{"id":"df9c189a-88b5-4c51-a02e-e1fac504d5c3","arxiv_id":"2603.19480","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Optimal regression adjustments for MRD marketplace estimators minimize asymptotic variance among linear imputation estimators, are data-estimable without outcome linearity, and improve inference via new CLTs.","lead":"This paper derives optimal linear covariate adjustments for total, direct, and spillover effect estimators under multiple randomization designs in two-sided marketplaces. The adjustments are model-robust, estimable from data, and deliver substantial efficiency gains over unadjusted or classical ANCOVA estimators.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The reader correctly isolates the strongest claim (estimability + model-robust asymptotic normality of the non-interacted plug-in) and the weakest supporting assumption (the \theta(I^{-1}) lower bound and Gram invertibility). The technical development—explicit quadratic forms for Var(\taû_c(eta)), the projection property of the within-cell estimators, the Wasserstein CLT that removes the earlier boundedness and variance-floor hypotheses, and the residual-plug-in conservative intervals—is careful and self-contained. Simulations confirm the no-harm property and substantial efficiency gains. The limitations (interacted class deferred, non-standard scalings only partially treated) are explicitly stated. No further load-bearing concern arises that would move the verdict from ACCEPT.","tokens_in":55287,"tokens_out":522,"duration_ms":10363,"concrete_test":"Reproduce the direct-effect simulation of Example 3.1 / Figure 2 under the vanishing-row-mean regime \theta(I^{-1}) for \theta_B(y(\theta)) while keeping \theta_BS invertible; if the Monte-Carlo coverage of the residual-plug-in interval (43) drops below 90 % or the efficiency gain of Opt over ANCOVA disappears, the special-case extension of Proposition 3.2 fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (min-variance non-interacted imputation coefficients are estimable from observed double-decentered moments alone, and the plug-in is model-robustly asymptotically normal for \tau_c) holds under the paper’s stated conditions. The only material soft spot is already flagged by the reader: Assumption 4 (Var(\taû_c(etã_c))=\theta(I^{-1}) plus invertibility of the limiting buyer/seller Gram matrices). When buyer- or seller-mean variation vanishes faster, or the Gram matrix is singular, both consistency of etâ_c and the conservative intervals can fail; the paper itself isolates the direct-effect \theta((IJ)^{-1}) case and notes that a full multi-regime theory remains open. No internal inconsistency, circularity, or unacknowledged gap appears in the derivation of Z̃_c, ũ_c, the plug-in CLT, or the improved Wasserstein CLT for doubly-randomized sums.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper develops model-robust regression adjustments for Multiple Randomization Designs (MRDs) in two-sided marketplaces. Among non-interacted linear imputation estimators of the form τ̂_c(β)=∑_γ c_γ (1/(I_γ J_γ)) ∑_{(i,j)∈γ}(y_ij−X_ij^\topβ), it derives the asymptotic-variance-minimizing coefficient β̃_c as the solution of an explicit positive-definite quadratic form whose Gram and cross-moment matrices are population moments of covariates and potential outcomes. These matrices are estimable from within-cell double-decentered sample moments, yielding a plug-in estimator that is asymptotically normal for the finite-population target τ_c without any linearity assumption on potential outcomes (Theorem 3.1, Propositions 3.1–3.2). For the direct effect the plug-in coincides with a weighted interacted two-way fixed-effects regression; for total and spillover effects the optimal adjustment is generally not a simple OLS form. The paper also supplies improved Wasserstein CLTs and consistent conservative variance estimators for doubly-randomized sums, and demonstrates efficiency gains over unadjusted and ANCOVA estimators in simulations.","tokens_in":55540,"tokens_out":1512,"duration_ms":22325,"significance":"If the results hold, the paper supplies the natural design-based analogue of Lin/ToM-style regression adjustment for marketplace experiments under local interference—an important practical gap, since MRDs often suffer from low power when one side of the market is small. The derivation that the optimal coefficient is identifiable from observed double-decentered moments alone, without unobservable potential-outcome contrasts, is non-obvious and useful. The improved finite-population CLT for doubly-randomized sums (Theorem 3.3) and the accompanying variance-consistency result (Theorem 3.4) are of independent interest and relax boundedness and θ(I^{-1}) lower-bound assumptions used in prior MRD theory. Simulations show material efficiency gains, especially in imbalanced designs. The work is carefully situated relative to Freedman, Lin, Li–Ding, and Lu–Liu, and the finite-population randomization framework is rigorous.","major_comments":[{"comment":"Assumption 4 (and the companion lower bound used for the improved CLT) is load-bearing for plug-in consistency of β̂_c and for validity of the conservative intervals (Propositions 3.1–3.2, Theorem 3.2). The paper correctly isolates the direct-effect θ((IJ)^{-1}) regime in Example 3.1 and flags multi-regime