I agree - bayesian estimation is often a pragmatic choice when MLE simply doesn’t converge in complex model situations. I wonder how many users of bayesian stats are truely « philosophical » users !

Although, tbf, for the total predictive validity (and not only the incremential one) of that manifest to manifest best case scenario, one would have to add .57 * .41 * .67 * .62 = 0.10. So .16, which still is pretty bad!

Yeah, this looks pretty bad…

Combine*, rather

U can use both with the model-implied instrumental variable (MIIV) framework 😎
scholar.google.com/scholar?hl=f...

It is surprising I agree. Also I wonder if some samples (e.g. n=140) can actually handle (well) 2nd order factor models

The cherry in top is the fact that the relashionship between 2 of the variables are likely bidirectionnal

I don’t know how the economists do it… I have been trying to find potential IVs for the M->Y path in a mediation model, but the exclusion restriction is so unlikely everytime, and i can’t think of practical partial IVs. For psych variables, I feel like there is very little room for IV method 😅

And decide a way (or ways) forward :)

My point was that it feels wrong to outright consider non-manipulable variable as non-causal ones. Saying X is not a cause is actually more speculative than saying it is cause, because it is a negative. But I 100% agree that we should have a discussion on the epistemological side of the matter

Very much agree!

one with « associationnal mediation » or no mediation at all

That is indeed questionnable... Regarding this issue and any causal claim (here, mediation and longitudinal effects), my opinion is that it is better to be (proven) wrong than not trying. I prefer a world with cross-sectionnal mediation with spelled assumptions (that can be challenged) than

I find the « not manipulable = not a cause » view of PO framework quite depressing (and anthropocentric). Then, what about sociology, evolutionnary biology or Astronomy, just to name a few ? I totally side with Pearl here