acceptodds
Under review as a conference paper at ICLR 2027

Partial Identification of Network Causal Effects under Structural Support Deficiency

Abstract

Repeated network experiments can leave policy contrasts unidentified even as sampling error vanishes: operational constraints assign zero probability to relevant exposures. We characterize this limitation through a compact completion polytope combining outcome bounds, scientific linear restrictions, and bidirectional transformation distances respecting one-way operational boundaries. Under the finite-population mean-response model, its target image is a sharp identified set: diameter implies worst-case squared-error risk at least for every point estimator under any number of unchanged-design repetitions. Optimal dual faces characterize when these bounds coincide and their local sensitivity to supported means and restrictions. To infer this set, we project a design-based confidence polytope for supported means, obtaining finite-sample full-set coverage; independent stable waves permit affine refinement through exact common-dual vertex certification. Matching upper and block-experiment lower bounds give minimax expected interval length of order for the specified fixed-dimensional classes with uniformly bounded endpoint sensitivity, identification-diameter bound , supported-cell probabilities , and dependency degree . In 600-repetition experiments, widths approach the identification plateau, while projection attains – full-set coverage versus for raw optimization. The predicted dependence scaling appears in a variance log–log slope of against ; across 4,000 end-to-end runs, dependence-calibrated full-set coverage is , versus – with a unit-iid radius. In a six-cell insurance-network audit, descriptive projection spans all held-out empirical means at average width , versus unrestricted width . These results connect the limits of causal learning to experimental design and specify when greater precision requires new identifying information.

Primary area:
causal reasoning
Keywords:
causality

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

The paper studies network experiments where operational constraints assign zero probability to some relevant exposures. It combines supported exposure means with scientific restrictions in a completion polytope, derives sharp bounds for causal contrasts, and constructs confidence intervals. The key theoretical finding in my opinion is that additional observations reduce sampling uncertainty but cannot eliminate uncertainty caused by unchanged support deficiencies.

Critical strengths

The paper makes a useful distinction between estimating observed exposure means accurately and identifying an unsupported causal contrast. Its main inferential objective is clearly specified and well motivated. The combination of dependence-aware inference and common-dual certification is the most promising methodological component: certification explicitly checks whether an affine endpoint representation remains valid throughout a confidence region, rather than merely at an estimated mean.

Critical weaknesses
  1. My main concern is the limited methodological novelty. Much of the theoretical framework appears to follow from established partial identification, linear programming duality, and confidence set projection techniques, making the contribution beyond existing work unclear. An external check also confirms that Khan et al. (ICML 2024) already develop sharp Lipschitz linear-programming bounds and minimax-rate estimation under non-overlap. So the paper needs to isolate what its network-specific inference enables beyond adapting this existing framework

  2. The practical usefulness also depends heavily on scientific restrictions whose validity is difficult to justify or verify, particularly for unsupported exposures. Furthermore, the experiments primarily confirm expected theoretical properties rather than demonstrate clear advantages over competitive existing methods.

  3. The real-data analysis provides only modest reductions in uncertainty and does not establish meaningful improvements in causal or policy decision-making.

Questions for authors

See the weakness section; I think the paper has to show a sufficiently distinctive methodological contribution and convincing empirical evidence of practical value to meet the expectations of ICLR.

4 comments
AuthorsOpenReview ↗

Thanks for the review and here is the response

W1 — What the method adds

Which policy contrasts can a constrained network experiment learn, and how much uncertainty must remain? Repeating the same design improves observed means but leaves excluded responses unresolved. The contribution characterizes the target's attainable range and the statistical cost of covering that entire range.

Khan et al. establish sharp Lipschitz policy-value bounds and minimax estimation under non-overlap. Wang et al. extend policy-value identification to network interference and develop network estimation rates, asymptotic inference and misspecification analysis. Both bound policy values using nonnegative policy weights and coupled restrictions. Our contribution concerns arbitrary signed contrasts and general joint linear restrictions, with design-based finite-sample coverage of the entire identified interval and exact regional certification of affine endpoints.

