acceptodds
Under review as a conference paper at ICLR 2027

Proximal Balancing for Causal Effect Estimation under Unmeasured Confounding

Abstract

Estimating causal effects from observational data is central to science and policy, but the effects are not identified when confounders are unmeasured. Proximal causal inference addresses this problem with proxies of the unmeasured confounders. However, existing proxy-based approaches either designate proxy roles and solve an inverse problem, which is ill-posed and hard to estimate with high-dimensional proxies, or use a latent-variable model, which assumes that the learned latent variable matches the hidden confounder and leaves bias when it does not. To address these challenges, we introduce proximal balancing. It carries the classical idea of covariate balancing to confounders that are observed only through proxies: it learns a low-dimensional summary of the covariates and proxies that makes the treatment groups comparable, and then adjusts for this summary. It needs no designated proxy roles, inverse problem, or latent model. We give identification theory, finite-sample guarantees, and a practical algorithm, PROBE. We demonstrate the method on low-dimensional, high-dimensional, and image proxies and on real-world data.

Primary area:
causal reasoning
Keywords:
causal inference, unmeasured confounding, proximal causal inference, high-dimensional proxies, covariate balancing, representation learning, average treatment effect

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_vP9xOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces proximal balancing, which learns representations and audits conditional balance using held-out proxies to identify average treatment effects under unmeasured confounding. PROBE handles unknown valid proxies through screening and aggregation. The paper provides identification and error analyses, supported by synthetic, semi-synthetic, and observational evaluations.

Critical strengths
  1. The method combines proxy balancing with representation learning without explicitly solving a bridge equation or recovering the latent confounder, offering a valuable new approach to proxy-based causal inference.

  2. The analysis covers exact identification, approximate balance, effect estimation, and aggregation with unknown valid proxies, forming a well-developed theoretical and algorithmic framework.

  3. The authors explicitly acknowledge limitations of the balance audit, image experiments, and uncertainty estimates, helping readers interpret the empirical evidence appropriately.

Critical weaknesses
  1. The comparison with standard proximal assumptions is useful, but I find the assumptions difficult to relate to real measurements. It is unclear when the retained proxies contain enough information while the held-out proxies can still reveal hidden differences that affect the outcome.

  2. The method relies on informative proxies and on correct estimates forming the largest group. The W4W_4 experiment tests one invalid candidate, but offers limited evidence about weaker proxies or several incorrect estimates agreeing with each other.

  3. The image experiment uses a frozen, supervised reader to recover scalar values before applying PROBE. This makes it hard to judge how well the method handles image proxies when useful information cannot be extracted accurately.

Questions for authors
  1. Can a concrete measurement example explain when outcome-relevant completeness and balance attainability hold together?

  2. How does PROBE perform when Assumption 4 fails, especially when proxies become less informative or several incorrect estimates agree?

  3. How does PROBE perform when the image reader makes errors or the images contain more complex proxy information?

3 comments
AuthorsOpenReview ↗

Response to Reviewer vP9x (1/2)

Thank you for the insightful review.

The comparison with standard proximal assumptions is useful, but I find the assumptions difficult to relate to real measurements. It is unclear when the retained proxies contain enough information while the held-out proxies can still reveal hidden differences that affect the outcome.

A held-out proxy is useful when it can be regarded as independent of treatment given the hidden confounder, for example a pretreatment measurement that the treatment decision does not use. In a chest X-ray, severity U shows in every block of the image (four lung quadrants and a central strip) and drives treatment. A region without treatment-related content, such as a device or a treatment response, is independent of treatment given U, yet it still shows severity, which affects the outcome. The other regions also show severity, so they carry enough information. In our chest X-ray study, the five region features explain 91% of the variance of U, and a region with a treatment-dependent mark breaks this condition and is flagged by the screen (Appendix D.9, pp. 38–41).


The method relies on informative proxies and on correct estimates forming the largest group. The W₄ experiment tests one invalid candidate, but offers limited evidence about weaker proxies or several incorrect estimates agreeing with each other.

