🤖 AI

Identification of Probabilities of Causation: from Recursive to Closed-Form Bounds

This paper extends Probabilities of Causation to multi-valued treatments and outcomes by introducing equivalence classes and a replaceability principle to derive sound, closed-form bounds that are empirically tighter and computationally simpler than existing recursive methods.

Original authors: Xin Shu, Shuai Wang, Ang Li

Published 2026-08-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xin Shu, Shuai Wang, Ang Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of looking for fingerprints, you are looking for "what if." In the world of science, this is called causality. Usually, we look at what happened: "The patient took the medicine, and they got better." But the real magic happens when we ask counterfactual questions: "If this same patient had taken a different dose, would they have gotten better, worse, or stayed the same?" This is the realm of Probabilities of Causation. It's like trying to see all the different versions of a story that could have happened in parallel universes, just to figure out which one is the "real" cause of an outcome. For a long time, scientists could only solve these mysteries when the choices were simple, like a light switch: "On" or "Off." But the real world is rarely that simple. We have many shades of gray, like choosing between five different flavors of ice cream or four different levels of homework help. Figuring out the cause when there are so many options has been a massive headache for researchers, often requiring complex, slow computer calculations that still didn't give a clear answer.

This paper is like a new, super-smart shortcut for those detectives. The authors, Xin Shu, Shuai Wang, and Ang Li, realized that while the "many-flavors" problem seemed impossible to solve with a simple formula, it actually follows a hidden pattern. They discovered that you don't need to calculate every single possibility from scratch. Instead, you can group these complex scenarios into "families" of similar questions. They introduced a clever trick called equivalence classes, which is like realizing that asking "What if I swapped chocolate for vanilla?" is mathematically the same as asking "What if I swapped strawberry for mint?" once you adjust the numbers correctly. By using this "replaceability principle," they turned a messy, recursive puzzle (where you have to solve a problem to solve another problem) into a clean, closed-form formula. Think of it as swapping a tangled ball of yarn for a straight, smooth line.

The team tested their new formulas using standard data from experiments and observations. They proved that their math is sound, meaning it never gives a wrong answer—it always stays within the true boundaries of what's possible. When they compared their new "straight line" formulas against the old, slow "tangled yarn" methods used by previous researchers (specifically Li and Pearl in 2024), they found something exciting: their new bounds were tighter. In the world of detective work, a "tighter bound" means you can narrow down the list of suspects much more effectively. In their simulations, which involved testing thousands of different scenarios with three to twenty different options, their method consistently squeezed the range of possible answers closer to the truth than the old methods did. For example, in a medical scenario with three treatments, they narrowed the probability of a specific patient response pattern from a wide range of [0.428, 0.588] down to a much more decisive [0.509, 0.588]. This tiny shift is huge because it tells us that a specific outcome is actually more likely than we thought, helping doctors make better decisions.

The paper also showed how this works in real-life stories. In one example, they looked at students and different levels of academic support. The old methods suggested that intense tutoring might be good for everyone on average, but the new, tighter bounds revealed a hidden risk: for a specific group of students who are already doing okay, too much help might actually make them fail. The new math caught this nuance where the old math was too fuzzy to see. While the authors are very confident their formulas work and are mathematically proven to be safe, they admit that proving these formulas are the absolute tightest possible for every single dimension is still a work in progress. They suspect it's true, but the final, formal proof for all cases is still open. However, for now, they have handed us a powerful, simpler tool that turns a confusing maze of "what ifs" into a clear path, helping us make smarter, more personalized decisions in medicine, education, and beyond.

Technical Summary: Identification of Probabilities of Causation: from Recursive to Closed-Form Bounds

Problem Statement
Probabilities of Causation (PoCs) are fundamental for counterfactual analysis and personalized decision-making, enabling the attribution of outcomes to specific treatments. While foundational work by Pearl, Tian, and others established tight bounds for binary treatments and outcomes (e.g., Probability of Necessity, Sufficiency, and Necessity and Sufficiency), existing analytical results have largely been confined to these binary settings. Real-world applications in healthcare, finance, and education often involve multi-valued treatments (e.g., dosage levels) and outcomes (e.g., severity grades). Recent attempts to address non-binary cases, such as the recursive-form bounds proposed by Li and Pearl [2024], suffer from computational complexity and a lack of sharpness (tightness), particularly in high-dimensional settings. Numerical methods based on nonlinear programming are also constrained by computational costs. This paper addresses the need for non-recursive, closed-form, and tighter bounds for arbitrary discrete PoCs within Structural Causal Models (SCMs) without imposing structural restrictions.

Methodology
The authors propose a framework that reduces arbitrary discrete PoC queries to a representative family of conjunctions, for which they derive closed-form bounds using standard experimental (interventional) and observational distributions.