theory as open (Introduction, §5). For a methods paper aimed at practitioners, however, the manuscript should give clearer operational guidance: how an experimenter can diagnose whether buyer/seller-mean variation is of order θ(1) versus o(1), and what fails (and what still works) when the Gram matrix I∑_γ a_γ,θ Z_θ is nearly singular. Without that, the scope of the main inferential guarantees is hard to assess from data alone.","section":null},{"comment":"Section F derives optimal interacted imputation coefficients for direct, total, and spillover effects and reports a simulation (Figure 6) in which the optimal interacted estimator substantially outperforms Lin-style separate OLS. The abstract and §1.3 list interacted estimators among the contributions, yet asymptotic normality, plug-in consistency, and inference for the interacted class are deferred. Either complete the parallel theory (the non-interacted arguments appear to extend) or clearly relegate interacted estimators to an exploratory appendix so that the main claims match the proved results.","section":null},{"comment":"Section 4 (Figures 3 and 5): the optimal adjustment reduces Monte Carlo variance more than it shortens the conservative confidence intervals, and all methods overcover. The text notes this briefly but does not quantify how much of the efficiency gain is lost to the Cauchy–Schwarz-style bound V_c of [MVR+24]. Because the paper’s practical selling point is “better inference when running MRDs,” a short analysis of when the plug-in residual variance estimator remains substantially conservative after optimal adjustment would strengthen the inferential claims.","section":null}],"minor_comments":[{"comment":"Table 1 is helpful but the “∼ (Direct Effect)” entry for non-interacted regression form is slightly cryptic; a one-sentence clarification in the table note would help.","section":null},{"comment":"Notation for group sizes I_γ, J_γ versus I_T, I_C is introduced gradually; a short notation paragraph early in §1.5 would reduce friction.","section":null},{"comment":"In the direct-effect WLS representation (Eq. 6 and Appendix B), the full list of identifiability constraints is long; stating that they force within-group double-decentering would make the equivalence more transparent.","section":null},{"comment":"Figure 1 and the simulation legends use “σ =” for estimated standard deviations; labeling them as Monte Carlo SDs would avoid confusion with the theoretical σ_Tot.","section":null},{"comment":"A few typos and typesetting issues remain (e.g., “/leftr↦g⊳tl↦ne” artifacts in Theorem 3.4 and related displays; “I ∑_γ a_γ,θ Z_θ” in Assumption 4).","section":null},{"comment":"The discussion of switchback and clustered MRD extensions (§1.4) is interesting but could be shortened or moved to the discussion to keep the main narrative focused.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Solid, carefully executed design-based paper that fills a genuine gap for marketplace experimentation. The central identification of the optimal non-interacted coefficient from observed moments is clean and non-circular. I would not block on the open multi-regime theory; the authors already flag it. Fit for a top methods / applied-statistics venue is good. Minor revision is appropriate mainly to align the abstract’s scope with what is fully proved and to give practitioners a clearer map of Assumption 4."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This paper does exactly what the title promises: it derives the minimum-asymptotic-variance linear adjustments for the four standard MRD estimands (total, direct, buyer/seller spillover) inside the non-interacted imputation class, shows the optima remain estimable from observed double-decentered moments alone, and supplies the accompanying model-robust CLT plus conservative intervals. That is new. The weighted TWFE representation for the direct effect is a clean, unexpected bonus; the improved Wasserstein CLT for doubly-randomized sums (relaxing the earlier variance lower bound and boundedness assumptions) is a genuine technical contribution that will be useful beyond this paper.\n\nThe math is careful and finite-population throughout. Variance formulas, quadratic-form derivation of the optimal beta, plug-in consistency, and the Stein-method CLT all check out under the stated conditions. Simulations are honest: Opt beats ANCOVA and unadjusted, sometimes substantially, and the no-harm property is visible. The authors flag the interacted class and the multi-regime scaling problem themselves rather than burying them.\n\nThe soft spot is real but already named: Assumption 4 (and its direct-effect cousin) requires Var = Theta(I^{-1}) (or Theta((IJ)^{-1})) plus invertible limiting Gram matrices. When buyer/seller mean variation vanishes faster, consistency of the plug-in beta and the intervals can fail. That is a genuine open regime, not a hidden flaw; the paper isolates the direct-effect case and leaves the rest for later work. No circularity, no invented entities, no load-bearing unacknowledged gaps.