We added a stress test in which the valid splits are not a majority. In ten SCM-1 datasets in which only W₁ and W₂ are valid and W₃, W₄, and W₅ are invalid, PROBE returned an estimate in 8 of 10, all from the valid group (MAE 0.0309). In 7 of these 8, the valid group held exactly half of the retained splits (group sizes 3:2:1), so the valid group only needs to be the largest group, not a majority. In the other 2, the valid and invalid groups had the same size, and PROBE declined to return an estimate (Appendix D.6, pp. 33–34).


The image experiment uses a frozen, supervised reader to recover scalar values before applying PROBE. This makes it hard to judge how well the method handles image proxies when useful information cannot be extracted accurately.

Our chest X-ray study addresses this. It uses no scalar reader: each of five image blocks is encoded by a network trained on that block alone, with cross-fitting, giving 80 features that explain 91% of the variance of severity U. When a treatment-response mark appears in one region, adjustment for the whole image has MAE 0.268, while PROBE hides that region and has MAE 0.138, close to the 0.132 of unmarked images. That 0.132 is the part of U the encoder cannot extract, which no choice of blocks can remove (Section 5, pp. 9–10, Figure 5; Appendix D.9, Table 6, p. 41).

AuthorsOpenReview ↗

Response to Reviewer vP9x (2/2)

Can a concrete measurement example explain when outcome-relevant completeness and balance attainability hold together?

In SCM-1 (Section 5, Figure 8(a)), W₁, W₂, and W₃ are independent noisy measurements of U. If W₁ is held out, the retained W₂ and W₃ carry information about U, so balance is attainable. W₁ is independent of treatment given U and is a measurement of U with independent noise, so outcome-relevant completeness holds. In contrast, W₅ also affects treatment and violates held-out independence, and W₄ is shifted by W₅ and violates completeness. Appendix B.1 proves that both conditions hold in a simplified Gaussian measurement model of this kind, in which Var(U | Z) = 1/6, so balance does not require recovering U (pp. 17–18).


How does PROBE perform when Assumption 4 fails, especially when proxies become less informative or several incorrect estimates agree?

Incorrect estimates exist, but valid splits form the largest group. PROBE returns a valid estimate, as in the stress test above with only W₁ and W₂ valid: 8 of 10 returned, all from the valid group.

Incorrect estimates agree and form the largest group. PROBE returns a wrong estimate. When W₂, W₃, and W₄ were invalid in addition to W₅, so W₁ was the only valid block, the splits holding out the invalid blocks W₃ and W₄ gave similar biased estimates and formed the largest group, and PROBE returned 0.553 against a true effect of 1 (Appendix D.6, Table 3, p. 33). This failure is not visible from the output.

Proxies become less informative. When we raised the measurement-noise standard deviation of every proxy in SCM-1 from 0.85 to 1.1, 1.4, and 1.7, PROBE returned in every dataset, and its MAE was 0.021, 0.024, 0.135, and 0.249 (Appendix D.15, Table 10, p. 46). Weaker proxies capture less of U, so the error grows. In terms of Assumption 4, the same hidden difference changes the held-out proxy less, which corresponds to a larger stability constant.


How does PROBE perform when the image reader makes errors or the images contain more complex proxy information?

Reader errors. We added Gaussian noise with standard deviation 0.15 to the Shapes3D pixels and kept the reader frozen. The reader's orientation-level error rose from 0% to 23.5%, yet PROBE's MAE was 0.0172, against 0.0194 on clean images, over ten datasets (Appendix D.14, Table 9, p. 45).

More complex images. In the chest X-ray study, when one region shows a mediator of the treatment effect, adjustment for the whole image has MAE 0.260, while PROBE hides that region and has MAE 0.149. When a treatment-response mark appears in every region, hiding a region cannot help (MAE 0.436, against 0.390 for the whole image) (Appendix D.9, Table 6, p. 41).

Thank you for taking the time to respond to each of my questions in detail. Your responses have helped clarify most of my concern. After considering the responses, I will keep my score unchanged.

Reviewer_c7B3OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes proximal balancing for average treatment effect estimation under unmeasured confounding using proxy variables. The method learns representations from retained proxy blocks and uses held-out blocks to assess residual treatment information, which avoids explicit proxy-role assignment and bridge-function estimation. What's more, the authors establish identification under outcome-relevant completeness, extend the framework to unknown valid blocks through a plurality condition, and derive bias and estimation guarantees under approximate balance.

