acceptodds
Under review as a conference paper at ICLR 2027

Temporal Eligibility Routing (TEAR) for Treatment-Effect Estimation under Structured Covariate Contamination

Abstract

Temporal provenance and target-relative causal eligibility encode different information in structured covariate exports. Raw inclusion can expose treatment descendants, while temporal deletion can discard delayed records of pretreatment measurements. Temporal Eligibility Routing (TEAR) learns metadata-guided continuous gates and routes feature contributions through a finite-dimensional aggregate before conditional average treatment-effect estimation. A perturbation analysis connects routing-mass errors to representation and prediction perturbations for fixed embeddings and a fitted host. Controlled experiments show that inclusion and deletion favor opposite eligibility regimes, while gated aggregation helps when eligible and ineligible features coexist. Adding TEAR to an otherwise identical preprocessing candidate pool lowers test error measured by precision in estimation of heterogeneous effect (PEHE) in all eight dataset–host cells, with all paired 95% confidence intervals excluding zero. Observed-data selection recovers most of the PEHE-based improvement on MIMIC-P3, while provenance corruption drives gate discrimination toward chance and shifts selection toward Deletion. Code is available in an anonymous repository.

Primary area:
causal reasoning
Keywords:
Treatment-Effect Estimation, Casual Inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SpUfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the problem that the temporal provenance of covariates may not match their causal eligibility. Either keeping all covariates or deleting all post-index covariates would impact the performance of the treatment effect estimation. So, the paper proposes a preprocessing framework(TEAR) to address the problem through a soft routing gate that attributes the contribution of all covariates and lowers the adverse impact of post-treatment covariates without losing important information. The main contributions are formulating covariate preprocessing as continuous feature-level routing and combining metadata-guided gates with finite-dimensional aggregation. The paper finally validates TEAR on synthetic controlled data and semi-synthesized MIMIC data using multiple CATE estimators such as DML, R-Learner, TARNet, and DragonNet. The results show that TEAR performs best when both eligible delayed information and treatment descendants exist in data.

Critical strengths

The paper has a clear and easy-to-follow story. The problem is well-defined, and the authors conduct relatively comprehensive experiments to validate TEAR under different settings, including different eligibility regimes, multiple CATE estimators, ablation studies, observed-data selection, and provenance corruption. These experiments help us understand under what scenarios TEAR could work better and when its advantage would become weaker.

Critical weaknesses
  1. The evidence of comparing the TEAR with naive inclusion or deletion of covariates is too weak. It would be better to compare it with careful causal work.

  2. The gate does not shrink the weight of post-treatment covariates completely to zero, meaning that some post-treatment information leakage is still present. This could convince us that bias would be minimized or controlled from post-treatment covariates.

Questions for authors
  1. The mismatch between recorded time and causal eligibility of covariates is plausible. But is it a common scenario in the real world that the clinically meaningful measurement time is missing or the recorded time is not rigorously monitored, so that this problem becomes important enough to require a learned routing framework?

  2. Only comparing TEAR mainly with these naive methods does not fully demonstrate the contribution of TEAR. I would like to see it compared with careful causal work. Here is the paper I found that might be used as a stronger baseline: Doutreligne et al. (2025), “Step-by-step causal analysis of EHRs to ground decision-making.” Another control could be to adjust the cutoff timestamps for inclusion or deletion instead of just including or deleting all post-index covariates.

  3. TEAR might be able to achieve a lower PEHE, but the gate does not shrink the weight of post-treatment covariates completely to zero, meaning that some post-treatment information leakage is still there. I would like the authors to clarify whether the remaining leakage is reduced to an acceptable level.

  4. Please discuss the following related work and consider comparisons where applicable:

    • Nagalapatti et al., Leveraging a Simulator for Learning Causal Representations from Post-Treatment Covariates for CATE (TMLR 2025).
    • Lin and Crespo, Variance reduction combining pre-experiment and in-experiment data (CLeaR 2026).
  5. Could you clarify which function computes the reported root-PEHE, defined as 1n∑i(τ^i−τi)2\sqrt{\frac{1}{n}\sum_i(\hat{\tau}_i-\tau_i)^2}? The supplied compute_caie_rmse averages over both the treatment-effect error and a zero reference column, yielding root-PEHE divided by 2\sqrt{2} for binary treatment—for example, 0.7071 instead of 1 for unit error. Was this function used for the reported tables? If not, please identify the implementation used.

