{"id":"873aca81-ea52-41a2-94c4-35f06515e13c","arxiv_id":"2607.04644","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new class of linear weighting estimators provides unbiased, asymptotically normal inference for causal effects in two-stage cluster randomized experiments with cross-cluster interference.","lead":"This paper develops a general framework for estimating causal effects in cluster-randomized trials where units influence each other across clusters, without assuming a specific exposure mapping. It proposes new weighting estimators with statistical guarantees, including under complete randomization, and recommends one that outperforms earlier methods in simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unbiasedness and all downstream estimators depend on correct interference neighborhoods (Assumption 3.3); the paper's simulations never test omitted-edge misspecification, so the practical 'exposure-mapping-agnostic' claim is only as strong as the network.","rationale":"The reader's weakest assumption—Assumption 3.3 (known network and neighborhood interference)—is indeed the most load-bearing condition for the paper's central characterization and estimation results. The paper is internally consistent: Theorem 3.6, Theorem 3.7, the variance-rate results, and the CLTs all explicitly condition on this assumption, and the proofs are elaborate and plausible. The simulations, however, always use the true interference network to define neighborhoods, so they provide no evidence about the practical cost of network misspecification. The paper claims improved robustness to misspecification through conservative neighborhoods, but this only covers the direction where the assumed network contains extra edges; omitted-edge misspecification is not addressed. Because the central practical selling point is 'without exposure mapping assumptions,' the boundary condition should be stated as clearly as the exposure-mapping assumption it replaces. This does not invalidate the paper's conditional theoretical contributions, so the reader's ACCEPT verdict with moderate confidence remains appropriate. A targeted misspecification simulation would settle whether the practical scope claim needs qualification.","tokens_in":67730,"tokens_out":21058,"duration_ms":194320,"concrete_test":"Run the §6 simulation with assumed neighborhoods N_ij^assumed equal to the geometric graph only, while the outcome-generating model uses the full geo ∪ ER graph (ρ = 0.5 or 1.0) for spillover terms. Compare bias, RMSE, and CI coverage of the MRN estimator against the correctly specified run (where assumed neighborhoods include the ER edges). If bias grows with ρ and coverage drops below nominal, Assumption 3.3 is genuinely load-bearing and the manuscript should explicitly state that unbiasedness claims require the true interference graph. If the MRN estimator remains approximately unbiased despite the omitted edges, the robustness claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The entire identification theory hinges on Assumption 3.3: each unit's potential outcome depends on treatment only through the known neighborhood N_ij. Theorem 3.6 characterizes weights that are unbiased for all Y_ij in L(W_Nij). If the true network has an unmodeled edge, Y_ij depends on treatments outside W_Nij, the change-of-measure identity E[Y_ij β_ij] = E_ϕ[Y_ij] can fail, and the iff condition E[β_ij | W_Nij] = α_ij(ϕ) h(W_Nij) no longer guarantees unbiasedness for the actual outcome. All asymptotic results (Theorems 4.15, 4.19), variance estimators (§5), and simulations (§6) take the network as fixed and correctly specified. The paper's stated robustness via 'conservative specifications' (adding extra connections) addresses only the case where the assumed network is larger than the true one; the harder and more common misspecification—omitted edges—is not analyzed or simulated. Since the abstract and introduction emphasize the absence of exposure-mapping assumptions, users may infer robustness to unknown interference structure that the theory does not provide. This is not an internal inconsistency, but it is the most load-bearing condition for the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a design-based causal inference framework for two-stage randomized experiments with interference on a known network. It defines population-level estimands under counterfactual uniform treatment regimes, characterizes the full class of linear weighted estimators that are unbiased for these estimands over all potential outcome functions of the neighborhood assignment (Theorem 3.6), and provides a complementary characterization of cluster-agnostic weights that are unbiased conditional on cluster assignments (Theorem 3.7). The paper then derives variance-rate bounds (Theorem 4.15) and central limit theorems (Theorem 4.19) for general and cluster-agnostic estimators under Bernoulli and complete randomization, using a novel Stein-method coupling for complete randomization. Conservative and bias-corrected variance estimators are developed in Section 5, and