Well the issue is that for any regression model, the default ML SEs are incorrect for IPW. One at least needs robust SE (conservative for the ATE). Even better is an SE that accounts for estimation of the weights. Bootstrapping makes this easy.
Comments
Log in with your Bluesky account to leave a comment
Noah, isn't it correct that these robust sandwich SEs are "conservative or anti-conservative"?... So isn't the characteristics of those SEs better described as 'unpredictable'?
Yes, like I said, conservative *for the ATE* ;) For the ATT it's a different story as explained in that article. See also this one: https://doi.org/10.1002/sim.10078, in particular figure 3. Robust SEs (first column) are generally conservative even for the ATT, but not always.
HC standard errors are unfortunately not available for glmmTMB. Would it be possible to model a dispersion parameter in `glmmTMB()` instead? Anyway, thanks for the feedback, we added a short section on standard error estimators.
I don’t think the issue is heteroscedasticity, so modeling the dispersion would not be enough. I don’t think including a mixed effects model in a vignette for IPW is a good idea because it is not well studied. I would focus on point treatments and outcomes instead and use the correct SEs.
Comments
https://pubmed.ncbi.nlm.nih.gov/35106534/