Convergence failure with glmer().
On 14-03-18 01:59 AM, Rolf Turner wrote:
This is kind of a follow-up to a question that I asked a few days ago with the subject line "Bernoulli glmm question". [The responses to my question more-or-less confirmed that I was basically on the right track with my model and [g]lmer() syntax. The suggestion (in retrospect "obvious") was made that I should include a "word" random effect in my model. It was also suggested that I should include some interaction terms. However no result "came back" after over an hour of waiting, so I killed the job and gave up on this.] However --- yesterday, after a few hiccups, I installed the latest version of lme4 (version 1.1-5) from CRAN. Now the model that I successfully fitted with my rather elderly version (cannot remember the version number; I *think* I could dig it up from backups if it is of any interest) of lme4 results in a warning:
Model failed to converge with max|grad| = 0.0922338 (tol = 0.001)
I tried switching the optimizer to "Nelder-Mead" and got an apparently "even worse" warning:
Model failed to converge with max|grad| = 0.150396 (tol = 0.001)
For the record, my call was:
fit <- glmer(y ~ sex+type+(1|student)+1|word),family=binomial,data=X,
control=glmerControl(optimizer="Nelder_Mead"))
This is just a warning, not an error, but it does seem to indicate that
the fit obtained is not (quite?) correct.
Is there anything that I can *do* about this? I've tried reading the
help files, but failed to discern anything useful in this regard. I
would appreciate suggestions.
If anyone is sufficiently interested, I can make the data set available
to them. It's a wee bit on the large side, but; 10314 records.
Thanks for any words of wisdom.
My guess is that this is a false positive. Some trouble-shooting thoughts: * are the parameters you're getting similar to those you got before? * if you install lme4.0 (from http://lme4.r-forge.r-project.org/repos), do you get similar parameters to those you got before? * if you extract the parameters from lme4.0, make a deviance function from lme4 using devFunOnly, and compare, do you get similar or better (lower) deviances for the new lme4-fit parameters? * is your fit singular, i.e. are any of the 'theta' parameters corresponding to diagonal elements (I think in your case you only have scalar random effects, so this would correspond to any of them) equal or close to zero? Do the components of the gradient that have large absolute value (which(abs(fit at optinfo#derivs$grad)>0.002)) correspond to the singular theta parameters? Ben Bolker