Causal Shapley Is Not Causal Identification

In two earlier notes—The Interpretive Boundary of SHAP: From Model Attribution to Causal Identification and Three Nature Papers, Three Boundaries for SHAP—I emphasized a simple point: SHAP does not, by itself, provide causality.

In actual research, however, we rarely stop at asking whether a variable is important to a model. We want to know whether it is a driver and whether it has an effect on the outcome. This is one reason methods with causal language in their names, such as Causal Shapley, are attractive.

The name creates an easy but dangerous inference: if the method is called causal, perhaps causal identification has already been solved. It has not.

Causal identification asks whether we have sufficient grounds to say that X has a causal effect on Y. It must distinguish that effect from confounding, proxy dependence, selection bias, and other alternative explanations. Its task is effect identification.

Causal Shapley answers a different question. If we accept a specified causal structure, how should an overall effect be allocated among variables? Which contribution is direct, and which is indirect? Its task is effect decomposition.

This distinction explains why Causal Shapley depends on causal assumptions without producing them. It can be defined on a causal graph, a causal ordering, or an intervention semantics, but it does not establish that those commitments are correct.

Causal Shapley should therefore be used with a precise claim. It is not, by itself, a causal-identification tool. It is a method for decomposing effects under explicit causal commitments.