Compensate for correlated observations

#1
Dear users
I have an issue that I am not to solve alone and I am here to ask for your help. I work in the ophthalmology field and I have a group of patient included in a clinical study and the goal of this study is to correlate some continous parametric variables and 1 nonparametric variable each other. However, half of the patients had included one eye only, while the others both eyes, which are correlated. How can I compensate this bias?
Thanks in advance
 

hlsmith

Omega Contributor
#2
You need to run a multilevel model that controls for this covariance. You will see if you look for it, this is what is used in many ophthalmology papers. I have helped on a coupe and it is what I did. They may use other terms like patient clusters or hierarchical model, but all get at controlling for this dependency.
 
#4
GEE's are probably the best choice when correlation structures are not exactly known. However, in your case, you have a clear nested design where eyes are nested within individuals. Thus, you should use a repeated-mesaures ANOVA or a (G)LMM, where the patient ID is your random coefficient.
 

hlsmith

Omega Contributor
#5
Agreed. Though, many people may say GLMM trump repeated-measure ANOVA, in that it has a lot of options, but most importantly it doesn't drop observations when data are missing.