{"id":"43405253-26d5-4436-a9fc-c436b87c2ed5","arxiv_id":"2607.29532","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"W-2SLS, a winsorized-mean version of 2SLS, attains the minimax-optimal error rate under adversarial contamination and preserves clean-sample Gaussian inference when sqrt(n) eta_n^{1-1/m} -> 0.","lead":"This paper introduces W-2SLS, a version of two-stage least squares that replaces ordinary averages with winsorized averages so that a few manipulated observations cannot dominate the estimate. It proves that under adversarial contamination this estimator achieves the best possible error rate, and that the price of robustness disappears asymptotically when contamination shrinks fast enough.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Upper-bound rate and inference rest on unverified companion-paper lemmas (Kock & Preinerstorfer 2026); if Lemma G.1's quantile-containment facts fail, the minimax-sharp rate for W-2SLS does not follow.","rationale":"The reader's weakest assumption correctly identifies the delegation of the core robust-estimation lemmas to a self-cited companion paper. My stress-test confirms this is the single most load-bearing concern: every positive result passes through Lemma G.1/G.2, and the proof of these lemmas uses companion results that are not stated. I do not see an additional internal gap; the lower-bound constructions (Lemma D.1–D.2) are self-contained and the algebra checks out. Thus the appropriate verdict remains CONDITIONAL, and my read does not move it. A concrete external verification of the companion lemmas, or a direct derivation of the quantile containment, would settle the concern.","tokens_in":37954,"tokens_out":19490,"duration_ms":190553,"concrete_test":"Obtain the published/forthcoming version of Kock and Preinerstorfer (2026) and verify: (i) Lemma B.5 indeed delivers the containment (G.9) for ε = λ1η + λ2 log(6/δ)/n when (G.6) holds; (ii) Lemma B.3 gives c1 ≥ c(λ1,λ2)>0 uniformly in n; (iii) Lemma C.1 implies |Q_p(S) − ES| ≤ σm / p^{1/m}. If any fails, the η^{1−1/m} rate collapses. As a rapid cross-check, re-derive (G.9) from the elementary counting argument: with ε ≥ 1.01η, the εn-th contaminated order statistic is at least the (ε−η)n-th clean order statistic, and this exceeds the c1ε clean quantile; if this derivation works, the companion dependency is benign and the main theorem stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The minimax-sharp estimation rate in Theorem 2.2 and the inference claims in Theorems 2.3–2.4 and 4.1 are all downstream of Proposition 2.1, which applies Lemma G.1 to the IV moment products. Lemma G.1's proof is not self-contained: it invokes Lemma B.5 (quantile containment of contaminated winsorization points), Lemma B.3 (uniform bounds on c1,c2), and Lemma C.1 (quantile/moment bound) from the self-cited companion paper Kock and Preinerstorfer (2026, to appear), none of which are reproduced. In particular, the key event (G.9) — that contaminated α-hat and β-hat lie between clean quantiles at indices c1ε and 1−c1ε — is exactly what converts the ηn contamination budget into the η^{1−1/m} rate through (G.10). Without a proof of (G.9) for the specific choice εn = 1.01ηn + λ log(n)/n, the decomposition (G.8) lacks its main term. The lower-bound side (Theorem 3.1) is self-contained, so the asymmetry means the upper bound is the vulnerable half: the headline claim could be false if the companion lemmas fail, and a reviewer cannot currently verify them.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes W-2SLS, a robustified 2SLS estimator that replaces the sample averages in the 2SLS moment conditions by quantile-winsorized means. Under an adversarial contamination model in which an adversary may alter up to an η_n fraction of observations, with full knowledge of the clean sample, the paper claims that W-2SLS attains the estimation error O_P(η_n^{1-1/m} + n^{-1/2}) under m-th moment assumptions, and that this rate is minimax sharp via matching lower bounds. It further claims that centered Gaussian inference with the same asymptotic covariance as clean-data 2SLS is possible when √n η_n^{1-1/m} → 0, and that this condition is necessary. The paper also constructs a PSD heteroskedasticity-robust covariance estimator, a winsorized Anderson-Rubin test valid under weak identification and adversarial contamination, and finite-sample uniform deviation bounds showing W-2SLS has sub-Gaussian concentration while ordinary 2SLS does not. The lower-bound arguments are largely self-contained, but the upper-bound and inference results rely on lemmas whose key inequalities are deferred to a companion paper by the same authors that is cited as 'to appear'.","tokens_in":38361,"tokens_out":4033,"duration_ms":41737,"significance":"If the companion-paper lemmas are correct, the paper makes a substantial contribution. It gives a simple closed-form estimator