{"id":"775fcc78-118f-486a-96ea-f18e1edebe18","arxiv_id":"1908.09237","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper derives a boundary-parameter GMM asymptotic distribution for a ridge-penalized IV estimator with sample-split tuning, and reports simulation MSE gains over 2SLS.","lead":"This paper studies a ridge-penalized instrumental variables estimator whose tuning parameter is chosen by splitting the data into training and test samples. The intended result is an asymptotic distribution for both the coefficients and the tuning parameter, with lower mean squared error than two-stage least squares in small, low-precision samples.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"First-order condition mismatch: row 3 of equation (14) has the wrong sign relative to equation (10), so Theorem 1's GMM estimator is not the ridge path estimator.","rationale":"The reader's verdict is REJECT, and the rationale already cites the sign inconsistency between equation (14) and equation (10). I agree with the rejection. I treat the GMM moment mismatch as the single most load-bearing concern because it breaks the link between the paper's definition of the ridge path estimator and the estimator appearing in Theorem 1. Even if every other step of the proof were correct, Theorem 1 would characterize a different estimator, so the main theoretical contribution would fail. The beta_p != beta_0 condition is a genuine unstated assumption, but it only creates a problem at a measure-zero point in the parameter space; the moment condition error is fatal for all nonzero alpha. The reader's weakest_assumption field points to beta_p != beta_0, while the sign issue is the more fundamental defect; hence partial agreement. A quick algebraic or numerical check can settle the mismatch.","tokens_in":22211,"tokens_out":10203,"duration_ms":89443,"concrete_test":"Independently re-derive the third block of equation (14) from the first-order condition (10). Substitute S_tau = Z'_tau X_tau/[tau n], R_tau = Z'_tau Z_tau/[tau n], and A = S'_tau R_tau^{-1} S_tau into the block, set it to zero, and compare to equation (11). If the resulting estimating equation is (A - alpha I) beta = X'_tau P_Ztau Y_tau/[tau n] - alpha beta_p instead of (A + alpha I) beta = X'_tau P_Ztau Y_tau/[tau n] + alpha beta_p, then Theorem 1 applies to a different estimator. Alternatively, for a k=1, m=1 design with alpha=1 and beta_p != beta_0, compute the solution of the moment condition and compare it to equation (11); any nonzero difference confirms the mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 1, which gives the asymptotic distribution of theta_hat = argmin H_n(theta)'H_n(theta) using the moment vector H_n in equation (14). For this theorem to describe the ridge path estimator, the third block of (14) must be an estimating equation equivalent to the training-sample first-order condition (10). Equation (10) is - (1/[tau n]) sum_i S'_tau R_tau^{-1} z_i (y_i - x_i' beta) + alpha (beta - beta_p) = 0, where S_tau = Z'_tau X_tau/[tau n] and R_tau = Z'_tau Z_tau/[tau n]. The third block of (14) is S'_tau R_tau^{-1} z_i (y_i - x_i' beta) + alpha (beta - beta_p), with the opposite sign on the score term. Setting this block to zero gives (A - alpha I) beta = X'_tau P_Ztau Y_tau/[tau n] - alpha beta_p, whereas the ridge solution (11) solves (A + alpha I) beta = X'_tau P_Ztau Y_tau/[tau n] + alpha beta_p, with A = S'_tau R_tau^{-1} S_tau. The equations coincide only at alpha = 0. Therefore the estimator whose asymptotics are characterized by Theorem 1 is not the ridge path estimator defined in Section 2. The sign error is not cosmetic: for any alpha > 0 the two estimators differ, and the limiting matrix M0 in Theorem 1 is computed for the wrong moment vector. Consequently, the paper's main theoretical result, including the claimed mixture distribution for the empirically selected tuning parameter, does not apply to the proposed estimator. A separate, secondary gap is that Assumptions 1-3 do not require beta_p != beta_0; the proof of Lemma 1 and the invertibility of D in Appendix B both fail when beta_p = beta_0. But the moment mismatch alone is sufficient to invalidate the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a ridge-type estimator for linear instrumental variables models in which the parameter of interest is shrunk toward a prior vector beta_p, and the