Skip to content
Prev 16527 / 20628 Next

Convergence problem example

Dear Ben

I understand. Help us think of ways we can help you.

I wonder why you don't take the easy route.  Have glmer do the
convergence test. If failed, then change the optimizer to bobyqa, and
IF that silences the convergence warning, report that result? That's
what I'd do, if I knew how :)


Dear everybody else

Lets see if we can help!

Questions/Ideas

1. How often does this come up?

Casual googling indicates the problem was widespread 2 years ago,
maybe not so many posts about it now.

Do you think so?

2.  What do you think about making a survey for lme4 users to find out
how often convergence warnings happen?  We could make a place to
attach files that have code and R objects.

If that survey existed, could we insert its address into the lme4
convergence warning along with instructions on what to do.  I'm
thinking something simple like "Help with diagnostics.  Run this:

saveRDS(the_data_frame, file = "your-last-name-20180425.rds")

and then paste in the trouble-causing function call into the following box...

Some cases have private data, I understand, but I doubt that all of them do.

3. Should we archive & categorize the reproducible examples?

What do you say if we (I) make a GitHub project for convergence
failure examples?  Self contained examples. Maybe the one I had today
is example 1. I suppose we need maybe  20 or 30 more, with a variety
of different warning messages.

I wish I had an example where the convergence diagnostic is correct
and the problem is not solved in a superficial way.  I'd be especially
enthusiastic if Stata or SAS don't notice and report non-converged
results on same model.

I found a post that hits the high points for me:
https://rstudio-pubs-static.s3.amazonaws.com/33653_57fc7b8e5d484c909b615d8633c01d51.html
(I wish the author would put his/her name in it!)

pj
On Wed, Apr 25, 2018 at 1:47 PM, Ben Bolker <bbolker at gmail.com> wrote: