Skip to content
Prev 5179 / 20628 Next

confidence intervals in lmer

Rachel,

I think you have understood quite well.  The second expression
_assumes_ no correlation between the random effects.  The same thing
happens with pvals.fnc() in the languageR package, I think because
mcmcsamp (in the lme4 package) cannot deal with correlated random
effects.

There has been some discussion of this recently in the R-lang mailing
list (which I just discovered):
https://mailman.ucsd.edu/mailman/listinfo/ling-r-lang-l

I thought it might be possible to fiddle with centering to make the
correlation go away.  (I'm still not convinced that this isn't
possible in the case of a random slope and a random intercept.)

I would imagine that, if the correlations are "small" and if the
coefficients of the two forms of your model are "close", then the
results you get from the second one are a "good" approximation to what
you would get from the first if it worked.

Jon
On 01/22/11 19:57, Rachel Cohen wrote:
[...]