regularization tuning parameter alpha is selected empirically by splitting the sample into training and test parts. The ridge path estimator is defined as the training-sample ridge solution evaluated at the alpha that minimizes the test-sample IV objective. The paper's central theoretical contribution is Theorem 1, which claims that under Assumptions 1-3 the joint asymptotic distribution of the parameter estimates and the empirically selected alpha is a nonstandard mixture distribution obtained by applying Andrews (2002) boundary GMM theory to a just-identified moment system H_n(theta)=0 given in equation (14). Simulation evidence compares the ridge path estimator with 2SLS across sample sizes, instrument strengths, and prior values, and reports the empirical distribution of the selected alpha.","tokens_in":22528,"tokens_out":10773,"duration_ms":102297,"significance":"If Theorem 1 were correct, the paper would fill a real gap: most regularized IV estimators either treat the tuning parameter as fixed or impose rate conditions on it, whereas this paper attempts to characterize the asymptotic distribution when the tuning parameter is itself estimated from a holdout sample. The GMM embedding and the use of the boundary-parameter framework of Andrews (2002) are sensible strategies, and the simulations cover a wide grid of sample sizes, signal strengths, and priors. The paper is also transparent about the bias-variance trade-off and about when the ridge path estimator does not dominate 2SLS. However, the central theorem as stated is not for the estimator defined in Section 2 because equation (14) contains a sign error in the beta moment condition, and the proof also requires an unstated assumption that the prior differs from the true parameter. These issues are load-bearing for the paper's main claim.","major_comments":[{"comment":"The third block of the moment vector in equation (14) is 1_{tau n}(i)(S'_tau R^{-1}_tau z_i(y_i - x_i' beta) + alpha(beta - beta_p)), whereas the first-order condition for the ridge estimator in equation (10) has the negative of the score term: - (1/[tau n]) sum_i S'_tau R^{-1}_tau z_i(y_i - x_i' beta) + alpha(beta - beta_p) = 0. Setting the third block of H_n(theta) = 0 gives (A - alpha I) beta = X'_tau P_{Z_tau} Y_tau / [tau n] - alpha beta_p, with A = S'_tau R^{-1}_tau S_tau, while the ridge path estimator in equation (11) solves (A + alpha I) beta = X'_tau P_{Z_tau} Y_tau / [tau n] + alpha beta_p. The two systems coincide only when alpha = 0. Consequently, Theorem 1 characterizes the asymptotic distribution of a different estimator, and the matrix M_0 in Theorem 1 is the derivative of the wrong moment vector. This is a load-bearing error because the mixture distribution for the ridge path estimator is the paper's main theoretical result.","section":"Equation (14) vs. Equation (10)"},{"comment":"The asymptotic theory requires beta_p different from beta_0, but this is not stated in Assumptions 1-3. In the proof of Lemma 1, the deterministic limit of Q_{n(1-tau)}(alpha) is uniquely minimized at alpha = 0 only when the quadratic form (beta_p - beta_0)'(Gamma'_0 R_z Gamma_0)(beta_p - beta_0) is positive; if beta_p = beta_0 the limit is identically zero in alpha and the argument fails. In Appendix B, the matrix D is invertible only if delta_tilde = (beta_0 - beta_p)'(S'_0 R^{-1}_z S_0)^{-1}(beta_0 - beta_p) > 0; when beta_p = beta_0, delta_tilde = 0 and D is singular, so M_0 is not invertible. The conclusion's discussion of priors 'at' the true parameter value is therefore inconsistent with the stated assumptions and proofs. The authors should add an explicit assumption beta_p != beta_0 and adjust the text accordingly.","section":"Assumptions 1-3, Lemma 1, and Appendix B"}],"minor_comments":[{"comment":"The first block of Table 3 uses delta = 0.01 while the text describes the precision parameter values as 0.1, 0.25, 0.5, and 1; this inconsistency should be corrected.","section":"Section 4, Table 3"},{"comment":"The claim that the histograms confirm the 'mass at alpha = 0 approaching 50%' predicted by Theorem 1 is not supported by an overlay or a numerical computation of the theoretical mixture distribution; adding such a comparison would make the simulation evidence quantitative.","section":"Section 4 and Figure 5"},{"comment":"Assumption 3 does not restrict the split fraction tau, although the