GEE Variance Component - An R-side variance component is confounded

[JonKetchup]

I built a GEE with a compound symmetry covariance structure using PROC GLIMMIX - observations are clustered by "doctor", and each doctor can see as many as 100 patients. I would like to compare to a GEE a variance component covariance structure. However, whenever I run the code, I get the note, "An R-side variance component is confounded with the profiled variance." What does this mean and is it ignorable? Even if I change the covariates in the model, I get this same note. Appreciate the help. Here is my code. The variable constructions are:

• Outcome: binary yes/no
• Location: binary, location1/location2
• Var2: binary, yes/no

proc GLIMMIX data = DATA ORDER=INTERNAL empirical=root ;
class Doctor Location Var2;
model Outcome(event='1')= Location|Var2     
	/ dist=bin link=logit covb solution ;
RANDOM _residual_ / SUBJECT=Doctor TYPE=VC vcorr  solution;
RUN;

NOTE: An R-side variance component is confounded with the profiled variance.

[sbxkoenk]

Hello,

See here :

Need Help Deciphering a Mysterious Note in the SAS Log When Running PROC GLIMMIX
https://communities.sas.com/t5/Statistical-Procedures/Need-Help-Deciphering-a-Mysterious-Note-in-the...

See also here (although this paper contains no direct mentioning of the NOTE you have bumped into) :

SAS Global Forum Proceedings 2018 : Paper SAS2179-2018
Insights into Using the GLIMMIX Procedure to Model Categorical Outcomes with Random Effects
Kathleen Kiernan, SAS Institute Inc.
https://www.sas.com/content/dam/SAS/support/en/sas-global-forum-proceedings/2018/2179-2018.pdf

Cheers,

Koen

[JonKetchup]

Thanks for the resources. I stumbled on the post previously. The model in that post had random effects such that both the g-sided and r-sided effects were defined; my model only has r-sided effects, but I'm not sure if that matters. I tried the adjusted code and got the exact same output with the same note. 

proc GLIMMIX data = DATA ORDER=INTERNAL empirical=root ;
class ID Doctor Location Var2;
model Outcome(event='1')= Location|Var2     
	/ dist=bin link=logit covb solution ;
RANDOM _residual_ / SUBJECT=ID(Doctor) TYPE=VC vcorr  solution;
RUN;

[StatsMan]

R-side structures are notoriously difficult to fit and interpret when you have a binary response. If you wish to correlate the observations from the same patient, then try fitting this as G-side instead with

random intercept / subject=id(doctor); 

Also, try adding an NLOPTIONS statement to improve convergence

nloptions tech=nrridg;

To understand the confounding message, consider a model with normal errors. An R-side fit of 

random _residual_ / subject=id(doctor) type=vc;

adds nothing to the model. The variance fit by this statement is the same as the residual error variance for this model. You can try adding a covariance structure to the R-side fit through the TYPE= option, but again these structures are hard to fit with a binary response. 

[JonKetchup]

I really appreciate the insights, but I feel like that is equivalent to saying that other multi-level models are better than GEE. Note that there is only one observation per patient; we are clustering by doctor since we expect there to be an association between patients that saw the same doctor. Our analysis is not interested in estimating the effect of the doctor, but we wanted to take into account the natural clustering construction of the data. It is for this reason we chose a GEE instead of say an HLM.