[R-sig-ME] Silent conversion problem in lmer - should have said '_convergence_ problem'

Douglas Bates bates at stat.wisc.edu
Wed Apr 20 19:53:13 CEST 2011

On Wed, Apr 20, 2011 at 9:43 AM, S Ellison <S.Ellison at lgc.co.uk> wrote:

>>>> Douglas Bates <bates at stat.wisc.edu> 19/04/2011 15:44 >>>
>>The problems are due to the optimizer used in the lme4 package, which
>>is the one from the nlminb function in R.  It has a habit of getting
>>stuck at the lower bound.  The lme4a package uses a different
>>optimizer from the minqa package and is more successful in this case.

> Very interesting; thanks for taking the time to run on the alternative
> package version.
> The alternative is indeed giving the same answer as
> manual calculation from the ANOVA table.

> But I note you said 'in this case'; does that mean it'll be a while
> before you settle on
> one or t'other?

The long term goal is to allow for different optimizers to be chosen
by the user.  Optimization software is an area where R is not as
strong as commercial software, such as Matlab, mainly because those
working in optimization have a tradition of either trying to
commercialize their efforts or of releasing it under vague or
undefined conditions (e.g. "free for non-commercial use").  We can't
incorporate such software in R.  Much of the optimization software
that is available under a suitable license is written in 1960's style
Fortran, meaning that creating interfaces for it is painful.  (Try
passing an R function through a Fortran calling sequence.)

This is changing but slowly.

All the lmer, glmer and nlmer do is to set up evaluation of the
deviance for the model/data combination and pass it to an optimizer.
Eventually it is the optimizer that determines the quality of the

The reason I said "in this case" is because I have seen cases where
nlminb does better than bobyqa from the minqa package (notice the
Fortran'ish names that, long ago, were restricted to 6 characters).
However, I have seen more cases where bobyqa does better than nlminb.

> Steve Ellison
> *******************************************************************
> This email and any attachments are confidential. Any u...{{dropped:9}}

More information about the R-sig-mixed-models mailing list