Causal inference in degenerate systems: An impossibility result

Yue Wang, Linbo WangAISTATS 2020

Introduction: The Challenge of Quantifying Causal Strength

Causal Directed Acyclic Graphs (DAGs) are powerful tools for identifying that one variable causes another. However, a significant challenge arises when researchers attempt to quantify how much influence exists. While metrics like Conditional Mutual Information (CMI) are effective in simple systems, they often fail in "degenerate" systems—environments where variables are redundant or perfectly correlated.

Defining Degeneracy through Markov Boundaries

The authors define a degenerate system as one where a variable possesses multiple Markov boundaries. A Markov boundary of a variable $Y$ is the minimal set of variables that, when known, makes $Y$ independent of all other variables in the system. In degenerate cases, different sets of variables provide identical information about the target, making it impossible to distinguish their individual causal contributions.

A causal DAG illustrating a potential degenerate system where $Z$ and $X$ may provide redundant information about $Y$.

A classic example is the "confounder triangle" where $Z \to X \to Y$ and $Z \to Y$. If $Z$ and $X$ are identical or perfectly correlated, determining whether the effect on $Y$ originates from $X$ or $Z$ becomes mathematically ambiguous. Most current measures either return nonsensical values or are undefined in these scenarios.

The Impossibility Theorem

The core contribution of this work is a formal "impossibility result." The authors propose three criteria (C1-C3) for a robust causal measure:

  • Continuity (C1): The measure should be stable under small data perturbations.
  • Faithfulness to Uniqueness (C2): If a variable is not in the unique Markov boundary, its causal strength should be zero.
  • Positivity (C3): If a variable is in the unique Markov boundary, its causal strength should be non-zero.

The paper proves that in degenerate systems, no identifiable measure can satisfy all three criteria simultaneously. The authors liken this behavior to Picard’s Great Theorem in complex analysis: at the point of degeneracy, the measure becomes "essentially singular," meaning it could take on almost any value depending on how the limit is approached.

Detecting Degeneracy: Algorithm 2-AF

Given that quantifying effect size is unreliable in degenerate systems, the authors focus on detection. They introduce Algorithm 2-AF (Assumption-Free), which tests for the uniqueness of a Markov boundary. By systematically removing variables and checking if a valid alternative boundary can still be formed, the algorithm identifies whether a system is degenerate.

Experimental Results and Real-World Prevalence

The authors validated their approach using synthetic Bernoulli-distributed datasets across various settings, including "pathological" cases where standard assumptions like the composition property fail.

Success rate of Algorithm 2-AF compared to baseline methods across varying sample sizes in a controlled simulation.

In simulations, Algorithm 2-AF maintained high success rates (near 1.0) as sample sizes increased, significantly outperforming existing methods like KIAMB, which often failed to converge on the correct answer in complex settings. Furthermore, a meta-analysis of 13 benchmark datasets (ranging from breast cancer studies to protein signaling) revealed that multiple Markov boundaries are common in high-dimensional real-world data, making degeneracy a feature rather than a bug of complex systems.

Conclusion

The impossibility result indicates that in degenerate systems, researchers should shift from seeking a single quantitative "strength" to providing a qualitative summary of all possible causal explanations (all Markov boundaries). While computational costs remain a factor for high-dimensional data, identifying degeneracy is a necessary step for robust and honest causal inference.

Summary updated

BibTeX
@inproceedings{wang2020causal,
  author = {Yue Wang and Linbo Wang},
  title = {{Causal inference in degenerate systems: An impossibility result}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2020},
}
Figure