4 comments

Thank you for recognizing the clarity of the causal-eligibility problem and the breadth of the controlled evaluations. We took your tunable-cutoff suggestion as a direct test of whether a stronger temporal rule can replace feature-specific routing; we also clarify residual descendant exposure, clinical provenance, and the metric implementation.

W1 & Q2: Stronger Controls—Including Your Proposed Cutoff

We evaluated a validation-tuned global timestamp cutoff over seven predeclared candidates, selected on validation data and assessed once on matched test splits (50 paired seeds; new App. H.2, Table 19). The same fixed TEAR configuration improves mean test PEHE in all three settings:

Dataset / hostTuned cutoffFixed TEARReduction
Controlled mixed / DML0.1390.12112.9%
MIMIC-P3-Causal / DML0.1640.1499.1%
MIMIC-P3-Causal / R-Learner0.1730.1579.2%

Thus tuning one global temporal boundary does not recover the gain from feature-specific routing in these settings, despite allowing the cutoff seven validation candidates versus a fixed TEAR operating point.

Existing controls also include L0L_0 selection (Table 3), Gate×Raw and random projection (Table 1C), and a 50-seed causal-representation comparison (Table 15). On Synth-High/DML, TEAR-B0 / TEDVAE / D2VD-nonlinear have mean PEHE 9.23 / 10.20 / 17.71, with P95 23.77 / 26.65 / 55.59. Together, these test temporal rules, learned sparsity, compression, and causal representations.

W2 & Q3: Residual Descendant Exposure and CATE Error

A post-index timestamp does not by itself imply treatment-induced content. The controlled mixed regime separates delayed-neutral records from genuine descendants. TEAR gives them mean gate weights 0.82 vs. 0.11, with normalized ineligible exposure 0.09 (Table 1B; App. D.1). Crucially, we also measure effect-estimation error rather than relying on gate weights alone (Table 10):

Mixed regimeOracleTEARDeletionRaw
PEHE ↓0.1030.1210.1560.193
Absolute CATE bias ↓0.0060.0110.0140.083

The controlled evidence shows lower ineligible exposure than Raw, and lower CATE estimation error than both Raw and Deletion. Nonzero descendant mass remains, and the oracle benchmark makes the remaining gap explicit; the claim is reduction in this setting, not guaranteed elimination of causal bias.

Q1: Why the EHR Timing Mismatch Is Operationally Real

MIMIC-IV chartevents contains distinct charttime (observation time) and storetime (entry/validation time), illustrating that when a value was observed and when it entered a structured export need not coincide (revised App. E.2). TEAR targets this distinction after the estimand and admissibility question have been specified. Our EHR-shaped tests use semi-synthetic outcomes; they test the mechanism, not its prevalence across hospitals.

Q4: Relation to the Suggested Causal Literature

We added all three references (revised Sec. 2). Doutreligne et al. address target-trial design and expert identification, which are necessary upstream decisions rather than a plug-in feature router. Nagalapatti et al. use simulator-generated supervision for post-treatment CATE representations; Lin & Crespo study pre-/in-experiment data for randomized-experiment ATE variance reduction. Their supervision and estimands differ from this static provenance-guided CATE setting. The directly comparable controls here are learned feature selection, causal representations, and the proposed tuned cutoff.

Q5: Root-PEHE Function and Reported Values

Thank you for the precise numerical check. We reproduced the public compute_caie_rmse helper's binary-treatment normalization error: including the zero-reference contrast returns 0.7071 rather than 1.0000 for unit effect error. Correct root-PEHE uses only the nonreference treatment contrast (Sec. 5.1; App. D.1). The available repository does not identify the implementation used to generate the manuscript tables, so their absolute scale is not verified against this helper. Uniform use of this helper would rescale PEHE and paired intervals by 2\sqrt{2} without changing relative improvements; which tables require correction depends on their actual calculation path.

The new cutoff test, clinical timestamp example, and comparisons with the cited literature are incorporated in the revised PDF (Sec. 2; Sec. 5.3; Apps. E.2, H.2).

