acceptodds
Withdrawn from ICLR 2027

Constraint-Based Framework for Bounding Probabilities of Causation under Partial Structural Assumptions and Auxiliary Statistical Information

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Probabilities of causation (PoCs) are fundamental to individual level explanation and decision making, yet they are inherently counterfactual and not point identifiable from data in general. Existing bounds either disregard available auxiliary information, require complete structural assumptions and auxiliary statistical information, or rely on restrictive binary settings, limiting their practical use. In real world applications, structural assumptions and auxiliary statistical information are often incomplete but still informative. This paper proposes a unified constraint based framework for bounding PoCs using partial structural assumptions and partial auxiliary statistical information. We show how partial structural assumptions and partial auxiliary statistical information can be systematically incorporated as constraints in an optimization formulation, even when such information is available only for subsets of variables or covariate values. The resulting framework yields tighter and formally valid bounds without full identifiability and naturally accommodates multiple sources of incomplete structural and statistical information within a single optimization program. This approach extends the applicability of PoCs to realistic settings where both structural assumptions and auxiliary statistical information are incomplete.

Primary area:
causal reasoning
Keywords:
Probabilities of causation, Structural Causal Models, Counterfactuals

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_humPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper aims to extend existing results for bounding probabilities of causation (PoCs), from the setting of binary treatments and outcomes, to non-binary treatments and outcomes, and to incorporate partial structural assumptions and covariate- or mediator-specific information as additional constraints.

Critical strengths

(S1) Probabilities of causation are of fundamental interest in many areas of science, and providing good bounds on them can have practical utility.

Critical weaknesses

(W1) The theoretical undertaking is modest.

Existing results show how to bound PoCs using linear programs for binary treatments and outcomes. Extending this to the setting of multi-valued, non-binary data, even though technical, does not seem to be a major contribution. Similarly for the covariate-specific approaches, compared to existing work.

In summary, it is unclear whether the paper proposes fundamental new ideas in the area.

(W2) Competing approaches are not really engaged with properly.

Previous literature introduces and discusses tools for generic bounding from combinations of observations and experiments [1]. Even though cited, this work is not really engaged with properly: what is the benefit of the proposed method compared to such approaches, if any? This really needs to be clarified, because otherwise the direction of research is not well justified.

[1] Zhang, Junzhe, Jin Tian, and Elias Bareinboim. "Partial counterfactual identification from observational and experimental data." International conference on machine learning. PMLR, 2022.

(W3) Lack of rigor in the paper.

The basic PN, PS, and PNS quantities are introduced, and given ample space. However, the key object of interest in the paper, PNS$(k)$, is simply cited from another paper, and appears only inline without a formal definition or interpretation. This is highly unusual, to not even write down the definition of the key object being investigated.

(W4) Notation could be much improved.

The choice of notation in the paper is really suboptimal. Theorems 2 & 4, Corollary 3 have massive subscripts in the summation, which makes the reading very difficult. The notation could be heavily compressed and improved, to foster readability.

(W5) The space in the paper is utilized modestly.

The paper has 9 pages, but these pages are utilized modestly. The expanded notation eats away huge amounts of space; preliminary work and known bounds are given a lot of space as well, whereas some of the load-bearing, key concepts for the paper are deferred to previous work.

(W6) Theorem 4 may have issues.

Firstly, saying that ZZ satisfies the backdoor criterion, and I assume this is with respect to the pair (X,Y)(X, Y). However, if so, what about a bidirected arrow between W and Y, which makes Yx⊥Wx’∣Wx,ZY_x \perp W_{x’} \mid W_x, Z untrue, since the bidirected W <--> Y can never be closed. This seems like it invalidates the theorem.

(W7) Experiments do not test the main claim.

The paper describes multi-valued X,YX, Y, but focuses on binary data, a setting already handled in previous work. This makes the experiments less convincing.

Questions for authors

See weaknesses.

Reviewer_krJyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a framework for bounding probabilities of causation (PoCs) from observational and experimental data. The proposed framework accommodates partial structural assumptions and partial auxiliary statistical information, even when these are available only for subsets of variables or covariate values. It formulates the bounding problem into a constrained optimization for two specific settings of NON-DESCENDANT COVARIATES and BACKDOOR SET WITH MEDIATORS. The paper validates the framework through simulation experiments, reporting that incorporating additional structural and statistical information consistently produces tighter bounds than existing baselines Tian–Pearl and Mueller–Li–Pearl bounds.

Critical strengths

The proposed framework subsumes existing Tian–Pearl bounds and the Mueller–Li–Pearl bounds as special cases.

Critical weaknesses

My main concern is the novelty and significance of the proposed framework. It looks to me that the paper over-claims its methodological novelty and the generality of its formulation. The general idea of formulating bounding as a constrained optimization problem is well established, and the methodology in Section 3.1 appears to be a relatively direct extension of Mueller et al. (2022). The paper does not explain why in the setting of Section 3.2, the model parameters should be joint probabilities P(Yx1,…,Wx1,…)P(Y_{x_1},\ldots,W_{x_1},\ldots). Why is this representation sufficient, and how does it accommodate other relevant counterfactuals, such as YwY_w or YxwY_{xw}? Without a clear justification, I cannot determine whether the formulation is capable of capturing all relevant observed constraints and counterfactual relationships reflected by the causal diagram. It looks to me, generalizing from the particular settings studied in the paper to arbitrary settings with partial structural information is a non-trivial problem, and the paper does not convincingly establish that its framework achieves this generality.

The problem formulation also lacks precision. The theorems state that a partial causal diagram G is given, but it is unclear how the information encoded in G is translated into optimization constraints, or whether the results rely on particular graph structures. Why specifying constraints (3) and (4) over every subset Z⊆Z1,…,ZmZ\subseteq {Z_1,\ldots,Z_m}, rather than over the full set jointly? The conditional independence assumptions introduced in Lines 288–289 come out of nowhere. Where do they come from?

Questions for authors

What are the computational complexities of the optimization programs in Sections 3.1 and 3.2?

For the parameterization in Section 3.2, why is the chosen joint distribution over counterfactual variables sufficient to represent the relevant constraints?

Reviewer_VgGQOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a unified constraint-based optimization framework for bounding PoCs under partial structural assumptions and auxiliary statistical information. The paper develops linear programming formulations for combining information from multiple non-descendant covariates. It further introduces a nonlinear optimization formulation that incorporates the backdoor variable and mediators.

Critical strengths
  • The paper aims to solve an interesting and important problem. The proposed method is applicable to general discrete variables. Combining information from multiple covariates without requiring the full joint distribution is useful and can be practical.
Critical weaknesses
  • Although the proposed framework supports multivalued treatments and outcomes, the experiments are restricted to binary settings. The experimental results in Figure 2 show little improvement over the Muller-Li-Pearl baseline.

  • The positioning and contribution of the paper should be stated more clearly. The paper claims that existing methods rely on a restrictive binary setting, which is inaccurate. The multivalued PoCs have already been studied. The paper's contribution should focus on utilizing partial knowledge or an auxiliary variable.

Questions for authors
  • Could the authors please clarify the key technical novelty of the proposed framework relative to existing approaches for bounding counterfactual probabilities?

  • Given that multivalued PoCs have been studied, adding additional marginal information as a constraint seems relatively straightforward. What would be the main nontrivial theoretical contribution of Theorem 2?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.