Ya 100%! Pre-registration is not a straightjacket that prevents you from doing other analyses. It is just a tool for increasing transparency so that we can know what you planned ahead of time vs what you found afterward

It's also basically impossible from a practical perspective bc there's no way to have a human expert available on the fly to answer any possible conspiracy theory a participant describes. I'm sure somebody will do something like this at some point, but it's not the direction I'm pursuing

We looked at what happens if you label the AI as an expert vs an AI (doesn't make it any less persuasive to call it an expert). But havent tried actual humans - almost certainly humans will do much worse bc they dont have easy access to all the relevant facts etc academic.oup.com/pnasnexus/ar...

For a deeper dive, check out this 20 minute talk youtu.be/qVjjcw4w6-Q and try the bot yourself at www.DebunkBot.com

Hats off to lead author Nat Rabb, joint senior @tomcostello.bsky.social and collabs @gordpennycook.bsky.social @adamberinsky.bsky.social Alexander Levontin

In Jan 2024, yes X was certainly more ideologically diverse than BlueSky

Well, this data is from Jan 2024. Unclear what it would look like today...

Interestingly, it's more about some notable high-quality outlets underperforming on engagement, rather than low-quality outlets overperforming:

It's actually more about some notable high-quality outlets underperforming, rather than low-quality outlets overperforming:

The paper, in figures:
F1=Cross-platforms Corr b/w partisan lean and quality
F2=Corr between lean and engagement varies cross-platform, s/t dominant-lean news gets more engagement
F3=Neg corr between quality and engagement across all platforms!
Led by @mmosleh.bsky.social w @jennyallen.bsky.social

I'm trying to understand what this means and I haven't succeeded yet. Any chance you'd want to break to down more for us bskyers?

For me, the point of Multiverse analysis is when there are multiple ways of analyzing the data that all seem reasonable. So it's not that you just throw in any possible model and drown out the reasonable models in the specified ones, but rather that you show robustness among reasonable models