The geometry is simple: for u,v∈[0,1]u,v\in[0,1] with ∣u−v∣≤0.1|u-v|\le0.1, each marginal range is [0,1][0,1], while the sharp range of u−vu-v is [−0.1,0.1][-0.1,0.1]. Joint optimization preserves the constraint linking the two means. This example explains the loss from marginal subtraction; the inferential contribution adds valid uncertainty quantification under the network assignment law.

The theory separates irreducible ambiguity from sampling error. Theorem 3.1 gives squared-error lower bound ω2/4\omega^2/4 for observationally identical worlds whose targets differ by ω\omega. Section 5 establishes matching interval-length order D0+M(1+Δˉ)/(Np⋆).D_0+M\sqrt{(1+\bar\Delta)/(Np_\star)}. Here D0D_0 bounds identification width, MM bounds outcomes, NN counts unit-wave observations, and Δˉ,p⋆\bar\Delta,p_\star bound dependence degree and exposure probability. The result holds on the declared fixed-dimensional classes with uniformly bounded endpoint slopes and the specified lower-bound submodels.

Projection is the default full-set inference procedure. The optional common-dual certificate verifies one exact affine representation over the whole pilot region (Theorem 4.2); independent evaluation or fallback preserves finite-sample coverage (Theorem 5.1). Conditions include valid exposure/scientific restrictions, bounded outcomes, known probabilities and stable independent waves. The basic guarantee permits nonunique optimizers and failed certificates.

The matched experiment includes an equality control. With identical observations, restrictions and confidence bands, joint LP equals the published-envelope calculation in all 32 nonnegative-target settings. Against marginal propagation of those envelopes to signed contrasts, it is narrower in 32/32 settings; both cover the full population interval in all 500 repetitions per setting. This finite-grid comparison isolates compatibility loss; it implements bound formulas with common calibration, separately from the original regression pipelines. The covariate formula also agrees with Khan's released code.

References: Khan et al. (2024); Wang et al. (2026).

Revision: Section 2; Appendix N.1 explains the geometry and existing rate results, and N.3 gives the matched comparison (blue).

AuthorsOpenReview ↗

Thanks for the review and here is the response (cont'd)

W2 — Scientific restrictions and competitive evidence

We now state the calibration input and show how the conclusion changes when it is relaxed. The field rule uses the largest difference-to-distance ratio among retained cells, excluding masked outcomes, with sensitivity multipliers 1, 1.5 and 2. Insurance uptake is binary, giving the outcome range [0,1][0,1]. The bound LL specifies the allowed response difference per unit distance, including unsupported exposures; the tighter interval is conditional on that scientific bound.

The hidden-group bound controls identification: two worlds can have exactly the same observed distribution but different unsupported responses. The identical observed distribution leaves their hidden responses distinguishable only through additional information. At L=0.25L=0.25, target containment is 2,000/2,000 in the valid world and 0/2,000 in the violating world; at L=0.45L=0.45, both are contained, with wider intervals.

We therefore report a decision frontier: the largest LL under which adoption survives. Its median is 0.2567 in the beneficial setting. The identical-observation violation world has the same frontier but requires L≥0.45L\ge0.45. The frontier gives the assumption strength required for adoption. A practitioner compares it with a domain-justified range for LL; the observed-cell diagnostic and this sensitivity threshold serve distinct roles.

The matched comparisons show where retaining joint structure improves precision. Each row compares interval width on its own fixed grid; all settings, including ties, remain in the corresponding appendix.

SettingComparisonStrictly narrower / totalTies
Signed targetsJoint LP vs marginal propagation of published envelopes32/320
Partial exposure supportJoint completion vs same-constraint hull6/60
Bounded omitted exposure featureShared nuisance vs rectangular inflation8/102
Stable carryover historiesEqual-block projection vs valid degree calibration12/186
Valid decision settingsJoint vs marginal completion, width metric60/688

These comparisons isolate the gain from joint constraints or dependence calibration under the stated information sets. The nonnegative-target controls recover envelope–LP equality in 32/32 settings. Appendix N.2 retains the named comparisons; N.7–N.11 give the extension grids.

