Causal Transportability with Limited Experiments

Elias Bareinboim, Judea PearlAAAI 2013 · OpenReview

Abstract

We address the problem of transferring causal knowledge learned in one environment to another, potentially different environment, when only limited experiments may be conducted at the source. This generalizes the treatment of transportability introduced in [Pearl and Bareinboim, 2011; Bareinboim and Pearl, 2012b], which deals with transferring causal information when any experiment can be conducted at the source. Given that it is not always feasible to conduct certain controlled experiments, we consider the decision problem whether experiments on a selected subset Z of variables together with qualitative assumptions encoded in a diagram may render causal effects in the target environment computable from the available data. This problem, which we call z-transportability, reduces to ordinary transportability when Z is all-inclusive, and, like the latter, can be given syntactic characterization using the do-calculus [Pearl, 1995; 2000]. This paper establishes a necessary and sufficient condition for causal effects in the target domain to be estimable from both the non-experimental information available and the limited experimental information transferred from the source. We further provides a complete algorithm for computing the transport formula, that is, a way of fusing experimental and observational information to synthesize an unbiased estimate of the desired causal relation.

The Challenge of Generalizing Causal Knowledge

In science and artificial intelligence, the ability to generalize knowledge across different environments is a fundamental hurdle. Often, researchers conduct experiments in a controlled source domain (like a laboratory or simulation) with the intent of applying those findings to a target domain (the real world). However, differences in populations or environmental mechanisms can invalidate direct application. The framework of transportability was developed to mathematically fuse experimental data from a source with observational data from a target to estimate causal effects.

Traditional transportability research assumes that any randomized controlled trial (RCT) can be performed in the source domain. In practice, this is rarely feasible due to financial, ethical, or technical constraints. For instance, while we might want to know the effect of cholesterol levels ($X$) on heart failure ($Y$), we cannot ethically force specific cholesterol levels on patients. We can, however, randomize their diet ($Z$). This paper addresses this gap by formalizing the $z$-transportability problem: determining if a causal effect in a target domain can be estimated when interventions in the source are restricted to a specific subset of variables $Z$.

Selection diagrams illustrating the z-transportability problem where interventions are restricted to set Z.

Methodology: Selection Diagrams and do-calculus

The authors utilize Structural Causal Models (SCMs) and Selection Diagrams to model these cross-domain differences. A selection diagram is a causal graph augmented with $S$-nodes (selection variables) that point to mechanisms where the source and target environments differ.

To solve the $z$-transportability problem, the research employs three key pillars:

  1. The do-calculus: A formal logic for manipulating interventional distributions, denoted by the $do()$ operator.
  2. The zs-hedge: A novel topological structure identified in the causal graph. The zs-hedge acts as a graphical "roadblock"; if it exists between the treatment and outcome, the effect cannot be transported using the restricted set of experiments.
  3. The TRz Algorithm: A recursive, divide-and-conquer algorithm that decomposes the causal graph into smaller components (sC-components). It checks if local causal relations are identifiable via the available $Z$-experiments and synthesizes a final "transport formula."
Examples of selection diagrams with varying causal structures between treatment X, outcome Y, and experimental set Z.

Key Results and Theoretical Insights

This work establishes several landmark theoretical results for causal inference:

  • Soundness and Completeness: The TRz algorithm is proven to be both sound and complete. If a valid transport formula exists for a given graph and set of experiments, the algorithm is guaranteed to find it. Conversely, if the algorithm fails, it constitutes a proof that the effect is not transportable.
  • Necessary and Sufficient Conditions: The paper proves that a causal effect is $z$-transportable if and only if the selection diagram does not contain a zs-hedge. This provides a clear graphical criterion for researchers to evaluate the feasibility of their studies.
  • Sufficiency of do-calculus: A major insight is that Pearl’s do-calculus is strictly sufficient for this task. No additional mathematical frameworks are required to derive the transport formula if the effect is indeed transportable.

Conclusion and Future Work

By providing a complete graphical condition and an algorithmic solution, this paper provides a rigorous foundation for transferring knowledge under experimental constraints. While the framework currently relies on the availability of an accurate causal graph (selection diagram), it opens the door for future research into "meta-transportability"—fusing data from multiple heterogeneous sources—and the integration of causal discovery algorithms to infer these diagrams directly from data.

Summary updated

Abstract

We address the problem of transferring causal knowledge learned in one environment to another, potentially different environment, when only limited experiments may be conducted at the source. This generalizes the treatment of transportability introduced in [Pearl and Bareinboim, 2011; Bareinboim and Pearl, 2012b], which deals with transferring causal information when any experiment can be conducted at the source. Given that it is not always feasible to conduct certain controlled experiments, we consider the decision problem whether experiments on a selected subset Z of variables together with qualitative assumptions encoded in a diagram may render causal effects in the target environment computable from the available data. This problem, which we call z-transportability, reduces to ordinary transportability when Z is all-inclusive, and, like the latter, can be given syntactic characterization using the do-calculus [Pearl, 1995; 2000]. This paper establishes a necessary and sufficient condition for causal effects in the target domain to be estimable from both the non-experimental information available and the limited experimental information transferred from the source. We further provides a complete algorithm for computing the transport formula, that is, a way of fusing experimental and observational information to synthesize an unbiased estimate of the desired causal relation.
BibTeX
@inproceedings{bareinboim2013causal,
  author = {Elias Bareinboim and Judea Pearl},
  title = {{Causal Transportability with Limited Experiments}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2013},
  url = {https://openreview.net/forum?id=RJbkLUPTKx2},
}
Figure