acceptodds
Withdrawn from ICLR 2027

Probabilities of Causation: Bounds from Incomplete and Subpopulation Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Probabilities of causation (PoCs) play a central role in individual-level explanation and decision making. Tian and Pearl showed that PoCs are in general not point-identifiable and derived tight bounds for three fundamental PoCs using observational and experimental data. However, in real-world applications, such data are often incomplete; for instance, data may only be available for a subpopulation. In this paper, we show how informative bounds on fundamental PoCs can be obtained from incomplete data. We further investigate whether informative bounds on the PoCs of one subpopulation can be derived given data from another subpopulation; that is, we study the transportability of PoC bounds.

Primary area:
causal reasoning
Keywords:
Probabilities of Causation, Incomplete Data, Structural Causal Models

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_T21eOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

Pearl and Tian gave bounds for PN, PS and PNS. But this requires that we have enough information about the population. This submission claims that in the real world, sometimes we can only have information about a subpopulation. Therefore, it raises 2 questions: 1. can we use a subpopulation bound to get a population bound. 2. can we use PoC information about subpopulation 1 to help us infer the PoC of subpopulation 2? Authors discuss these two questions without assuming any cross-subpopulation invariance assumptions.

Critical strengths

This paper shows a worthy concerning problem that in the real world sometimes we can only have the observational/interventional information of a subpopulation, and they give a very realistic example related to adults and teenagers.

Critical weaknesses

Because the authors do not assume any cross-subpopulation invariance assumptions, this means that if we only know the distribution of subpopulation c, the distribution of the other subpopulation c′ still can be arbitrary. Under this setup, most of this paper's theorems (theorems 3-11) are straightforward consequences of existing Tian–Pearl bounds and elementary probability inequalities (Theorems 9-11). They give limited novel theoretical insights. Most of this paper's theorems (theorems 3-11) address two bounding tasks: one is how to use subpopulation data to infer the whole population bound, and the second one is how to use subpopulation 1 to infer subpopulation 2's bound:

  1. subpopulations infer population: in this part, most of the theoretical results can be summarized into this pattern: Take PNS as example, let the two subpopulations have proportions π\pi and 1−π1-\pi, and let their respective Tian–Pearl bounds be [L1,H1][L_1,H_1] and [L2,H2][L_2,H_2]. Then author's theorems mainly tell us: we can take the weighted sum of these two intervals and we can get the whole population bound:
\leq\ \mathrm{PNS} \leq\ \pi H_1+(1-\pi)H_2.$$ Of course this is correct in math, but the theoretical contribution appears limited. 2. subpopulation 1 infer subpopulation 2: The authors also give negative results here: "However, the results show that either no informative bounds can be obtained, or the resulting bounds are equivalent to directly applying the Tian–Pearl bounds to the new subpopulation." (lines 468-470) These negative results are obvious, because the author does not assume any similarity/invariant property between subpopulation 1 and subpopulation 2. In this case of course information of subpopulation 1 can not help with subpopulation 2 by itself.
Questions for authors

What is the reason that authors do not assume any cross-subpopulation invariance assumptions?

Reviewer_dMKcOpenReview ↗
Rating
4
Confidence
1
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The area chair has excused me from reviewing this paper due to a lack of relevant expertise.

Critical strengths

The area chair has excused me from reviewing this paper due to a lack of relevant expertise.

Critical weaknesses

The area chair has excused me from reviewing this paper due to a lack of relevant expertise.

Questions for authors

The area chair has excused me from reviewing this paper due to a lack of relevant expertise.

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

The paper studies bounds on probabilities of causation (PN, PS, PNS) and the effect of treatment on the treated (ETT) when data are available only for part of the population. A binary covariate CC, not a descendant of the treatment, splits the population into subpopulations cc and c′c'. The observational and experimental distributions are fully known for cc, together with P(c)P(c), while the data for c′c' are absent or partial. The authors formulate the problem as a linear program over counterfactual response types in the manner of Balke and Tian–Pearl, and they report closed-form bounds in three cases: no data for c′c', additional observational data for c′c', and additional experimental data for c′c'. They give bounds both for the population quantities and for the c′c'-specific quantities. They conclude that data from cc yield informative population bounds, but provide no information about c′c' beyond what the c′c' data provide on their own.

The paper’s contribution is a clear formulation of the subpopulation data setting and a catalogue of closed-form bounds for it. The bounds themselves follow from applying the Tian–Pearl bounds to each subpopulation separately, so the paper offers no new bounding principle.

The authors illustrate the results with a vaccine example, in which adult data alone imply 0.306≤PNS≤0.6970.306 \le \mathrm{PNS} \le 0.697 for the whole population. They also report simulations over 10,000 compatible distributions under three sampling schemes. In these, the proposed bounds are narrower than the baseline bounds in every reported setting.

