# Convergence sensitive to choice of baseline category?

While running a simulation in R, I noticed that in fitting one particular model, R spits out a warning message, but if I simply change the baseline category in the response variable, it converges without complaining. The results produced do not differ appreciably, except that one value in the random effect correlation matrix is changed. I am curious why this occurs, and whether it suggests something interesting about the data set (the response variable is binary valued, and the explanatory variables are continuous-valued).

> library(lme4)
> fit1.MS.eE <- glmer(label ~ zSpec * zF1 + zF2 + (1 + zSpec + zF1 + zF2 | part), data = MSall.eE, family = binomial())
Warning message:
In mer_finalize(ans) : singular convergence (7)
> MSall.eE$label <- relevel(MSall.eE$label, "e")
> fit1.MS.eE <- glmer(label ~ zSpec * zF1 + zF2 + (1 + zSpec + zF1 + zF2 | part), data = MSall.eE, family = binomial())
>(no warning message)


I have uploaded the data set referred to in the code for interested persons.

-
This question seems to be only about how something about R works, and not any related statistical content (although that doesn't make it a bad question). As such, it is better asked (& probably answered) on Stack Overflow. – gung Jul 19 '12 at 13:37
@gung Thanks, can it be migrated? – jlovegren Jul 19 '12 at 14:08
This question looks like it has statistical content to me: after all, with the same data and similar procedures, it is plausible that any statistical package might encounter the same problem. It might help if you could post a sample dataset (or a means to simulate a dataset) so that interested readers could reproduce this behavior. – whuber Jul 19 '12 at 14:21
@whuber I've uploaded a data set which can be used to reproduce the result. – jlovegren Jul 19 '12 at 17:53
@whuber I just installed R 2.15.1 on my Windows Vista/32 computer, and got the same false convergence (8) warning as you report. Anyway, playing with other models based on the same data set it looks like the issue is that there are 3 groups for the purpose of fitting random slopes (similar data sets with 4 and 5 groups converged fine), so I suppose that my question is going to end up as an academic issue. – jlovegren Jul 19 '12 at 21:46