simulations in Section 6 compare the proposed estimators with difference-in-means, IPTW, and cluster-agnostic alternatives. The central claims are that the framework avoids exposure-mapping assumptions and that a subclass of estimators attains the parametric root-N rate.","tokens_in":68041,"tokens_out":18507,"duration_ms":162868,"significance":"If the results hold, this is a substantial contribution to the causal-inference-under-interference literature. The Riesz-representation/change-of-measure characterization of unbiased weights is clean and unifies several existing estimators, including Leung (2025), as special cases. The Stein-coupling argument for complete randomization appears genuinely novel and may be of independent interest; the paper is also unusually honest about the limits of its techniques (Remark 4.22 explicitly flags where complete randomization at the unit level cannot be handled). The proof sketches are detailed and, for the main identification theorems, rigorous. The simulation study is extensive and demonstrates tangible gains for the marginal Radon–Nikodym estimator. The main scope limitation—correct specification of the interference network—is acknowledged in the introduction, but the paper would benefit from a more prominent statement that the framework is network-dependent, not network-agnostic.","major_comments":[{"comment":"There is an apparent internal inconsistency between Theorem 4.19(ii) and Remark 4.21. The theorem states that the randomly normalized statistic Σ(C)^{-1/2}(ˆτ−τ) converges unconditionally to N(0,I), and the proof integrates conditional Wasserstein bounds to obtain exactly that. Remark 4.21, however, says that unconditionally the asymptotic distribution is a normal mixture and can be multimodal. These two statements cannot both refer to the same normalized object; the proof shows the mixture components are all N(0,I), so the unconditional limit is also N(0,I). The remark should be reworded to specify that the mixture claim applies to the unstandardized estimator (or to the finite-sample distribution), not to the Σ(C)^{-1/2}-standardized statistic. As written, this is a load-bearing point because it directly concerns the interpretation of the main CLT result.","section":"Theorem 4.19(ii) and Remark 4.21"},{"comment":"The entire identification, asymptotic, and inference theory is conditional on Assumption 3.3, which requires the interference neighborhoods N_ij to be known and correctly specified. The paper's robustness discussion in §1 properly mentions conservatively adding edges, but it does not analyze or simulate omitted-edge misspecification, which is the more practically concerning failure mode: if the true network has an unmodeled edge, Y_ij depends on treatments outside W_Nij and the key identities E[Y_ij β_ij | W_Nij] = α_ij h no longer guarantee unbiasedness. Given the abstract's emphasis on 'without relying on exposure mapping assumptions,' readers may infer robustness to unknown interference structure that the theory does not provide. I recommend adding an explicit scope statement and, ideally, a simulation or sensitivity discussion for omitted edges. This is not an internal error, but it","section":"Assumption 3.3 and §6"}],"minor_comments":[{"comment":"The displayed adjustment condition is typeset incorrectly: 'sgn(Δ) ϵ_i = |Δ|' should read '∑_{i∈I_k\\N^cl} (2C_i−1) ϵ_i = sgn(Δ) |Δ|'. The surrounding text suggests the intended meaning, but the current display is confusing.","section":"Proposition 7.3, coupling construction"},{"comment":"The notation L(R) defined as 'Lebesgue-integrable functions of R' is nonstandard when R is discrete with finite support, where every function is integrable. Consider replacing with 'the space of all real-valued functions on the support of R' to avoid confusion.","section":"Section 3, notation"},{"comment":"For the CRN estimator under dense interference (ρ=1.0), empirical coverage using the oracle SE is extremely low (e.g., 0.082 at N=2000). The paper explains this via heavy-tailed weights, but since the CRN estimator is a headline theoretical contribution, a one-sentence recommendation against its use in dense-interference finite samples would help practitioners.","section":"Section 6, Table 2"},{"comment":"The paper states it develops the 'general cluster exppackage' but provides no link, repository, or pseudocode. If this is part of the contribution, please provide details; otherwise, it can be omitted to avoid an unverifiable claim.","section":"Section 1, contributions"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is technically strong and the main theorems appear sound. The two issues I raise—the Remark 4.21 ambiguity and the scope caveat on network specification—are both fixable without new technical work. The Remark 4.21 point is worth fixing before publication because it directly concerns the interpretation of the central CLT theorem. