[R] ordered logistic regression with random effects. Howto?
Prof Brian Ripley
ripley at stats.ox.ac.uk
Tue May 8 07:45:18 CEST 2007
On the definitional question, some texts do indeed consider multi-category
logistic regression as a glm. But the original definition by Nelder does
not. I've never seen polr considered to be a glm (but it way well have
been done).
Adding random effects is a whole different ball game: you need to
integrate over the random effects to find a likelihood. That integration
is tricky, and I am not sure we yet have reliable software for it in the
binary ('dichotomous dependent variable') case: SAS's NLMIXED certainly is
not reliable. I've had students run real problems through a variety of
software, and get quite different results. (It is possible that the shape
of the likelihood is a problem but it is not the only one.)
MCMC approaches to that integration are an alternative not mentioned
below.
On Mon, 7 May 2007, Paul Johnson wrote:
> I'd like to estimate an ordinal logistic regression with a random
> effect for a grouping variable. I do not find a pre-packaged
> algorithm for this. I've found methods glmmML (package: glmmML) and
> lmer (package: lme4) both work fine with dichotomous dependent
> variables. I'd like a model similar to polr (package: MASS) or lrm
> (package: Design) that allows random effects.
>
> I was thinking there might be a trick that might allow me to use a
> program written for a dichotomous dependent variable with a mixed
> effect to estimate such a model. The proportional odds logistic
> regression is often written as a sequence of dichotomous comparisons.
> But it seems to me that, if it would work, then somebody would have
> proposed it already.
You need to combine all the binary comparisons to get the likelihood, and
the models have parameters in common.
> I've found some commentary about methods of fitting ordinal logistic
> regression with other procedures, however, and I would like to ask for
> your advice and experience with this. In this article,
>
> Ching-Fan Sheu, "Fitting mixed-effects models for repeated ordinal
> outcomes with the NLMIXED procedure" Behavior Research Methods,
> Instruments, & Computers, 2002, 34(2): 151-157.
>
> the other gives an approach that works in SAS's NLMIXED procedure. In
> this approach, one explicitly writes down the probability that each
> level will be achieved (using the linear predictor and constants for
> each level). I THINK I could find a way to translate this approach
> into a model that can be fitted with either nlme or lmer. Has someone
> done it?
>
> What other strategies for fitting mixed ordinal models are there in R?
>
> Finally, a definitional question. Is a multi-category logistic
> regression (either ordered or not) a member of the glm family? I had
> thought the answer is no, mainly because glm and other R functions for
> glms never mention multi-category qualitative dependent variables and
> also because the distribution does not seem to fall into the
> exponential family. However, some textbooks do include the
> multicategory models in the GLM treatment.
>
>
>
--
Brian D. Ripley, ripley at stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595
More information about the R-help
mailing list