Meta-Transportability of Causal Effects: A Formal Approach

Elias Bareinboim, Judea PearlAISTATS 2013 · OpenReview

Abstract

This paper considers the problem of transferring experimental findings learned from multiple heterogeneous domains to a different environment, in which only passive observations can be collected. Pearl and Bareinboim (2011) established a complete characterization for such transfer between two domains, a source and a target, and this paper generalizes their results to multiple heterogeneous domains. It establishes a necessary and sufficient condition for deciding when effects in the target domain are estimable from both statistical and causal information transferred from the experiments in the source domains. The paper further provides a complete algorithm for computing the transport formula, that is, a way of fusing observational and experimental information to synthesize an unbiased estimate of the desired effects.

Introduction

In scientific research, it is rarely possible to conduct a perfect experiment on the exact population of interest. Instead, trials are often performed in specific settings—such as a controlled laboratory environment—with the hope of "transporting" those findings to a broader target population. However, traditional meta-analysis and external validity guidelines often lack a formal mathematical language to handle the inherent heterogeneity between domains. A target domain might differ from a source domain in age distribution, genetics, or environmental factors.

Previous work by Pearl and Bareinboim (2011) addressed transportability between a single source and a single target. This paper tackles the more complex meta-transportability problem: how can we estimate a causal effect when no single source study is sufficient, but a combination of several disparate studies—each differing from the target in different ways—could provide the answer?

Methods and Selection Diagrams

The authors leverage Causal Diagrams and Selection Diagrams to represent the problem formally. Selection diagrams are causal graphs augmented with special "S-nodes" (represented as square nodes). An S-node pointing to a variable $V$ indicates that the mechanism determining $V$ differs between the source and the target domains.

Selection diagrams representing causal mechanisms and domain differences (S-nodes). Diagram (b) illustrates the 's-bow arc', a fundamental non-transportable structure.

The core strategy involves:

  • Do-calculus: Utilizing Judea Pearl’s mathematical language for interventions to manipulate causal expressions.
  • Decomposition: Breaking a complex causal query into smaller sub-relations. If each sub-relation is "directly transportable" from at least one source, the entire effect can be reconstructed.
  • The $\mu sID$ Algorithm: A recursive algorithm that systematically attempts to decompose the target query. It checks if components of the causal chain can be "borrowed" from different sources and combined with observational data from the target.

Key Insights: Synergy of Sources

The most striking insight is the "synergy of sources"—the idea that causal effects can be "stitched together." A source domain $A$ might be ineffective for understanding a drug's direct effect but excellent for understanding a specific biomarker. Conversely, source $B$ might provide the link between that biomarker and the final outcome. Meta-transportability allows researchers to use $A$ for the biomarker and $B$ for the outcome to solve the target problem.

Multiple selection diagrams for the same causal chain, demonstrating how different source domains can be combined to achieve meta-transportability.

The authors also identify the $\mu s$-hedge, a specific graphical structure that serves as a mathematical "certificate" of impossibility. If this structure exists within the collection of diagrams, it proves that the causal effect cannot be estimated from the available data, regardless of sample size.

Key Results and Conclusion

The paper provides the first complete algorithmic treatment for synthesizing experimental results from multiple heterogeneous sources. It establishes the necessary and sufficient graphical conditions for meta-transportability and proves that the rules of do-calculus are sufficient for establishing these conditions.

The algorithm outputs a specific transport formula—a mathematical instruction for weighting and combining various data streams to produce an unbiased estimate. Because the algorithm is complete, if it fails to find a transport formula, it is mathematically certain that no such formula exists under the given assumptions. This work reinforces the power of the do-calculus framework and provides a rigorous foundation for future research into partial transportability and computational complexity in causal inference.

Summary updated

Abstract

This paper considers the problem of transferring experimental findings learned from multiple heterogeneous domains to a different environment, in which only passive observations can be collected. Pearl and Bareinboim (2011) established a complete characterization for such transfer between two domains, a source and a target, and this paper generalizes their results to multiple heterogeneous domains. It establishes a necessary and sufficient condition for deciding when effects in the target domain are estimable from both statistical and causal information transferred from the experiments in the source domains. The paper further provides a complete algorithm for computing the transport formula, that is, a way of fusing observational and experimental information to synthesize an unbiased estimate of the desired effects.
BibTeX
@inproceedings{bareinboim2013meta,
  author = {Elias Bareinboim and Judea Pearl},
  title = {{Meta-Transportability of Causal Effects: A Formal Approach}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2013},
  url = {https://openreview.net/forum?id=R6DByORCViS},
}
Figure