Statistical Models Answer the Fundamental Clinical Question and Provide Clinical Trial Estimands

[isagonas]

Frank just posted an excellent blog post Statistical Models Answer the Fundamental Clinical Question and Provide Clinical Trial Estimands – Statistical Thinking .

I am sitting here with a glass of wine after a hard day of clinical work, and I would like to present some points on which I hope the experts in this forum could weigh in.

The core argument as I see it and which I fully endorse, is that model-based, rich, covariate-conditional outcome distributions are ultra useful for translating RCT results into clinically relevant absolute effects.

Since I feel that disagreement can instigate some interesting and fruitful discussion, I would like to offer some points were I would push back.

1. Randomization established treatment group exchangeability with respect to potential outcomes. It does not imply or require prognostic homogeneity among trial participants. I think, this distinction is important, since if not careful with the wording we may risk to invalidate marginal randomization based inference just because prognostic heterogeneity exists.

2. Marginal estimands do not require population sampling. A trial can just target a sample average treatment effect (SATE). I think this point has been extensively discussed in this forum.

3. I do not believe that the marginal estimands are irrelevant, since we can use them to make policy decisions.

4. CPMs are semi-parametric, not assumption free.

5. Maybe I did not get it correctly and it’s getting late here, so bare with me. I did not quite understand why omitting U is equal to setting tau=0. Are we sure that after marginalizing over U, the resulting model belongs to the same parametric family or retains the same additive linear predictor? By comparing the conditional RD to an X-speficic conditional target, dont we evaluate the marginal estimate against a target it was never designed to judge?

Thanks

Ioannis

[f2harrell]

@isagonas

Randomization established treatment group exchangeability with respect to potential outcomes. It does not imply or require prognostic homogeneity among trial participants. I think, this distinction is important, since if not careful with the wording we may risk to invalidate marginal randomization based inference just because prognostic heterogeneity exists.

Thanks very much for starting this discussion. Yes that distinction is important and the article said nothing to contradict it. Marginal randomization inference is valid, just not as useful or flexible.

@isagonas

Marginal estimands do not require population sampling. A trial can just target a sample average treatment effect (SATE). I think this point has been extensively discussed in this forum.

The problem with that is the majority of uses actually state that the SATE is estimating PATE which is far from true.

@isagonas

I do not believe that the marginal estimands are irrelevant, since we can use them to make policy decisions.

That is true in a world where decisions have to be oversimplified, e.g. a drug formulary has to include or exclude a new drug. It ignores variation in absolute benefit that is needed to be taken into account for individual patient decision making.

@isagonas

CPMs are semi-parametric, not assumption free

It’s true, although I never made that statement.

@isagonas

I did not quite understand why omitting U is equal to setting tau=0

What I meant there was this: Failing to include a term in the model is equivalent to including it but finding that it’s coefficient is zero. The only possible difference is a technical one: including a term, even when its estimated coefficient is 0.0, will increase standard errors of other terms (but not change their coefficients unless using lasso).

I appreciate the discussion and look forward to more.