Joe Bak-Coleman

Apologies for not tagging! What a brilliant piece. Would love to chat with the team in the new year about some findings we’re dropping soon that are very relevant.

I was wondering that, as well as if they turn off or tamp down predation during flowering to avoid selecting against pollination.

Christmas came early this year! Very happy to see our paper out in Science Advances. Led by @lfoswaldo.bsky.social, we ran a unique collective field-experiment on Reddit, to better understand who is participating in online debates and why.

Paper: www.science.org/doi/10.1126/...

And more below 👇

I can be picky about stats and methodology but this is just a fantastic on platform experiment top to bottom.

The authors push back saying they don’t want to deviate, because if they’re do (in their mind) they can’t make strong claims. But the claims they make can be confounded by including things like post treatment effects.

Really great thread to read if you use regression. A key point of awkward in peer review for me is pointing out that a preregistered choice to make causal salad yields and analysis that can’t answer the questions.

If you want to be able to interpret those relationships (even shy of causal claims) I’d keep covariates tight and justified.

I should add if your goal is solely prediction, penalized regression with whatever in it is a reasonable starting point but you wouldn’t want to take too much stock in the individual coefficients in isolation.

We need transparency standards for social media research that involves companies | PNAS We need transparency standards for social media research that involves companies

Fantastic and blistering comment about the need for transparency in industry-academic collaborations.

From @georgiaturner.bsky.social, Ian Anderson, and Luisa Fassi

www.pnas.org/doi/10.1073/...

Counter intuitively you can include variables that confound findings and bias effects of others. It’s all so much worse when things are nonlinear. In those cases, everything is implicitly interacting anyways.

Regression, Fire, and Dangerous Things (1/3) It isn't my job to disappoint people, but I'm good at it.

Pretty good run down here by @rmcelreath.bsky.social, also power considerations from @statmodeling.bsky.social

elevanth.org/blog/2021/06...

statmodeling.stat.columbia.edu/2023/11/09/y...

Two modes: hungry and horny

Flytrap flowers might as well be saying “sorry I ate your buddies”

The llm made the step function right? It goes up and then it’s flat. I wouldn’t be terribly shocked if the llm whiffed something here.

I can’t find the code the llm wrote but I’d wager it’s a bug.

Also, I can’t think of a good mechanism by which going from one to two authors pumps citation rates which plateau regardless of increasing authors beyond that. Certainly big team science should be cited more, no?

If the finding is generalizable you’d expect the same general functional shape. If the data really are qualitatively distinct, then I’d start to
worry about the conserved green curve being a statistical artifact from variance not being well accounted for as author N increases.

The authors suggest it’s a different time period or some such, which fair enough, but you’d think the shape of the purple function wouldn’t have entirely changed with a bit more data.

I’ve given them a whirl for analysis and while they absolutely can code up a half decent model, they also sometimes just replace your data with test data.

One is concave up, the other is a step function.

My brother in Christ those are such very different graphs. Is the original work wrong? Is the llm wrong?

This a pretty funny encapsulation of ai scientist work. The authors hand it some softball software tasks, then it comes back with very different results than what it’s replicating and the authors shrug at the data being different

www.nature.com/articles/s43...

Algorithmic monoculture and social welfare | PNAS As algorithms are increasingly applied to screen applicants for high-stakes decisions in employment, lending, and other domains, concerns have been...

There’s also this whole thing. Even if it improves decision making it can make everyone wind up with worse hiring decisions.

www.pnas.org/doi/10.1073/...

I checked one that failed (200) and one that succeeded (1100)

…

The sample sizes are wild for those.

I honestly would have expected the retraction of High Replicability paper would have taught the metascience community some things abt design and inference. But ofc due to the authors reframing it as "a minor honest mistake of forgetting to preregister", important lessons have not been learned.

What Year Is It GIF ALT: What Year Is It GIF

But bayes factors?

Meanwhile I’ve never found a use for either. Arguably did my first “objective” Bayesian stats trick recently and it was really just to be able to reuse the same model across 6 of the same type of distribution that varied in scale.

I love the absolute bogeyman that Bayesian stats has become for some folks.

Exclusive: Trump administration orders enhanced vetting for applicants of H-1B visa An internal State Department memo said that anyone involved in "censorship" of free speech should be considered for rejection.

Heads up for colleagues from abroad planning travel to the us.

www.reuters.com/world/us/tru...