[R-sig-ME] Dropping fixed factor from a converged glmer model leads to singular fit

[Ben Bolker]

  I couldn't say exactly what happened, but I don't find this shocking.

  I might find time to a look at code & data if you send it.

  If you wanted to look into this in detail you could profile over the
interaction term, i.e. fit the model with the interaction parameter
fixed to values ranging from zero to the estimated value in the full
model, in order to understand how the random effects parameters change
as you go from the full to the nested model.

  I would go ahead and compare the full and nested model even though the
nested model is singular.

  cheers
   Ben Bolker



On 2020-03-03 6:23 a.m., Noam Tal-Perry wrote:
> Hey all,
> 
> I was fitting the following model to test whether condition (categorical, 4
> levels) and reaction time (continuous, scaled) predict trial's result
> (binomial, correct/incorrect).
> 
> model.acc.main_re = glmer(Result ~ 1 + Condition*scale(RT) + (1 +
> Condition+scale(RT) | Subject_ID), data=behavioral_df.response_only, family
> = binomial, control = glmerControl(optimizer ="bobyqa", optCtrl =
> list(maxfun = 100000)))
> 
> This model successfully converged.
> I was interested in doing an NHST for the interaction term, by comparing
> this model with the same model sans the interaction (i.e. compute type-II
> SS). So I dropped the interaction and fitted the following model:
> 
> model.acc.main_re.no_int = glmer(Result ~ 1 + Condition+scale(RT) + (1 +
> Condition+scale(RT) | Subject_ID), data=behavioral_df.response_only, family
> = binomial, control = glmerControl(optimizer ="bobyqa", optCtrl =
> list(maxfun = 100000)))
> 
> Surprisingly, I got boundary (singular) fit error. I tried playing around
> with the maxfun parameter but it didn't lead anywhere.
> 
> So I was wondering,
> (1) How could this be? I was under the impression that with fewer factors,
> it should be easier to fit the model, such that a sub-model of a converged
> model should ought to converge as well.
> (2) What is the recommended approach in such case? Should I drop re from
> the main model (with the interaction) even though it converges, to reach a
> point where the model without the interaction converges too, before
> comparing with anova? Would it make more senes to drop re just from the
> main effects model and preserve the re for the fuller model?
> 
> Would gladly share code and output if that may help.
> 
> Cheers,
> NTP
> 
> Noam Tal-Perry
> PhD student
> Shlomit Yuval-Greenberg's Eye-Movement Research Lab
> School of Psychological Sciences, Tel-Aviv University
> 
> 	[[alternative HTML version deleted]]
> 
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>