02-07-2018 01:10 PM - edited 02-07-2018 01:11 PM
Hi, I am having a problem with a GLIMMIX model of zero-inflated data (count data) not converging in SAS 9.4. I have tried many things to get convergence and it appears that the convergence issue is related to a single interaction term between two of my variables. One variable has 3 levels and the other has 2 levels. Of the six levels of the two-way interaction, one has all zeros. This is the only thing that I can think of that is causing the model to not converge.
I have tried increasing MAXITER, using MAXFUNC, different TECH options, using a simple model to establish starting values for the PARMS statement. None of those things work when the interaction is in the model. The model runs and converges IF that one two-way interaction is not included. The problem is that interaction is biologically interesting so I would like to test it.
Does anyone have any suggestions for how to get around this issue? I do have random effects which limits my options somewhat. I have read that a preliminary GLIMMIX model can be used before going to NLMIXED for a ZIP model, but if I can't include the interaction in the preliminary GLIMMIX model then I can't obtain the starting value I need for that interaction in the NLMIXED model.
Here's the general code but not showing all possible options I've tried. I've also trimmed out all the other interaction terms
to keep it simple.
proc glimmix data = behav order=data initglm MAXOPT=100 IC=PQ; class round context morph size; model totcops = context morph size logitb motionless size*context/solution ddfm=kr dist=poisson; nloptions TECH=NRRIDG MAXITER=100; random morph*round / sub=femid type=TOEP RESIDUAL; covtest / wald; run;