Yeah, the double accounting question can be a problem. It is inherent to counterfactual impact. Imagine a production chain X → Y → Product. Then counterfactually, X can call 100% dips of the product; as can Y. So together, they have 200%, which does not make sense.
However, there are alternative impact metrics. For example, Shapley values have some nice properties. In particular, they guarantee that they sum up to one. Intuitively, they calculate the mean counterfactual impact for each player over all possible configurations of players. This can be useful to assess important predictors in statistical modles. But it is also the reason why I don’t find them partucularly useful for decision making. After all, you are not interested in your impact in hypothetical worlds, but just in your impact in the current constellation of the world, i.e. your counterfactual impact.
So in summary, I’d say use counterfactuals for decision making and Shapley values for determining bragging rights ;)
Yeah, the double accounting question can be a problem. It is inherent to counterfactual impact. Imagine a production chain X → Y → Product. Then counterfactually, X can call 100% dips of the product; as can Y. So together, they have 200%, which does not make sense.
However, there are alternative impact metrics. For example, Shapley values have some nice properties. In particular, they guarantee that they sum up to one. Intuitively, they calculate the mean counterfactual impact for each player over all possible configurations of players. This can be useful to assess important predictors in statistical modles. But it is also the reason why I don’t find them partucularly useful for decision making. After all, you are not interested in your impact in hypothetical worlds, but just in your impact in the current constellation of the world, i.e. your counterfactual impact.
So in summary, I’d say use counterfactuals for decision making and Shapley values for determining bragging rights ;)