Causal Conclusions that Flip Repeatedly and Their Justification

Kevin T. Kelly, Conor Mayo-WilsonUAI 2010

TL;DR

This paper proves that any consistent method for inferring causal relationships from observational data will inevitably produce "causal flips" – reversals of inferred causal directions – as more data becomes available. This unavoidable flipping suggests that the goal shouldn't be to eliminate flips, but to minimize them, leading to a new criterion for evaluating causal discovery algorithms.

What Problem Does This Paper Address?

Inferring causal relationships from observational data is a fundamental challenge in many scientific fields. Unlike experimental data, where interventions can directly reveal cause-and-effect, observational data only shows correlations. While consistent statistical methods can eventually converge to the true causal structure, this paper highlights a critical limitation: even when a method is "consistent" (meaning it will eventually find the truth with enough data), the inferred causal directions can repeatedly flip back and forth as the sample size increases. This means that a conclusion reached with high confidence at one sample size might be reversed with equal confidence at a larger sample size, making it difficult to trust any single conclusion. The paper aims to understand this phenomenon of "causal flipping" and propose a new way to justify and improve causal discovery methods.

Methods

The paper introduces the concept of "empirical approximation" to formalize the idea that some theories are "closer" to others in terms of statistical indistinguishability. This is defined using total variation distance between probability measures. A key lemma (Proposition 2) establishes that for any consistent method and any "empirical approximation chain" of theories, the method can be forced to flip between these theories with arbitrarily high probability as sample size increases. This lemma is then used to prove the "Causal Flipping Theorem" (Theorem 2), which states that for dependency-driven causal questions (like edge orientation in linear Gaussian or discrete Bayes networks), any consistent method will produce causal flips. The paper also discusses "retractions in chance," quantifying the magnitude of these flips, and proposes "retraction minimization" as a principle for designing better causal discovery algorithms. Simulations are used to illustrate the flipping behavior of existing algorithms like PC, CPC, and FCI.

Datasets

Not discussed in the paper. The paper focuses on theoretical results and uses simulated data generated from known causal structures for its empirical evaluation.

Key Insights and Ideas

  • Unavoidable Causal Flipping: The core insight is that causal flips are not a bug, but a feature of consistent causal discovery methods when dealing with observational data under certain assumptions (linear Gaussian or discrete Bayes networks).
  • Empirical Approximation as the Root Cause: The paper formalizes why flipping happens: it's due to the existence of chains of statistically indistinguishable theories. As more data arrives, the method can be "fooled" into switching between these indistinguishable theories.
  • Retraction Minimization as a New Goal: Since flips are unavoidable, the paper argues that the best causal discovery methods are those that minimize these flips or "retractions in chance." This shifts the focus from absolute certainty to a more pragmatic goal of stability.
  • Ockham's Razor Connection: Methods that minimize retractions tend to be those that favor simpler models (fewer edges), aligning with Ockham's razor. This provides a theoretical justification for why simpler causal networks are often preferred.

Key Results

  • Causal Flipping Theorem: Proves that for dependency-driven causal questions (e.g., edge orientation in linear Gaussian or discrete Bayes networks), any consistent causal discovery method will produce causal flips any number of times as sample size increases.
  • Quantification of Retractions: Introduces a measure called "retractions in chance" to quantify how often and how severely a method flips its conclusions.
  • Simulation Results: Demonstrates that algorithms like PC and FCI exhibit significant flipping behavior, while CPC and GES, which tend to favor simpler models (adhering more to Ockham's razor), show fewer retractions. For example, figures 3-6 show the percentage of "truth reversed" (flipped conclusion) over increasing sample sizes for different algorithms.

Limitations and Future Work

  • Assumptions: The theoretical results rely on assumptions about the underlying causal structure (linear Gaussian or discrete Bayes networks) and the dependency-driven nature of the causal questions. The paper notes that non-dependency-driven questions might be more common than initially thought.
  • Practical Implications of Flipping: While the theorem proves flipping is possible, the practical impact depends on the magnitude of the flips and the strength of the causal connections. The paper suggests that the "potential for future flips is, essentially, endless" due to unmeasured variables.
  • Future Work: The paper suggests that retraction minimization provides a concrete standard for progress in causal discovery algorithm design. Future work could explore how to best implement and evaluate algorithms based on this principle, and investigate its applicability to a wider range of causal models beyond linear Gaussian and discrete Bayes networks.

Summary updated

BibTeX
@inproceedings{kelly2010causal,
  author = {Kevin T. Kelly and Conor Mayo-Wilson},
  title = {{Causal Conclusions that Flip Repeatedly and Their Justification}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2010},
}
Figure