[R-sig-ME] Update: Easiest way to install lme4 on Mac OS X

[Reinhold Kliegl]

As to point one:
> * I've only observed it with slightly-pathological cases and not with real
> experimental data (there will of course be real experimental data where the
> bug is exposed)
I have observed it for experimental data (non pathological, I hope).
The effect was one decimal point in the t-value (1.9 vs. 2.0). (Not
that I do care too much about 0.1-differences in t-values, but I know
that some do.). Anyway, 32-bit R for MacOS X should be "out", in my
opinion. Could one change the default installation to 64-bit?

Reinhold Kliegl.

On Sat, Aug 7, 2010 at 12:26 AM, Daniel Myall <daniel.lists at zeno.co.nz> wrote:
> Hi,
>
> I've had several off-list emails inquiring about installing lme4 on Mac OS X
> and the progress in determining the cause of the non-deterministic
> behaviour. Hence, a quick update:
>
>
> Easiest way of installing lme4:
>
> The package is available from R-forge as package checks are not run for Mac
> OS X on R-forge (hence the non-deterministic behaviour under Mac isn't
> flagged). However, I would only run lme4 under R 64 bit on Mac (where no
> issues have been observed).
>
> Download and install gfortran from
> http://cran.r-project.org/bin/macosx/tools/ (available as a dmg which
> installs the required libraries for the builds from R-forge into
> /usr/local/)
>
> In R run: install.packages(c("Matrix","lme4"),
> repos="http://R-Forge.R-project.org")
>
>
> Summary of clues to what may be the cause of non-deterministic results of
> identical calls:
>
> * I've only observed it with slightly-pathological cases and not with real
> experimental data (there will of course be real experimental data where the
> bug is exposed).
> * Has only been observed in 32 bit R and not 64 bit R.
> * Occurs with both vecLib and R BLAS on 10.5.8 but only with vecLib BLAS on
> 10.6.4.
> * Occurs with all versions of lme4/Matrix in the past two years when run on
> R 2.11.1.
> * lme4a has a slightly different behaviour in that, instead of only two
> potential results, every run has a different result.
>
>
> When I have some more time I'll dig deeper into debugging each run to try to
> figure out exactly where the non-deterministic behaviour is creeping in. I
> want to get this resolved so that the package appears on CRAN again.
>
> Cheers,
> Daniel
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>