Thank you! Link fixed and page updated!

Do we have any information about the study design? The publicly available deactivation RCTs I've seen do also show that deactivation improves well-being, e.g.

www.nber.org/papers/w33697

They might read this one: psycnet.apa.org/doiLanding?d...

I think we have a duty to teach students that we now know that "how the field has thought" about mediation leads to theoretically convenient answers that are prone to bias.

Now we have to 'unlearn' those answers and reject that procedure.

Thanks for responding; not sure if that reg. is right.
how about comparing:
1D v 1S
(1D + 1D) v (1D + 1S)
(1S + 1D) v (1S + 1S)
(2D + 1D) v (2D + 1S)
(2S + 1D) v (2S + 1S)
etc
This analysis would deal with @adamzelizer.bsky.social's good points about "bad control" and that there's just 1 effect

If that's the design,

(2 Daughters) compared to (2 Daughters + 1 son),

it's not relying on the as-if random assignment of child sex, right?

There's obviously plenty of selection into a third child.

can we compare

(2 Daughters + 1 son) v (2 Daughters + 1 Daughter)

instead?

seriously hit refresh like what is happening to my browser

so well put. I esp. agree with turning off the damn in-line AI and copy-pasting like a sketchy SO answer. I like also how you talk about why the AI for advanced users is different from the AI for beginners

what do you think your modal referee comment is in those 47 (!) reports in 2025?

Absolutely, I often use reanalysis to estimate estimands that the original experimenters didn't consider (a different outcome, a different subset of subjects). In the probit / OLS example, I'm imagining that we're both after the same estimand (the ATE).

Definitions are hard, thanks for engaging!

Here's a common one in my work:

Someone analyses their experiment with a probit regression. That's fine, but I prefer OLS; I reanalyze the data because I want the OLS estimate, not because I want to assess the "robustness" of the probit estimate.

(This is often on the way to meta-analysis)

yes Matt, this is my view too. At least there's an estimand! Maybe you don't care about the value of that estimand (i.e., I think Cyrus cares most about estimands that inform policy decisions).

I agree with you that survey exps (lists, conjoints, etc) are often used for descriptive, not causal, inference (we write about it here: book.declaredesign.org/library/expe...). Also useful.

There's no need, though, to write survey experiments for causal inference out of the "revolution."

I especially like your definitions for replication and reproduction.

One word I use that's not here is "reanalysis." I think some of its meaning is covered by "robustness testing" but sometimes (usually?) my reanalysis isn't about the "robustness" of the original.

thoughts on "reanalysis"?

That's great to hear! Congrats on an amazing paper

Should have read the actual paper first! Thanks for your engagement on the viz side; v. interesting example here.

I was surprised @ 90% cis and the "predicted" values" in the viz and prefer your viz (95%CIs and the raw data) for sure.

I'll say that the viz is ironically v. useful in this case because it raises a tension between the picture and the design (wait, I thought the outcome was binary???)

Also, friends, reviewers, editors -- please be willing to publish experimental estimates of the effects of anti-corruption messages on vote choice even if the "theory isn't novel" or "so and so already published on the effects of anti-corruption."

This is an impt. estimand; we need many estimates!