Revision: Section 6 and Appendices N.2 and N.5 (blue).

W3 — Practical and decision value

We added an explicit decision rule, adoption frequencies and regret. The semi-synthetic study uses real covariates from 1,010 households and 505 friendship pairs, with simulated assignments and responses; other network edges are omitted. Historical assignments treat both or neither member; the target policy treats one randomly selected member.

At L=0.25L=0.25, cost 0.05 and 400 independent waves (404,000 unit-wave observations), true net benefit is 0.10:

MethodMean widthAdopt / 2,000Mean regret
Outcome box1.013200.10
Marginal completion0.20642,0000
Joint completion0.19332,0000

Abstention means no deployment. Structural information enables the decision gain over the box; joint optimization adds a 6.38% width reduction. Across all 544 valid cost combinations, marginal and joint completion have identical decisions and regret.

The held-village audit gives mean single-cell width 0.4382, with all six masked means contained. The expanded pooled sample gives 0.3793; the original processed sample gives 0.9219. Different samples preclude attributing this change to the method. Contrast diagnostics contain 12/15 pooled and 13/15 fixed-split differences, with no sign declarations. These descriptive checks assess observed cells; cross-village edges also preclude an independent-split interpretation. They do not establish repeated-sampling coverage for unsupported exposures.

Revision: Section 6 and Appendix N.5 distinguish field diagnostics, conditional decisions and the full cost grid (blue).

We are eager to know whether we have addressed your questions, and we would be happy to discuss further at any time to assist with your re-evaluation.

AuthorsOpenReview ↗

Thanks for the review and here is the response

We sincerely appreciate the time, care, and thoughtfulness you have devoted to reviewing our work. As the deadline for this stage is now less than two days away, we would be more than happy to address any further questions and will respond promptly.

We sincerely hope that our clarifications and additional results have addressed your concerns. If so, we would be truly grateful if you would consider the re-evaluation and rating.

Thank you again for your time, expertise, and invaluable feedback. Your contribution has been instrumental in helping us improve this work.

Thank you for the detailed response and additional comparisons, which have helped me better understand the paper’s contribution and addressed several of my concerns. However, the reported gains from joint optimization remain modest and do not improve decisions or regret over marginal completion in the evaluated settings, so I remain unconvinced that the methodological and practical contribution is sufficient for acceptance. That said, I will raise my rating to 'weak reject' and my confidence score to 4.

Reviewer_qo7MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses network causal inference when randomization constraints prevent certain exposures. It defines a "completion polytope" to establish sharp bounds for the identified set and constructs confidence intervals that cover this entire set. The authors also propose a hybrid method using pilot data to certify estimates, validating the approach through simulations and an insurance network analysis.

Critical strengths
  1. The paper provides a coherent treatment of sharp identification, finite-sample full-set coverage, and matching minimax interval-length bounds.

  2. The common-dual certificate accounts for changes in optimal dual solutions, while independent evaluation preserves coverage after branch selection.

Critical weaknesses
  1. The main framework combines established tools from partial identification, LP duality, IPW, and confidence-region projection. The incremental contributions relative to [1] and [2] need clearer justification.

  2. Global assignment constraints may induce dense score dependency graphs even on sparse networks, making the proposed variance bounds uninformative as the number of units grows.

  3. The experiments evaluate projection and certification largely separately, without demonstrating that the complete hybrid improves interval width or computational efficiency over full-sample projection.

  4. The scientific restrictions require credible calibration, and the field-data projection intervals remain relatively wide (0.922 versus the unrestricted 1.000).

[1] Off-policy Evaluation Beyond Overlap: Sharp Partial Identification Under Smoothness.

[2] Partial Identification of Policy Values under Network Interference.

