I’ve said it before (as have others) and I’ll say it again, there’s no such thing as a non-causal mediation analysis. journals.sagepub.com/doi/10.1177/... - ThreadSky | a Reddit-style client for Bluesky

dingdingpeng.the100.ci • 171 days ago

I’ve said it before (as have others) and I’ll say it again, there’s no such thing as a non-causal mediation analysis.

https://journals.sagepub.com/doi/10.1177/25152459221095827

Comments

seanpmackinnon.bsky.social•170 days ago

Ooo, we're doing mediation discourse again!

Mediation is just a null hypothesis test for the difference between two slopes in multiple regression (total - direct), so I guess if every mediation analysis is causal then every multiple regression analysis is also causal.

seanpmackinnon.bsky.social•170 days ago

I guess if you're saying "every time somebody specifies a variable as the outcome Y in an equation, it is implicitly assuming a causal model" then yeah, I kinda agree.

Big difference between "the model assumes thing" and "the model proves thing" I guess.

marcuscrede.bsky.social•170 days ago

Disagree on that. I can use regression models to make predictions (e.g., in selection settings) without making causal claims about the IV-DV relations.

zheer.bsky.social•170 days ago

If there is anything in the multiple regression, that has been designated “a confounder” or “a mediator”, then yes.

And if there isn’t, it’s often not entirely clear what is being estimated — but it often turns out to be an effect of X on Y.

dingdingpeng.the100.ci•170 days ago

Then why is it called mediation test and not confounding test or collider bias test? Because that’s statistically the same, isn’t it.

dingdingpeng.the100.ci•170 days ago

And “it’s just the null hypothesis test for the difference between two slopes” is very much *not* how it’s used in the literature. This seems to be trying to redefine mediation analysis as a statistical analysis completely divorced from what people mean when they talk about mediation.

seanpmackinnon.bsky.social•170 days ago

Well, people in the literature also frequently use a lot of statistics incorrectly (e.g., that large p-values mean that the null hypothesis is true).

However, it wouldn't follow logically that we should re-define p-values to mean the incorrect assumption people hold.

seanpmackinnon.bsky.social•170 days ago

If your point is that virtually all people using a mediation analysis *assume* a causal relationship in their theory when they run the analysis, then sure I agree.

If you're just saying that the word "mediation" is a bad name because it's statistically equivalent to confounding, I also agree.

dingdingpeng.the100.ci•170 days ago

??? Are you talking about mediation analysis or are you talking about "the practice of checking whether a coefficient changes when another variable is included in the model"?

dingdingpeng.the100.ci•170 days ago

But mediation analysis was developed for the express purpose of testing mediation analyses which are explicitly phrased in causal terms by, for example, Baron and Kenny? How on earth is that comparable to people misinterpreting p-values?>

dingdingpeng.the100.ci•170 days ago

Seriously, this is taking a particular analysis approach developed with a clearly causal purpose, reducing it to the statistical essentials, and then claiming that people who use it for the purpose for which it was expressly developed are misinterpreting what it's good for.

boryslaw.bsky.social•170 days ago

Direct & indirect effects have known, clear causal definitions, the test you talk about is a *special case* of mediation analysis, & there is no other reason to call this test "mediation" unless one has mediation in mind. The test you mention is the only thing that is *just* the test you mention :-)

marcuscrede.bsky.social•171 days ago

If we were honest with ourselves we would retract all of these idiotic mediation papers from the literature en masse.

aidangcw.bsky.social•170 days ago

I’ll say it again too, there’s no such thing as a mediation analysis, there’s only a mediation inference, the analysis and model are ignorant of anything you might want to conclude

dingdingpeng.the100.ci•170 days ago

You think it’s wrong to name an analysis after the intended inferences?

aidangcw.bsky.social•170 days ago

I think it contributes to the false impression that if you set the model up a certain way the inference necessarily follows

