Longitudinal data analysis using proc GLIMMIX for binary outcome

[UcheOkoro]

Hello,

Please I need help with using proc GLIMMIX for binary outcome. My outcome variable is Adherence to safety guidelines (Adherence) which is binary. This was measured  weekly over a 15-week period. My independent variables include job title , race, ethnicity etc. which are all categorical. I am interested in seeing if adherence changed  over time. I am using the following codes for my analysis but  the model did not converge. Please find also the SAS log information.

 PROC GLIMMIX DATA=COVERED.Demographics;

CLASS  primary_job_cat week record_id;

MODEL complete_adherence = primary_job_cat week primary_job_cat*week/SOLUTION DIST=bin LINK=Logit DDFM=BW;

RANDOM week/SUB=record_id TYPE =AR(1) RESIDUAL;

lsmeans primary_job_cat week primary_job_cat*week/  diff e cl;

  run;

Thank you.

[jiltao]

You might try the PARMS statement to specify different starting values; Also try a different TYPE= option for the R-side random effect to see if that helps.

Did you get the table  - covariance parameter estimates at the last iteration? If so, what does it look like?

Jill

[UcheOkoro]

Below is what the covariance parameter estimate shows

[jiltao]

How many levels for week? 

What if you try a G-side random effect model --

random int / subject=record_id;?

You could also add method=quad in the PROC GLIMMIX statement for this G-side random effect model to see if that helps.

[UcheOkoro]

Week has 20 levels.

I tried the G-side effect model but got an error message but it did not converge.

I also added the method=quad but got an error message

 PROC GLIMMIX DATA=COVERED.Demographics;

CLASS  primary_job_cat week record_id ;

MODEL complete_adherence = primary_job_cat week primary_job_cat*week/SOLUTION DIST=bin LINK=Logit oddsratio;

RANDOM int/SUB=record_id TYPE =AR(1) RESIDUAL;

lsmeans primary_job_cat week primary_job_cat*week/  diff e cl;

  run;

    PROC GLIMMIX DATA=COVERED.Demographics method=quad(qpoints=19);

CLASS  primary_job_cat week record_id ;

MODEL complete_adherence = primary_job_cat week primary_job_cat*week/SOLUTION DIST=bin LINK=Logit oddsratio;

RANDOM int/SUB=record_id TYPE =AR(1) RESIDUAL;

lsmeans primary_job_cat week primary_job_cat*week/  diff e cl;

  run;

[jiltao · Accepted solution]

That is not a G-side random effect model. Use this instead --

random int / subject=record_id;

And the METHOD=QUAD option must be used with G-side model, not the R-side model as you did.

[UcheOkoro]

Thank you so much for your assistance.