{"id":"0b8ec17b-615b-41b9-a3f8-5ae0ef1825cb","arxiv_id":"2506.23834","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A self-normalized, random-matrix-based test for the structural parameter in IV regressions with K proportional to N and general error dependence, with the central proof deferred to the authors' prior unpublished work.","lead":"This paper proposes a test for instrumental variable regressions that remains valid when there are more instruments than observations and when errors are correlated across units. If it works, it would give applied economists a way to test causal effects without relying on strong instruments or restrictive error assumptions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof is omitted and deferred to an unpublished overlapping preprint, so the central null-distribution claim is unsupported as submitted.","rationale":"The reader's weakest assumption correctly identifies the same load-bearing concern: Theorem 1 is not proven in the preprint but deferred to an unpublished manuscript by overlapping authors. This is the foundation for Corollary 1 and is also used as the J11 term in Theorem 2, so the entire distributional theory rests on an external, unverifiable proof. The condition sum_i |u_i|^3 = o_P(1) is also not derived from the primitive network/spatial dependence assumptions, and the simulations, which use N=400 and 1,000 replications, are consistent with the claimed limit but cannot validate an asymptotic theorem. The proof of Theorem 2 is included and appears internally consistent conditional on Theorem 1, but it does not rescue the central claim. If the omitted proof were supplied and verified, the REJECT verdict should be revisited; as submitted, the reader's REJECT remains appropriate.","tokens_in":20791,"tokens_out":6482,"duration_ms":75393,"concrete_test":"Obtain Feng et al. (2024) and check whether its Theorem 1 actually proves Q_N -> N(0,1) under the present Assumptions 1-3, including a derivation of condition (ii) for the network and SAR error processes used in Section 3; if it does not, re-derive Theorem 1 directly from Lemma 1 and verify that sum_i |u_i|^3 = o_P(1) for these processes, e.g., by computing the statistic over 1,000 replications at N=400 and N=1600 and checking that it decays.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—Corollary 1's N(0,1) limit for the feasible statistic—depends entirely on Theorem 1, whose proof is not in the preprint: Appendix A.1 states 'The proof is identical to Theorem 1 in (Feng et al., 2024) so we omitted the proof,' and A.2 does the same for Corollary 1. That referenced paper is an unpublished archive preprint by overlapping authors; neither its proof nor a version of it is included or verified here. The same theorem is also the J11 building block in Theorem 2. In addition, condition (ii) of Theorem 1, sum_i |u_i|^3 = o_P(1), is asserted for general network and spatial dependence without being derived from Assumption 3; the simulations only report size at N=400 and do not establish the asymptotic regime. The concern is not an internal contradiction, but an unverifiable external dependency: the paper's main distributional result cannot be checked from the submitted material.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a weak-identification-robust test for linear instrumental variable regressions with high-dimensional instruments, allowing K to exceed N, and claims robustness to general error dependence such as network and spatial dependence. The test statistic is a self-normalized quadratic form in the instruments, normalized by tr(Σ²), and its asymptotic null distribution is stated in Theorem 1 (oracle) and Corollary 1 (feasible). Theorem 2 states a nondegenerate power property under local alternatives. The simulation section reports finite-sample size and power for N=400 across network, spatial, and multiplicative heteroskedastic error designs.","tokens_in":20991,"tokens_out":7125,"duration_ms":76683,"significance":"If the stated results were fully proven, the paper would fill a genuine gap by combining high-dimensional-instrument asymptotics (K comparable to or larger than N) with general error dependence, an area where the existing AR-type tests mostly assume independence or cluster structure. The literature review is broad and the simulation design covers plausible dependence structures. However, the central null-distribution claim is not established in the submitted manuscript: the proofs of Theorem 1 and Corollary 1 are explicitly deferred to an unpublished overlapping preprint, and the proof of Theorem 2 inherits that unproven result. The paper therefore cannot be verified from its own content.","major_comments":[{"comment":"The proofs of the two central limit theorems are not included. Appendix A.1 states 'The proof is identical to Theorem 1 in (Feng et al., 2024) so we omitted the proof' and A.2 states the same for Corollary 1. The referenced manuscript is an unpublished preprint by overlapping authors and is not reproduced or independently verified. This is load-bearing because Corollary 1 is the feasible statistic used in the simulations, and because the proof of Theorem 2 in Appendix A.3 explicitly invokes the asymptotic normality of J11 from Theorem 1. The main null-distribution claim of the paper therefore cannot be checked from the submitted material. The authors should supply complete, self-contained proofs or clearly state that the results are conditional on an external unpublished manuscript.","section":"Appendix A.1 and A.2 (Theorem 1 and Corollary 1)"},{"comment":"Theorem 1 condition (ii) requires \\sum_{i=1}^N |u_i|^3 = o_P(1) for the self-normalized errors u_i = \\epsilon_i / \\|\\epsilon\\|. This is a delocalization condition on the error vector, but the paper does not derive it from Assumption 3 or from any primitive network/spatial dependence condition. Corollary 1 further requires \\sum_{i=1}^N u_i = O_P(1) without any argument. Since Assumption 3 only imposes mean independence and an unidentified functional form, the paper does not establish that the error processes used in the