Questions for authors
  1. How do the identification and inference results extend when an exposure is supported for some units but has zero probability for others?

  2. How sensitive are the results to violations of the assumed exposure mapping, particularly when outcomes depend on assignment features not captured by the exposure labels?

  3. The framework assumes stable, independent experimental waves. Can the inference guarantees be extended to settings with treatment carryover or dependence across waves?

  4. Can the target-specific identification and dual sensitivity results guide the redesign of randomization schemes to reduce identification width under operational constraints?

4 comments
AuthorsOpenReview ↗

Thanks for the review and here is the response

W1 — Contribution relative to existing work

We characterize which policy contrasts remain identifiable under support constraints and the precision attainable for their full identified intervals.

Khan et al. establish sharp Lipschitz policy-value bounds and minimax estimation under non-overlap. Wang et al. extend policy-value identification to network interference and develop network estimation rates, asymptotic inference and misspecification analysis. Both bound policy values using nonnegative policy weights and coupled restrictions. Our contribution concerns arbitrary signed contrasts and general joint linear restrictions, with design-based finite-sample coverage of the entire identified interval and exact regional certification of affine endpoints.

The geometry is simple: for u,v∈[0,1]u,v\in[0,1] with ∣u−v∣≤0.1|u-v|\le0.1, each marginal range is [0,1][0,1], while the sharp range of u−vu-v is [−0.1,0.1][-0.1,0.1]. Joint optimization preserves the constraint linking the two means. This example explains the loss from marginal subtraction; the inferential contribution adds valid uncertainty quantification under the network assignment law.

The theory separates irreducible ambiguity from sampling error. Theorem 3.1 gives squared-error lower bound ω2/4\omega^2/4 for observationally identical worlds whose targets differ by ω\omega. Section 5 establishes matching interval-length order D0+M(1+Δˉ)/(Np⋆).D_0+M\sqrt{(1+\bar\Delta)/(Np_\star)}. The terms quantify identification width and sampling uncertainty under dependence and limited support. The matching result fixes dimension and confidence, bounds endpoint slopes uniformly, and requires the stated lower-bound submodels; Section 5 defines the parameters.

Projection supplies default full-set inference. Theorem 4.2 certifies regional affine exactness for the optional branch; Theorem 5.1 preserves coverage after evaluation or fallback under valid restrictions, bounded outcomes, known probabilities and stable independent waves, allowing nonunique optimizers and failed certificates.

Matched-information tests recover the published envelopes in all 32 nonnegative controls. Joint LP improves on their marginal propagation in all 32 signed-target settings; both contain the full set in all 500 repetitions per setting. This finite-grid bound-formula comparison uses common calibration and is distinct from the original regression pipelines. The covariate formula agrees with Khan's released code.

References: Khan et al. (2024); Wang et al. (2026).

Revision: Section 2 and Appendices N.1–N.3 (blue).

W2 — Dense dependence under global constraints

The assignment law determines whether dense dependence reflects conservative calibration or limited information. On BB disjoint six-node rings and RR independent waves, a half-treated quota with fixed binary response schedules gives variance μ(1−μ)/[(B−1)R]\mu(1-\mu)/[(B-1)R]; one common assignment gives μ2/R\mu^2/R. Only the former improves with BB.

Wave-level Hoeffding projection permits arbitrary within-wave dependence under bounded outcomes, known probabilities and independent stable waves. In the 27-setting study, the quota case at B=1600,R=100B=1600,R=100 gives:

CalibrationWidthFull-set containment
Score-degree bound1.00002,000/2,000
Wave-level bound0.74212,000/2,000
Design-specific bound0.16582,000/2,000

The degree bound is uninformative in this case; wave-level and design-specific calibration give informative intervals. The latter uses the stated quota/fixed-response assumptions. Under common assignment, ignoring dependence gives 68.45% coverage.

Revision: Appendix N.6 gives the variance calculation and complete grid (blue).