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a substantial, serious paper. The characterization of unbiased linear weighted estimators for two-stage designs under interference is a genuine advance over Harshaw et al., and the complete-randomization central limit theorem via the new Stein coupling is the kind of technical contribution that will be read independently. The reconciliation of the n^-1/2 vs N^-1/2 rates is clean and useful. Credit where earned: they prove a full characterization, handle complete randomization where previous techniques stalled, give conservative and bias-corrected variance estimators with closed-form corrections, and their simulations back up the theoretical comparisons. They also flag where they can't deliver (Remark 4.22). This is good work.\n\nThe soft spot is the network. The stress-test note is right: Assumption 3.3 is load-bearing. Unbiasedness fails if the true network has omitted edges, and the 'conservative specification' argument only covers the case where you assume a larger network than the true one—the harder case of omitted edges is never analyzed or simulated. The abstract's emphasis on being 'exposure-mapping-agnostic' may mislead readers into thinking the method is also network-agnostic; it isn't. The assumption is standard and stated, so this is not an internal inconsistency, but it should be surfaced more prominently. A simulation with an omitted-edge misspecification would strengthen the practical message.\n\nThe proofs are detailed but not machine-checked; the Stein coupling is intricate and should get careful referee scrutiny. That said, the proof structure is transparent and Remark 4.22 honestly admits a case the technique can't handle. The lack of code is a minor issue given the technical content.\n\nBottom line: the reader's ACCEPT is defensible. This deserves a serious referee and, assuming the complete-randomization proofs survive scrutiny, publication. I would cite it if I work on interference. I'd maybe bring it to a reading group — it's long, but the coupling section is worth the effort.","headline":"A serious, valuable theory paper; the load-bearing caveat is the known-network assumption, and the complete-randomization CLT needs careful referee scrutiny.","tokens_in":68483,"tokens_out":2778,"would_cite":true,"duration_ms":29894,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D05","62G20","60F05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under a known interference network and mild measure transportability, the complete class of unbiased linear weighted estimators for causal effects under interference is characterized by a single moment condition on the weights — no exposure","keywords":["causal inference","interference","two-stage randomized experiments","exposure mapping","design-based inference","complete randomization","linear weighted estimators","central limit theorem"],"falsifier":"Run a two-stage randomized experiment on a network where one hidden edge connects a unit to a treatment outside its declared neighborhood; compute the marginal Radon–Nikodym estimator over many randomizations and check whether the average estimate equals the true potential outcome. Any systematic bias exceeding sampling error would refute the neighborhood-interference assumption, which the characterization relies on. Alternatively, in an empirical trial with recorded cross-cluster ties, compare the proposed estimator with one using a richer network: disagreement beyond the estimated standard e","tokens_in":67647,"feed_emoji":"🕸️","tokens_out":4861,"duration_ms":43678,"temperature":0.7,"pith_summary":"The paper establishes that, in two-stage randomized cluster experiments with possible cross-cluster interference, population-level causal effects can be estimated without specifying an exposure mapping. Its central result is a complete characterization of all unbiased linear weighted estimators: a weight on a unit's outcome is unbiased exactly when its conditional expectation given the neighborhood treatment vector equals the marginal Radon–Nikodym derivative of the counterfactual assignment law. A subclass of cluster-agnostic weights removes the dominant cluster-level dependence and attains the root-N convergence rate, and the paper supplies conservative variance estimators and central limit theorems, including under complete randomization. If correct, the framework turns a previously model-dependent problem into a design-based one, with estimators that outperform standard inverse-probability weighting in simulations.","feed_headline":"One moment condition fully characterizes unbiased effect estimators","feed_subtitle":"Exposure-mapping-free estimators for interference achieve root-N rates and valid inference.","key_machinery":"The central object is the class of linear weighted (LW) estimators with weights that may depend on the local neighborhood treatment vector W_Nij and the local cluster assignment vector C_Nij. The load-bearing identity is the moment condition E[β_ij | W_Nij] = α_ij(φ) h(W_Nij), which characterizes exactly when multiplying an outcome by a weight reweights the observed randomization distribution to a counterfactual regime. The asymptotic analysis for complete randomization relies on a new