topic PROC GLMMIX no converge and p value in Statistical Procedures

PROC GLMMIX no converge and p value

oohsienoo — Mon, 16 Dec 2024 06:23:52 GMT

Hi SAS experts,

My research involves simultaneously measuring IBP and NIBP in 157 dogs.Some discrepancies occur during the measurements, we define a discrepancy as a difference >10 mmHg (match_S = 0), and no discrepancy as a difference <10 mmHg (match_S = 1). The factors I want to analyze include:

catagorical variables: gender, breed, asa, bcs, disease, surgerytype, duration, sirs, qsofa, na, alb, cuffsite, IBPICsize, IBPsite, perianalgesia, dexcri, local, anticholinergic, inotrope, other

continuous variables: weight, age, pcv

The current code is:

proc glimmix data=one order=data;
class gender breed asa bcs disease surgerytype duration sirs qsofa na alb cuffsite IBPICsize IBPsite perianalgesia dexcri local anticholinergic inotrope other;
model match_S(event='0') = gender breed asa bcs disease surgerytype duration sirs qsofa na alb cuffsite IBPICsize IBPsite perianalgesia dexcri local anticholinergic inotrope other weight age pcv / solution dist=binary link=logit;
random intercept / subject=No solution;
run;

But the results do not converge, and no p-values are generated. However, when No is not set as a random variable, p-values are generated. How can I improve my results?

Re: PROC GLMMIX no converge and p value

Ksharp — Mon, 16 Dec 2024 07:35:22 GMT

I think you have too many independent variables, especially too many category variable, that make design matrix is way too wide .
and the number of obs/records in dataset is too small for such big/wide design matrix. you could try LAPLACE method:
proc glimmix data=have method=laplace ;

For more troubleshooting this question ,check this paper(see attachment) at the end .

Re: PROC GLMMIX no converge and p value

SteveDenham — Fri, 03 Jan 2025 17:33:13 GMT

In addition to @Ksharp 's observation about model complexity, I would suggest that you look into using the NLOPTIONS statement in your GLIMMIX code. If the iteration history is relatively well behaved (i.e., it is close to monotonic throughout), simply increasing the number of iterations from the default of 20 using the MAXITER= option may result in convergence. I almost always set this at MAXITER=1000, especially when fitting binary data.

SteveDenham