And sorry, didn’t want to detract from your actual point about TTE which I agree with.

My preference is always to think of them as assumptions because so many things could go wrong. Non-adherence, loss to follow-up, competing events, not double blinded, imperfect allocation concealment then. We‘re on the same page though that most often the assumptions in the RCT are (much) weaker.

I agree with the first point. It’s unnecessary. RCTs nice when possible. TTE can sometimes be a substitute when not.

But RCTs rely on the same assumptions. Their advantage is that the design can make them much more believable.

Or you would need very rich datasets measured a very fine time scales.

Makes sense. So if some effect goes mostly througt such and such mediator it supports theory X and not theory Y?

I hope this isn’t coming off as me being hardheaded! I used to make the same argument you’re making and now I can’t see the logic anymore. Maybe you can help see what mediation estimands tell me that is more useful than the condition or joint estimands above!

But why am I spending money specifically on researching things that mediate the PPD-outcome relationship if my goal is just to reduce the outcome? Why not study the effect of attachment among people with PPD? Or the effect of joint interventions on attachment and PPD on the outcome?

Thanks for the clarification! But if my goal is to reduce the outcome, why is knowing how much attachment mediates the PPD effect more useful than just knowing how much intervening on the attachment reduces the outcome?

Maybe you can clarify (or maybe I’m misreading) because i read your wording as implying you’re just interested in intervening on the mediator or possibly jointly with the exposure. mediation analyses don’t answer those questions.

I know that feeling well…

What do psych people want to do with info on a mechanism?

Genuine question because at one point in my life I thought knowledge of mechanisms was obviously very useful but the more I think about it I wonder whether it really helps that much (at least in some contexts)

H/t to Vanessa didelez who got me thinking more about this

I went through a bunch of social epi papers recently that did mediation and this is pretty common:

Intro: need to know about mechanism
Conclusion: the mediator is another place to intervene to reduce Y

So the question is rarely really about mechanisms to begin with.

I remember someone came up with that in grad school but I can't even remember if it was me or someone else?

Is this many worlds bayes?

I would like this comment twice if I could.

Also interesting is this applied example of this: academic.oup.com/aje/article/...

Very fun to work with Andreas Stang and Charlie Poole on this. The latter of which was part of the original debate on this topic in the 1980's: academic.oup.com/aje/article/...

And presenting evidence or arguments for why the assumptions are to be believed or at least approximately so.