- 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
-
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
-
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.
-
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
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 . The bound 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 , target containment is 2,000/2,000 in the valid world and 0/2,000 in the violating world; at , both are contained, with wider intervals.
We therefore report a decision frontier: the largest under which adoption survives. Its median is 0.2567 in the beneficial setting. The identical-observation violation world has the same frontier but requires . The frontier gives the assumption strength required for adoption. A practitioner compares it with a domain-justified range for ; 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.
| Setting | Comparison | Strictly narrower / total | Ties |
|---|---|---|---|
| Signed targets | Joint LP vs marginal propagation of published envelopes | 32/32 | 0 |
| Partial exposure support | Joint completion vs same-constraint hull | 6/6 | 0 |
| Bounded omitted exposure feature | Shared nuisance vs rectangular inflation | 8/10 | 2 |
| Stable carryover histories | Equal-block projection vs valid degree calibration | 12/18 | 6 |
| Valid decision settings | Joint vs marginal completion, width metric | 60/68 | 8 |
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 , cost 0.05 and 400 independent waves (404,000 unit-wave observations), true net benefit is 0.10:
| Method | Mean width | Adopt / 2,000 | Mean regret |
|---|---|---|---|
| Outcome box | 1.0132 | 0 | 0.10 |
| Marginal completion | 0.2064 | 2,000 | 0 |
| Joint completion | 0.1933 | 2,000 | 0 |
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.
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.
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] with ∣u−v∣≤0.1, each marginal range is [0,1], while the sharp range of u−v is [−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 for observationally identical worlds whose targets differ by ω. Section 5 establishes matching interval-length order D0+M(1+Δˉ)/(Np⋆). Here D0 bounds identification width, M bounds outcomes, N counts unit-wave observations, and Δˉ,p⋆ 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).