with a sharp minimax rate under a very flexible contamination model, identifies the exact threshold for root-n consistency and for clean-data Gaussian inference, and provides feasible inference. The lower-bound constructions are elegant and self-contained, and the paper is honest about the dependence of the upper bounds on external results. The finite-sample concentration comparison between 2SLS and W-2SLS is also a valuable contribution. However, the central positive claims are not currently verifiable from the manuscript alone, because the key step converting contamination into the η^{1-1/m} rate is delegated to a not-yet-available companion paper. This is a transparency and verifiability problem, not circularity.","major_comments":[{"comment":"The proof of Lemma G.1 is not self-contained. The key inequality (G.9) — that the contaminated winsorization points α̂ and β̂ lie between clean quantiles Q_{c1ε}(S1) and Q_{1−c1ε}(S1) — is asserted by reference to Lemma B.5 of Kock and Preinerstorfer (2026, to appear). This containment is exactly what converts the contamination budget η into the ε^{1-1/m} bound through (G.10). The lower bounds on c1 and c2 are also delegated to Lemma B.3 of the same companion paper. Since Proposition 2.1, and hence Theorems 2.2, 2.3, 2.4, 4.1, and 5.2, all funnel through Lemma G.1, these results are currently unsupported unless the companion-paper lemmas are reproduced or supplied in full. The manuscript should either include complete proofs of these lemmas or state them as assumptions/conditions with proofs available in the appendix.","section":"Appendix G, Lemma G.1, Eq. (G.9)"},{"comment":"The consistency proofs for the PSD covariance estimator Ω̂ and the winsorized Anderson-Rubin covariance matrix rely on Theorem 2.1 of Kock and Preinerstorfer (2026, to appear) and on Lemma G.2, which itself invokes Lemma B.5, Lemma B.3, and Lemma C.1 of the same companion paper. These lemmas are not reproduced. In particular, Lemma G.2's bound (G.12) depends on the same quantile-containment facts and on bounds for c1 that are not derived here. Given that Theorem 2.4 is the basis for feasible inference and Theorem 4.1 is the basis for weak-identification-robust inference, the same external dependency applies. The authors should provide the missing proofs or otherwise make the companion results available for verification.","section":"Appendix C, Theorem 2.4; Appendix E, Theorem 4.1"},{"comment":"The finite-sample sub-Gaussian deviation guarantee for W-2SLS is obtained by applying Theorem 2.1 of the companion paper with m=2. The constant B in Eq. (F.7) is defined in terms of companion-paper quantities A(1.01,1), B(1.01,1), l(1.01,1), and u(1.01,1), whose derivation is not included. Thus Theorem 5.2 is also conditional on the companion paper. Since this theorem drives the striking Corollary 5.3 comparison between 2SLS and W-2SLS, the proof should be completed or the companion results should be stated in full.","section":"Section 5, Theorem 5.2 and Eq. (F.7)"}],"minor_comments":[{"comment":"The probability bound in Lemma G.1 reads 'with probability at least 1 − 4/6 δ − 1/M', which is awkwardly written and could be misread as 1 − (4/6)δ − 1/M. Please clarify, e.g., '1 − (2/3)δ − 1/M', and ensure the intended meaning is transparent.","section":"Appendix G, Lemma G.1"},{"comment":"The choice ε_n = 1.01 η_n + λ log(n)/n uses a fixed 1.01 factor. The paper motivates this as 'slightly more than η_n', but it may help to state explicitly that the 1.01 factor is arbitrary and can be replaced by any constant strictly greater than 1; the proofs appear to only need λ_1 > 1.","section":"Section 2.2, Eq. (7)"},{"comment":"The lower-bound construction in Lemma D.1 is a useful concrete example, but the notation Q_{1,n} in Lemma D.1 and Q_1 in the proof is introduced with a slight inconsistency. Please harmonize the notation.","section":"Section 3, Theorem 3.1"},{"comment":"The bound in (G.14) is stated after a chain of inequalities; for readability, please indicate which lines use Cauchy-Schwarz, Hölder, and Lemma C.1 respectively, since these steps are important for verification.","section":"Appendix G, Eq. (G.14)"}],"recommendation":"major_revision","confidential_remarks":"The central positive results are conditional on a to-appear companion paper by the same authors. This is not circularity, and the lower-bound side is self-contained, but it makes the current version impossible to verify. I recommend requesting that the companion-paper lemmas be either fully stated with proofs in this manuscript or made available as a supplement, before considering acceptance. The contribution is potentially significant, but the external dependency is load-bearing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is good and the paper does real work: it proposes a winsorized-mean 2SLS estimator that is a genuine drop-in replacement for ordinary 2SLS, and it gives matching lower bounds showing the contamination rate is minimax sharp. The lower-bound constructions are self-contained and careful, and the finite-sample concentration gap in Section 5 is a genuinely nice contribution — it shows ordinary 2SLS has a uniform deviation radius that can grow like delta^{-1/2} in the failure probability, while W-2SLS scales like sqrt(log(1/delta)), even without contamination. That speaks directly to Young's (2022) fragility results and is a point I haven't seen made cleanly before.