Cameron PatrickLearned recently that if you add a collider to your multiple imputation model it will break everything, even if you don’t condition on it in the analysis. Just in case you wanted to hear about new ways to ruin your statistical day.
Peter Ellis@cameronpat this surprises me (not saying it's wrong, just that i didn't know). Is it well known and I've just not been paying attention?
@peter_ellis I don’t think it’s well known - my supervisor mentioned it off hand and me and the postdocs in the group weren’t aware. This blog cites what I think is the article my supervisor sent us: https://tompepinsky.com/2019/04/29/multiple-imputation-with-colliders/
Zheer Kejlberg Al-Mashhadi@cameronpat @peter_ellis So only if missingness depends on both X and C in this case. I wonder if the same is the case for missing data in a confounder Z1 with the following collider (Z1 -> C <- Z2/X/Y). I’m guessing again missingness would have to be dependent on both of the other nodes before causing bias if the collider is used for imputation. Although I wonder if it would be as extreme as it is with missing data in Y…
@zheer_kejlberg @peter_ellis I don’t have a good answer to that. You could simulate to find out I guess!
@cameronpat @peter_ellis oh, I’m gonna! Will post results later.