[R-sig-ME] Fwd: lme4 and sample size

[Douglas Bates]

After saying I would cc: the list on this reply, I didn't.


---------- Forwarded message ----------
From: Douglas Bates <bates at stat.wisc.edu>
Date: Fri, Jan 13, 2012 at 9:51 AM
Subject: Re: lme4 and sample size
To: "Bouwmeester, W." <W.Bouwmeester at umcutrecht.nl>


I have taken the liberty of copying the reply to the
R-SIG-Mixed-Models at R-Project.org mailing list so that it will be
available in a searchable archive.

On Fri, Jan 13, 2012 at 9:37 AM, Bouwmeester, W.
<W.Bouwmeester at umcutrecht.nl> wrote:
> Dear professor Bates,
>
> I will incvestigate the required sample size to develop a prediction model in hierarchical data, using a simulation study. One of the model evaluation criteria will be whether the model converged. R will give warning messages if no or singular convergence appeared.
> Is it possible to evaluate model convergence from the cvg column in attr(mix.model,"dims")? Has this "cvg" anything to do with convergence?
>
> I saw this "cvg" output from attr(mix.model, "dims") in your publication on October 4, 2011, Linear mixed model implementation in lme4, page 6 and 20.

I would strongly recommend using lmer instead of lme to fit
heirarchical linear models in a simulation study.  The lmer function
in the lme4 package is much faster and more reliable than the lme
function from the nlme package.

The current version of lme4 on CRAN can sometimes encounter a warning
about "false convergence".  The version named lme4Eigen on the R-forge
site is, in our preliminary tests, more reliable and usually faster
than the released version.  You do need to be able to build an R
package from source to be able to use the lme4Eigen at present.