coupling construction: for each unit, a coupled copy of the full treatment array that matches the original distribution, is independent of that unit's neighborhood, and differs on only a vanishing fraction of","core_discovery":"The paper proves that, for any treatment regime, an estimator of the form Y_ij β_ij(W_Nij, C_Nij) is unbiased for the average potential outcome for all possible outcome functions in L(W_Nij) if and only if the weight satisfies E[β_ij | W_Nij] = α_ij(φ) h(W_Nij), where α_ij(φ) is the Radon–Nikodym derivative of the counterfactual neighborhood assignment law with respect to the observed one (Theorem 3.6). It further characterizes a cluster-agnostic subclass whose weights equal the complete Radon–Nikodym derivative and which remove the dominant cluster-level dependence, yielding variance of order N^{-1} and asymptotically normal, conditionally-on-cluster, limits (Theorems 3.7, 4.15, 4.19). For","pith_inferences":["The characterization suggests a specification test: comparing exposure-mapping-agnostic estimators with exposure-mapping-based ones could reveal misspecification of the mapping, akin to a Hausman test; the paper hints at this but does not develop it.","If the root-N cluster-agnostic rate holds generally, then under regimes where the number of clusters grows slowly relative to units, cluster-agnostic weights should be preferred; this offers a design heuristic not explicitly stated.","The coupling technique could transfer to other finite-population sampling problems with complete randomization and non-symmetric summands, such as network interference in observational studies with known assignment mechanisms.","The strongest practical caveat is the known-network assumption; conservative network specification (adding plausible edges) is suggested to mitigate misspecification, but the paper does not quantify how much conservativeness is needed."],"forward_implications":["Researchers can estimate population-level direct, indirect, total, and overall effects in two-stage randomized trials with cross-cluster interference without committing to an exposure mapping, using weights derived solely from the known network and assignment probabilities.","A subclass of cluster-agnostic weights removes the dependence induced by cluster-level treatment, achieving root-N convergence rates even when cluster sizes grow; practical guidance is given for when to prefer them over general weights.","Conservative variance estimators, including a bias-corrected variant under complete randomization, make valid confidence intervals available for the proposed estimators across a wide range of designs.","The coupling-based central limit theorem provides a general tool for asymptotic normality of sums of non-symmetric, densely dependent statistics under complete randomization, beyond the specific estimators here.","In simulations, the marginal Radon–Nikodym derivative estimator consistently dominates inverse-probability-of-treatment weighting and difference-in-means, with the advantage growing under strong cross-cluster interference and complete randomization."],"fun_headline_variants":["One weight condition unlocks unbiased interference estimates","Unbiased interference estimators: one Radon-Nikodym condition settles it","Root-N rate achieved by cluster-agnostic weights under interference","Characterizing all unbiased linear estimators under interference"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire identification collapses if the interference network is misspecified: if any unit's outcome depends on treatments outside its declared neighborhood, the change-of-measure identity E[Y_ij β_ij] = E_φ[Y_ij] no longer holds, and unbiasedness is lost.","fun_headline_variants_meta":{"raw":{"variants":["One weight condition unlocks unbiased interference estimates","Unbiased interference estimators: one Radon-Nikodym condition settles it","Root-N rate achieved by cluster-agnostic weights under interference","Characterizing all unbiased linear estimators under interference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000788,"raw_usage":{"total_tokens":3344,"prompt_tokens":812,"completion_tokens":2532,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2467}},"tokens_in":556,"tokens_out":2532,"duration_ms":14494,"temperature":1.0,"reasoning_tokens":2467,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T00:47:05.846923+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a two-stage randomized experiment on a network where one hidden edge connects a unit to a treatment outside its declared neighborhood; compute the marginal Radon–Nikodym estimator over many randomizations and check whether the average estimate equals the true potential outcome. Any systematic bias exceeding sampling error would refute the neighborhood-interference assumption, which the characterization relies on. Alternatively, in an empirical trial with recorded cross-cluster ties, compare the proposed estimator with one using a richer network: disagreement beyond the estimated standard e","supporting_citations":[],"review_version":2}