[R-sig-ME] Wrong convergence warnings with glmer in lme4 version 1.1-6??

[Ben Bolker]

UG <uriel.gelin at ...> writes:

> 
> Tom Davis <tomd792 <at> ...> writes:
> 
> > 
> > Dear lme4 experts,
> > 
> > Yesterday, I ran the code for two published papers 
> (de Boeck et al.,2011;
> > de Boeck and Partchev, 2012) on psychometric modeling with glmer in lme4
> > version 1.1-6 and the vast majority of the models I 
> ran produce convergence
> > warnings (even the simple ones).

 [major snippage]

> > I had no warnings using version 1.0-5 and 
> version 1.0-6 so this seems to be
> > a recent problem of lme4?
> > 
> > Is it best to ignore all these convergence warnings for now? Should I
> > switch back to an older version of lme4 to avoid this problem? Should I
> > generally avoid using large datasets with lme4?
> > 
> > Many thanks in advance,
> > Tom

  Re-iterating what I may have said earlier:

 * these warnings are a result of changes in the WARNING behaviour, not a
change in the fitting algorithm.  The only major recent change in the
optimizer has been to move in 1.1-6 from Nelder-Mead to bobyqa for
lmer fits <http://cran.r-project.org/web/packages/lme4/news.html>; this
should isn't relevant to glmer fits and we believe should _improve_
results in almost all cases.
 * You should probably ignore the convergence warnings; we now believe
that they are (almost entirely) the result of checking the gradients,
rather than the scaled gradients (i.e. the results of solve(Hessian,grad).
 * Version 1.1-7, now on Github (available via devtools::install_github)
and with a macos binary on http://lme4.r-forge.r-project.org/repos ,
tests the scaled gradients and should make the problem go away.
 * If you want to double-check, either 
   * test with lme4.0 (although we
have had reports  <http://stackoverflow.com/questions/23662589/
r-reverting-to-lme4-0-and-still-getting-inconsistent-results>
(broken URL!) of 'Downdated XtX' problems -- still chasing that down ...
or * use the code at
https://github.com/lme4/lme4/blob/master/misc/issues/allFit.R
to test with a wide range of optimizers.

  Ben Bolker