Counterfactual Identification Under Monotonicity Constraints

Aurghya Maiti, Drago Plecko, Elias Bareinboim

Introduction

In causal inference, counterfactuals represent the highest level of Judea Pearl’s "Causal Hierarchy." They allow researchers to answer "what-if" questions: given an observed outcome, what would have happened under a different treatment? While interventional queries (Level 2) are often identifiable from data, counterfactual queries (Level 3) are typically non-identifiable in non-parametric settings without restrictive global assumptions like linearity.

This paper introduces M-ID, an algorithmic framework that bridges this gap. By leveraging local monotonicity constraints—the assumption that a causal relationship only moves in one direction—the authors demonstrate that many previously unidentifiable counterfactuals can be uniquely computed from observational and experimental data.

Bridging the Identification Gap

Traditional causal analysis often forces a choice between completely non-parametric models (which offer few guarantees) and rigid linear models. This research targets the "middle ground" of the identification spectrum. By adding qualitative shape assumptions, such as "treatment always helps or does nothing, but never hurts," researchers can unlock Layer 3 insights without sacrificing the flexibility of non-parametric modeling.

The identification spectrum and various causal graph settings (IV, LATE+, Fairness) where monotonicity constraints enable counterfactual identification.

The M-ID Framework

The core of the M-ID approach relies on two theoretical innovations:

  1. Causal Graph with Monotonicity Assumptions (CGMA): Standard causal diagrams are enhanced with labels indicating positive or negative monotonic relationships between variables.
  2. Monotonicity Reduction Lemma (MRL): This provides formal rules—the Simplification Rule and the Difference Rule—to decompose complex counterfactual expressions into simpler interventional quantities.

The M-ID algorithm systematically applies these rules to reduce counterfactual queries into Layer 2 quantities, which are then solvable via standard causal calculus (do-calculus). This allows for conditioning on post-treatment variables, like mediators, without the typical risk of selection bias.

Algorithmic Identification of Complex Effects

The framework is particularly powerful for identifying effects in complex structures where traditional estimators fail. For instance, in settings involving Local Natural Direct Effects (LNDE), monotonicity allows for the resolution of logical inconsistencies that usually prevent identification.

Causal graphs for LNDE and LNDE+ settings, illustrating the complex structures handled by the M-ID algorithm.

Empirical Validation

The authors validated M-ID using both real-world and synthetic data:

  • 401(k) Dataset: In analyzing the impact of 401(k) eligibility on wealth, M-ID remained accurate in complex graph structures where classic Local Average Treatment Effect (LATE) estimators produced biased or sign-reversed results.
  • Fairness Scenarios: Using synthetic models of hiring bias (Sex → Education → Hiring), the authors proved that quantities like the Probability of Necessity (PN) and Probability of Sufficiency (PS)—usually only boundable by ranges—become uniquely identifiable under local monotonicity.

Conclusion and Future Work

M-ID demonstrates that global parametric assumptions are not always necessary for counterfactual identification. By identifying specific "bottleneck" edges that follow monotonic patterns, researchers can compute complex causal effects in arbitrary graphs. Future work aims to extend these reduction rules to purely continuous variables and explore other shape constraints like convexity or concavity.

Summary updated

BibTeX
@inproceedings{maiti2025counterfactual,
  author = {Aurghya Maiti and Drago Plecko and Elias Bareinboim},
  title = {{Counterfactual Identification Under Monotonicity Constraints}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i25.34888},
}
Figure