\n\nThis is for people who already run or design marketplace MRDs and want better power without inventing a new design. It is also for anyone working on finite-population regression adjustment under interference. A serious editor should send it to referees; the contribution is clear, the proofs are there, and the practical payoff is immediate. I would cite it and bring it to reading group.","headline":"Solid, usable extension of Lin/ToM-style adjustment to MRD marketplace designs; the optimal coefficients are estimable and the theory is careful.","tokens_in":56125,"tokens_out":488,"would_cite":true,"duration_ms":6656,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62K99","62J05","62F12"],"pacs":[],"model":"grok-4.5","headline":"Optimal regression adjustments for two-sided marketplace experiments can be estimated from observed data alone and beat classical ANCOVA without assuming a linear model.","keywords":["covariate adjustment","randomization inference","design-based inference","spillovers","marketplaces","multiple randomization designs","two-way fixed effects"],"falsifier":"In a large-scale MRD with known ground-truth effects, compute the optimal plug-in estimator and ordinary ANCOVA on the same covariates; if the plug-in fails to reduce empirical variance relative to ANCOVA (or the unadjusted estimator) once treatment fractions become unbalanced, the optimality claim is falsified.","tokens_in":56211,"feed_emoji":"📊","tokens_out":715,"duration_ms":7099,"temperature":0.7,"pith_summary":"Two-sided marketplaces need special experimental designs called multiple randomization designs (MRDs) that randomize both buyers and sellers so that direct effects and spillovers can be separated under local interference. Power is often low because of the double randomization. This paper shows how to bring covariates into the analysis optimally. Inside a large class of linear imputation estimators that subtract a common regression adjustment from every group average, the coefficient that minimizes asymptotic variance is a quadratic form built from row, column, and double-decentered moments; that form can be estimated from the observed data alone. The resulting plug-in estimator is consistent and asymptotically normal for the finite-population total, direct, or spillover effect without any linearity assumption on the potential outcomes. For the direct effect the optimal adjustment is exactly a weighted two-way fixed-effects regression whose weights up-weight the smaller cells, analogous to tyranny-of-the-minority. Simulations confirm large efficiency gains over unadjusted estimators and over ordinary ANCOVA, especially when treatment fractions are unbalanced.","feed_headline":"Optimal marketplace regression adjustments beat ANCOVA","feed_subtitle":"The variance-minimizing coefficients are estimable from data and need no linear model","key_machinery":"The non-interacted imputation estimator τ̂_c(β) together with the estimable quadratic form Z̃_c, ũ_c built from buyer-mean, seller-mean, and double-decentered residual moments; minimizing that form yields the optimal β̃_c that can be plugged in from data.","core_discovery":"Among the class of non-interacted linear imputation estimators for any linear combination of the four cell means, the asymptotic-variance-minimizing coefficient is estimable from observed data by plugging empirical double-decentered moments into the quadratic form that defines the population variance; the resulting plug-in is model-robustly consistent and asymptotically normal for the true finite-population effect.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Optimal MRD adjustments are data-estimable and model-robust","Double-decentered moments yield min-variance marketplace estimators","Linear imputation beats ANCOVA for two-sided double randomization","Variance-minimizing coefficients for MRD total direct spillover effects","Plug-in optimal adjustments improve efficiency in marketplace MRDs"],"cache_read_input_tokens":49280,"weakest_assumption_plain":"The variance of the optimally adjusted estimator must scale like one over the number of buyers (or one over the product of buyers and sellers for the pure direct-effect case) and the associated Gram matrices must stay invertible; if row or column variation vanishes faster, consistency and the confidence intervals can fail.","fun_headline_variants_meta":{"raw":{"variants":["Optimal MRD adjustments are data-estimable and model-robust","Double-decentered moments yield min-variance marketplace estimators","Linear imputation beats ANCOVA for two-sided double randomization","Variance-minimizing coefficients for MRD total direct spillover effects","Plug-in optimal adjustments improve efficiency in marketplace MRDs"]},"model":"grok-4.5","effort":"low","cost_usd":0.005426,"raw_usage":{"total_tokens":1433,"prompt_tokens":697,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":54260000,"prompt_tokens_details":{"text_tokens":697,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":647,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":697,"tokens_out":89,"duration_ms":5806,"temperature":1.0,"reasoning_tokens":647,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T22:01:21.511401+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"In a large-scale MRD with known ground-truth effects, compute the optimal plug-in estimator and ordinary ANCOVA on the same covariates; if the plug-in fails to reduce empirical variance relative to ANCOVA (or the unadjusted estimator) once treatment fractions become unbalanced, the optimality claim is falsified.","supporting_citations":[],"review_version":1}