Simplification and Representative Family: The paper simplifies the definitions of complex non-binary PoCs (such as Probability of Necessity and Sufficiency(k), Substitute(k,p), Replacement(k,q), and Necessity(k,p,q)) into concise forms. For instance, the Probability of Necessity and Sufficiency(k) is simplified to $P(y_1x_1, \dots, y_kx_k)$ , where values are ordered.
Equivalence Classes and Replaceability: Two key principles are introduced to handle arbitrary PoC queries:
- Equivalence Classes (Theorem 6): This principle demonstrates that bounds for probabilities involving mismatched cardinalities of treatment and outcome variables (e.g., $|X| \neq |Y|$ ) can be mapped to a standard form by padding with zero-probability events, ensuring the bounds remain invariant.
- Replaceability (Theorem 7): This principle allows for the transfer of bounds across value permutations. If a bound is derived for a specific conjunction (e.g., $P(y_1x_1, y_2x_2)$ ), it can be directly applied to permuted versions (e.g., $P(y_2x_1, y_1x_2)$ ) by substituting the corresponding terms.
Derivation of Closed-Form Bounds: Using Fréchet inequalities and logical decomposition, the authors derive explicit lower and upper bounds for the representative family of PoCs. These bounds rely solely on observable quantities: $P(x, y)$ , $P(y_x)$ , and $P(x)$ .
Verification:
- Soundness: A formal proof is provided establishing that the derived formulas define valid lower and upper bounds for all dimensions.
- Tightness: The tightness of these bounds is empirically verified in low-dimensional settings (specifically where treatment and outcome take 3 or 4 values) by comparing the closed-form results against solutions obtained via Balke's linear programming method. For the general $n$ -dimensional case, the authors conjecture that the bounds are tight but note that a formal proof for all dimensions remains an open problem.

Key Contributions

Closed-Form Bounds: The derivation of non-recursive, closed-form bounds for multi-valued PoCs in general SCMs, utilizing only experimental and observational data.
Reduction Framework: The introduction of equivalence classes and a replaceability principle that reduces arbitrary discrete PoC queries to a small, representative set of conjunctions, facilitating the transfer of bounds across value permutations.
Improved Tightness and Efficiency: Demonstration that the proposed closed-form bounds consistently tighten the recursive bounds previously established by Li and Pearl [2024] while offering significantly lower computational complexity.
Soundness and Empirical Verification: Formal proof of soundness for all dimensions and empirical verification of tightness in low-dimensional cases via linear programming.

Results

Theoretical Bounds: The paper presents specific theorems (Theorems 2–5) providing max/min expressions for $PNS(k)$, $PSub(k,p)$, $PRep(k,q)$, and $PN(k,p,q)$. These expressions involve sums and differences of joint and interventional probabilities.
Simulation Performance: In simulations sweeping dimensions $n \in [3, 20]$ , the proposed bounds consistently narrowed the gap compared to Li and Pearl's recursive method. The average improvement in the lower bound was positive across all dimensions. In specific sample distributions, the bounds were narrowed in approximately 10%–20% of cases.
Case Studies:
- Medical Example: In a scenario with three treatments and three outcomes, the proposed method yielded a lower bound of 0.509 for a specific cross-treatment response pattern, whereas the recursive method yielded 0.428. This tighter bound resolved ambiguity, indicating that a majority of patients exhibit this specific response pattern.
- Educational Example: For a query regarding the risk of intensive intervention on learners who currently pass without support, the proposed method provided a non-vacuous lower bound (0.344), whereas the recursive method resulted in a vacuous bound of [0, 1].

Significance and Claims
The paper claims to advance the theory of non-binary probabilities of causation by providing a toolkit that is both analytically derived and practically computable. The significance lies in the ability to capture "cross-world response patterns" that are invisible to marginal outcome comparisons, thereby enabling risk-aware and heterogeneity-aware treatment selection.

The authors maintain a modest stance regarding the completeness of their results:

They explicitly state that while soundness is proven for all dimensions, the proof of tightness (completeness) for the general $n$ -dimensional case remains open, though they conjecture it holds.
They acknowledge that incorporating covariates and explicit causal graphs may further tighten bounds, identifying this as a future direction.
They note that while monotonicity assumptions yield point identification in some binary cases, non-binary counterparts typically do not, and identifying structural conditions for sharper quantities remains challenging.

Ultimately, the work provides a unified, compact framework for non-binary PoCs that facilitates efficient counterfactual attribution in multi-valued settings, moving beyond the limitations of binary assumptions and recursive computational methods.

Technical Summary: Identification of Probabilities of Causation: from Recursive to Closed-Form Bounds

More like this