Running multinomial models with random effects

Alternatively, you could fit a "tree-based" mixed logit model
based on continuation ratio logits. See a recent JSS paper by De
Boeck and Partchev (2012, http://www.jstatsoft.org/v48/c01/).

The idea is to convert the three-level response into a binary one
using a decision tree. In one possible tree, the first node is
indifferent w.r.t. size (response both) vs. picky (response small
or big). The second node is small given picky vs. big given
picky.

You need to extend your data by a two-level node variable. The
new response is then binary. The model can be fit using glmer().

Best, Florian

---
Florian Wickelmaier
Department of Psychology
University of Tuebingen
Schleichstr. 4, 72076 Tuebingen, Germany

Hi Florian,
thanks a lot for your help. I have been reading De Boeck and  
Partchev paper and I have some questions.
First of all I am having some problems when applying dendrify function.
My dataframe has the following structure:
  RISK       Area CHOICE
1 MONTE      Sopie      2
2 MONTE      Sopie      2
3 MONTE      Sopie      2
4 MONTE Anchurones      3
5 MONTE Anchurones      1
6 MONTE Anchurones      2

being 1 (both acorns choice, 2 chosing the big one and 3 chosing the  
small one)
Following your instructions I mapped my tree as:
mapping <- cbind(c(0, 1, 1), c(NA, 0, 1))
Then I remove the first two columns in order to implement dendrify function
dendrify(jay[,-(1:2)],mapping)
The following error arise:
Error: is.matrix(mat) is not TRUE

I have not been able to find the problem. I have  followed tutorial  
package and paper instructions. mat=(jay[,-(1:2)], has only one  
column since I only have one item and each row corresponds to one  
choice event.
Any suggestions?




Quoting :

Alternatively, you could fit a "tree-based" mixed logit model
based on continuation ratio logits. See a recent JSS paper by De
Boeck and Partchev (2012, http://www.jstatsoft.org/v48/c01/).

The idea is to convert the three-level response into a binary one
using a decision tree. In one possible tree, the first node is
indifferent w.r.t. size (response both) vs. picky (response small
or big). The second node is small given picky vs. big given
picky.

You need to extend your data by a two-level node variable. The
new response is then binary. The model can be fit using glmer().

Best, Florian

---
Florian Wickelmaier
Department of Psychology
University of Tuebingen
Schleichstr. 4, 72076 Tuebingen, Germany