Critical strengths

This paper has the following advantages that I appreciate very much:

  • I think using held-out proxies to audit a learned representation provides a useful alternative to bridge-based proximal inference.

  • In addition, the paper provides a thorough theoretical analysis and shows when the method identifies the causal effect and how imperfect balance and finite samples affect estimation accuracy. It also clearly shows that the proposed method and traditional proximal inference rely on different assumptions with neither applying to all settings covered by the other, which is definitely a plus.

Critical weaknesses

This paper can also benefit from the following clarification or improvements:

  • I think the robustness of plurality aggregation could benefit from more stress testing. I would like to see how plurality aggregation performs when fewer proxy blocks are valid. The current experiments show that an invalid block can pass the balance screen, but aggregation still recovers the correct effect. What happens if several invalid splits produce similar biased estimates and form the largest group?

  • I appreciate that the paper discusses the bias caused by imperfect balance. However, the bias bound depends on a structural constant that cannot be determined from the observed data alone. Could the authors clarify how practitioners should assess this bias when reporting uncertainty for the estimated causal effect? A sensitivity analysis over plausible values of this constant would be helpful.

  • Finally, I would be interested in seeing whether the results hold under other proxy-generating mechanisms. Twins, IHDP, and ACIC retain real covariates and potential outcomes, but their proxies and treatment are generated using a shared recipe based on the synthetic construction. RHC provides an observational application, although the true effect is unknown. Varying how informative the proxies are, how they depend on one another, or how strongly outcome-relevant completeness is violated would help assess the method’s applicability. That said, I knew that empirical evaluation in causal inference is difficult. Therefore, I view this as a suggestion rather than a major concern, and it does not lower my assessment of the paper.

Questions for authors

Please refer to my summary of weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer c7B3 (1/2)

Thank you for the insightful review.

I think the robustness of plurality aggregation could benefit from more stress testing. I would like to see how plurality aggregation performs when fewer proxy blocks are valid. The current experiments show that an invalid block can pass the balance screen, but aggregation still recovers the correct effect. What happens if several invalid splits produce similar biased estimates and form the largest group?

This is a very good question. To answer it, we ran the following stress tests and updated the paper.

Fewer valid blocks. To simulate fewer valid blocks, we used ten SCM-1 datasets (n = 24,000, true effect 1) in which only W₁ and W₂ are valid and W₃, W₄, and W₅ are invalid. The detailed simulation design is in Appendix D.6 (pp. 33–34). PROBE returned an estimate in 8 of 10 datasets, all from the valid group, with MAE 0.0309. In the other 2, the valid and invalid groups had the same size, and PROBE declined to return an estimate. Even with fewer valid blocks, PROBE returns a valid estimate if the valid splits form the largest group.

When invalid splits agree and form the largest group. To simulate this case, we made W₂, W₃, and W₄ invalid in addition to W₅, so W₁ was the only valid block (one dataset; Appendix D.6, Table 3, p. 33). The splits holding out the invalid blocks W₃ and W₄ gave similar biased estimates and formed the largest group, and PROBE returned 0.553 against a true effect of 1. In a second dataset with only W₁ valid, single linkage chained valid and invalid estimates into one group, and PROBE returned 0.768. Taken together, the simulations show that PROBE returns a valid estimate when the valid blocks, even if few, form the largest group, and a wrong estimate when invalid splits do.


I appreciate that the paper discusses the bias caused by imperfect balance. However, the bias bound depends on a structural constant that cannot be determined from the observed data alone. Could the authors clarify how practitioners should assess this bias when reporting uncertainty for the estimated causal effect? A sensitivity analysis over plausible values of this constant would be helpful.

Thank you for this question. Γ measures how strongly the mean outcome can change with differences in the hidden confounder U that the held-out proxy cannot detect. Theorem 3 bounds the bias of the estimate by B(Γ) = ΓD/[η(1−η)].

