Skip to content
Prev 132866 / 398502 Next

logistic mixed effects models with lmer

On Dec 28, 2007 9:35 AM, Sharon Goldwater <sgwater at stanford.edu> wrote:
It can happen that the significance levels reported by the different
tests will be different.  The test in the summary is based on a local
approximation to the conditional mean and assumes that the variances
of the random effects will not change substantially when fitting the
model with and without that term.  Did they?

One thing I noticed is your output (and thank you for including that)
is that it reports that corpus has only two levels.  Is that correct?
If so, I would not advise using a random effect for corpus.  It is
very difficult to estimate a variance from only two levels of a
factor.  I suggest using a fixed effect for corpus instead.
The difference could be due to approximations or due to poor
convergence.  For some models the Laplace approximation can be
generalized to a higher-order approximation, called adaptive
Gauss-Hermite quadrature (AGQ), but that isn't feasible when you have
crossed random effects (and, besides, I  still haven't written the
code for AGQ even in the case where you don't have crossed random
effects).

First I suggest that you check the value stored as the log-likelihood

dput(logLik(m17))
dput(logLik(m18))

to see how different they really are.

You could, if you are feeling brave, try the development version of
the lme4 package from http://r-forge.r-project.org.  To do that,
however, you will need to install a new version of R (R-2.6.1 is
preferred) and a new version of the Matrix package.  Then you can
install the new lme4 with

install.packages("lme4", repos = "http://r-forge.r-project.org")

If you prefer, you can contact me off-list and I can try to fit your
models for you using the development version.