08-14-2012 07:19 PM
I'm trying proc glimmix where the outcome is dichotomous and the model does not converge. Are there any tips for setting up the proc that will make it more likely to converge? The model does converge when the outcome is continuous.
08-15-2012 08:36 AM
First, take a look at the paper by Kathleen Kiernan, Jill Tao and Phil Gibbs from SGF 2012:
Some of what I say will follow their recommendations. First, when you say it doesn't converge, do you mean that the convergence history is "smooth" in the objective function, or does it jump around? If it is smooth, then it may be just a matter of increasing the number of iterations (NLOPTIONS statement), or controlling the outer loop convergence (pconv= option in the PROC GLIMMIX statement). However, if it is jumpy, then it is time to look at section II in the Kiernan paper. The first thing I think of is changing the technique to NRRIDG or NEWRAP. These work better than the default QUANEW for Bernoulli data (I assume that the response for each individual is 0/1).
I encourage to share pertinent parts of the output and program with us, so that we might give a more complete answer.
08-15-2012 04:37 PM
Thank you Steve for steering me to that paper. The model converged when I increased the iterations, and changed the technique to NEWRAP.