Also, not for nothing, the lack of a dichotomous criteria for making claims helps cut any urge. Who cares if the whatever credible interval overlaps zero?

Also in practice bad priors manifest in ways that aren’t your inference. Like the posterior predictive goes out an order of magnitude beyond your data.

And if you have enough data, any prior that isn’t awful and contrived winds up yielding the same inference.

Loves them!

WOW a loggerhead shrike without intense heat distortion, what afind!

Agree. I rarely see it.

I measured my leg to the nearest kilometer and my arm to the nearest nanometer. The latter was a much better predictor of height. This calls into question notions that legs, not arms, make you tall.

More broadly if you’re making a predictive kitchen sink model with varying metrics you really can’t make claims about those metrics relative impacts on your outcomes.

By analogy, I have little doubt that which stores you go to would better predict purchases than what adds run in your zip code… but it’s a stretch to say physical advertising is more effective than digital.

Effectively any influencer and media effects are shunted to zero by this measure which we *know* is simply not the case. Ditto for groups, viral content etc. All if the means by which we know/beleive the internet influences opinion are shunted to zero.

But there’s a much more obvious problem we can find by thinking about mechanism.

What are the odds a voter in 43402 is friends with Trump or Biden on Facebook? Nearly zero as measured. Does this approximate the probability they saw their posts?

Truly nothing in this analysis supports any causal claims, even in the hedged language they use. Arguably not even predictive as described because online connections aren’t actually measured here.

First and foremost, it estimates online friend ties using data that predicts those ties at the zip code level. It’s one value per county pair. By contrast, their offline networks use 5-minute physical proximity to define a tie.

It’s unsurprising that coarser measure are less predictive.

Just vibes lol

I think that view is common. Here’s the center for open sciences take on why you need to preregister. It distinguishes your work as confirmatory which they assert a lot about.

www.cos.io/initiatives/...

It happens a lot, so someone needs to unfortunately point out that it simply isn’t the fact science has been lost and finally figured it out.

Which is indirect evidence that preregistration doesn’t really reduce publication bias all that much.

We dig into that a little here. The tldr is that bias seems a lot larger if your model of publication is that false positives are true zeros. Our model maps very well to empirical estimates of reformed research, even parameterized by the open literature.

osf.io/rkyf7_v1/

But I think he’d agree with you on the tail end. Some folks mistake preregistration merely being associated with a paper as quality.

@devezer.bsky.social discusses it somewhat here.

direct.mit.edu/qss/article/...