W3 — Complete hybrid at the same budget

We compare default projection with the optional certified branch at equal budgets. Complete hybrid costs include pilot construction, all-vertex certification, evaluation and fallback, including failures.

At 400 rings, 2,500 total waves, mean gap 0.30 and 20% pilot, widths are 0.1721 hybrid, 0.1678 full-sample Chebyshev projection and 0.1558 full-sample Hoeffding projection; full-set containment is respectively 1,000/1,000, 1,000/1,000 and 999/1,000. Across all 48 settings and 48,000 repetitions, the hybrid is 1.05%–42.17% wider than full-sample Chebyshev projection. Complete LP times are 15.27 versus 2.95 ms. Full-sample projection gives the better width–cost tradeoff in this benchmark.

The common-dual certificate verifies that one affine formula remains exact throughout the pilot region, and fallback preserves validity when it fails. The scaling grid varies 8–64 cells, 14–240 rows and pilot dimensions 2–8, counting certification failures and fallback.

Section 4 identifies projection as default and certification as optional; Appendix N.4 reports full costs and scaling (blue).

AuthorsOpenReview ↗

Thanks for the review and here is the response (cont'd)

W4 — Calibration and field precision

We provide the calibration rule, expanded diagnostics and a decision-sensitivity threshold. Calibration uses retained cells' maximum difference-to-distance ratio, excludes masked outcomes, and varies multipliers 1, 1.5 and 2. Insurance uptake is binary, giving the outcome range [0,1][0,1]. The smoothness bound LL specifies how far unsupported responses may differ per unit distance.

The held-village test gives mean cell width 0.4382. Pooled expansion gives 0.3793, versus 0.9219 in the original sample; these different samples cannot establish a method improvement. Each contains all six masked means. Paired holdouts contain 28/30 means; contrast checks contain 12/15 pooled and 13/15 fixed-split differences, with no sign declarations. These are descriptive observed-cell diagnostics. Cross-group friendship edges also preclude an independent-split interpretation.

Observationally identical worlds with different hidden responses show why observed-cell checks cannot establish the restriction for unsupported exposures. We report the largest LL under which adoption survives, so a practitioner can compare that threshold with a domain-justified range for LL.

Revision: Section 6 and Appendix N.5 give calibration, diagnostics and decision sensitivity (blue).

Q1 — Partial support and the original population

Partitioning by design-known support retains the original target. Estimate supported stratum–exposure means, complete missing coordinates and aggregate with original population weights. Appendix N.7 proves sharpness for the declared aggregate class and full-set coverage by projecting a simultaneous primitive region. Cross-stratum restrictions are explicit inputs.

With two of six strata supporting target exposures and budget 14,400, widths are 0.3448 joint, 0.4331 same-constraint hull, and 0.5000 after discarding partial cells; all contain the full set in 1,000/1,000 repetitions. Joint beats the hull in all six settings. Separate-arm restrictions alone give envelope–LP equality, isolating the role of coupling. Original population weights are retained.

Revision: Appendix N.7 gives the proposition and proof; N.11 retains all settings (blue).

Q2 — Omitted exposure features

A shared bounded nuisance propagates the omission coherently. Write z=Dv+Bγz=Dv+B\gamma and θ=c⊤v+d⊤γ\theta=c^\top v+d^\top\gamma, with supplied feature-frequency matrices and a declared nuisance set. Appendix N.8 proves coverage and containment within the corresponding rectangular relaxation when the true nuisance belongs to that set.

Across ten valid settings, shared-nuisance intervals improve in eight and tie in two, with full-set containment in all 1,000 repetitions per setting. At true effect 0.10, bound 0.30 and budget 16,000, widths are 0.5227 versus 0.6000. Splitting labels by the observed omitted feature uses richer summaries and is more precise here. Both underestimated-budget cases remain in the table: true effect 0.30 with bound 0.15 gives 21.1% target coverage. The bound defines the covered omission magnitude.

