Convergence issues LME4 version 1.1-11

Question

I was working with a student, running some models with GLMER and we found that, using the same code, the models would converge for me and not for him. i.e. he would get an error message like the following: Warning message: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, : Model failed to converge with max|grad| = 0.00261244 (tol = 0.001, component 1)

After checking, we realised that because I was running an older version of R, when I updated my packages LME4 only went as high as 1.1-7 while he was running 1.1-11. So I updated R and LME4, and now we get the same results. I thought it might have been specific to the models we were running, but now I am working with another student, who is running LME4 1.1-7 and once again her models converge for her, but not for me. Is there something changed in 1.1-11 that would cause it to be more likely to give these warnings? If so, could the nature of the change give us a hint as to why we get the warnings with 1.1-11? And finally, which results do we believe, and is there anything we can do to deal with the convergence warnings?

Answer 1

The news page on CRAN for LME is probably the best place to look for changes.

I noticed this for example.

1.1-8 gradient scaling for convergence checks now uses the Cholesky factor of the Hessian; while it is more correct, this will lead to some additional (probably false-positive) convergence warnings

For avoiding convergence warnings, there's a reasonable amount you can do.

• Rescale continuous predictors
• Try different optimisers
• Look for sparse cells in combinations of predictors
• Consider whether your experiment design needs nested random effects

See also this cross-validated discussion of convergence failures.