formulas using [tau n] and n - [tau n] require 0 < tau < 1; this should be stated explicitly.","section":"Assumption 3"},{"comment":"There are several typographical errors, including 'prioir' in Section 4, 'Assumtions' and 'Anderws' in Appendix B, and the conclusion's reference to priors 'at' the true value despite no such simulation being reported.","section":"Conclusion and Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The sign error in equation (14) appears to be a typo rather than a fundamental flaw in the overall approach, but it invalidates Theorem 1 as stated and requires a full re-derivation of the moment system, the derivative matrix M_0, and the resulting mixture distribution. I would want to see the corrected proof and a quantitative comparison of the theoretical mixture with the simulation histograms before reconsidering the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea here is genuinely interesting: shrink a linear IV estimator toward a nonzero prior with a ridge penalty, select the tuning parameter on a holdout sample, and then derive the joint asymptotic distribution of coefficients and tuning parameter as a boundary-GMM mixture. That combination is new, and the paper engages seriously with the post-model-selection literature. The simulation design is thoughtful — four sample sizes, four precision levels, three priors, 10,000 reps — and the MSE gains in low-precision small samples are plausible. Credit where it's due: the path interpretation from 2SLS to prior is clear, and the use of Andrews (2002) is appropriate if the moment system were correct.\n\nBut the central theorem is built on a sign error. The first-order condition for the ridge path estimator, equation (10), is the negative of the score plus the penalty: -S' R^{-1} z(y - x'b) + alpha(b - beta_p) = 0. The third block of the moment vector in equation (14) has the positive score plus the penalty: +S' R^{-1} z(y - x'b) + alpha(b - beta_p). Setting that block to zero gives (A - alpha I)b = X'P_ZY - alpha beta_p, whereas the actual ridge solution solves (A + alpha I)b = X'P_ZY + alpha beta_p. These coincide only at alpha = 0. The GMM estimator whose asymptotics are characterized by Theorem 1 is therefore not the ridge path estimator defined in Section 2 for any positive tuning parameter. Since the limiting matrix M0 is computed for the wrong moment vector, the claimed mixture distribution does not apply to the proposed estimator. This is not a cosmetic slip; it invalidates the paper's main theoretical contribution.\n\nThere is also a secondary gap: Assumptions 1–3 do not require beta_p != beta_0, and the proof of Lemma 1 plus the invertibility of D in Appendix B both fail when the prior equals the true parameter. The authors do not flag this. The abstract also overclaims when it says the asymptotic distribution captures the sampling distributions — the simulations only examine the proportion of zero tuning parameter values, not the full distribution of the coefficients. That is a minor overstatement relative to the sign problem.\n\nThe paper deserves a referee's attention because the idea is worth salvaging, and the flaw is the kind a careful referee can pinpoint. But as it stands, the central result is wrong. I would tell a serious editor: send it out if the authors can first verify the moment conditions; otherwise the paper should not proceed. My own recommendation is reject in current form, with a clear path to resubmission if the sign is fixed and the theory is reworked.","headline":"Novel shrinkage-IV idea, but Theorem 1's moment conditions have a sign error and characterize a different estimator, so the central asymptotic claim does not hold as written.","tokens_in":23110,"tokens_out":2839,"would_cite":false,"duration_ms":28279,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A data-selected ridge penalty makes the tuning parameter part of the asymptotic law and improves IV estimation in low-precision samples.","keywords":["ridge path estimator","instrumental variables","ridge regression","tuning parameter selection","training and test samples","GMM","boundary parameter","mean squared error"],"falsifier":"Simulate the paper's design with $\\beta_p$ set equal to $\\beta_0=(0,0)'$ and let $n$ grow. If the central claim is right, the limiting objective is flat in $\\alpha$ and the claimed mixture law for $\\hat\\alpha$ should