RT @fchollet: We're building a world-class research team. If you're excited about our ideas and you'd like to join us, let's chat! We're a…

@Chanabassarah these results suggest that @yudapearl is not shadowbanned, right? Except for 2 of the tests, which couldn't be run due to "technical reasons".

The point was that NNT, as defined, should be 1/PNS. If you're interested in 1/ATE, it should be defined differently. Here's a scenario where 1/PNS is useful. Imagine being harmed is not that problematic. Maybe a migraine goes away without treatment and remains with treatment. But the remaining migraine is very mild. Now, 1/ATE may give us a very large NNT, whereas we really only need to treat a few people to prevent 1 person's migraine.

@soboleffspaces Good luck! I’d love to see the talk and read the paper if and when they’re available.

@soboleffspaces @rafaeljmsvieira you may enjoy our paper on monotonicity, which explores probability of benefit and associated probabilities of causation:

New paper by Kerwash & Johnston ( shows Casgevy as a very effective treatment for transfusion-dep β-thalassemia, but it was a single-arm trial. They avoided issues with conv statistical analysis through Causes of Effects analysis with PN=100%, PS=PNS=93%.

RT @yudapearl: Readers ask whether we, Daniel Pearl's family, were consulted, or whether KSM's murder of Daniel Pearl was a consideration i…

A more accurate phrase might be: #causality cannot be guaranteed from data analysis alone, instead it can be assumed or demonstrated by argument outside the statistical analysis. Casual discovery is a branch of causality addressing this issue:

If you're asking if dropping the mussel from a height breaks its shell is a counterfactual statement, I think it is since breaks implies not dropping the mussel doesn't break its shell. But if you're asking if this is a good test for counterfactual reasoning, I don't think it is.

But did the crow's decision to drop the mussel and break its shell require counterfactual reasoning?

Had the pleasure and honor of spending time with @AleksanderMolak at @Conf_CLeaR 2024. I can't promise that I'm at all interesting, but he also interviewed me for the Causal Bandits Podcast:

@stevekrouse some startups allow semicolons?

@eyeslasho The parenthetical statements might be true, but not at all because of that graph or the statements that came before them. “Produce” and “Earn” are causal words. “Correlate” and “predict” are statistical words. Wealthier & healthier environment → intelligence of parents & kids.

@soboleffspaces Ilya Shpitser presented that quote at the end of his presentation to express the benefits of having multiple styles of CI (PO, SCM). "Advocating a single paradigm approach to causal inference is, inherently, intellectually impoverishing." I'm not convinced regardless of the quote

@nathanb12345 @yudapearl Correct, that substitution doesn't directly work. I think I understand what you're getting at. The linear program that yielded the upper bound of Eq. 11.41 should collapse to a point if it included inputs p₀₁₁ and p₀₁₀, or P(yₓ', y'ₓ), but it doesn't include those.

@soboleffspaces @yudapearl Interesting diagram. I assume P¹ is P(yₓ) and P⁰ is P(yₓ'). Then the top left pink area minus the bottom right pink area is P¹-P⁰=ATE=PNS-P(harm). If we knew P(harm)=0, then ATE=PNS. But we usually can't know P(harm)=0 unless we have information beyond what Eq. 11.41 uses.