The critique of unmeasured confounding is often levied in a lazy/broad way. It is trivially true in any observational study. But if the critic can't think of a plausible such confounder and posit a reasonable direction/magnitude of its bias then they're not doing productive science.
I agree with you that looking at frequentist properties (ie frequencies over repeats of experiments) of Bayesian procedures is great, whether your dgp is Bayesian or not.
Simulating their trial several times sounds a lot like frequentism, interesting to hear from you that we should be doing this more (I agree).
Non-overlap Average Treatment Effect Bounds by Herbert P. Susmann, Alec McClean, and Iván Díaz
New preprint out on a way to handle structural and practical violations of the overlap (also known as positivity) assumption in causal inference -- as long as the outcome is bounded, we derive simple partial identification bounds on the ATE. With @alecmcclean.bsky.social and @idiaz.bsky.social
He did it before Double Machine Learning
I met with professor Mark van der Laan because I think his work is pretty incredible and it sometimes feels like a secret that only a few people know about, especially in industry.
1/
#CausalSky #StatSky #CausalInference
But I do agree that the addendum is progress from the previous state of things!
But more seriously, the addition would have been a sentence and a couple of references, not a complete framework, which already exists.
The better analogy is a bus that drops you off midway to your destination, not even in a bus stop, and steals your phone so you can’t get home 😉
I think the addendum fell short and should have described a need and approaches to mathematically define and identify estimands.
Thanks for sharing. Tangential to the thread but I have to say that I disagree with the statement of the abstract that causal inference estimands correspond to an effect in an ideal trial. Many valid scientific questions are expressed in terms of parameters that cannot be identified in trials!
What I do not follow about both TTE and addendum is why reinvent the wheel. I get it if it is about communicating to new audiences, but not if they are presented as new methods when they are so clearly inferior to what's already out there.
bsky.app/profile/idia...
TTE is certainly useful (I use it). But it is not a replacement for formal causal models + causal estimands + identification + optimal estimation etc. I recommend reading Maya Petersen and colleagues' papers on the "roadmap for causal inference."
Presenting TTEs as a method rather than as a communication tool had the unintended consequence of folks slapping the moniker in studies as a quality signifier without doing the actual leg work required to address the issues, discussing them, or even understanding them.
Right, IMO target trials are a great *communication tool* to talk with folks who do not have in-depth causal inference training (hence their success), but to really understand the issues one has to rely on standard theory (causal models, identification, estimands, optimal estimation, etc.)
Curious to hear if the gripes are substantive or if they are attribution-type (e.g., causal inference was using estimands long before the addendum).
I have a new paper out on a simple way to do causal inference with left-censored outcomes. This comes up with environmental data because measurements often have a lower limit of detection -- e.g. a chemical is undetectable below a certain level
www.tandfonline.com/doi/full/10....
I wrote something about statistics under authoritarianism
It should be telling that it is the field of CI that has given folks the tools to understand the conditions under which observational studies deliver causal effects. You may argue whether those conditions are ever achievable, but criticizing CI for achieving its goal seems silly.
Underlying this there is a valid and worrisome criticism of causal inference in practice, but most comments criticizing CI as a field miss the fact that “x methodology is being abused in practice” can be correctly said about almost anything.
Clinician: How do I make sure Y(a) is independent of A conditional on covariates
Statistician: You measure all common causes of A and Y…
🤷🏻
This is why I prefer causal assumptions in terms of exogenous vars in structural causal models rather potential outcomes. Sure, the former is often mathematically stronger, but the latter is inscrutable by subject matter experts.
I guess I would agree that Borges was pretentious:
www.youtube.com/watch?v=NJYo...
I think missing data is a CI problem (counterfactual is "would have observed the data") but not the opposite. E.g., recasting mediation analyses etc as a missing data problems seems contrived.
Cheap non alcoholic cava in a tetra pack container. How come the outcomes have to be “potential”?
At some point we’ll have to give these guys some agency 😉
Re: the OP, it also seems right to me that stats should focus on the properties of estimators, but the stubborn rejection of the language that links stats to science (CI) by some “trad” statisticians seems very odd to me
On this, I agree with Pearl and others who have emphasized that estimators aren’t causal. A causal estimand equals a statistical estimand (usually) under assumptions; estimators target statistical estimando which may or may not have a causal interpretation.
Excited to present on Thursday @eurocim.bsky.social on new work with @idiaz.bsky.social on (smooth) trimming with longitudinal data!
"Longitudinal trimming and smooth trimming with flip and S-flip interventions"
Prelim draft: alecmcclean.github.io/files/LSTTEs...
For ordinal outcomes you need to make sure the logit models either include treatment-specific intercepts, or fit the models separately for each treatment arm.
Agree, but would just add a caveat for the unsuspecting that this works for binary but not ordinal outcomes. In general it only works because the estimator is doubly robust, so a more robust ;) piece of advice would be to just use doubly robust estimation.