\n\nThe soft spot is structural and cannot be waved away: the upper-bound rate, the inference theorems, the covariance estimator, and the Anderson-Rubin test all depend on Lemma G.1 and Lemma G.2, whose key inequalities (quantile containment, bounds on c1 and c2, quantile/moment bounds) are delegated to a companion paper by the same authors, Kock and Preinerstorfer (2026, to appear). The decomposition in (G.8) and the event (G.9) are the load-bearing pieces, and without a proof for the specific choice epsilon_n = 1.01 eta_n + lambda log(n)/n, the central rate claim does not follow. The lower-bound side (Theorem 3.1) is self-contained, so the asymmetry is clear: the upper bound is the vulnerable half. This is not a fatal flaw if the companion lemmas are correct, but the current manuscript is not independently verifiable. A referee cannot check the main theorem without access to a paper that may or may not be in the pipe.\n\nMinor issues: no code or simulations, which would help calibrate the constant and the practical behavior of the winsorization tuning; and the choice of 1.01 in epsilon_n is arbitrary, though that's not a technical problem.\n\nWho is this for? Empirical econometricians and any theorist working on robust IV or adversarial contamination. It deserves a serious referee, but the editor should make sure the companion paper is available or demand that the key lemmas be reproduced in the appendix before final acceptance.","headline":"Promising robustification of 2SLS with sharp worst-case rates, but every positive result leans on unverified companion-paper lemmas; worth refereeing once those are in hand.","tokens_in":38802,"tokens_out":2167,"would_cite":false,"duration_ms":25091,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F35","62F12","62J05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A winsorized-mean modification of 2SLS is minimax-optimal under adversarial contamination, and keeps clean-data Gaussian inference whenever sqrt(n) eta^(1-1/m) vanishes.","keywords":["instrumental variables","two-stage least squares","winsorized mean","adversarial contamination","minimax rates","robust inference","weak identification","finite-sample concentration"],"falsifier":"Use the two distributions Q0 and Q1 from Lemma D.1 in the just-identified location model, contaminate through the map (11) with fixed eta_n and n, and evaluate any candidate estimator's worst-case error over the class Q_m. The theorem predicts the worst-case error is at least 0.25 eta_n^{1-1/m} with probability at least 0.5(1-(e/4)^{eta_n n/2}); an estimator with uniformly smaller worst-case error, or numerical evidence that the total-variation bound (e/4)^{eta_n n/2} is violated, would refute the sharpness claim.","tokens_in":37880,"feed_emoji":"🛡️","tokens_out":7802,"duration_ms":77063,"temperature":0.7,"pith_summary":"This paper asks how much data contamination 2SLS can survive before its estimates and inference break, under a worst-case model in which the identities and values of altered observations may depend on the clean data. It proposes W-2SLS, a drop-in replacement that substitutes quantile-winsorized means for every sample average in the 2SLS formula. The paper proves that W-2SLS's estimation error is of order eta_n^{1-1/m} + n^{-1/2} under finite m-th moments, and that no estimator can improve on the dependence on the contamination fraction. It also draws the exact boundary for first-order-free robustness: when sqrt(n) eta_n^{1-1/m} -> 0, W-2SLS has the same limiting Gaussian law as clean-sample 2SLS, so standard confidence intervals and tests remain valid. A winsorized Anderson-Rubin test keeps weak-instrument validity, and a finite-sample analysis shows W-2SLS concentrates much better than 2SLS even on clean data.","feed_headline":"A drop-in fix makes 2SLS robust up to the theoretical limit","feed_subtitle":"Under finite moments, W-2SLS keeps the usual rate and clean-data inference whenever sqrt(n) eta^(1-1/m) -> 0.","key_machinery":"The central object is the quantile-winsorized mean: replace observations below the empirical epsilon-quantile by that quantile and observations above the empirical (1-epsilon)-quantile by that quantile, then average. W-2SLS inserts this estimator in place of every sample average in the 2SLS formula, with a winsorization level set to epsilon_n = 1.01 eta_n + lambda