acceptodds
Withdrawn from ICLR 2027

Simultaneous Inference for Proximal Causal Learning with Weak Proxies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Proximal causal learning uses the treatment-inducing proxy and the outcome-inducing proxy through the outcome bridge and the treatment bridge to provide a way to address unobserved confounding in observational studies. This paper studies simultaneous inference for the marginal counterfactual distribution and quantile treatment effect under weak proxies. First, we propose a measure of weak-proxy strength and tightens the population-bias bound for the counterfactual CDF estimate. The baseline Cauchy–Schwarz bound introduces the inverse of weak-proxy strength twice, whereas the proposed bound introduces the inverse of weak-proxy strength only once and is sharp for the bridge estimators, thereby relaxing the restriction on the decay rate of weak-proxy strength. Second, we prove that, even when outcome thresholds vary continuously, the number of distinct estimated outcome bridges is at most one greater than the number of observations in the training sample, thereby establishing a uniform expansion across treatment arms and outcome thresholds. Finally, we prove that the critical value remains valid after bridge estimation, thereby obtaining simultaneous CDF bands, and constructs quantile treatment effect bands through confidence-band inversion. The proposed method provides deterministic band-width bounds and does not require counterfactual density estimation. In the numerical study, the baseline Cauchy–Schwarz bound overestimates the true population bias by a factor of approximately 50, whereas the proposed bound is very close to the true population bias. The quantile treatment effect bands constructed by the proposed method achieve coverage of 99.8%–100.0% while being only approximately 20% wider than the oracle 95% bands, compared with baseline coverage of 68.4%–89.2%.

Primary area:
causal reasoning
Keywords:
Proximal causal inference, Quantile treatment effect, Efficient influence function

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_JmoKOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies simultaneous inference for proximal causal learning with weak proxies and proposes a procedure for constructing confidence bands in this setting. The authors motivate their approach through three challenges and establish asymptotically valid simultaneous confidence bands for counterfactual distributions and quantile treatment effects. The authors also present numerical experiments assessing their method's performance.

Critical strengths

The paper clearly describes three challenges: quantifying bridge estimation errors under weak proxies, keeping these errors small across treatment arms and outcome thresholds, and constructing reliable confidence bands with estimated bridges. It provides targeted solutions to each challenge and establishes theoretical guarantees through detailed derivations.

Critical weaknesses
  1. Insufficient evaluation in Table 1 and sensitivity to proxy strength. Table 1 demonstrates a coverage advantage over the selected baselines. However, the current results demonstrate conservative coverage but do not clearly strike a balance between efficiency and robustness. Furthermore, I would like to see Table 1 extended to show how both coverage and bandwidth change as proxies become progressively weaker.

  2. Incomplete specification of the constructed baselines. The construction and parameter settings of the baselines are insufficiently documented. In particular, the Estimated-density delta method requires a more explicit description of its implementation.

  3. Insufficient explanation of the theoretical results. The main text presents many equations but provides limited explanation, particularly for the main theorems. Clearer discussion of their meaning and implications would make the theoretical results easier to follow.

Questions for authors

The coverage rates in Table 1 reach 99.8%–100%, substantially exceeding the nominal 95% level. What accounts for this conservatism?

Reviewer_uwA7OpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents proximal causal learning, a technique for handling unobserved confounding in observational studies, with reliable statistical inference even when the available proxy variables are "weak." The authors introduce a new population-bias bound that scales much better, depending on weak-proxy strength only once instead of twice.

Critical strengths
  1. Meaningful directions in causality with unobserved confounding,
  2. The theoretical aspects in the article are considered in great detail.
Critical weaknesses

see Qs.

Questions for authors
  • The three theoretical contributions in the article constitute a complete inference framework. There is a theoretical dependency between these three improvements. Has the author evaluated the overall performance degradation in numerical simulations when one of the theoretical improvements is missing? Can you quantify which improvement has the greatest impact on the final inference performance?
  • Why is this method used to solve matrices that are close to singular in formulas (7) and (8)? Has the author compared other correction schemes (such as truncated singular value decomposition )? How is the penalty parameter determined in formula (8)?
  • Section 3.3 used multiple approximations when constructing estimates and confidence bands. Has the author derived upper bounds on the errors caused by these approximations to ensure that the entire theory still holds true?
1 comment

Thanks for the thoughtful response. The authors thoughtfully added substantial supplements and appendices, which effectively address my questions. Furthermore, the authors acknowledged my critique, and I will now carefully review these revisions to reach a final conclusion.

However, I still find the paper and rebuttal somhow AI-slope, which is one of the main concern, so I maintain a negative view, and the noverty is limited.

