The 4 part individual effect of an intervention proposed by the causal inference community is based on counterfactuals. For example, if a group of 10 people are treated and 6 survive and then we go back in time and don’t treat, 4 survive. However 2 individuals would have survived with or without treatment (always survivors), 4 would have survived with treatment but not without (benefited), 2 would have survived without treatment but not with treatment (harmed) and 2 would not have survived with or without treatment (never survivors).
In order to discover what happened to each individual above we would need a Time Machine to treat, go back in time and not treat and then compare what happened to each individual. However Pearl & Muller calculated the above proportions (but not what happened to each individual) using various inequalities from a combination of RCTs and observational studies.
There is also a question of stochastic processes. In the messy real world if the above counterfactual study was repeated a few days later the above 2 individuals ‘harmed’ in the first study might appear in the benefit group the second time and 2 of those in the ‘benefit’ group in the first study might appear in the ‘harm’ group during the second study. The overall proportions of 6/10 and 4/10 would stay the same suggesting that the treatment was beneficial on the whole. Individuals from all 4 groups would probably jump around leaving the overall proportions the same.
The problem is that Pearl and Muller don’t explain how knowing the above 4 proportions changes the decision of how to advise an individual when making a decision about whether to accept or decline a treatment. @Stephen and @phildawid have written a paper recently explaining why "the approach is dangerously misguided and should not be used in practice” https://arxiv.org/pdf/2301.11976.pdf. I agree that the 4 proportions are of theoretical interest only and have no place in practical decisions including those made using established decision theory.
In my latest post 220 Individual response - #224 by HuwLlewelyn I suggest that the 4 proportions (for what they are worth) can be arrived at by using traditional diagnostic reasoning from RCT results alone using covariants (e.g. those that represent disease severity or other information such as genetic markers). Observational studies are not necessary. Also the 4 proportions provide less information than that of diagnostic reasoning as explained in my ‘P Maps’.
The only way that I can envision individuals really being harmed and also benefiting from a single treatment are via two different causal mechanisms. For example a drug might benefit by killing cancer cells but harm by killing bone marrow cells. You would then have 2x4 theoretical proportions, 4 for each of the 2 causal mechanisms for what they are worth.