Hi all,
I am using PROC GLIMMIX to fit a 3-level model (clustered by hospital and then by physician, for example).
SAS automatically computes the overall covariance parameter estimate for hospital(physician).
How can I obtain the individual ones? i.e. just by hospital and just by physician? (this is asked of me to justify the use of clustering modelling techniques)
Thanks in advance!
The parameter for hospital should not be difficult--just adding it to the RANDOM statement should give a value. The request for "just by physician" kind of begs the question--if the physicians are truly nested in hospital, what does this mean? The only approach I can think of is to fit the data without hospital as a random effect. Since the fixed effects are identical, you could look at the Information Criterion of your choice (AIC, AICc, etc.) as a guide as to whether the model fits your data "better". So, I see three runs using the following RANDOM statements:
1. Just hospital
RANDOM intercept/subject=hospital:
2. Just physician
RANDOM intercept/subject=physician;
3. Nested
RANDOM intercept hospital/subject=physician;
Since these are all subject-based, I would start with adaptive quadrature (METHOD=QUAD) as the estimation method.
Steve Denham
The parameter for hospital should not be difficult--just adding it to the RANDOM statement should give a value. The request for "just by physician" kind of begs the question--if the physicians are truly nested in hospital, what does this mean? The only approach I can think of is to fit the data without hospital as a random effect. Since the fixed effects are identical, you could look at the Information Criterion of your choice (AIC, AICc, etc.) as a guide as to whether the model fits your data "better". So, I see three runs using the following RANDOM statements:
1. Just hospital
RANDOM intercept/subject=hospital:
2. Just physician
RANDOM intercept/subject=physician;
3. Nested
RANDOM intercept hospital/subject=physician;
Since these are all subject-based, I would start with adaptive quadrature (METHOD=QUAD) as the estimation method.
Steve Denham
Hi Steve. (love your work by the way)
Is it possible that some physicians are observed at more than one hospital in hypermonkey's data? In some health systems this is fairly common within a region or city, or if, for example, some specialists are hard to find.
If that is the case (physicians not truly nested), then two RANDOM statements will be needed, each with a different subject. Ordering the RANDOM statements then becomes critical as well. I would go with:
4. Physicians not nested
RANDOM intercept/subject=hospital;
RANDOM intercept/subject=physician;
If METHOD=MSPL is adequate (rather than adaptive quadrature), these can then be combined into a single, non-subject defined, statement:
RANDOM hospital physician;
Steve Denham
Message was edited by: Steve Denham, who can't spell this morning
Hi Damien,
Interesting question! Not a problem in this particular case, but definitely something to consider. Thanks!
Hi Steve,
This seems to be exactly what I am looking for. The estimates in case 3 (nested) will help justify why a clustering modelling approach is necessary for this analysis.
Many thanks!
I'll take a "Correct Answer" for that then
Steve Denham
Are you ready for the spotlight? We're accepting content ideas for SAS Innovate 2025 to be held May 6-9 in Orlando, FL. The call is open until September 25. Read more here about why you should contribute and what is in it for you!
ANOVA, or Analysis Of Variance, is used to compare the averages or means of two or more populations to better understand how they differ. Watch this tutorial for more.
Find more tutorials on the SAS Users YouTube channel.