CODEWITHSUNDEEP
rmixed-modelsconvergence
CatM
CatM

Reputation: 284

glmer - inconsistent convergence issues

I am starting to build the model that looks like this:

model_sim <- glmer(Accuracy ~ x*y*z_scaled + (1 |Participant), 
          binomial(link = "logit"), data = Data)

And it failed to converge so I ran the allFit function:

(model_sim <- allFit(model_sim, maxfun = 1e+05))

to see if there were actual reasonable reasons to be concerned, it converged with 5 out 6 optimizers, all with the same value, so I selected the one I always select - bobyqa with 1e+05 iterations but it failed to converge again. Could anyone explain why this happened? Shouldn't it converge based on the all fit results? What would you do in this situation?

$fixef


 $llik
      bobyqa                   Nelder_Mead                    nlminbwrap 
      -24286.15               -24286.15                     -24286.15 
 nloptwrap.NLOPT_LN_NELDERMEAD     nloptwrap.NLOPT_LN_BOBYQA 
                -24286.15                     -24286.15 

  $theta
                                     Participant.(Intercept)
  bobyqa                                      0.6872716
  Nelder_Mead                                 0.6872370
  nlminbwrap                                  0.6872456
  nloptwrap.NLOPT_LN_NELDERMEAD               0.6872732
  nloptwrap.NLOPT_LN_BOBYQA                   0.6872732

Upvotes: 0

Views: 395

Answers (1)

Ben Bolker
Ben Bolker

Reputation: 226761

You might be misunderstanding the meaning/purpose of allFit() (and maybe the convergence warnings themselves). allFit() doesn't actually change anything/make anything converge "better" - it just tests whether a range of different optimization algorithms converge on the same solution (or sufficiently similar solutions).

Your output shows that all of the available optimizers are converging to the same coefficients, at least to several decimal places (hopefully a difference between a Probability coefficient of 0.449 and 0.451 will not make a substantive difference to your conclusions), and that the difference in log-likelihoods is less than 0.005 log-likelihood units (which is small). So my conclusion in this case is that the fit is OK, and it doesn't actually matter which optimizer you use.

Upvotes: 1

Related Questions