I do think there’s much more to be said on this topic though, and it probably depends on things like your model of the distribution of impact and various counterfactuals. Shapley values: Better than counterfactuals could also be a useful way to think about these things (in some specific cases, I personally don’t think they’re usually better).
See also Triple counting impact in EA (and comments) on the problem of “double counting”.
I do think there’s much more to be said on this topic though, and it probably depends on things like your model of the distribution of impact and various counterfactuals. Shapley values: Better than counterfactuals could also be a useful way to think about these things (in some specific cases, I personally don’t think they’re usually better).