The forecast of a hurricane definitely requires causal inference, as factors such as the hurricane's strength, population density, and city size can act as confounding variables, influencing both its projected landfall location and its potential category at landfall. @yudapearl

@yudapearl and I have also had two papers accepted for the AAAI-24 main track, focusing on the non-binary cases of the probabilities of causation and the unit selection problem. Here is the link to the papers:

@yudapearl FYI its the workshop W31 at AAAI24 (today), the presentation will be in Room 214 at 3:10PM, here is the link of the paper,

If you are interested in this opportunity, please send me a brief email with your curriculum vitae (C.V.) and any other relevant information. I look forward to hearing from you! @yudapearl

@soboleffspaces Thank you Boris : )

@yudapearl Thank you, Judea. I couldn't have completed this journey without your wonderful support and encouragement. I hope that our unit selection model and counterfactual reasoning will elevate personal decision-making to a whole new level.

I was fortunate to have had the guidance of Professor Judea Pearl @yudapearl , who led me to the fascinating field of causal inference and taught me how to become a skilled researcher.

@yudapearl Happy birthday, Judea, can't wait to work with you for another 86 years : )

@RWJE_BA @soboleffspaces @artistexyz @yudapearl @VC31415 @causalinf @smueller In my point,if you want bounds of PNS, or benefit function, then here it is the case. If there is causal structures that are sufficient to determine the exp data, then things are easier, otherwise, we got to find a way to obtain expdata as expdata play major role in those bounds

@artistexyz @soboleffspaces @yudapearl @HL327 @MatheusFacure @causalinf @VC31415 @smueller um, here is the link , but i ll email you one

@artistexyz @soboleffspaces @yudapearl @HL327 @MatheusFacure @causalinf @VC31415 @smueller lol, thats true, especially for counterfactual things, sometimes when I looked my old notes, I just couldnt believe i had that....

@artistexyz @soboleffspaces @yudapearl @HL327 @MatheusFacure @causalinf @VC31415 @smueller our new aaai22 paper have an example to partial mediator case, prob gonna help?

@artistexyz @soboleffspaces @yudapearl @HL327 @MatheusFacure @causalinf @VC31415 @smueller Even it is pure mediator, Z could be affect by exogenous variables u_z. Or the population specific characteristics C may affect Z. So Z still can have multi values, as in theory, I summed over all z and z' where z != z'

@artistexyz @soboleffspaces @yudapearl @HL327 @MatheusFacure @causalinf @VC31415 @smueller Just to be clear. We just assumed treat and effect (x and y) to be binary, but not z. The population specific characteristics c and covariates z are not binary.

Another intuition why our unit selection is advanced : our objective func is a linear combination of complier,always-taker,never-taker,and defier, while A/B test is a linear combination of (complier+always-taker)(i.e.,treated) and (always-taker+defier)(i.e.,controlled) @yudapearl

@soboleffspaces @yudapearl @smueller @stephensenn @RWJE_BA @PavlosMsaouel @f2harrell @ChristosArgyrop @AndrewPGrieve @HL327 The only assumption is that z is not affected by X, otherwise, p(y_x|z) is counterfactual rather than causal effects. But when we are using RCT data to obtain the z-PNS bounds, we might need the assumption that the RCT data represent the true distribution of P(y_x|z).

@totteh @smueller @yudapearl @Uber lol, it was you : )