1
$\begingroup$

I see the question of model selection for GEE come up and answers seem to usually involve QIC/MLIC for non-nested models, or LRT for nested. That's all fine, but when I have 50 predictors it's a bit tedious to write all 50-factorial model statements.

Some have floated stepwise selection using QIC/MLIC but my understanding is that stepwise tends to create a lot of problems and is generally considered outdated. What are better processes for screening initial variables that handle longitudinal data considerations well? This sample is with subjects being measured 1-4 times at varying intervals. I'm hoping for something I can implement in R, ideally.

$\endgroup$
1

1 Answer 1

2
$\begingroup$

I think "it's a bit tedious to write all 50 factorial model statements" should get some award for "biggest understatement on Crossvalidated"!

I would argue (as I often do) against any automatic method. The reasons for opposing this in ordinary multiple regression are, if anything, amplified in GEE. A good discussion of this is at this thread.

Instead, I suggest using your substantive knowledge and your intelligence to do this. 50 variables is a lot! In most situations, there are surely some near duplicates. In the GEE context (or any longitudinal context) you might try removing ones that are missing a lot of data. (There are some areas where 50 variables is fairly common, but ... are you working in one of those? You don't say what your subject is.)

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.