GENMOD or GLIMMIX for 3-level random effects models

[halladje]

Hi There!

I am currently doing an analysis using 3-level data (students in classrooms in schools) with continuous, binary, and oridnal independent variables dependent variables. I am planning to do a random intercept only model for the analyses, but I am modeling independent variable effects at upper levels (both class and school level) in addition to individual level. 

I am currently using Proc GLIMMIX and altering the link and distribution parameters for the various outcomes. Currently, my ICC for my binary outcome does not seem to be accurate when using GLIMMIX. My supervisor stated that GLIMMIX is not a good method for binary data and I should use PROC GENMOD and obtain the ICC from t the “Exchangeable Working Correlation” in the sas output.

However, I am having a difficult time understanding the differences for GLIMMIX and GENMOD. When looking at documentation, it states that GLIMMIX handles random effects and variance estimates at different levels while GENMOD does not (opposite to what my supervisor stated). But I am not sure if "random effects" in the SAS documentation is referring to random slopes. 

Therefore, if I am using a random intercept only model, using 3 level clustered data, a binary outcome, and modeling upper level fixed effects what is the most appropriate statistical method? What would be the pros and cons fo GLIMMIX versus GENMOD and vice versa? 

Thanks so much, 

Jillian 

[PaigeMiller]

Certainly PROC GLIMMIX can handle random intercept models

https://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#statug_glimmix_a00...

using 3 level clustered data, a binary outcome, and modeling upper level fixed effects

I really don't know what "3 level clustered data" means, and I really don't know what "modeling upper level fixed effects" means. Please explain in more detail.

[halladje]

Hello there, 

• 3 level explanation: students (level 1) clustered in classrooms (level 2) clustered in schools (level 3). I have ~30,000 students in ~2,000 classrooms in ~200 schools.
• upper level effects explanation: using teacher reported variables (class level) and student aggregated variables to the school level (school level variables) as independent variables in the model. 

Can GENMOD also model random intercept models? What would be the benefit of using one over the other? Can both calculate an accurate ICC for continuous and binary outcome data? 

Thanks!

[PaigeMiller]

---

@halladje wrote:

Hello there, 

• 3 level explanation: students (level 1) clustered in classrooms (level 2) clustered in schools (level 3). I have ~30,000 students in ~2,000 classrooms in ~200 schools.

---

I would use the word "nested", not "clustered". Yes, GLIMMIX can handle nesting.

• upper level effects explanation: using teacher reported variables (class level) and student aggregated variables to the school level (school level variables) as independent variables in the model. 

Yes, you can aggregate the data to any level you think is appropriate.

Can GENMOD also model random intercept models? What would be the benefit of using one over the other? Can both calculate an accurate ICC for continuous and binary outcome data? 

I don't know.

[StatDave]

GENMOD does not fit random effects models. Instead, it fits a Generalized Estimating Equations (GEE) model when the REPEATED statement is specified. The GEE model is a population-averaged model and it could be used to model your data, but it does not provide correlation structures for multilevel data as discussed in this note. Multilevel (also called hierarchical) models can be fit in PROC MIXED (for normal responses) or PROC GLIMMIX and are subject-specific models.