A New Characterization of the Experimental Implications of Causal Bayesian Networks

Jin Tian, Judea Pearl

Introduction

In causal inference, Directed Acyclic Graphs (DAGs) serve as the standard framework for representing causal relationships and predicting the outcomes of interventions. While much research has focused on predicting effects given a known graph, a fundamental inverse problem remained: given a set of experimental probability distributions, how can we determine if they are compatible with any underlying causal DAG?

This paper addresses this gap by providing a complete mathematical characterization of the distributions generated by interventions in a Causal Bayesian Network (CBN). The authors move beyond graph-theoretic definitions to identify the essential properties that experimental data must satisfy to be considered "causal."

The Three Norms of Coherence

The authors define a set $P_*$ as the collection of all possible interventional distributions $P_t(v)$, where $T$ is a set of variables forced to value $t$. They prove that for a model to be Markovian (a DAG without unobserved confounders), the experimental data must satisfy three "norms of coherence":

  1. Effectiveness: Forcing a variable to a specific value ensures that the probability of it taking that value is 1.
  2. Interventional Markov Property: This novel property shows that joint distributions under specific interventions can be decomposed into products of distributions under separate, simpler interventions.
  3. Recursiveness: This property ensures that causal influences are acyclic, preventing feedback loops in the stochastic model.

The authors establish both Soundness (all CBNs obey these rules) and Completeness (any data obeying these rules can be represented by a unique DAG).

The Interventional Markov Property

The paper's most significant contribution is the redefinition of the Markov property. Traditionally viewed as a relationship between a distribution and a graph, the authors redefine it as a relationship between different experiments. Specifically, they show that: $$P_{v \setminus (s_1 \cup s_2)}(s_1, s_2) = P_{v \setminus s_1}(s_1)P_{v \setminus s_2}(s_2)$$ This allows researchers to test for causal structure by comparing results across different experimental conditions without ever needing to see the underlying graph.

Detecting Hidden Variables

In real-world scenarios, we often encounter "Semi-Markovian" models where some variables are unobserved. In these cases, the standard Markov property fails. The authors introduce Property 8, an inequality that serves as a diagnostic tool: $$P_{tr}(s) \ge P_t(r, s) + P_r(t, s) - P(t, r, s)$$ If this inequality is violated, it confirms the presence of unobserved confounders or feedback loops. This provides a mathematical "test" for the existence of hidden variables influencing the system.

Key Results and Structural Learning

The core result of the paper is Theorem 2 (Completeness), which proves that Effectiveness, Markov, and Recursiveness are sufficient to fully characterize the experimental implications of a CBN. This has profound implications for structural learning: violations of these properties can be used to "uncover" hidden variables. If the interventional Markov equality does not hold, the existence of unmeasured confounders is mathematically guaranteed.

Figure 1: A causal graph with an unobserved confounder $U$ affecting $X$ and $Y$.

As shown in Figure 1, the authors use a classic confounding case to demonstrate the power of their approach. By applying the interventional Markov property, they can derive the causal effect of $X$ on $\{Z, Y\}$ in significantly fewer steps than traditional do-calculus manipulations. This demonstrates how these properties can be used as symbolic tools to simplify complex causal derivations in models with hidden variables.

Limitations and Future Outlook

While the paper provides a complete characterization for Markovian models, it only offers necessary properties for semi-Markovian models. Identifying a sufficient set of properties for models with unobserved variables remains a key area for future research. Additionally, the current framework assumes "atomic" interventions and does not yet account for the statistical challenges of estimating these distributions from finite data samples.

Summary updated

Abstract

A New Characterization of the Experimental Implications of Causal Bayesian Networks Jin Tian and Judea Pearl, University of California, Los Angeles We offer a complete characterization of the set of distributions that could be induced by local interventions on variables governed by a causal Bayesian network. We show that such distributions must adhere to three norms of coherence, and we demonstrate the use of these norms as inferential tools in tasks of learning and identification. Testable coherence norms are subsequently derived for networks containing unmeasured variables. This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.
BibTeX
@inproceedings{tian2002new,
  author = {Jin Tian and Judea Pearl},
  title = {{A New Characterization of the Experimental Implications of Causal Bayesian Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2002},
}
Figure