Reviewer_SpUfOpenReview ↗

reply to updates

Thanks for all the updates to improve our understanding of the paper. I have read both the reply to me and the replies to the other two reviewers to capture all details. I appreciate the additional baseline comparisons, including the validation-tuned temporal cutoff. But I still have some confusion that needs clarification:

  1. Metadata vs. routing: It is still unclear whether TEAR's improvement comes from learned routing or its use of richer metadata. Could the authors clarify what the existing ablations establish?

  2. Residual leakage: Lower PEHE and CATE bias in controlled settings do not necessarily guarantee valid causal estimation. Could the authors clarify the limitations of retaining post-treatment information?

  3. PEHE calculation: The authors acknowledge the normalization bug but have not confirmed whether it affected the reported results. Could they clarify which results have been verified?

These clarifications would help me better assess the contribution and reliability of the paper.

And I know the rebuttal period is closing. So, I would not expect running more experiments. Maybe for some questions, I just misunderstood or missed something in the manuscript. So, it is fine to just use your existing work or understanding to remove my concerns.

AuthorsOpenReview ↗

Follow-up Response to Reviewer

Thank you for your thoughtful follow-up, for reading our replies across all three reviews, and for acknowledging the tuned-cutoff baseline. We also appreciate your patience as we revisited the evidence and its limitations. As you suggested, we address your three points using existing results.

1. Metadata versus Learned Routing

Both contribute, and the existing ablations separate their roles. In the controlled mixed regime (DML, 50 matched seeds), App. H.1 (Tables 17–18) varies only gate training, with the same covariates and timestamp/schema inputs:

Gate configurationPEHE ↓Post-index AUROC ↑Delayed / descendant h
Full TEAR0.1210.900.82 / 0.11
No prior loss or prewarming0.1390.760.61 / 0.29
Frozen prior-only gate0.1500.510.06 / 0.05

With the same metadata, the frozen gate closes both post-index feature types, whereas joint training separates delayed-neutral records from descendants and lowers PEHE by 19.3%. Even without the gate-prior loss or prewarming, joint training achieves a post-index AUROC of 0.76. The post-index AUROC of 0.90 (Table 18) is computed among post-index features; the overall gate AUROC is 0.95 (Table 1B).

App. H.4 (Table 21) holds the architecture fixed and varies metadata availability: PEHE is 0.121 with full metadata, 0.136 with schema only, and 0.148 with time only (Deletion: 0.156). Metadata richness therefore sets how much the gate can separate, and joint learning converts that metadata into routing. Because Eq. (4) also feeds metadata into the embeddings, we report these as complementary ablations rather than an additive decomposition. Table 1C further separates coordinate-wise gating, random projection, and eligibility-guided aggregation.

2. Scope of Retaining Post-Treatment Information

We agree that lower PEHE and valid causal identification are distinct claims. In the controlled mixed regime, TEAR assigns descendants a mean gate of 0.11 (normalized ineligible exposure 0.09) and reaches PEHE / absolute CATE bias of 0.121 / 0.011, compared with 0.156 / 0.014 for Deletion, 0.193 / 0.083 for Raw, and 0.103 / 0.006 for the oracle (Table 10). This shows reduced error under the evaluated generator, and the oracle provides a reference for the remaining estimation gap.

The scope of this result follows from our theory. Proposition 1 bounds representation and fixed-predictor perturbations by the sum of exposed ineligible mass and omitted eligible mass, under bounded embeddings and a fixed Lipschitz host; a nonzero gate on a genuine descendant can still admit bias in a refitted estimator. Lemma B.1 identifies CATE when its eligible-representation, balancing, and heterogeneity-sufficiency assumptions hold, and for a learned gate these remain assumptions. Known descendants belong outside the adjustment set by causal design; TEAR is designed for features whose eligibility recorded time leaves uncertain. Its boundary is also measurable: at 100% provenance corruption, fixed TEAR reaches PEHE 0.159 versus 0.156 for Deletion, and observed-data selection then chooses Deletion in 87% of seeds (Fig. 1; Table 9), so the selector falls back to the conservative rule when provenance carries no signal.