Critical strengths
  1. Data restricted to a subpopulation are common in medicine and policy, and the paper states this setting clearly. It enumerates the cases of data availability for four quantities and reports closed-form bounds for each. The vaccine example shows that subpopulation data can still yield an informative lower bound on population PNS. This matters for my assessment because it gives practitioners usable expressions for a common situation.
  2. The bounds can be checked against the linear program, and the vaccine example reproduces exactly. The closed forms require only the reported marginal and conditional probabilities. This matters for my assessment because, whatever their novelty, the results can be used as stated.
Critical weaknesses

The main theoretical results follow from stratum-wise applications of the existing Tian–Pearl bounds. This substantially limits the novelty of the paper’s principal contribution.

For PNS, the population quantity decomposes as PNS=P(c) PNSc+P(c′) PNSc′\mathrm{PNS} = P(c)\,\mathrm{PNS}_c + P(c')\,\mathrm{PNS}_{c'}. Analogous decompositions hold for ETT, PN, and PS, with weights determined by their respective conditioning events. The formulation introduces no assumptions that relate the counterfactual distributions of the two subpopulations. The only constraints that involve both strata are total probability and the denominator constraints (13)–(15), which fix known constants. The linear programs therefore separate into stratum-specific problems, and the population bounds are weighted combinations of the bounds obtained independently for each stratum.

This structure reproduces the main results. In Theorem 3, the lower bound equals P(c)P(c) times the Tian–Pearl lower bound for subpopulation cc, and the upper bound equals P(c′)P(c') plus P(c)P(c) times the corresponding upper bound. The vaccine example illustrates this directly. The adult Tian–Pearl bounds are 0.481≤PNSc≤0.5230.481 \le \mathrm{PNS}_c \le 0.523, so 0.636×0.481=0.3060.636 \times 0.481 = 0.306 and 0.636×0.523+0.364=0.6970.636 \times 0.523 + 0.364 = 0.697, which are exactly the reported population bounds. The same construction reproduces Theorems 4, 5, 8, and 9, and the remaining theorems appear to have the same structure.

This observation also explains the negative transportability results. Without assumptions that link the two subpopulations, information from one cannot constrain the counterfactual distribution of the other. The paper does not introduce the structural relationships that transportability typically requires.

Finally, the simulation baseline is weak for evaluating theoretical novelty. In the absence of data for c′c', the authors compare their bounds with the trivial interval [0,1][0, 1]. A more informative baseline would apply the existing Tian–Pearl bounds separately to the available strata and combine them with the appropriate weights. This construction reproduces the proposed bounds exactly.

These concerns are important because the closed-form bounds constitute the paper’s main claimed contribution. Their derivation and systematic presentation may be useful, but the results follow from established partial-identification arguments rather than from a new bounding principle.

Questions for authors
  1. I verified that Theorems 3, 4, 5, 8, and 9 coincide with P(c)P(c) times the Tian–Pearl bounds for cc, plus P(c′)P(c') times the bounds that the c′c' data allow. Is there any result in the paper that cannot be obtained this way? If so, which one, and what in the linear program produces it?
  2. The negative transportability result follows from the absence of any constraint that links the two strata. Do the authors have results under a linking assumption? Examples would be equal response-type distributions across strata, invariance of P(yx,yx′′∣c)P(y_x, y'_{x'} \mid c) for some subset of response types, or monotonicity in one stratum. Under such assumptions, data from cc could genuinely tighten the bounds for c′c'.
Reviewer_H9UZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates bounds on probabilities of causation when observational and experimental information is incomplete across subpopulations. Building on the framework of Tian and Pearl, the authors derive tight bounds for PNS, PN, PS, and ETT under several levels of data availability. The paper also examines whether information from one subpopulation can sharpen bounds for another. A vaccination example and synthetic simulations are used to illustrate the results.

Critical strengths
  1. Important problem. In many applications, experimental and observational data are not equally available across demographic groups, so partial identification under such limitations is worth studying.
  2. Systematic treatment. The paper considers multiple data availability settings and provides explicit bounds for several fundamental causal quantities.
  3. Useful conceptual distinction. The discussion distinguishes bounding causal quantities for the overall population from transporting causal information to a different subpopulation.
Critical weaknesses

Main Concerns

  1. Notation and readability. The use of x, x′, y, y′, c, and c′ makes some derivations harder to follow, especially in the linear programming formulation. Since the variables are binary, a consistent 0 and 1 notation, perhaps accompanied by a compact table of indices, could make the formulas more intuitive and easier to verify.
  2. Practical motivation and real data evaluation. This is my primary concern. The vaccination example highlights a discrepancy between observational associations and experimental effects, but it is a constructed illustration rather than an application to actual data. The remaining experiments are also synthetic. Since the paper is motivated by limitations in real subpopulation data, I would strongly encourage a real application, ideally one that exhibits such a discrepancy. It would help demonstrate how informative the proposed bounds are in practice and whether they support meaningful conclusions.
  3. Technical novelty and depth of evaluation. The optimization builds on an established linear programming framework, and introducing subgroup constraints appears to be a relatively direct extension. I appreciate the effort required to obtain explicit tight bounds, but the paper could explain more clearly which steps involve new technical insights. The simulations are also fairly straightforward and mostly demonstrate tighter intervals from additional information. A more substantive analysis of when the gains are meaningful, or when the bounds remain too wide, would strengthen the evaluation.
Questions for authors

See Critical Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.