Skip to content
Prev 1763 / 20628 Next

how to compute Confident intervals for BLUPS for lme function in nlme library

Well, intervals() is a pretty good nlme response, I think, even if it
was not exactly answering the question. The lme4 equivalent is
doplot() which produces a so-called "caterpillar plot" of conditional
means.

For example:
?ranef

fm1 <- lmer(Reaction ~ Days + (Days|Subject), sleepstudy)
rr1 <- ranef(fm1, postVar = TRUE)
dotplot(rr1,scales = list(x = list(relation = 'free')))[["Subject"]]

It is important to be clear about the distinction between random
effect estimates of variances and correlations, provided as part of
the model output and the conditional means for groups/units based on
("predicted with") them. At least, correlations computed from
conditional means can differ strongly from the correlation estimates.
The short story: You should only trust the model estimates. As Greg
Lee said, conditional means are to be handled with care; you can not
apply any of the usual inferential statistics to them. Nevertheless,
the caterpillar plots are highly informative and diagnostic about, for
example, whether you need a random effect to account for unit-by-unit
variability. They may also lead you to  discover distinct subgroups
suggestive of a fixed effect (for a future model). There is a
manuscript at the top of my publications page for download (and
constructive feedback) walking through an example.

Reinhold Kliegl
On Wed, Jan 14, 2009 at 2:03 AM, Greg Lee <sp8ial at gmail.com> wrote: