CausalSmith — CSWP-2026-006 · ← back to the paper · arXiv · GitHub

AI referee report

on “Graph-Adaptive Bernoulli Design for Bipartite Interference”

Overall score 6.2/10 8 findings — 3 major · 4 minor · 1 nit

The formal results appear internally sound and carefully caveated, but the paper does not yet demonstrate enough econometric value or practical relevance for a leading journal.

The paper may have been revised since this report. One automated referee’s estimate — not peer review. Produced by a single automated referee in the pipeline's review stage. It can be wrong, and it is not peer review.

Summary

The paper develops a graph-only variance-envelope criterion for heterogeneous independent Bernoulli assignment in bipartite interference settings, proves convexity/KKT characterization, CLT and conservative Wald coverage under bounded graph conditions, and gives a bounded-degree surrogate certificate with a negative unbounded-degree result. The manuscript is unusually careful in matching prose to formal claims, and most earlier overclaiming has been corrected. The main remaining weakness is not mathematical soundness but contribution: as written, it proves a conservative design rule without evidence that the rule improves realized precision, coverage, exposure support, or implementability in empirically relevant graphs.

Strengths

The core formal statements are coherent, finite-population/design-based, and mostly represented faithfully in the prose.
The paper is honest about the envelope not being a sharp minimax or schedule-specific variance criterion.
The heterogeneous-probability extension, convex design characterization, separation diagnostic, and surrogate certificate form a clean theoretical package.
The discussion now gives a useful design-stage checklist and makes the graph-scope limitations explicit.

Findings

major abstract; Introduction; Discussion and extensions

The paper still asks readers to care about a new experimental design rule without showing that it improves any decision-relevant object in realistic bipartite experiments. The manuscript says the criterion is useful because it is observable, convex, and conservative, but gives no numerical evidence on realized variance, interval width, exposure frequencies, coverage, sensitivity to the positivity floor, or computation time.

Suggested fix

Add a simulation and computational section comparing homogeneous Bernoulli, the exact envelope optimizer, the surrogate, and at least one clustered or graph-cluster benchmark on representative bipartite graph classes. Report envelope values, realized design variance under several potential-outcome schedules, exposure-probability diagnostics, Wald coverage/length, runtime, and sensitivity to epsilon and budget.

major Introduction and positioning

The novelty case remains too narrow for a leading econometrics journal. The manuscript now correctly states that it does not prove a sharp minimax design theorem, but that leaves the central contribution as optimizing a conservative upper bound whose tightness and practical ranking properties are unknown.

Suggested fix

Either add a theorem or examples characterizing when the envelope is sharp/near-sharp for relevant bounded-outcome classes, or explicitly reposition the paper as a methodological note on a conservative graph diagnostic rather than as a full design solution. The introduction should state what econometric decision problem the envelope solves and what it deliberately does not solve.

major Discussion and extensions

The design checklist is helpful but still not operational enough for applied use. It tells readers to prespecify acceptable floors, tolerances, and exposure-probability minima, but gives no default diagnostics or calibration guidance for when the envelope is too loose or when clustering should dominate.

Suggested fix

Add a reproducible protocol with concrete reported quantities and decision thresholds, even if framed as recommended diagnostics rather than universal constants. At minimum, define an envelope-looseness diagnostic, budget/floor sensitivity table, KKT tolerance, exposure-support rejection rule, and a rule for declaring the surrogate certificate vacuous.

minor Setup and assumptions, Definition 10; Main results, Wald theorem

The object is still titled “Graph-only conservative variance scale” but labeled as Definition “Conservative variance estimator,” even though the definition is exactly $V_{cons} (G_{n}, p) = V_{env} (G_{n}, p)$ , a deterministic pre-randomization scale. This terminology can mislead readers into expecting a realized data-based variance estimator.

Suggested fix

Rename the object throughout as “conservative graph scale” or “deterministic conservative Wald scale,” reserve “estimator” for random data-dependent objects, and update theorem titles and interval discussion accordingly.

minor Appendix, Dependency-graph normal approximation and convex-program provenance

The appendix now names Baldi–Rinott and Boyd–Vandenberghe, but the imported mathematical inputs are still too compressed for a conventional econometrics reader. The phrase “under the usual measurability and integrability conditions” is especially imprecise in a theorem-like lemma.

Suggested fix

State the exact measurability, probability-space, boundedness, dependency-neighborhood, and variance-growth hypotheses in ordinary mathematical terms, and cite the precise dependency-graph CLT version being specialized. For the KKT argument, cite the exact constraint qualification or normal-cone theorem used for affine equality plus box constraints.

minor global

The caveat that the envelope does not rank realized variances is accurate but repeated too often, including in the abstract, introduction, main-results preamble, CLT discussion, and discussion. The repetition makes the manuscript read defensively.

Suggested fix

State the limitation once in the abstract, once in the introduction, and then refer back to it where necessary. Use the saved space for diagnostics, examples, or sharper positioning.

minor Related work and bibliography

The paper’s design-based lineage is underdeveloped. The estimator and inference discussion invoke Hájek normalization, Horvitz–Thompson weighting, Neyman finite-population logic, and interference relaxations of SUTVA, but the main text cites mainly recent network and bipartite papers.

Suggested fix

Add concise references to the foundational randomization-inference and interference literature, including Neyman, Horvitz–Thompson, Hájek, Cox/Rubin, Hudgens–Halloran, and Liu–Hudgens where relevant. Keep this short and use it to clarify what is new relative to established design-based inference.

nit Appendix

The appendix structure is somewhat bulky: “Appendices,” “Proofs and auxiliary lemmas,” and “Proofs of the main results” create unnecessary layering.

Suggested fix

Use cleaner appendix headings, for example “Auxiliary Probability Results,” “Convex-Program Details,” and “Proofs of Main Results.”

Questions for the authors

How tight is the graph envelope in graph classes you expect applied researchers to encounter, and can you report worst-case or empirical looseness diagnostics?
When the exact envelope optimizer puts many coordinates on the positivity bounds, how should practitioners interpret or constrain such designs beyond the formal feasibility condition?
Can the method be compared directly with the homogeneous Bernoulli design of Lu et al. and with at least one bipartite clustering design on the same graphs and estimands?