2 levels, and the
> independent variables are a hodge-podge of binary tests, numerical
> variables, etc...)
>
> [No explicit linguistic link in the question, but this is to analyze
> some lexico-semantic data, and I've seen that this list features many
> experts on mixed model categorical data analysis.]

Hi Marco,

This is a good question, and the answer to the best of my understanding
is "not directly, right now."  Here's a link to a query (unanswered) of
mine on this topic on R-sig-ME, which was a follow-up to a question by
Austin Frank:

https://stat.ethz.ch/pipermail/r-sig-mixed-models/2007q3/000261.html

What *is* possible is to fit a surrogate Poisson mixed-effects
regression and to re-interpret the fitted model as a multinomial model.
  Let me know if it's not obvious what that might mean.  (You might also
take a look at Agresti 2002, Section 8.5.)  As I recall, however, there
are some issues I never was able to work out on how to do certain types
of model comparisons within this approach.

Also, it is supposedly possible to fit Bayesian multinomial
mixed-effects regression using WinBUGS (but this is only for Windows :().

Hope this helps.  There is clearly a need for this type of model in the
analysis of linguistic data...

Roger

--

Roger Levy                      Email: rlevy at ucsd.edu
Assistant Professor             Phone: 858-534-7219
Department of Linguistics       Fax:   858-534-4789
UC San Diego                    Web:   http://ling.ucsd.edu/~rlevy