Inequality Constraints in Causal Models with Hidden Variables

Changsung Kang, Jin Tian

Abstract

This paper presents a systematic framework for deriving inequality constraints in causal models containing unmeasured variables. By leveraging Causal Bayesian Networks and the concept of c-components, the authors develop the FindIneqs algorithm to identify mathematical signatures that falsify incorrect causal structures and bound unobserved causal effects.

Introduction

In causal inference, Directed Acyclic Graphs (DAGs) are standard tools for representing variable dependencies. However, real-world scenarios often involve "hidden variables"—latent confounders that influence multiple observed variables. While conditional independence serves as a test for fully observed models, models with hidden variables often lack such simple relationships. This research addresses the challenge of identifying inequality constraints, which define the valid bounds of a probability distribution under a specific causal structure, enabling researchers to falsify models and estimate unknown effects.

Methodology

The proposed approach focuses on semi-Markovian models where hidden variables lack parents. The framework is built upon several technical pillars:

  1. C-components: The graph is partitioned into confounded components (c-components), which are groups of variables linked by paths of hidden variables. This decomposition allows researchers to break complex global problems into smaller, local ones.
  2. The Truncation Formula: Utilizing do-calculus logic, interventional distributions (representing the effect of forcing a variable to a specific value) are expressed as truncated versions of the original probability distribution.
  3. General Inequality Derivation: The core contribution, Proposition 1, uses the inclusion-exclusion principle to establish that specific combinations of interventional distributions must remain non-negative in a valid model.
  4. The FindIneqs Algorithm: This procedure automates the discovery of these inequalities, filtering them to include only variables that a researcher has actually measured or experimented on.

Case Studies

The authors validate their framework using synthetic causal structures to demonstrate the algorithm's utility and scalability.

The Instrumental Variable (IV) model (Figure 1) serves as a foundational example. The proposed method successfully recovers the classic instrumental inequality (Pearl, 1995) and demonstrates that it is a specific instance of a much broader class of structural constraints.

A complex 5-variable model with overlapping hidden influences.

To test the algorithm's capability on more intricate structures, the authors apply FindIneqs to a 5-variable model (Figure 3). This model features overlapping hidden influences ($U_1, U_2, U_3$), demonstrating how the framework handles complex dependencies that were previously computationally prohibitive for standard quantifier elimination methods.

Results and Discussion

The framework provides a systematic, automated method for constraint discovery with several key advantages:

  • Falsification Power: The derived inequalities are necessary conditions; if observational data violates these bounds, the underlying causal model is definitively incorrect.
  • Bounding Unobserved Effects: These constraints allow for the calculation of lower and upper bounds for causal effects that have not been directly tested in experiments, using only observational data and unrelated experimental results.
  • Computational Efficiency: While the complexity is $3^{2^m}$ (where $m$ is the size of the largest c-component), it represents a significant improvement over previous methods that were often doubly exponential and failed on simple 3-variable models.
  • State-Independence: Unlike methods restricted to binary variables, these constraints are purely structural and hold regardless of the number of states for each variable.

Conclusion and Future Work

While the current set of inequalities is not exhaustive (incomplete), it provides a powerful tool for causal discovery and model validation. Future research will focus on improving scalability for dense graphs and developing robust statistical tests to distinguish between model violations and sampling noise in finite datasets.

Summary updated

BibTeX
@misc{kang2012inequality,
  author = {Changsung Kang and Jin Tian},
  title = {{Inequality Constraints in Causal Models with Hidden Variables}},
  year = {2012},
  eprint = {1206.6829},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1206.6829},
}
Figure