[R-sig-ME] GLMMTMB false convergence

[Ben Bolker]

   I don't know if I've included this in the troubleshooting vignette
yet, but: the "false convergence" message (ultimately from nlminb())
is pretty obscure.
https://stackoverflow.com/questions/40039114/r-nlminb-what-does-false-convergence-actually-mean

   I can think of the following strategies:

 * restart the fit from the optimized parameters (the false
convergence may not occur because nlminb can reset its iterative
procedure) and see if you get the same fit without the warning

* scale and center predictors (always worth a try)

* as of a *very* recent version of the glmmTMB master branch, you can
specify a different optimizer (see ?glmmTMBControl) and see if you get
the same answer.


On Wed, May 29, 2019 at 3:02 PM Michael Whitby <michael.whitby using gmail.com> wrote:
>
> Hello,
>
> I have built a zero-inflated GLMM (truncated_nbinom2 distribution) with an
> ar1 covariance structure using GLMMTMB (latest development version as of
> yesterday) and MRAN 3.5.1 on a server running windows server 2016.
>
> The model produces a warning about false convergence and points to the
> troubleshooting vignette.
> However, that page does not specifically mention false convergence. Is the
> advice for handling this warning the same as another section in the
> vignette?
>
> There are no NA's in the parameter estimate, the AIC values are reported,
> and sdr$pdhess is TRUE.
>
> Is this model useable? If not, any advice on determining where the problem
> exists? (I assume it is overparameterized despite a large dataset)
>
> I apologize for not including a reproducible example, I haven't been able
> to reproduce the false convergence warning with other datasets
>
>
> Michael Whitby
> michael.whitby using gmail.com
>
>         [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models