So yeah, it causes problems IMO

dingdingpeng.the100.ci•170 days ago

I'm not quite sure how you'd report a mediation analysis without implying that the corresponding inferences follow. What I see a lot, however, is that people try to weasel their way out of inferential responsibilities ("tis but a correlation") while still implying strong claims.>

aidangcw.bsky.social•170 days ago

I prefer something like, “when adding M into the model, it accounted for 80% of the association between X and U

dingdingpeng.the100.ci•170 days ago

I dunno but I feel that already implies that the DAG is either X <- M -> U or X -> M -> U, discarding the possibility of X -> M <- U which could lead to literally the same pattern. So I don't think this is causally neutral at all actually.

dingdingpeng.the100.ci•170 days ago

In general I feel that people are *very* eager to claim that their statistics are just that, statistics that don't imply any inferences resting on strong assumptions. And then magic happens and in the discussion section the inferences are drawn anyway.>

dingdingpeng.the100.ci•170 days ago

So I think what is causing a lot more issues is that people try to deny they're trying to draw certain inferences while simultaneously drawing them. And "the statistics are just what they are, they don't mean anything on their own" is part of that logic.

michelnivard.bsky.social•171 days ago

I subscribe to a “All single headed arrows are causal claims” policy for all DAGs, path diagrams and SEMs

dingdingpeng.the100.ci•170 days ago

Uuuh do you include factor analysis…

michelnivard.bsky.social•170 days ago

Hard question, like AI image generating models have a latent space with latent variables, those are clearly causal (they cause changes to the image) but their not “real” there just convenient low dimensional representations. So I’d say yes, but… 1/2

michelnivard.bsky.social•170 days ago

I allow for a latent variable being an abstraction, a part of the model where the researcher isn’t to worried about getting it right or wedded to that solution. When a factor solution becomes ingrained (big-5) have to be able to set it aside if it impedes progress or make a case for it being real.

carolinefiennes.bsky.social•170 days ago

Do you ever use a flow-chart?
How does that typically work out for you? 😜

michelnivard.bsky.social•170 days ago

🤣

michelnivard.bsky.social•170 days ago

For real though…

Comment image

bachl.bsky.social•170 days ago

Is this your "reintroduction to all the new Bluesky users" :D

dingdingpeng.the100.ci•170 days ago

lol maybe!

jimmycoan.bsky.social•170 days ago

Interesting. But I always think about them conceptually as, at best, failing to rule causal meditation out.

dingdingpeng.the100.ci•170 days ago

Ideally, for that, you don’t just have a simple mediation model (which is saturated and thus untestable) but an overidentified model which would allow you to discard a model.>

dingdingpeng.the100.ci•170 days ago

Otherwise, I guess you can conclude that “Assuming the mediation model is actually true, the indirect effect wouldn’t be zero.”

jimmycoan.bsky.social•170 days ago

I think I was trying to get at the idea that in typical, non-overidentified mediation models, while we can estimate the indirect effect, we’re still in a space where causal interpretations are super hard without strong experimental designs.

jimmycoan.bsky.social•170 days ago

I definitely see how an overidentified model allows us to more rigorously test assumptions and discard incorrect models, which gets us closer at least to justifiable causal inference.

jimmycoan.bsky.social•170 days ago

Yes

ianhussey.bsky.social•171 days ago

exhibit number 2, your honor

https://www.cambridge.org/core/journals/behavioural-and-cognitive-psychotherapy/article/abs/effects-of-brief-behavioural-activation-on-approach-and-avoidance-tendencies-in-acute-depression-preliminary-findings/937187B4C4F92AFC36287805FA1FC33E

Comment image

dingdingpeng.the100.ci•170 days ago

“Potential role of…conveying” is definitely one for the “I can’t believe it’s not causal” list 😭 apart from that, mediation analysis is not supporting the assumption. It’s only valid under the assumption…

ianhussey.bsky.social•171 days ago

Whenever I see pairs of qualifiers and then strong claims like this between two sections in a manuscript I think of this Nelly line

Comment image

ianhussey.bsky.social•171 days ago

It's like Rorty's "Truth is what your contemporaries let you get away with saying" in action.

It's not causal... Unless? No, .. but maybe?