How to report the bias. D and η can be estimated; only Γ must be assumed. D (Definition 3, p. 5) is the treatment information the held-out proxy still carries given the representation. D̂² is the drop in the Brier score for predicting treatment when the held-out proxy is added to the representation, which our screen computes on a separate fold (Section 4, p. 6). η̂ is the smallest fitted propensity given the representation, or one minus the largest. For each assumed Γ, we recommend reporting [L − B(Γ), R + B(Γ)], where [L, R] is the usual 95% confidence interval of the estimate. This interval accounts for both sampling error and the largest bias that Γ allows. If it includes zero, a null effect can no longer be ruled out; the smallest Γ at which this happens shows how strongly the outcome would have to depend on undetected confounder differences to overturn the conclusion.

Sensitivity analysis. In a synthetic example (η = 0.1, D ≈ 2.3 × 10⁻⁴), B(Γ) = 0.0025Γ. Γ = 0.5, 1, and 2 give population regions [1.0014, 1.0039], [1.0001, 1.0052], and [0.9975, 1.0077] (Table 2, p. 21). The region includes zero only when Γ exceeds about 390, so the conclusion that the effect is positive is overturned only if the outcome depends very strongly on confounder differences that the held-out proxy cannot detect.

AuthorsOpenReview ↗

Response to Reviewer c7B3 (2/2)

Finally, I would be interested in seeing whether the results hold under other proxy-generating mechanisms. Twins, IHDP, and ACIC retain real covariates and potential outcomes, but their proxies and treatment are generated using a shared recipe based on the synthetic construction. RHC provides an observational application, although the true effect is unknown. Varying how informative the proxies are, how they depend on one another, or how strongly outcome-relevant completeness is violated would help assess the method’s applicability. That said, I knew that empirical evaluation in causal inference is difficult. Therefore, I view this as a suggestion rather than a major concern, and it does not lower my assessment of the paper.

Thank you for this suggestion. We ran new experiments for each of the three directions and added them to Appendix D.15 (pp. 45–46).

1. How informative the proxies are. In SCM-1, we raised the measurement-noise standard deviation of every proxy from 0.85 to 1.1, 1.4, and 1.7, keeping treatment, outcome, and blocks fixed, with ten datasets each. Each proxy measures U plus independent noise, so larger noise makes each proxy reveal less about U; this lowers how informative the proxies are. PROBE returned in every dataset. MAE was 0.021 at 0.85 and 0.024 at 1.1, then rose to 0.135 at 1.4 and 0.249 at 1.7. The split holding out the valid block W₁ showed the same rise before aggregation, so the loss comes from a representation that captures less of U, not from aggregation. See Table 10 (p. 46).

2. How the proxies depend on one another. We correlated the measurement errors of the valid blocks W₂ and W₃ at r = 0.4 and 0.8. With the blocks given, the splits holding out one of them and keeping the other became biased and tied with the valid splits, so PROBE declined to return an estimate in all 20 datasets. Our block-finding step (Appendix D.7) detected the dependence and merged W₂ and W₃ in all 20 datasets. With these blocks, PROBE returned within 0.1 of the true effect in every dataset (MAE 0.026 at r = 0.4 and 0.021 at r = 0.8). See Table 10 (p. 46).

3. How strongly outcome-relevant completeness is violated. We extended the measurement example of Appendix B.1 with a second latent component V that affects treatment and, with coefficient β, the outcome, but that no proxy measures; β sets the strength of the violation. The held-out proxy stays exactly balanced, yet the population adjustment value is 1 + 0.72β: 1.00, 1.36, and 1.72 for β = 0, 0.5, and 1, against a true effect of 1. That is, when outcome-relevant completeness fails, PROBE cannot remove the resulting bias, and no balance screen can detect it.

I thank the authors for the detailed response. My concerns have been well-addressed.

Reviewer_E67ROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces proximal balancing, a new framework designed to estimate causal treatment effects from observational data in the presence of unmeasured confounding by utilizing high-dimensional proxies. Traditional proxy-based causal methods either require pre-designating strict proxy roles to solve an ill-posed inverse problem or rely on restrictive latent-variable models that leave residual bias if the latent factor is misspecified. Proximal balancing circumvents these challenges by extending classical covariate balancing to unobserved confounders: it splits a multi-dimensional proxy into a held-out block and remaining blocks to learn a low-dimensional summary that makes the treatment groups directly comparable. The authors formalize this framework via a practical algorithm called PROxy Blockwise Exclusion (PROBE), prove its identification theory and finite-sample guarantees, introduce a plurality vote strategy for when the valid proxy structure is unknown, and empirically demonstrate its success across scalar, vector, and image-based data.