not appear; in particular, the asymptotic mass at $\\hat\\alpha=0$ should depart from the predicted 50\\% limit and the matrix that must be inverted in the proof becomes singular.","tokens_in":21942,"feed_emoji":"📉","tokens_out":11008,"duration_ms":102186,"temperature":0.7,"pith_summary":"This paper proposes a ridge-penalized estimator for linear instrumental variables models in which the shrinkage tuning parameter is chosen by the data rather than fixed or forced to satisfy an arbitrary rate condition. The estimator works by tracing a one-dimensional path in the parameter space from the IV estimate to an economically meaningful prior, then selecting the point on that path that minimizes the IV objective on a holdout sample. The paper's central theoretical claim is that the empirically selected tuning parameter must be treated as an estimated parameter: it converges to zero at root-$n$ speed and has a nonstandard, boundary-type asymptotic distribution jointly with the coefficients. This matters because in low-precision or small-sample IV settings the estimator can achieve materially lower mean squared error than two-stage least squares, and because it gives practitioners a way to report uncertainty that accounts for having chosen the tuning parameter from the data.","feed_headline":"Data-selected ridge penalty trims IV error in low-precision samples","feed_subtitle":"The chosen tuning parameter is estimated, not fixed, and its boundary-mixture law matters for inference; simulations show gains over 2SLS.","key_machinery":"The central object is the ridge path itself. For each value of $\\alpha$, the training subsample defines $\\hat\\beta_{IV,\\tau n}(\\alpha) = (X'_{\\tau n}P_{Z_{\\tau n}}X_{\\tau n}/[\\tau n] + \\alpha I_k)^{-1}(X'_{\\tau n}P_{Z_{\\tau n}}Y_{\\tau n}/[\\tau n] + \\alpha\\beta_p)$, a smooth curve in the coefficient space that runs from the IV estimate at $\\alpha=0$ to the prior $\\beta_p$ as $\\alpha\\to\\infty$; the estimator is the point on this curve that minimizes the IV objective on the test subsample. The asymptotic machinery is a GMM reparameterization whose moment conditions include the first-order conditions of both the training ridge problem and the test-sample selection problem. Applying a boundary-parameter GMM theorem to this just-identified system yields the cone projection and, therefore, the mixture limiting distribution with mass at $\\alpha=0$.","core_discovery":"On the paper's own terms, the central discovery is that a ridge-penalized IV estimator with a data-selected tuning parameter can be embedded in the standard GMM framework and its limiting behavior characterized exactly. Under the paper's Assumptions 1-3, the tuning parameter $\\hat\\alpha$ converges to 0 and $\\sqrt{n}\\hat\\alpha = O_p(1)$; the asymptotic distribution of $\\sqrt{n}(\\hat\\theta - \\theta_0)$ is the distribution of the minimizer $\\hat\\lambda$ of $(Z-\\lambda)'M_0'M_0(Z-\\lambda)$ over the cone $\\Lambda = \\{\\lambda : \\lambda_{\\alpha} \\ge 0\\}$, where $Z$ is the limiting normal vector $(\\lim_{n\\to\\infty} E[-\\partial h_i(\\theta_0)/\\partial\\theta'])^{-1}\\sqrt{n}H_n(\\theta_0)$. Because the population value $\\alpha_0=0$ lies on the boundary of the parameter space, the limiting law is a mixture: a discrete mass at $\\alpha=0$ and a continuous part over $\\alpha>0$. The paper further claims, from Monte Carlo evidence, that in strongly identified models with low finite-sample precision this ridge path estimator has lower MSE than two-stage least squares, and remains competitive when 2SLS is best.","pith_inferences":["Because $\\hat\\alpha$ converges to zero, the estimator is asymptotically equivalent to 2SLS; its value is a finite-sample MSE improvement, so it is best read as a small-sample method rather than a new asymptotic estimator.","The same split-sample, cone-projection logic should extend to other penalties (Lasso, bridge, elastic net) and to $k$-fold cross-validation when the selected penalty converges to a boundary; the paper leaves these extensions open.","A practical diagnostic follows directly: plot the ridge path $\\hat\\beta(\\alpha)$ and the test-sample objective as functions of $\\alpha$; the shape of these curves reveals whether the data can discriminate between the prior and the IV estimate."],"forward_implications":["Inference must treat the tuning parameter as part of the parameter