[R-sig-ME] [Lme4-authors] lmer/glmer false positive in convergence problems

Ben Bolker bbolker at gmail.com
Fri May 30 00:50:52 CEST 2014


 [cc'ing to r-sig-mixed-models]

  To be frank, we don't really know.  The tolerance for this warning
will be set to 0.002 in the next release, but that just seems to us (OK,
me) to be for a fit that's supposed to be a decent numerical fit to the
minimum of a smooth, deterministic objective function.

  I would normally try the fit with
control=glmerControl(optimizer="bobyqa") and see if that helps ...

On 14-05-29 05:38 PM, Miren Arantzeta wrote:
> Sorry for disturbing you again. I keep getting the warning message in
> each model I fit, I would like to know under what value should give the
> max(abs(relgrad)) to consider the warning a false positive. The highest
> value I got was 0.08.
> 
> Thank you again.
> 
> /Miren
> /
> 
> 
>> From: bbolker at gmail.com
>> Date: Thu, 29 May 2014 16:34:57 -0400
>> To: arantzetam at hotmail.com
>> Subject: Re: [Lme4-authors] lmer/glmer false positive in convergence
> problems
>>
>>
>> Yes, that indicates a false positive (i.e. convergence is actually OK).
>>
>> On 14-05-29 01:36 PM, Miren Arantzeta wrote:
>> >
>> >
>> > Dear all,
>> >
>> > I write you in relation to the post of Ben Bolker about the lmer/glmer
>> > convergence warnings. I am modeling binomial data with glmer and I get
>> > this warning message (I did not get it previously while using the
>> > previous version of r). As mentioned by Ben, I applied the next to test
>> > if it is a false positive.
>> >
>> >> relgrad <- with(mixedModelACC4 @
> optinfo$derivs,solve(Hessian,gradient))
>> >> max(abs(relgrad))
>> > [1] 0.0001807051
>> >
>> >
>> > It is not clear to me if this value indicates that the convergence
>> > warning is a false positive, and therefore I can continue using glmer
>> > with this dataset. I would appreciate your feedback.
>> >
>> > Thank you in advance.
>> >
>> > /Miren/
>> >
>> >
>> > _______________________________________________
>> > Lme4-authors mailing list
>> > Lme4-authors at lists.r-forge.r-project.org
>> >
> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/lme4-authors
>> >
>>



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