Warning message in mixed model lme4

Question

What is the meaning of the following warning message when fitting a 'glmer' model using package lme4?

Warning messages:
1: glm.fit: fitted probabilities numerically 0 or 1 occurred 
2: In mer_finalize(ans) : false convergence (8)

The model I'm trying to fit is like this:

glmer(dummy ~ constituency.coa + I(governat.part) + I(district2) + gdp.cap + lula.power + ifdm + bf.cap + year + (1 | munname), data=pool, family=binomial(link = "logit"), REML=T, verbose=T)

Thanks

Answer 1

Warning 1: the fitted values became 0 or 1 for one or more observations, but this should not be possible under a logistic regression. The causes are many; one is discussed on the help page for ?glm but that is little more than a pointer to some other documentation. It is only a warning so may not be a problem, but it is a warning that something is not quite right with the fit.

Warning 2: I don't know the exact meaning, but the code is telling you that the optimisation routine declared the fitting procedure to have converged to the estimated values but that this claim is false and that the fitting did not really converge.

One thing to look at is whether there is a separability problem, where one predictor or a linear combination of predictors can perfectly split 0 and 1 events.

I suggest you follow this up on the R-SIG-Mixed mailing list where there are real experts who can help further. You may need to provide further details of the fitting process (turn on verbose mode) or even the data so the problem can be diagnosed.

Answer 2

For Warning 2, you can up the number of iterations, the default is 300, to see if it converges when you add more iterations. try:

glmer(dummy ~ constituency.coa + I(governat.part) + I(district2) + gdp.cap + lula.power + ifdm + bf.cap + year + (1 | munname), data=pool, family=binomial(link = "logit"), REML=T, verbose=T, control = list(maxIter = 600))

this changes it from 300 iterations to 600, but you could try more if that doesn't work.