[R-sig-ME] Bleeding edge lme4 (or lme4a) plus DF estimation

[Ben Bolker]

Joehanes, Roby (NIH/NHLBI) [F] <roby.joehanes at ...> writes:

> I learned about the impending release of the new version of lme4 
> (or lme4a) from Dr. Bates' post here:
> https://stat.ethz.ch/pipermail/r-sig-mixed-models/2012q1/007499.html
> 
> Firstly, just to make sure, is this based on lme4a? 
> I have the source tarball of lme4a 0.999375-65 and found
> that a new optimizer, BOBYQA, is now in use, instead of the nlminb. 
> The announcement above also mentioned
> the same change. 
> So, I suspect that the older lme4 is supplanted with lme4a. Is this true?

  I'm not quite sure what you mean.  There are three versions of lme4:

* lme4 "classic" (version 0.999375-42 is the latest) lme4, built on nlminb
  (as was Bates's previous mixed-models package, nlme)

*** lme4 will be preserved on CRAN and (probably) renamed "lme4.0" after
the new version (see below) is released as lme4, for users who
need backward compatibility (we thought this was a reasonable
name: there are still a few issues with package names containing
dots, so we may need to change this ...). 

* lme4a (0.9996875-1) uses bobyqa

* lmeEigen (to be renamed lme4 when released, sometime soon ...)  uses
a mixture of bobyqa and a new Nelder-Mead implementation that allows box 
constraints, adapted from the nloptr package, which in turn wraps the 
NLopt open-source optimization codes 
(http://ab-initio.mit.edu/wiki/index.php/NLopt_Introduction).

> Secondly, I would like to get a source tarball of the latest 
> bleeding edge release to play with (or even
> read-only SVN access). I found from some sniff tests the lmer outputs of 
> lme4a to be closely matched with
> those of SAS than those of the lme4 (with nlminb optimizer), except for 
> the lack of p-values. I would love to
> play with the new version and even give you comparisons with the old 
> version. The problem I am facing with
> lme4a 0.999375-65 is that it sometimes crashes (core dumps).

  That's easy: just go here for instructions on SVN access:

http://r-forge.r-project.org/scm/?group_id=60

(the package is *so* bleeding edge that if you get it right now,
you should roll back to SVN release 1618; releases 1619+ are
currently broken ...)

> Thirdly, I also would love to see the Satterthwaite or
> Kenward-Rogers DF estimation. I would like to try to add these
> features into lme4, if you will. I don't know much about the
> formulas to compute the DFs from quantities output by lmer /
> glmer. Any pointers?

 Doug Bates is on record as saying that adapting the Kenward-Roger
formulation to work with his code would be difficult
<https://stat.ethz.ch/pipermail/r-help/2008-February/155372.html>, 
but Ulrich Halekoh and Søren Højsgaard have an implementation in 
the pbkrtest package
<http://cran.r-project.org/web/packages/pbkrtest/index.html>
which you could look at.  Perhaps that would give you a hint about
a Satterthwaite implementation as well ...

[PS there is no "s" in "Kenward-Roger" -- this is a very common mistake,
Kenward-Roger gets 1.03 million google hits while Kenward-Rogers
gets 438K ...]


> 
> Thank you.
> 
> Sincerely,
> Roby
>