Revision: Sensitivity proof and full grid, Appendices N.8 and N.11 (blue).

Q3 — Carryover and dependent waves

History-indexed completion with a valid covariance bound extends projection. Appendix N.9 derives variance-minimizing linear weights and proves full-set coverage for stable history means. The original hybrid retains its independent pilot/evaluation condition.

The reset two-period experiment excludes history 11 (treated in both periods), targets μ11−μ00\mu_{11}-\mu_{00} and bounds the history interaction by 0.05. At carryover 0.30, 2,000 independent blocks and groups of (1,1,1,64) perfectly correlated copies, equal-block width is 0.3768 versus 0.6410 for valid degree calibration, both with 1,000/1,000 full-set containment. Treating copies as independent gives 54.1% full-set coverage; ignoring carryover gives zero scalar-target coverage. Equal-block weighting improves in 12/18 settings and ties in six under the same records and restrictions. The extension requires stable history means and a justified covariance envelope.

Revision: Weighting proof and experiments, Appendices N.9 and N.11 (blue).

Q4 — Constrained redesign

Target ambiguity yields an exact finite-library design criterion. For completion set CC, measurement map HdH_d and simultaneous radii rdr_d, Appendix N.10 proves

W(d)=max⁡u,v∈C{c⊤(u−v):∣Hd(u−v)∣≤2rd}.W(d)=\max_{u,v\in C}\{c^\top(u-v):|H_d(u-v)|\le2r_d\}.

The two completions remain indistinguishable: zero radii measure identification width; positive radii add sampling uncertainty.

Under cost cap 1.75 and a mandatory 25% quota, minimax allocation improves on the direct-target heuristic in 9/9 settings and on quota-uniform allocation in 2/9, with seven ties. At proxy bound 0.04 (the allowed target–proxy mean difference) and budget 480, worst-case width falls 0.6863→0.5653 relative to quota-uniform allocation.

AuthorsOpenReview ↗

Thanks for the review and here is the response (cont'd)

Both designs are feasible: quota-uniform allocation costs 1.75 and meets the 25% quota. Its population identification width is zero, versus 0.08 for the selected design. Concentrating a finite budget trades persistent ambiguity for less sampling uncertainty. Correct-sign declarations change from 3.0% to 0.9%; the optimized objective is worst-case interval width.

A separate capacity-two calculation permits at most two non-quota exposures. At proxy bound 0.12, two proxies, one direct plus one proxy, and both direct targets give identification widths 0.48, 0.24 and zero. Actual deployment requires realizable network randomization.

Revision: Design proof and full grid, Appendices N.10 and N.11 (blue).

We are eager to know whether we have addressed your questions, and we would be happy to discuss further at any time to assist with your re-evaluation.

AuthorsOpenReview ↗

Thanks for the review and here is the response

We sincerely appreciate the time, care, and thoughtfulness you have devoted to reviewing our work. As the deadline for this stage is now less than two days away, we would be more than happy to address any further questions and will respond promptly.

We sincerely hope that our clarifications and additional results have addressed your concerns. If so, we would be truly grateful if you would consider the re-evaluation and rating.

Thank you again for your time, expertise, and invaluable feedback. Your contribution has been instrumental in helping us improve this work.

Reviewer_M1i4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper characterizes the identifiable range of network policy effects when some exposure states are unobservable under the current experimental design. It constructs confidence intervals with finite-sample coverage guarantees and examines what additional information can narrow these bounds.

Critical strengths

1 The paper provides extensive and rigorous theoretical proofs. In particular, it offers a comprehensive and interpretable characterization of identification under structural support deficiency.

2 The paper provides both theoretical and experimental support for the statistical reliability of the proposed method in finite samples.

Critical weaknesses

1 The authors prove that their approach of “checking first, then choosing how to calculate the interval” is theoretically valid. However, the experiments do not clearly show how much narrower the intervals become compared with the basic method when both use the same amount of data. They also do not show how much additional computation this requires.

