old and new lme4

(This is a guest post by Klinton Bicknell.)

update 2014-06-24: Using lme4.0 probably isn’t necessary anymore. See post here.

The lme4 package‘s major 1.0 release was back in August. I and others have noticed that for typical psycholinguistic datasets, the new >=1.0 versions of lme4 often yield models with substantially poorer fits to the data than the old pre-1.0 versions (sometimes worse by many points of log likelihood), which suggests that the new lme4 isn’t as reliably converging to the actual maximum likelihood (or REML) solution. Since unconverged models yield misleading inferences about model parameters, it’s useful to be able to fit models using the old pre-1.0 lme4.

Happily, the lme4 developers have created a new package (named “lme4.0”), which is a bugfix-only version of the old pre-1.0 lme4. This allows for the installation of both old and new versions of lme4 side-by-side. As of this posting, lme4.0 is not yet on CRAN, but is installable by performing the following steps:

1. install the latest version of the new lme4

2. run this command in R:

install.packages("lme4.0", type="both",
                 repos=c("http://lme4.r-forge.r-project.org/repos",
                         getOption("repos")[["CRAN"]]))

3. If you’re on a Mac, you probably also need to install the gfortran package from here.

After that, you should just be able to load library(lme4.0) and have access to the pre-1.0 version of lme4. If you’re planning to load both versions at once, note that the latest-loaded package takes precedence on the search path, so if you do library(lme4) first and then library(lme4.0) next, the lmer command will use the lme4.0 package. Regardless of search path order, you can specify a particular package’s version of a function with :: syntax, e.g., lme4::lmer and lme4.0::lmer.

The new lme4 is under active development, so hopefully it will soon be able to consistently produce models that are as good as those produced by lme4.0. Until then, I recommend comparing crucial models across both versions, and putting more trust in the model with the higher log likelihood or lower deviance. (For comparing linear models, note that the “REML criterion” in new lme4 model summaries corresponds to “REMLdev” in lme4.0, and lower is better. For comparing binomial models, note that higher is better for log likelihoods, and in the usual case that these are negative, higher means closer to zero.)