Causal Inference by Surrogate Experiments: z-Identifiability

Elias Bareinboim, Judea Pearl

TL;DR

This paper tackles the problem of estimating the causal effect of an intervention when you cannot directly manipulate the target variables, but can manipulate a surrogate set of variables. The authors introduce the concept of "z-identifiability," providing a complete graphical condition and an algorithm to determine if these auxiliary experiments are enough to identify the desired causal effect, thereby extending the completeness of Pearl's do-calculus.

What Problem Does This Paper Address?

In causal inference, the standard "identifiability" (ID) problem asks whether a specific causal effect—denoted as $P(y|do(x))$—can be computed using only passive observational data and the theoretical assumptions encoded in a causal graph. Often, the answer is no; hidden confounders can make the true causal effect impossible to isolate.

However, in the real world, researchers might have the option to run auxiliary experiments on a different set of variables, $Z$, that are easier, cheaper, or more ethical to manipulate than $X$. For example, a researcher might want to know the effect of cholesterol levels ($X$) on heart disease ($Y$). Forcing a specific cholesterol level in patients is impossible, making the direct effect unidentifiable. But what if the researcher can run an experiment randomizing the patients' diets ($Z$)?

The paper formalizes this as the z-identifiability (zID) problem: Can we compute the causal effect of $X$ on $Y$ given observational data, the causal graph, and experimental data on a surrogate set of variables $Z$?

Methods

The authors build on Pearl’s structural causal models and the do-calculus. They extend existing graphical concepts used in ordinary identifiability—specifically "C-components" (confounded components) and "hedges" (graphical structures that prove a causal effect is unidentifiable).

To solve the zID problem, they:

  1. Formulate a graphical condition: They establish a necessary and sufficient graphical condition for z-identifiability (Theorem 3).
  2. Develop the IDz algorithm: They modify the standard ID algorithm to create IDz. This algorithm uses a divide-and-conquer strategy, breaking the causal graph into smaller C-components. Crucially, IDz tracks the available $Z$-interventions and uses them to simplify the graph during its recursive steps.
  3. Prove Completeness: If the algorithm hits a specific "hedge" structure that the $Z$ experiments cannot resolve, it fails, formally proving the effect is non-z-identifiable.

Datasets

Not discussed in the paper. (This is a theoretical and methodological paper focused on non-parametric causal graphs, so no empirical datasets or benchmarks are used).

Key Insights and Ideas

  • Surrogates can rescue unidentifiable queries: Surrogate experiments ($Z$) can make an otherwise unidentifiable causal query computable, but only if they are topologically situated correctly within the causal graph.
  • The "Intercept" Rule: For $Z$ to be a valid surrogate, the intervention target $X$ must intercept all directed paths from the surrogate $Z$ to the outcome $Y$. Intuitively, surrogate experiments should not disturb the causal paths of the variables being analyzed.
  • Completeness of do-calculus: The do-calculus, which was previously proven complete for ordinary identifiability, is shown here to also be complete for z-identifiability. This means if a causal effect can be found using surrogate experiments, there is guaranteed to be a finite sequence of do-calculus rules that will derive it.

Key Results

  • Theorem 3: The authors successfully established a formal, necessary, and sufficient graphical condition for z-identifiability for arbitrary sets of variables $X$, $Y$, and $Z$.
  • A Complete Algorithm: The proposed IDz algorithm is proven to be both sound and complete. It will either return the correct mathematical formula for the causal effect (fusing observational and $Z$-experimental data) or correctly output FAIL if the effect is mathematically impossible to determine from the available information.
  • Theoretical Unification: The paper proves that proofs of non-z-identifiability are at least as hard as proofs for non-identifiability, requiring the construction of two distinct causal models that agree on both the observational distribution and the interventional distributions over $Z$.

Limitations and Future Work

  • Non-parametric limitation: The entire framework is developed in a non-parametric setting. The authors explicitly note that applying these concepts to linear models is an open and significantly harder problem, largely because even ordinary identifiability is not yet fully characterized for linear systems.
  • Future Directions: The authors suggest extending this work to "meta-synthesis" tasks. This involves combining data from multiple heterogeneous sources to increase effective sample sizes, handling "transportability" (transferring causal information between entirely different environments), and controlling for selection bias.

Summary updated

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 = empty and, like the latter, can be given syntactic characterization using the do-calculus [Pearl, 1995; 2000]. 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.
BibTeX
@inproceedings{bareinboim2012causal,
  author = {Elias Bareinboim and Judea Pearl},
  title = {{Causal Inference by Surrogate Experiments: z-Identifiability}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2012},
  url = {https://openreview.net/forum?id=OetzFsDnRNw},
}

Loading related content...

Figure