[R-sig-ME] mcmcsamp error message...

[Douglas Bates]

On Wed, Aug 12, 2009 at 12:27 PM, Petar Milin<pmilin at ff.uns.ac.rs> wrote:
> Hello!
> I am puzzled with an error message that mcmcsamp for models with random
> correlation parameters is not implemented yet. However, this would be the
> case only with the model:
> lmer(rt ~ FACTOR1 + COVARIATE1 + COVARIATE2 + COVARIATE3 +
> (1+FACTOR1|subject) + (1|item) + (0+COVARIATE3|item), data=dat)
> And same with:
> lmer(rt ~ FACTOR1 + COVARIATE1 + COVARIATE2 + COVARIATE3 +
> (1|subject) + (0+FACTOR1|subject) + (1|item) + (0+COVARIATE3|item),
> data=dat)

> If I run just:
> lmer(rt ~ FACTOR1 + COVARIATE1 + COVARIATE2 + COVARIATE3 +
> (1|subject) + (1|item) + (0+COVARIATE3|item), data=dat)
> mcmcsamp ends fine.

> I guess that the problem is in the fact that subjects were assigned
> (randomly) to only one level of the FACTOR1. Am I right?

I'm not sure what you mean by "the problem".  If FACTOR1 is a
non-trivial factor (i.e. it has more than one level) then the
random-effects terms (1 + FACTOR1|subject) and (0+FACTOR1|subject)
generate correlated random effects and currently mcmcsamp does not
handle models with correlated random effects.

If, as you say, each subject is assigned to only one level of FACTOR1
then neither of the terms above make sense.  You can't expect to
estimate an interaction of FACTOR1 and subject when FACTOR1:subject is
equivalent to subject.

> I think, previously, mcmcsamp handled this kind of nesting, but I might be
> wrong.
>
> What is strange to me that much more complex model (like the one where I got
> significant interaction between non-linear covariate and random-effect of
> participants -- subject) worked perfectly fine. Also, leaving only
> (0+COVARIATE3|item) is okay with mcmcsamp.
>
> Best,
> Petar Milin
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>