3. Root-PEHE Implementation and Evaluation Scope

Thank you for following up on the metric implementation. We traced the public code and confirmed the source of the normalization issue. The compute_caie_rmse helper includes the zero-reference column when computing binary-treatment RMSE, introducing a factor of 1/√2. Its only call site in the released package is an optional diagnostic within OrbitTrainer.fit(), evaluated on the adapter's training inputs.

For any binary-treatment comparison evaluated uniformly with this helper, correcting the normalization multiplies absolute PEHE values, paired differences, and confidence-interval endpoints by √2. The common factor leaves method rankings, relative improvements, validation-PEHE selection, and scale-invariant paired tests unchanged. Within each comparison, candidates share the downstream host and evaluation protocol, differing in their preprocessing. If their binary-treatment PEHE values were computed through a common scoring path, the normalization correction preserves the tuned-cutoff and routing comparisons, including their relative gains and paired statistical conclusions. The actual table-generation path remains to be verified. Gate diagnostics, routing weights, exposure measurements, and observed-data selection scores do not depend on this PEHE normalization.

The released repository does not establish the scoring path for each manuscript table, so we distinguish the confirmed helper defect from table-level verification. We will correct the released helper and document any table-level numerical corrections supported by the verified evaluation paths.

Thank you again for the opportunity to clarify these points.

Thank you again for your thoughtful engagement throughout the discussion and for taking the time to consider our additional evidence!

We have now provided detailed clarifications addressing your three follow-up questions, and we sincerely hope they help resolve your remaining concerns.

As the discussion period draws to a close, we would be grateful for any final feedback you may have. If our responses have strengthened your assessment of the paper, we would also greatly appreciate your consideration of updating your overall rating accordingly.

Thank you once again for your time and constructive feedback.

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

This paper looks at the preprocessing step for data to be used in causal inference. Real world tabular data (the covariate data) may contain a mix of information where it is not clear what is admissible to estimate effects (e.g., pre-treatment features that are reported at the treatment). Common approaches include dropping all ambiguous features (loses information) or keeping everything (temporal contamination), but the paper introduces a new approach that learns how to admit these features along a continuum as a function how likely they are to be admissable from the metadata. In experiments with both synthetic data and with variants of MIMIC three, the paper's approach, TEAR, corrects for both issues in common approaches and allows multiple causal methods to achieve much lower error.

Critical strengths
  • Interesting approach to learning how to partially recover information covariates without over-admitting

  • Strong results that demonstrate performance improvements with multiple methods

  • Well structured paper that is easier to read in most places (the results sections get a bit dense a times)

  • Helpful robustness test showing the method is generally still helpful in the face of metadata corruption, except at extreme levels.

Critical weaknesses
  • While I like the approach, its success is dependent on the presence of metadata to infer eligibility. The authors do use multiple benchmarks but it would be useful to have a sense of how broadly TEAR might be applicable and to measure how sensitive it is to the amount of metadata.
Questions for authors
  • I think TEAR is being refit within every fold, if I've understood Line 204 and C.2. How fast is this if you're using a high dimension covariate matrix like one might get from an EHR structured dump?
3 comments

Thank you for recognizing the cross-host gains and the metadata-corruption test. Your two questions identify where TEAR's practical usefulness should be quantified: how much provenance is needed, and what fold-wise fitting costs.

W1: How Much Metadata Is Needed?

We added a field-availability study, distinct from the original provenance-permutation test, with 50 matched seeds in two DML settings (App. H.4, Table 21):

Metadata availableMixed PEHE ↓MIMIC-P3-Causal PEHE ↓
Full0.1210.149
25% fields missing0.1290.159
50% fields missing0.1370.168
Schema only0.1360.165
Timestamp only0.1480.176
Deletion0.1560.187

At 25% field missingness, TEAR retains 73.7–77.1% of its full-metadata improvement over Deletion; at 50%, it retains 50.0–54.3%. Schema-only outperforms timestamp-only in both settings. Together with the provenance corruption test (Fig. 1) and frozen-hyperparameter metadata-shift test (App. G.5), these results distinguish missing, misassigned, and shifted provenance. TEAR remains useful with partial informative schema fields; the benefit declines as those signals disappear.