Reviewer_k3rgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies simultaneous inference for the marginal counterfactual CDF F a ​ (y) and the quantile treatment effect ΔQ(τ) in proximal causal learning, when the treatment bridge q a ​ (Z,X) and outcome bridge h a,y ​ (W,X) are estimated from data under weak proxies. Three results:

(Bias) The population bias of
F ^

a ​ (y) is the bilinear form ( α ^ −α) ⊤ Σ a ​ ( θ ^ −θ) . Using a tested (weak) norm on the treatment side and a strong norm on the outcome side, the leading term becomes O p ​ (d n ​ /(nκ a,n ​ )) instead of the Cauchy–Schwarz O p ​ (d n ​ /(nκ a,n 2 ​ )) , where κ a,n ​ =σ min ​ (Σ a ​ ) . An explicit two-point construction shows the κ −1 is attained, so it cannot be improved for these estimators. (Uniformity) Since
θ ^ (y) is a fixed linear map applied to a step function
γ ^ ​ (y) , the estimated outcome-bridge class has at most N k ​ +1 distinct members, giving a covering-number bound min{N k ​ +1,(1+4/ϵ) d n ​

} and a uniform expansion
n ​ { F ^

a ​ −F a ​ }=n −1/2

i ∑ ​ φ a,y ​ (O i ​ )+o p ​ (1) over a∈{0,1} , y∈Y . (Bands) A Rademacher multiplier bootstrap with cross-fitted
φ ^ ​

(−k) gives simultaneous CDF bands; inversion yields QTE bands with no counterfactual density estimation. Simulations: synthetic binary Z,W,U,A , d n ​ =8 , K=3 , n≤5400 ; three studies covering bias, uniform expansion, and QTE coverage.

Critical strengths

S1. The main technical claim is correct, and I checked the mechanism. I verified (74) (the bias identity, including the cancellation E[ h ^ −h]−E[1{A=a}q a ​ ( h ^ −h)]=0 that relies on the true q a ​ ), (78)–(83), and the sandwich collapse Σ a ⊤ ​ (Σ a ​ Σ a ⊤ ​ +λ) −1 Σ a ​ (Σ a ⊤ ​ Σ a ​ +λ) −1 Σ a ⊤ ​ ≈I , which is exactly why one κ −1 disappears. Eq. (82)'s point — that Σ a ⊤ ​ ( α ^ −α) carries no κ −1 in its sampling term — is the crux and it is right.

S2. The sharpness construction (92)–(94) is real and honestly bounded. A binary W,Z design with κ a,n ​ =n −1/4 /2 yields N k ​ κ 1,n ​ ⋅bias⇝G q ​ G h ​ , a non-degenerate limit. The authors then state explicitly that "this example does not establish optimality of the factor d n ​ when d n ​ grows, or a lower bound over all estimators." That is the correct scope, and it is rare to see it stated.

S3. Theorem 3.1's proof is properly structured. The decomposition (115)–(116) into population bias + estimated-bridge fluctuation is clean; the estimated-bridge term is handled by a conditional local maximal inequality with the multiplier 1−1{A=a} q ^ ​

(−k) absorbed into a reweighted measure (B.3), which avoids a spurious product of the two bridge complexities. The observation that "no estimated outcome bridge complexity multiplies the treatment bridge estimation error" is a genuinely useful structural point.

Critical weaknesses

W1 (must fix). κ a,n ​ is not scale-invariant, so the headline "weaker proxy requirement" is not a well-posed statement. Σ a ​ =E[1{A=a}b Z ​ b W ⊤ ​ ] and κ a,n ​ =σ min ​ (Σ a ​ ) are defined on the raw basis, with no normalization. Rescale b Z ​ →cb Z ​ : then Σ a ​ →cΣ a ​ , κ a,n ​ →cκ a,n ​ , while α→α/c and θ is unchanged. The actual bias ( α ^ −α) ⊤ Σ a ​ ( θ ^ −θ) is invariant, but the bound in (83) improves by a factor c , because the second factor ∥ θ ^ −θ∥ 2 ​ is invariant while its bound
d n ​ /n ​ /κ shrinks. So the condition "κ a,n ​ may decay as fast as n −1/2 " can be satisfied for any proxy strength by rescaling the basis, and the claimed relaxation of the decay-rate restriction is then vacuous. The simulation does standardize the basis ("center and scale each function using equal weights on those eight values", D.2), which implicitly fixes a normalization — but the theory never says so, and Prop. 3.1 / Cor. 3.1 / the sharpness example are all stated in an unnormalized way. Please state the normalization (e.g. b Z ​ ,b W ​ orthonormal in L 2 (P Z∣A=a ​ ) , L 2 (P W∣A=a ​ ) ) and restate the bounds under it.

W2. The delta-method baseline is constructed to fail, and there is no neutral benchmark. All four noise laws in Table 2 — Double Weibull(1/2), Beta(1/4,1/4), Beta(1/4,3/4), χ 1 2 ​ — have an unbounded density at a boundary. Unbounded/singular densities are precisely where kernel density estimation degrades, so the 68.4%–89.2% coverage is not evidence that density-free bands are broadly necessary; it is evidence that they help in the regime the benchmark was chosen to stress. A Gaussian or uniform setting, where the delta method should be near-nominal, is never run. Nor is any tuning of the density estimator reported.

Questions for authors

Q1. Take b Z ​ →2b Z ​ . Does κ a,n ​ double? Does the bound (83) halve? Does the true bias change? If the answers are yes/yes/no, what normalization makes "weak-proxy strength" an intrinsic property of (Z,W) rather than of the basis? This is the single question that determines whether the paper's headline relaxation is substantive.

Q2. What is the proposed band's width when
c ^ is calibrated to give exactly 95% coverage (or when the oracle's critical value is used)? And what is
c ^ / c ^

oracle ​ ? If the ratio is ≈1, the efficiency claim is stronger than reported; if not, the 20% is over-coverage.

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

Related papers

Discussion (0)

Sign in to comment.