We indeed make the assumption of a superpopulation (you can definitely argue that’s not always relevant). Under that framework, we recommend using the variance estimator in the paper of Ting Ye er al (paper you mentioned earlier). This variance estimator is based on the efficient influence function.

It also applies to standardization. So, we show that simple models without random effects are not guaranteed sufficient to account for clustering in non-linear models, and especially when estimating counterfactual means (even in linear models).

Our paper focuses on AIPW. We show when there will be low coverage due to clustering (by center). You can account for it using fixed effects or random effects. However, fixed effects have bad performance due to overfitting when there are many small centers, which was the case in our example

It was an absolute honor to chair EuroCIM2025! Huge thanks to everyone who made it happen - especially Oliver Dukes and @svansteelandt.bsky.social for their incredible support. A big thanks also to all the participants, without whom this conference wouldn’t be possible.

Looking forward to EuroCIM next week, where I'm going to talk about how we can make rigorous causal inference more mainstream.

Anybody else who will be there 😊? www.eurocim.org

Only ONE WEEK left to register for EuroCIM. Don't miss your chance to join top experts in causal inference for insightful talks & networking. Secure your spot now! www.eurocim.org/registration...

The EuroCIM program is live! Explore the sessions, speakers, and schedule here: www.eurocim.org/program.html. Get ready for an exciting conference! 🎉