Q1: Fold-Wise Refitting Cost

Yes. DML and R-Learner refit the adapter inside each cross-fitting fold (App. C.2), preserving the training/held-out split. On an A100 80GB with 10,000 MIMIC-P3 admissions and 25 covariates, measured costs are 4.6 s per fit, 24.8 s for five folds, and 2.1 GB peak GPU memory. The adapter contributes 14.2% of the DML pipeline time and 8.7% of DragonNet's (App. G.5). These are measured costs at the evaluated dimensionality; they are not extrapolated to thousands of features.

The availability experiments now appear in the revised PDF (Sec. 5.6; App. H.4), with the fitting protocol and measured runtime in Apps. C.2 and G.5.

Thanks you for the follow up and the experiments. I think this response and the updates to the paper address my concerns and show that the approach is robust.

Thank you for your encouraging response. We apologize for our late reply, as we wanted to carefully complete the additional experiments and revisions before responding. We are glad to hear that our updates have addressed your concerns.

As the discussion period comes to an end, we would greatly appreciate your consideration of revisiting your overall rating in light of the additional evidence, if you feel it is warranted.

Thank you again for your time and constructive feedback!

Reviewer_Ee5fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a preprocessing method called Temporal Eligibility Routing (TEAR) for treatment-effect estimation that addresses the structured covariate contamination problem. TEAR learns continuous eligibility gates from temporal provenance and routes feature embeddings through a finite-dimensional aggregate before passing them to a downstream CATE estimator. A perturbation bound links routing-mass errors to representation or prediction errors. Experiments on synthetic and semi-synthetic benchmarks show that adding TEAR to a candidate preprocessing pool improves validation-selected PEHE across all dataset–host combinations.

Critical strengths
  • Structured covariate contamination is a genuine challenge in EHR and other temporally structured exports.

  • The evaluation spans controlled eligibility regimes, multiple CATE hosts, semi-synthetic EHR settings, observed-data selection, and provenance corruption.

  • In controlled experiments, TEAR consistently outperforms Raw and Deletion baselines in mixed-eligibility regimes where delayed-neutral features and treatment descendants coexist.

Critical weaknesses
  • Proposition 1 is a direct application of the triangle inequality and Lipschitz continuity. It provides no insight into learning dynamics, identifiability, or optimization.

  • The Bayesian model in Eq. 1 derives optimal routing weights from costs and beliefs, but TEAR never uses this formula, i.e., the gate is a neural network trained end-to-end. The theory and implementation are unrelated.

  • The selector rewards factual RMSE, balance MMD, and low risky exposure, i.e., the exact same quantities that TEAR's training objective (LY\mathcal{L}_Y, Lb\mathcal{L}_b, Lg\mathcal{L}_g) optimizes. TEAR being frequently selected is therefore unsurprising and does not constitute independent evidence of superiority.

  • Table 14 shows that removing zihidz_i^{hid} changes PEHE by −0.0006 (p=0.91p = 0.91) on MIMIC-P3 and −0.0743 (p=0.72p = 0.72) on Synth-High. The entire two-branch gating architecture is thus called into question.

  • On the Twins dataset (Appendix G.4), TEAR's improvement over Deletion ranges from 0.1% to 2.2% across hosts. This suggests the method's benefits may not transfer beyond synthetic settings designed to showcase it.

Questions for authors
  • The gate uses only metadata and timestamps, never feature values (hj=σ(MLPgate([mj,tj]))h_j = \sigma(\text{MLP}_{gate}([m_j, t_j]))). Two features with identical provenance but radically different relationships to treatment receive the same gate weight. Why should eligibility routing be purely a function of provenance rather than also incorporating feature content?

  • The temporal prior Eˉj=1(tj<0)\bar{E}_j = \mathbb{1}(t_j < 0) is exactly the Deletion rule that the paper criticizes (Section 5.2 shows Deletion fails in delayed-neutral regimes). Yet removing Lg\mathcal{L}_g causes the largest performance drop. This raises a concern: is TEAR learning something beyond the prior, or is the prior doing most of the work? Could you report gate weights for delayed-neutral features when Lg\mathcal{L}_g is removed to clarify what the factual losses alone learn?

2 comments

We thank you for recognizing both the EHR contamination problem and TEAR's advantage in mixed-eligibility regimes. The central contribution is feature-specific routing before CATE estimation. The controlled diagnostics below isolate what the prior, selector, and routing architecture each contribute.