2 In the insurance data experiment, how do the authors determine the maximum possible difference between the unobserved group and the observed groups? Small differences among the observed groups do not guarantee that the hidden group follows the same pattern. The real-data experiments do not sufficiently explain how to set these restrictions in practice so that the intervals include the true value while remaining narrow enough to be useful.

Questions for authors

1 The full method uses one batch of data for the initial check and another independent batch to calculate the final interval. Does the paper compare this approach with using all the data for the basic method, in terms of interval width and coverage? If not, what evidence currently supports the practical advantage of the full method?

2 The paper mainly tests this check in a simplified example. How does its runtime increase as the number of groups and restrictions grows? Do the existing experiments or computational analysis show whether the narrower intervals justify this additional cost?

3 The authors use differences among the observed groups to restrict the possible outcomes of an unobserved group. However, small differences among observed groups do not guarantee that the unobserved group follows the same pattern. Does the paper provide any other evidence supporting this choice?

2 comments
AuthorsOpenReview ↗

Thanks for the review and here is the response

W1 / Q1 — Optional certified branch versus all-data projection

Projection is the framework's default; certification is an optional branch for exact affine endpoint evaluation. We compare both at identical total budgets and nominal 95% full-set coverage. The complete hybrid includes pilot construction, all-vertex certification, independent evaluation and fallback, retaining failures.

Across 48 settings and 48,000 repetitions, hybrid width exceeds full-sample Chebyshev width by 1.05%–42.17%. At 400 rings, 2,500 total waves, mean gap 0.30 and 20% pilot:

MethodMean widthFull-set containment
Complete hybrid0.17211,000/1,000
Full-sample Chebyshev0.16781,000/1,000
Full-sample Hoeffding0.1558999/1,000

Median complete LP times are 15.27 versus 2.95 ms for hybrid and full projection, including estimation, certification and evaluation/fallback, excluding generation and I/O. The 95% Monte Carlo intervals are [99.617%,100%] for 1,000/1,000 and [99.436%,99.982%] for 999/1,000.

The certificate verifies one affine endpoint formula across the whole pilot region; independent evaluation and fallback preserve coverage. The measured splitting and error-allocation costs favor full-sample projection for width and speed in this benchmark.

Revision: Section 4 clarifies the default and optional branches; Appendix N.4 reports complete cost and coverage (blue).

W2 / Q3 — Restrictions for the hidden group

We state the scientific input and report how each decision changes as the restriction is relaxed. The field rule uses the largest difference-to-distance ratio among retained cells, excluding masked outcomes, with sensitivity multipliers 1, 1.5 and 2. Insurance uptake is binary, giving the outcome range [0,1][0,1]. The bound LL specifies the allowed response difference per unit distance, including unsupported exposures; the tighter interval is conditional on that scientific bound.

A controlled construction makes the restriction's role explicit. Two worlds have identical observed distributions but different hidden responses. At L=0.25L=0.25, target containment is 2,000/2,000 in the valid world and 0/2,000 in the violating world; L=0.45L=0.45 covers both, with wider intervals. Observed-cell evidence is therefore identical despite different valid restrictions: the hidden-group bound requires scientific information beyond that observed distribution.

We added the largest LL for which the lower benefit bound exceeds cost. In the beneficial 400-wave, cost-0.05 setting, its median is 0.2567, with 5th–95th percentiles [0.2537,0.2600]. The identical-observation violation world has the same threshold but requires L≥0.45L\ge0.45, outside the adoption range. This gives a concrete reporting procedure: supply a scientifically justified range separately and show exactly which decisions survive it. The outcome-range analysis applies when that range supplies no further information.

The held-village audit gives mean cell width 0.4382; the expanded pooled sample gives 0.3793. Both contain all six masked means. Paired holdouts contain 28/30 means; contrast checks contain 12/15 pooled and 13/15 fixed-split differences. These are descriptive observed-cell diagnostics, with different samples and cross-village dependence.

