So there are some answers I’ve heard to this, but none of them great to me, so I would ask this question in a rhetorical way! (always difficult to know whether a question is rhetorical on Twitter!) Thread alert… 1/9 https://twitter.com/W_O_Johnson/status/1295664224027512832

One advantage of GEE: fewer distributional assumptions of the random effects. Sure, but surely that’s outweighed by the fact that GEE doesn’t tell you anything at all about those random effects. 2/9

One key difference: GEE gives you the population average (marginal) effect; MLM gives you the cluster specific (conditional) effect. 3/9

That’s the difference between estimating the effect of changing X by 1 unit for everyone, on the population, vs the effect of changing X by 1 unit, for an individual. Those two things are the same with an identity link function (continuous Y) but they aren’t for a logit model 4/9

So they are different research questions, and your model choice should be defined by the research question, not the other way round (see this paper) https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3085136/ 5/9

But: we can actually get the pop average effect from a MLM as well – by simulating a prediction distribution across clusters on probability scale, and taking the mean of those predictions. We can’t do the reverse: we can’t get the cluster specific estimate from GEE 6/9

And there’s all sorts of other stuff we can’t do with GEE – random slopes, variance components, etc etc https://twitter.com/W_O_Johnson/status/1295766959477460993 7/9

This tweet is correct. GEE is trying to 'correct' for something that really isn’t a technical problem – it’s of substantive interest and should be explicitly modelled as such https://twitter.com/W_O_Johnson/status/1295673827712802822 8/9

This is also good on the topic from @StatModeling https://twitter.com/0xeinar/status/1295665694906908672 9/9