[R-sig-ME] post hoc comparisons in mixed logit models

[fernando barbero]

Dear Angel, I think that the contrasts you are trying to fit can be done
with the glht function (multcomp package) there are some pdf files (look for
Hothorn, Bretz and Westfall"Simultaneous Inference in General Parametric
Models", but there are several more) in the web that have plenty of examples
that can help you
Best regards
Fernando

-----Mensaje original-----
De: r-sig-mixed-models-bounces at r-project.org
[mailto:r-sig-mixed-models-bounces at r-project.org] En nombre de Angel Tabullo
Enviado el: miércoles, 10 de agosto de 2011 05:33 p.m.
Para: r-sig-mixed-models at r-project.org
Asunto: [R-sig-ME] post hoc comparisons in mixed logit models

Hi everyone!

I'm analizing data from a study where subjects were trained by four
different methods and exposed to three different kinds of stimuli in a
forced-choice task. The subject's responses could be either correct or
incorrect. I'm considering "response" as a categorical depedent variable and
running a mixed logit regression with lmer. My independent variables are
"group of training", with four levels (between-subjects) and "type of
stimuli", with three levels (within subjects). I introduced group, type of
stimuli and their interaction as fixed factors, and subject ID as a random
factor. According to lmer output, both the interaction  and the fixed
effects are significant. But the output table compares all the levels within
a fixed factor to a reference category, as well as the level combinations of
the interaction (for instance, I can tell that stimulus types 1 and 2
significantly differ from stimulus type 0, but I can't tell if they are
different from each  other). 

Does anyone know if there's a way to run this kind of contrasts with lmer?


As I'm new to this kind of test, I imagine that there might be some mistake
in the codification of the variables. My depedent variable "response" is
coded as 0 for errors and 1 for correct answers in the database. Stimulus
type and training group are codified with categorical labels (I also tried
with numbers, but it made no difference).

Furthermore, I tried to run a generalized mixed model with a logit
transformations in SPSS19. It offered an option to compare estimated
marginal means with pairwise comparisons (applying a Bonferroni correction)
and provided F-tests of significance for the fixed effects and the
interactions. Is it correct to apply this procedures in a mixed logit model?
Because I couldn't find any reference of it outside the SPSS software.


I have read Florian Jaeger's 2008 paper and Baayen's book for language data
analysis with R, but I still can't figure out what am I doing wrong.

I apologise for my ignorance and thank you in advance for your kind
attention. Even though I'm quite new to R, I find the software quite
helpful, and I'd really like to learn this procedure and get it right.

Thanks again!
	[[alternative HTML version deleted]]