First, I don’t have a dog in this fight but I have a lived experience of failed (nontrasportable) DT for over 3 decades. It’s very bad for the public health. So I’m going to be blunt about assumptions and consequences.
Certainly this post is brilliantly phrased as always, but this exposes the deeper vulnerability of idealized DT framing.
“Concurrent control” doesn’t remove the need for exchangeability; it merely constructs a temporary, synthetic version within the trial’s own P(X). The RCT replaces an assumed invariance across studies with an imposed balance inside one.
P(Y \do(T), X) may be stable, but P(X) rarely is.
When transportability is implied without explicitly comparing Xrct to Xicu the very covariates that define the structure of causal effect modification. Without that comparison, there is no bridge between P(X)_trial and P(X)_ICU the ATE is unanchored.
So yes while exchangeability across studies is not assumed between studies. it’s silently re-assumed when we pretend our RCT result applies beyond the study, at which point we assume it all over again, just without saying so.
That hidden assumption of non-assumption marks the epistemic fault line between the RCT(DT) and Causal Inference frameworks.
I’m working on some educational material regarding Causal Bounds, starting from observational data. Later on I’m planning on updating slides for experimental data, combined data and bounds with DAGs.
"There are several aspects of Baby K.J.’s story that define the FDA’s plausible mechanism pathway.
…
Fifth, there is an improvement in clinical outcomes or course. In conditions with progressive deterioration, consistent improvements will be viewed favorably by the FDA."
Here, this plausible-mechanism piece seems to describe N-of-1 scientific reasoning that proceeds without need for statistics to glean a faint signal out of masses of data.
Its vagueness and superficiality notwithstanding, the piece at least does not describe approvals of marginally-effective drugs such as I addressed in “ Where are the exceptional responders? ”.
