[R-sig-ME] lmer: no warning messages in dev version

[Ben Bolker]

Gustaf Granath <gustaf.granath at ...> writes:

> 
> Hi,
> When I use the stable CRAN version for parametric bootstrap I often get 
> warnings messages (e.g. false convergence and "singular convergence") 
> for some runs. Happily I installed the gitHub version that can catch 
> these warnings. However, to my surprise I dont get any warnings with 
> this version. My guess is that most warnings before were associated with 
> variances=0, and that the new version doesnt give a warning about that. 
> Is this a likely explanation?
> 
> Gustaf

   Reasonable, but I think incorrect.
   The previous version did *not* warn on positive semidefinite
(i.e. variance=0 or correlation=+/-1) variance-covariance estimates;
nor does this one.
   The proximal explanation is that we switched from the nlminb optimizer
to two derivative-free optimizers in this case (BOBYQA from the minqa
package and an implementation of Nelder-Mead from the nlopt library by
Steven Johnson).  While optimization is a bit of a black art, we're
hoping that these optimizers will be 'better' (more robust) than nlminb.
The new version also offers a lot more flexibility for users who
need to tinker with the optimization method.

  Ben Bolker