[R-sig-ME] Convergence warnings in lme4

Ben Bolker bbolker at gmail.com
Tue Oct 21 15:39:49 CEST 2014

On 14-10-21 09:21 AM, adam.hayward at ed.ac.uk wrote:
> Dear list users,
> This is a general question, rather than a specific example of my data. I
> have recently upgraded my computer and to R ver 3.1.1 from 3.0.2. Models
> (glmer) which previously ran perfectly are now producing warning
> messages regarding failure to converge, for example:
> Warning messages:
> 1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
>   Model failed to converge with max|grad| = 0.0238312 (tol = 0.001,
> component 13)
> 2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
>   Model failed to converge: degenerate  Hessian with 1 negative eigenvalues
> and this is true with a range of optimizers. Running identical models on
> my old and new versions of R produces identical output (estimates, AIC
> etc) but the new version of lme4 produces the warnings.
> My questions are thus (1) has lme4 recently been updated with
> significant changes in what constitutes acceptable convergence, and (2)
> can my models now be trusted?
> Many thanks for any comments or help,
> Adam

  Presumably you're using version 1.1-7 (the latest version on CRAN); it
would be helpful to know that.  You can see changes to lme4 at
http://cran.r-project.org/web/packages/lme4/news.html . The significant
changes you're referring to started in version 1.1-3.

  We are still working to figure out how to get the right level of
sensitivity and specificity on these tests; we haven't come up with
anything definitive yet.  I've got an example of various attempts at
troubleshooting (restarting the fit from the same point, trying
different optimizers, etc.) posted at
http://rpubs.com/bbolker/lme4trouble1 .

The bottom line is that your models are no more nor less trustworthy
than they used to be.  If you have tried them with a variety of
optimizers, and if the results make sense, it's *probably* the case that
they're OK, and just a little unstable.

  Ben Bolker

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