Critical strengths
  • the framework eliminates the need to pre-designate strict proxy roles, solve unstable and ill-posed integral equations, or fit complex latent-variable models. This makes proxy-based causal inference significantly more accessible and robust for empirical researchers.
  • by framing the issue as a dimensionality reduction and balancing problem rather than an explicitly generative or inversion task, the framework scales naturally to modern high-dimensional variables, such as raw medical images or radiographs, where domain expertise cannot easily partition the proxies a priori.
  • the authors do not merely offer a heuristic; they support their framework with a solid identification theory, finite-sample guarantees for the PROBE algorithm, a systematic plurality vote approach to handle structural uncertainty, and formal bounds on causal bias under approximate balance.
Critical weaknesses
  • the foundational identification theory relies on the assumption that a valid held-out block WjW_{j} exists (i.e., satisfying the required conditional independence structure). If the overall proxy matrix cannot be cleanly decomposed into blocks that satisfy these conditional independence criteria, the learned summary function ϕ\phi may fail to guarantee true balancing.
  • When the valid proxy block is unknown, the framework relies on a plurality vote mechanism. Calculating separate balancing pathways and voting weights across a combinatorial or highly granular set of potential block configurations could introduce substantial computational complexity and statistical noise.
Questions for authors

When dealing with high-dimensional vectors or images where natural "blocks" are not predefined by the data layout, what concrete strategies or heuristics do you recommend for partitioning W into the J distinct components needed for the PROBE algorithm?

1 comment
AuthorsOpenReview ↗

Response to Reviewer E67R

Thank you for the insightful review.


the foundational identification theory relies on the assumption that a valid held-out block exists (i.e., satisfying the required conditional independence structure). If the overall proxy matrix cannot be cleanly decomposed into blocks that satisfy these conditional independence criteria, the learned summary function ϕ\phi may fail to guarantee true balancing.

You are right that identification requires at least one valid held-out split. Without a valid split, PROBE can return a biased value. However, if valid blocks exist, even if they are unknown, the proposed algorithm can detect and use them to yield correct estimates.

Failure when no valid split exists. In SCM-2 with random coordinate sets used as held-out sets, every set contained an invalid coordinate, no split passed the screen, and PROBE returned nothing (Appendix D.7, pp. 35–36). When two invalid splits agreed and formed the largest group, PROBE returned 0.553 against a true effect of 1 (Appendix D.6, Table 3, p. 33).

When valid blocks exist but are unknown. In this case, PROBE can find them from W alone. With 16-dimensional blocks in SCM-2 and 256-dimensional blocks in SCM-3, the mean absolute error (MAE) over 100 datasets was 0.0195 and 0.0213, slightly above 0.0184 and 0.0178 with the blocks given. We added this to Appendix D.7, Table 4 (p. 35).


When the valid proxy block is unknown, the framework relies on a plurality vote mechanism. Calculating separate balancing pathways and voting weights across a combinatorial or highly granular set of potential block configurations could introduce substantial computational complexity and statistical noise.

Thank you for raising the cost of the search. PROBE searches over splits of blocks, not of coordinates. In our high-dimensional settings, W has 80 coordinates (SCM-2) or 1,280 coordinates (SCM-3). PROBE first groups them into J = 5 blocks from W alone, before it sees A or Y (Appendix D.7, p. 34), so the search covers only 2ᴶ−2 = 30 splits. In the image settings, each block is one or two Shapes3D images, or one anatomical region of a chest X-ray. Algorithm 1 can further fit and screen only m of these splits, sampled uniformly without replacement. Each retained split casts one vote, and PROBE returns the median of the unique largest group of nearby estimates. One full search over 30 splits took about 4–6 minutes per dataset of 24,000 units. See Appendix D.3 (p. 30) and D.6 (p. 33).

Theory. Theorem 6 shows that sampling m splits at random suffices with high probability. Its error bound holds except for two sampling terms, the chance that no sampled split passes the screen and the chance that sampling reverses the plurality, and both terms decrease as m grows. See Theorem 6 (pp. 7–8) and Appendix C.3 (pp. 27–28).

