Abstract
We address the problem of estimating the effect of intervening on a set of variables $X$ from experiments on a different set, $Z$, that is more accessible to manipulation. This problem, which we call z-identifiability, reduces to ordinary identifiability when $Z = \emptyset$ and, like the latter, can be given syntactic characterization using the do-calculus. We provide a graphical necessary and sufficient condition for z-identifiability for arbitrary sets $X, Z,$ and $Y$ (the outcomes). We further develop a complete algorithm for computing the causal effect of $X$ on $Y$ using information provided by experiments on $Z$. Finally, we use our results to prove completeness of do-calculus relative to z-identifiability, a result that does not follow from completeness relative to ordinary identifiability.
Introduction: Beyond Passive Observation
In causal inference, the standard identifiability (ID) problem asks whether a causal effect $P(y|do(x))$ can be computed from passive observational data and a causal graph. Often, hidden confounders make this impossible. However, researchers frequently have access to auxiliary experimental data on a surrogate set of variables $Z$ that are easier or more ethical to manipulate than the target variables $X$.
The authors formalize this as z-identifiability (zID): determining if the causal effect of $X$ on $Y$ is computable given observational data, the causal graph, and experimental data on $Z$. This framework allows researchers to "rescue" unidentifiable queries by leveraging more accessible experiments.
Graphical Conditions and the IDz Algorithm
To solve the zID problem, the paper introduces a necessary and sufficient graphical condition (Theorem 3). A key insight is the "Intercept Rule": for $Z$ to be a valid surrogate, $X$ must intercept all directed paths from $Z$ to $Y$. This ensures that the surrogate experiments do not introduce "side effects" that bypass the target intervention.
The authors developed the IDz algorithm, a complete procedure that extends the standard ID algorithm. IDz employs a divide-and-conquer strategy, decomposing the graph into confounded components (C-components). It tracks available $Z$-interventions and uses them to simplify the graph recursively. If the algorithm encounters a "hedge"—a specific graphical structure that $Z$-experiments cannot resolve—it correctly identifies the effect as non-z-identifiable.
Completeness of the do-calculus
A major theoretical contribution of this work is proving the completeness of Pearl’s do-calculus for z-identifiability. While do-calculus was already known to be complete for ordinary identifiability, this paper shows it is also sufficient for any case where surrogate experiments can identify a causal effect. This means if a solution exists, a finite sequence of do-calculus rules is guaranteed to derive it.
Conclusion and Future Directions
This framework provides a rigorous foundation for using surrogate experiments to resolve otherwise unidentifiable causal queries. While currently limited to non-parametric models, future work aims to extend these results to linear models and "meta-synthesis" tasks, such as combining data from heterogeneous sources or handling transportability across different environments.
Summary updated