Increasing iteration limit / lmer bug

Hello,

this same issue has been reported quite recently (29 March 2010), Ben Bolker 
found that there is indeed a bug and provided also a fix for it. Here is the 
link to the mail archives:

https://stat.ethz.ch/pipermail/r-sig-mixed-models/2010q1/003547.html


Cheers,

Luca


----- Original Message ----- 
From: "Adam D. I. Kramer" <adik at ilovebacon.org>
To: <r-sig-mixed-models at r-project.org>
Sent: Wednesday, March 31, 2010 8:14 PM
Subject: Re: [R-sig-ME] Increasing iteration limit / lmer bug

With apologies, after a lot of grepping around and reading pages on the
internet, yes, as you expected, my model is to blame.

That said, I think that it is quite appropriate for commands listed in the
?help page to actually work or do what they say--in the below case, it is
clear that maxIter is being ignored.

If it is the case that a "valid" or "good" or "non-stupid" model will 
easily
be fit with maxIter=300 and maxFN=900, then there is no reason to allow
users to think they have changed this value when they have not.

Cordially,
Adam

On Wed, 31 Mar 2010, Adam D. I. Kramer wrote:

Dear colleagues,

I am attempting to fit a model in this manner:

l3 <- glmer(mug ~ condition*time +
(time|cafe),family=binomial,data=data)

...the model fails to fit, however, noting:

In mer_finalize(ans) : iteration limit reached without convergence (9)

...so, of course, I read the man page for glmer and added
control=list(maxIter=3000). However, the program ran for the same amount 
of
time (about 10 minutes) and produced the same error (and, incidentally, 
the
same output).

So, I believe there to be a bug in glmer such that maxIter is not
functioning.

I then upped maxFN, too, (also to 3000) in case that was the problem, but
found no meaningful difference in the model produced or the time taken to
fit the model.

Could somebody recommend a workaround? I would like to fit this model.

(running R 2.10.1, lme4 v. 0.999375-32)

Also, in case it matters, 'time' here is a categorical variable
reperesenting several within-subjects timepoints--there are some other
(unanswered) posts to this list in which people appear to have fairly
complex within-subjects effects. Perhaps the issue is that the defaults
which are in place are insufficiently high for moderately complex
within-subjects models?

Cordially,
Adam D. I. Kramer