Empirical checks. We replayed the stored candidate libraries of 100 datasets per setting in SCM-1 to SCM-4 without refitting. With 20 of 30 splits, 99.7–99.8% of draws returned an estimate within 0.1 of the true effect (MAE 0.0191–0.0201, against 0.0178–0.0188 with all 30).

Splits without blocks. We recommend grouping related coordinates before sampling splits, because in high-dimensional settings PROBE is accurate with grouped blocks and fails without them. With every coordinate treated as its own block, single-coordinate splits gave MAE 0.50 in SCM-2 and 0.42 in SCM-3, because one noisy coordinate is too weak a check. With the blocks found from W, MAE was about 0.02 (Table 4, p. 35). We added this comparison to Appendix D.7 (pp. 35–36).


When dealing with high-dimensional vectors or images where natural "blocks" are not predefined by the data layout, what concrete strategies or heuristics do you recommend for partitioning W into the J distinct components needed for the PROBE algorithm?

Thank you for asking for a concrete procedure. Use measurement or acquisition groups when available. Otherwise, use a partition computed from W alone for vectors and a fixed grid for images.

Vectors. Cluster coordinates into J groups by their loadings on the leading signal directions of the correlation matrix of W. Reassign each coordinate to the nearest group subspace. Merge two groups when a cross-group pair has partial correlation above 0.5 given the rest. The analyst sets J (J = 5 in our designs), and merging can lower it; in SCM-1 it joined W₄ and W₅. We fixed the 0.5 threshold after a first run on Twins and confirmed it on new data. See Appendix D.7 (pp. 34–35).

Images. In our chest X-ray study, a 4×4 grid matched anatomical regions with a response mark in one region (MAE 0.128 vs 0.138). It did worse when a mark crossed a cell boundary (0.225 vs 0.149) or appeared in two regions (0.289 vs 0.171), so we kept the anatomical layout. A grid is a simple default, but a signal spanning cells can escape it. We added this to Appendix D.9, Table 7 (p. 41).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions

We thank all reviewers for their careful reviews. We have uploaded a revised PDF; the individual replies cite its pages. The main text is now 10 pages, the limit for the discussion phase.

Changes to the main text

  • Section 5, synthetic and real-world analyses. PROBE is updated so that it finds the blocks from W alone instead of being given them (Appendix D.7). Figure 3 is updated accordingly.
  • Section 5, Shapes3D analysis. The image setting SCM-4 now has its own paragraph and a diagram (Figure 4).
  • Section 5, chest X-ray analysis (new). A semi-synthetic study on 5,856 chest X-rays, in which PROBE excludes image blocks that carry treatment-dependent content (Figure 5; Appendix B.4 and D.9).
  • RHC. The RHC analysis moved from Section 5 to Appendix D.10, because the ground-truth is unknown in this experiment.
  • Introduction and conclusion. The introduction is condensed without changing its claims, and the conclusion now also summarizes the chest X-ray study.

New material in the appendix

  • B.1: a measurement example in which outcome-relevant completeness and balance attainability hold together (vP9x).
  • B.3, Table 2: how the bias allowance scales with the stability constant Γ, with reporting guidance (c7B3).
  • B.4: blockwise exclusion within a single image, used by the chest X-ray study.
  • D.6, Table 3: stress tests of plurality aggregation, a paired comparison with two valid blocks, and a replay of smaller split budgets (E67R, c7B3, vP9x).
  • D.7, Table 4: finding the blocks from W, and splits without blocks (E67R).
  • D.9, Tables 6–7: all chest X-ray scenarios and a 4×4 grid in place of anatomical blocks (E67R, vP9x).
  • D.14, Table 9: reader degradation and weaker proxy measurements (c7B3, vP9x).
  • D.15, Table 10: proxy informativeness, dependence between proxies, and completeness violations (c7B3).

The new stress tests also show where PROBE fails. When invalid splits agree and form the largest group, PROBE returns a biased estimate, and passing the balance screen does not certify validity (Appendix D.6 and D.14). We report these cases in the replies and in the appendix. The supplementary code package is unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.