The SVN version of lme4, available at https://svn.r-project.org/R-packages/trunk/lme4 has a new version of lmer2 that should require less memory on large model fits. It creates the entire lmer2 object at once rather than creating the mer2 object then later creating the lmer2 object from that. This should avoid a copy operation involving potentially large objects. If you have a very large model fit to create please try out the new version. It does require R-2.5.0 so you will need to build that first. If that is not feasible and you can allow me access to the data then I would be pleased to run the tests. The largest test for which I have real data is a model fit to the star data (about 25000 observations and approximately the same number of random effects) and that ends up taking too little memory (12 MB) to be a real test of the memory growth. The optional argument frame = FALSE (don't save the model frame) cuts down on the size of the returned object, sometimes dramatically.
SVN version of lme4 has a space-saving version of lmer2
1 message · Douglas Bates