Revision: Section 6 and Appendix N.5 give the calibration rule, identical-observation construction and sensitivity frontier (blue).

Q2 — Growth with groups, restrictions and pilot dimension

The new complete scaling benchmark varies all three dimensions. It covers stable and switching faces, eight paired trials per setting and randomized execution order. The common Hoeffding calibration isolates scaling; the original hybrid comparison above retains its original calibration.

Stable-face configurationFull projection, msComplete certificate, ms
8 cells, 14 rows, pilot dimension 24.9623.92
64 cells, 126 rows, pilot dimension 27.0435.87
16 cells, 30 rows, pilot dimension 25.5526.22
16 cells, 240 rows, pilot dimension 28.1242.09
16 cells, 30 rows, pilot dimension 85.721,274.81

An ss-dimensional pilot box has 2s2^s vertices; the generic implementation may require 2(2s+1)2(2^s+1) certificate LPs before evaluation. Switching cases exercise fallback. These timings include certification and fallback. The 2s2^s vertex count explains the growth with pilot dimension; certification supplies a verified affine representation, while projection has the lower computational cost in this grid.

Revision: Appendix N.4 includes every setting, pass rate, width, residual check and complete timing (blue).

Reviewer_M1i4OpenReview ↗

Thanks for your rebuttal.

Thanks for your rebuttal. My concerns are generally addressed. I will keep my rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers

General Response to the Area Chair and Reviewers

We thank the reviewers for their careful feedback. Our responses and manuscript revisions are organized into four modules. References such as qo7M-Q1 identify the corresponding sections of our reviewer-specific replies.

  1. Clarification of the scientific question and contributions:

    • Clarified how repeated constrained experiments reduce sampling error while leaving uncertainty about unsupported exposures. We distinguish target-specific signed-contrast identification, finite-sample coverage of the entire identified interval, and exact regional common-dual certification from prior policy-value bounds and asymptotic inference (swaj-W1, qo7M-W1, Section 2 and Appendix N.1).
  2. Matched comparisons and complete-procedure evaluation:

    • Added comparisons using identical observations, restrictions and confidence bands: joint optimization gives narrower intervals in 32/32 signed-target settings, while matching the published envelopes in all 32 nonnegative-target controls (swaj-W1–W2, qo7M-W1, Appendices N.2–N.3).
    • Evaluated the complete hybrid against all-data projection at equal total budgets, reporting interval width, full-set containment and runtime, including pilot construction, certification and fallback. Scaling tests vary cells, restrictions and pilot dimension, clarifying the roles of default projection and optional certification (M1i4-W1/Q1, M1i4-Q2, qo7M-W3, Section 4 and Appendix N.4).
  3. Extensions to broader designs and exposure uncertainty:

    • Added theoretical arguments and experiments for dense assignment dependence, unit-specific exposure support, omitted exposure features, carryover and dependent waves. Joint completion improves on the same-constraint hull in 6/6 partial-support settings, while preserving the original population target (qo7M-W2, qo7M-Q1–Q3, Appendices N.6–N.9 and N.11).
    • Derived a target-specific criterion for comparing feasible randomization designs under cost and quota constraints (qo7M-Q4, Appendices N.10–N.11).
  4. Scientific calibration, field diagnostics and decision value:

    • Specified the restriction-calibration rule, expanded the held-village diagnostics, and reported the largest smoothness bound under which adoption remains supported. This makes the scientific input and the decision's sensitivity explicit (M1i4-W2/Q3, qo7M-W4, swaj-W2, Section 6 and Appendix N.5).
    • Added an explicit deployment rule, adoption frequencies and regret in the semi-synthetic insurance study, alongside joint-versus-marginal width comparisons. The replies distinguish field diagnostics, conditional decision gains and precision gains from joint optimization (swaj-W3, Section 6 and Appendix N.5).

Thank you again to the Area Chair for your time and thoughtful consideration. We warmly welcome further questions and discussion, and would be happy to clarify any aspect of the theory, experiments, or revisions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.