log(n)/n, deliberately slightly larger than the contamination budget. The load-bearing estimates are the companion lemmas controlling the distance between the winsorized mean of contaminated data and the arithmetic mean of clean data, and the distance between coordinatewise winsorization and joint winsorization; these make the whol","core_discovery":"The central claim is that replacing the fragile sample averages in 2SLS by quantile-winsorized means yields an estimator that is simultaneously minimax sharp under adversarial contamination with finite m-th moments, achieving an estimation error of order eta_n^{1-1/m} + n^{-1/2}; asymptotically first-order equivalent to clean-data 2SLS, with the same N(0, Omega) limit, whenever sqrt(n) eta_n^{1-1/m} -> 0; and feasible for inference via a positive-semidefinite winsorized covariance estimator and a winsorized Anderson-Rubin test that is also valid under weak identification. Matching lower bounds show the conditions on eta_n are necessary: no estimator can be uniformly consistent if eta_n does","pith_inferences":["Editorial extension: the explicit thresholds give practitioners a contamination-budget check: if a bound on the fraction of possibly manipulated observations satisfies the stated condition, standard 2SLS output remains trustworthy without redesigning the specification.","Editorial extension: the coordinatewise-versus-joint winsorization device used for the covariance estimator is not specific to IV; the same idea should carry over to other moment-based estimators, such as GMM, where averages also enter nonlinearly.","Editorial extension: the finite-sample concentration gap is empirically testable by rerunning just-delete-two-observations sensitivity analyses with W-2SLS; the theory predicts far fewer flipped conclusions than reported for ordinary 2SLS.","Editorial extension: the non-existence of honest adaptive confidence sets suggests that any practitioner claiming robustness to an unknown contamination level must be relying on assumptions beyond finite moments, not on the estimator alone."],"forward_implications":["W-2SLS has estimation error of order eta_n^{1-1/m} + n^{-1/2}; matching lower bounds show no estimator can uniformly improve the dependence on the contamination fraction.","When sqrt(n) eta_n^{1-1/m} -> 0, W-2SLS has the same N(0, Omega) limit as clean-sample 2SLS, so robustness is first-order free for t-tests and confidence intervals.","The winsorized Anderson-Rubin test is asymptotically chi-squared and remains valid under weak identification, heteroskedasticity, and adversarial contamination.","A positive-semidefinite winsorized covariance estimator makes feasible inference possible without sample-average constructions.","Even with no contamination, W-2SLS's uniform finite-sample deviation radius grows only like sqrt(log(1/delta)), while 2SLS's grows like sqrt(1/delta)."],"fun_headline_variants":["2SLS gets a winsorized upgrade: robust to targeted outliers","Quantile-winsorized 2SLS: minimax-sharp under contamination","Drop-in robust 2SLS: same inference, even with contaminated data","W-2SLS: beats 2SLS under adversarial data, matches theory bound","Robustifying 2SLS: quantile-winsorized means hit sharp limits"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The engine behind every positive result is a set of bounds on quantile-winsorized means under heavy tails and adversarial contamination that the paper imports from the authors' companion work; if those companion lemmas are not correct, the rate, inference, and concentration claims collapse.","fun_headline_variants_meta":{"raw":{"variants":["2SLS gets a winsorized upgrade: robust to targeted outliers","Quantile-winsorized 2SLS: minimax-sharp under contamination","Drop-in robust 2SLS: same inference, even with contaminated data","W-2SLS: beats 2SLS under adversarial data, matches theory bound","Robustifying 2SLS: quantile-winsorized means hit sharp limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000866,"raw_usage":{"total_tokens":3619,"prompt_tokens":804,"completion_tokens":2815,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2707}},"tokens_in":548,"tokens_out":2815,"duration_ms":20508,"temperature":1.0,"reasoning_tokens":2707,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:09:04.091313+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the two distributions Q0 and Q1 from Lemma D.1 in the just-identified location model, contaminate through the map (11) with fixed eta_n and n, and evaluate any candidate estimator's worst-case error over the class Q_m. The theorem predicts the worst-case error is at least 0.25 eta_n^{1-1/m} with probability at least 0.5(1-(e/4)^{eta_n n/2}); an estimator with uniformly smaller worst-case error, or numerical evidence that the total-variation bound (e/4)^{eta_n n/2} is violated, would refute the sharpness claim.","supporting_citations":[],"review_version":1}