W1 & W2: What the Theory Establishes—and What Is Learned

The theoretical components have distinct roles (Sec. 3; App. B): Eq. (1) derives an interior routing decision under uncertain eligibility and asymmetric exposure/omission costs; Lemma B.1 establishes conditional CATE identification when the eligible representation balances covariates and preserves heterogeneity; Proposition 1 bounds the representation and fixed-host prediction perturbation from routing-mass errors. Eqs. (4)–(6) then learn continuous weights, rather than computing Eq. (1)'s posterior optimum. Thus the proposition is a perturbation result, while eligibility learning is tested empirically below (not claimed as an optimization or identification theorem).

W3: Independent Evidence Beyond the Observed-Data Selector

The primary eight-cell gain does not use Eq. (7). Table 2 applies the same validation-PEHE selection rule to Base = {Raw, Deletion, L0} and Base + {TEAR-B0, B1}, then evaluates on matched test seeds. All 8/8 paired 95% CIs exclude zero; for MIMIC-P3/DML, selected PEHE falls 0.1983 → 0.1500, paired Δ = −0.0483 [−0.0639, −0.0327]. This isolates the value of TEAR as an additional candidate.

We separately ablated the observed-data selector on the same five candidates (20 seeds, App. H.3, Table 20):

MIMIC-P3/DML selectorTest PEHE ↓Recovery vs. Deletion
Full Eq. (7)0.20678.1%
Without risky-exposure term0.22265.6%
Without MMD0.21968.0%
Factual-only0.23555.5%
Deletion0.3060%

The benefit persists even without either auxiliary selection penalty, while the full score performs best. Selection frequency is not the independent performance claim.

W4: Visible Routing, Not the Non-Exported Branch, Is the Mechanism

Table 14 removes the non-exported branch while retaining the eligibility-gated visible aggregate: its negligible PEHE change therefore tests the former, not routing itself. Table 1C tests the latter on Synth-High/DML: Gate×Raw 21.60, random projection 15.24, and two-role TEAR 9.29 PEHE. A 10-seed exploratory control fixes every gate to 1 while retaining balance/orthogonality losses (Table 11): 27.16 vs. 17.48 for TEAR-B0 (10 seeds). These tests separate gated aggregation from auxiliary-branch effects: the non-exported branch adds little in the reported DML comparisons.

Q2: The Learned Gate Goes Beyond Its Temporal Prior

We ran your requested diagnostic within post-index features, comparing 50 matched seeds in the controlled mixed regime (App. H.1, Tables 17–18):

ConfigurationDelayed hhDescendant hhPost-index AUROCPEHE ↓
Full TEAR0.820.110.900.121
No LgL_g, no prewarm0.610.290.760.139
Frozen prior-only gate0.060.050.510.150

Full TEAR reduces PEHE 19.3% relative to frozen prior-only. Removing prior supervision and prewarming weakens, but does not erase, post-index discrimination. The prior alone cannot explain the learned separation; joint optimization adds measurable value.

Q1 & W5: Why Provenance, and Where It Helps

Treatment association cannot itself establish eligibility: both a pretreatment confounder and a treatment descendant can strongly predict treatment. TEAR therefore gates at the column/provenance level, while feature values enter learned, feature-specific embeddings and the joint objective (Sec. 4.1). Identical provenance indeed yields identical gate weights; this is a defined scope of the metadata-only design.

Twins is the weak-provenance boundary: TEAR improves only 0.1–2.2% over Deletion, but 2.4–14.7% over the timestamp rule across four hosts (Table 16). In the controlled corruption test, AUROC falls 0.95 → 0.50 and selection moves from TEAR to Deletion (74% → 7% TEAR selections; Fig. 1). The gains thus track the information available in provenance rather than assuming unrestricted transfer.

These mechanism and selector studies are included in the revised PDF (Secs. 5.2–5.5; Apps. H.1, H.3), alongside the existing theory and boundary controls.

Thank you again for your careful review! We apologize for our late response, as we wanted to complete the additional analyses and carefully revise the manuscript to address your concerns.

As the discussion period comes to an end, we would greatly appreciate your feedback on whether our responses have addressed your main concerns. If you find the additional evidence convincing, we would also be grateful for your consideration of revisiting your overall rating.

Thank you for your time and thoughtful feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.