glmer and influence.me - complaining about nAGQ==0

[Replying to list]

   Good meta-question.

   You can check the archives at 
https://stat.ethz.ch/pipermail/r-sig-mixed-models/ to make sure you 
haven't missed a reply.

   Unfortunately, since this is an all-volunteer forum, there's no 
particular guarantee that people will remain interested in a thread/get 
around to answering your questions. IMO it's reasonable etiquette to 
"bump"/remind readers about a question that has been languishing for 
some time (once you've allowed a reasonable time for an answer, e.g. a 
few days); it may be helpful to include a link, in this case 
https://stat.ethz.ch/pipermail/r-sig-mixed-models/2021q2/029477.html , 
to help people re-locate the question.

   It would be OK to try your luck in another forum such as 
CrossValidated; it would be polite and useful if/when you post there to 
provide a link to this thread so people can see what has already been 
discussed (although you should also provide a good summary there, so 
readers don't *have* to come trawl through the archives).

   A few more points since I'm responding:

* From your _original_ problem; I've updated the influence.merMod() 
method in the *development* version of lme4 so that it doesn't choke 
when nAGQ=0 (the starting value can be set manually, but it also checks 
and does the right thing by default when nAGQ is 0). If you're able to 
installed packages with compiled code from source 
(remotes::install_github("lme4/lme4")), you should be able to run 
influence() on your original model.

* The gold standard for whether convergence warnings are really a 
problem or not is allFit().  If you run allFit(), and the results from a 
range of optimizers **are sufficiently similar to each other for your 
scientific purposes**, then you can feel free to disregard convergence 
warnings (at that point it doesn't really matter how nasty the warnings 
sound).

* You _might_ try simplifying the model a little bit, e.g. removing the 
Session : Probability interaction in the random effects term (the 
standard deviation for that component is 5 times smaller than the 
intercept variation, and 29 times smaller than the residual variation term).


  One tiny final point: please don't use the salutation "Dear Professor 
Bolker" when responding to the list - even though I might be answering 
your question at the moment, the query is still to the whole list.

   cheers
     Ben Bolker

Hello,

I am sorry for bothering you but I have asked a follow-up question and I 
haven't obtained a response, I wonder if it will ever be responded to as 
I don't know if there is any way of keeping track of which questions get 
or not answered. If not, I may need to try my luck on crossvalidated?as 
it is quite a timely problem that I am experiencing with the models. 
After removing the "logRT" from the glmer model I got a lot more 
convergence issues, so I am left wondering if having the logRT would be 
problematic. If so, would there be other options for me on what to do to 
be able to model this data.

Best wishes,

Catia

-- 
C?tia Margarida Ferreira de Oliveira
Psychology PhD Student
Department of Psychology, Room B214
University of York, YO10 5DD