Aren’t you just calculating the residualized pseudo-outcomes (so this all fits nicely in the semiparametric noncentered influence function literature)?

I have a similar take but for stats/data handlers who don’t have and don’t want to gain a familiarity for the process by which the data are made. E.g. engaging with the EMR signified and signifier doom loop.

Yes, this is a great example!

I have an interesting case study on this actually. So at the beginning of the semester, I was preparing a bit on the (mis)use of LLMs for the course I co-teach

One of things I did was have it summarize one of my own papers, since people say "it's so good at it"
arxiv.org/abs/2503.02789

And as a species we’re so badly wired for finding needles in the haystacks of systems we didn’t design and barely understand.

More complex stuff is doubly dangerous because some of the models (ahem Claude code) will generate so much code for you but then will make dumbest mistake (or nefarious stuff like functions with the right name that do nothing or generating fake data) and now you have a needle in haystack problem.

✅ Derive model performance statistics (IP weighting, standardization, doubly-robust) that allow the candidate prediction model to be misspecified.
✅ Theoretical results: identifiability conditions and efficiency.
✅ Simulation & applied examples showing where naïve models fail.

Our contributions:
✅ Formal definition of “counterfactual prediction estimands.”
✅ Derive estimators combining causal inference (IP weighting, standardization) with predictive modeling that allow for separation between covariates for confounding control and covariates for prediction.

Examples:

- Differences in post-baseline treatment policies between training and target population.
- Clinical decision support tools that are meant to inform treatment adoption.
- Removal of undesirable events or features in training data that are unreflective of the target population.

Why does this matter?

Many common tasks in clinical prediction modeling target outcomes or performance statistics under hypothetical interventions (either explicitly or implicitly).

Examples:

- Differences in post-baseline treatment policies between training and target population.
- Clinical decision support tools that are meant to inform treatment adoption.
- Removal of undesirable events or features in training data that are unreflective of the target population.

Why does this matter?

Many common tasks in clinical prediction modeling target outcomes or performance statistics under hypothetical interventions (either explicitly or implicitly).

Yes just got it!

Thank you!

Also frontier seems to be concurrent prospective trial plus observational study (although obviously this has been around forever eg WHI) pmc.ncbi.nlm.nih.gov/articles/PMC...

Not to say that both use cases don’t still have their issues of course!

Increasingly think that some of the best uses of TTE are complimentary, e.g. 1) postmarket in a space where you have premarket RCT that has helped set bounds of reasonable effects and maybe identified negative controls, or 2) in an emerging area to help argue for potential equipoise for an RCT

Depends on whether you’re interpreting it as SATE or PATE. If former then your sample has no power users and repeated randomizations using same procedure over fixed sample produces no bias. Now if this error raises suspicion of other randomization errors then maybe not?