vector; conditioning on the selected $\\hat\\alpha$ ignores the discrete mass at zero in the limiting distribution.","In low-precision or small-sample settings, the ridge path estimator achieves lower combined MSE than 2SLS; in high-precision, large-sample settings 2SLS is better but the ridge path estimator stays competitive.","The probability that the data choose 'no regularization' ($\\hat\\alpha=0$) approaches roughly one-half as $n$ grows, so the boundary mixture is a first-order feature rather than an edge case.","Closer priors improve the estimator's relative performance, and even a prior several standard deviations from the truth can beat 2SLS in small, low-precision samples."],"supporting_citations":[{"why":"Supplies the boundary-parameter GMM theorem used to derive the cone-projection mixture distribution for the estimator.","marker":"Andrews (2002)"},{"why":"Supplies the consistency and root-n argument (Lemma 5.4) used to prove the tuning parameter converges to zero at root-n speed.","marker":"Ichimura (1993)"},{"why":"Establishes that empirically selected tuning parameters change the final asymptotic distribution, motivating the joint treatment.","marker":"Leeb and Pötscher (2005)"},{"why":"Represents the existing approach requiring tuning parameters to satisfy rate restrictions, which this paper replaces with empirical selection.","marker":"Caner (2009)"},{"why":"Defines the strong-instrument, low-precision setting used to motivate the finite-sample MSE comparisons.","marker":"Sanderson and Windmeijer (2016)"},{"why":"Provides the regularization and bias-variance framework used to explain why shrinking toward a prior can lower MSE.","marker":"Hastie, Tibshirani, and Friedman (2009)"}],"fun_headline_variants":["Ridge IV with data-driven penalty beats 2SLS in low precision","Data-selected ridge IV: limiting law is a boundary mixture","Ridge-penalized IV: estimated tuning, mixture asymptotics","IV with data-chosen ridge: better MSE when 2SLS struggles","Ridge path IV: boundary mixture law, gains over 2SLS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, not stated among Assumptions 1-3, is that the researcher's prior $\\beta_p$ differs from the true coefficient $\\beta_0$; when they coincide, the limiting test-sample objective is flat in $\\alpha$, so the tuning parameter is no longer root-$n$ estimable and the proof of Lemma 1 breaks down.","fun_headline_variants_meta":{"raw":{"variants":["Ridge IV with data-driven penalty beats 2SLS in low precision","Data-selected ridge IV: limiting law is a boundary mixture","Ridge-penalized IV: estimated tuning, mixture asymptotics","IV with data-chosen ridge: better MSE when 2SLS struggles","Ridge path IV: boundary mixture law, gains over 2SLS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000639,"raw_usage":{"total_tokens":2939,"prompt_tokens":936,"completion_tokens":2003,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1907}},"tokens_in":552,"tokens_out":2003,"duration_ms":13374,"temperature":1.0,"reasoning_tokens":1907,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:19:08.493793+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the paper's design with $\\beta_p$ set equal to $\\beta_0=(0,0)'$ and let $n$ grow. If the central claim is right, the limiting objective is flat in $\\alpha$ and the claimed mixture law for $\\hat\\alpha$ should not appear; in particular, the asymptotic mass at $\\hat\\alpha=0$ should depart from the predicted 50\\% limit and the matrix that must be inverted in the proof becomes singular.","supporting_citations":[{"cited_title":"APACrefauthors \\ 2002","cited_arxiv_id":null,"evidence_quote":"Supplies the boundary-parameter GMM theorem used to derive the cone-projection mixture distribution for the estimator."},{"cited_title":"APACrefauthors \\ 2009","cited_arxiv_id":null,"evidence_quote":"Represents the existing approach requiring tuning parameters to satisfy rate restrictions, which this paper replaces with empirical selection."},{"cited_title":"\\ Windmeijer, F","cited_arxiv_id":null,"evidence_quote":"Defines the strong-instrument, low-precision setting used to motivate the finite-sample MSE comparisons."},{"cited_title":", Tibshirani, R","cited_arxiv_id":null,"evidence_quote":"Provides the regularization and bias-variance framework used to explain why shrinking toward a prior can lower MSE."}],"review_version":1}