PROC GEE does not converge

[niam]

Hi I have a binary outcome variable and a binary predictor. Observations are grouped at DRIVERID level. 

When I run the model with using DIST=BIN and LINK=LOGIT, the model does not converge, however, if I just run the model without these two options, the model runs and gives me results. 

Can you please let me know why the model fails to converge when I specify that the DV is binary?

The following model runs:

proc gee data=ANALYSIS2 descend;
   class  DRIVERID ;
   model effct=TREATMENT ;
   repeated  subject=DRIVERID  / CORR=IND;
run;

This one does not run:

proc gee data=ANALYSIS2 descend;
   class  DRIVERID ;
   model effect=TREATMENT /dist=bin link=logit;
   repeated  subject=DRIVERID  / CORR=IND;
run;

Here is the error message I get:

The generalized Hessian matrix is not positive definite. Iteration will be terminated.

[niam · Accepted solution]

I found the answer myself. I explain it below and hope it can help someone else with the same problem:

In my case, the dependent variable is binary, so is the independent variable. However, the DV is equal to one only when the IV is one. In other words, I do not have any variation in my DV when the IV is equal to one. This will make it impossible to have a maximum likelihood estimator, and therefore the algorithm does not converge. The fix is to manually change the DV from 1 to 0 in just one observation in which the IV is 1.

[Tolawak]

Dear Paige,

Thanks for your reply. Mine is a bit different as my independent variable (physical activity) has 3 categories (1=active, 2=moderate, 3=nactive).

Would please help me with this?

All the best,

Tola