simulations satisfy these conditions. The asymptotic regime for general error dependence is therefore unsupported.","section":"Theorem 1, condition (ii); Corollary 1, additional condition"},{"comment":"The proof of Theorem 2 is not self-contained: it states 'We already show the asymptotic normality of J11 in Theorem 1' and thus inherits the unproven Theorem 1. In addition, the included proof has notational and algebraic inconsistencies that prevent verification, including the interchangeable use of T and N, the mixing of p and K, an undefined sigma-algebra F_{T,*} in the L1/L2 discussion, and a garbled decomposition involving J221, J222, J223, and J224. Correcting these issues is not merely cosmetic because the proof as written does not allow the reader to check the steps.","section":"Appendix A.3 (Theorem 2 proof)"},{"comment":"The feasible statistic replaces tr(\\Sigma^2) with the estimator \\widehat{\\mathrm{tr}(\\Sigma^2)} = [N(N-1)]^{-1}\\sum_{i\\ne j}(z_i'z_j)^2, citing Li and Chen (2012). No proof is provided that this estimator is ratio-consistent under Assumption 2 with K/N \\to c \\in (0,\\infty) and the factor structure used in the simulations. Since the proof of Corollary 1 is omitted, the validity of the feasible statistic is doubly unsupported.","section":"Corollary 1 and \\widehat{\\mathrm{tr}(\\Sigma^2)}"}],"minor_comments":[{"comment":"The displayed definition 'Y := Y^* / (Y^{*\\prime}Y^*)^{1/2} \\in (0,1)' should state that Y is a unit-norm vector in \\mathbb{R}^N; as written it appears to define a scalar.","section":"Section 2, after Eq. (5)"},{"comment":"The phrase 'for all i = 0, \\ldots, N' should read 'i = 1, \\ldots, N'.","section":"Assumption 3"},{"comment":"The labels J211, J212, J213, J221, J222, J223, and J224 are used inconsistently; in particular, J211 appears to denote different quantities in different lines of the proof.","section":"Appendix A.3"},{"comment":"The simulations report only N=400 and use 1,000 replications with MATLAB's default seed; no standard errors for the rejection probabilities are given. The results are illustrative, but they do not by themselves demonstrate the claimed asymptotic regime.","section":"Section 3"},{"comment":"The entry for Feng et al. (2024) lists only 'archive' with no arXiv identifier or publication status; this makes the deferred proofs particularly difficult to locate or check.","section":"References"}],"recommendation":"reject","confidential_remarks":"The central results of the paper are not proven in the manuscript; they are deferred to an unpublished overlapping preprint. This is a fundamental verification problem, not a request for minor revision. If the authors can provide full self-contained proofs in a future submission, the underlying idea may be worth reconsidering. The relationship to Feng et al. (2024) also needs to be clarified, since the present paper appears to import the technical core from that manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: a weak-identification-robust AR-type test for IV regression when the number of instruments is proportional to or larger than the sample size, and the errors are only assumed to satisfy a general network or spatial dependence structure. No paper I know covers all three features together. The self-normalized quadratic form plus random matrix theory is inherited from the authors' own earlier work, but the application to IV and the local power theory under many instruments are new. Theorem 2's proof is in the appendix and actually engages with the dependence structure, so there is substance here beyond relabeling.\n\nThe soft spot, however, is load-bearing. Theorem 1, which gives the oracle statistic's N(0,1) limit, is not proven. Appendix A.1 says the proof is identical to Theorem 1 in Feng et al. (2024), an unpublished archive preprint by overlapping authors, and A.2 does the same for Corollary 1. The feasible statistic's validity, the actual object used in the simulations, is an immediate corollary of that unproven theorem. You cannot verify the central claim from the submitted material. The power theory in Theorem 2 also depends on the same J11 term. This is not an internal contradiction, but it is an unverifiable external dependency, and that is a serious problem for a paper whose main contribution is asymptotic validity.\n\nTwo smaller issues. Condition (ii) of Theorem 1, sum |u_i|^3 = o_P(1), is asserted for general network and spatial dependence without being derived from Assumption 3; the simulations do not show how this delocalization condition behaves. And the Monte Carlo study, while covering several dependence structures and K/N up to 3, reports only N=400, includes no comparison with existing many-instrument robust tests, and provides no code or data. These are fixable.\n\nThe paper is not incoherent, and the authors are honest about the deferred proofs, but as submitted the main result rests on an unavailable derivation. I would send it to a serious referee anyway: the question is important, the approach is plausible, and the power analysis suggests the authors know how to handle the dependence. But the referee must be asked to obtain and check the proof of Theorem 1 or a complete version of the prior work, and the simulations should be compared with the existing dimension-robust AR tests. Without that, the paper should not be accepted.","headline":"A plausible and genuinely novel combination—weak-ID-robust IV inference with K larger than N under general error dependence—but the main null-distribution theorem is deferred to an unpublished self-cited preprint, so the central claim cannot be checked as submitted.","tokens_in":21473,"tokens_out":1467,"would_cite":false,"duration_ms":18269,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes a weak-identification-robust test whose self-normalized statistic converges to standard normal under the null even when instruments outnumber observations and errors are network- or spatially dependent.","keywords":["high-dimensional instruments","weak identification","instrumental variables","random matrix theory","self-normalization","network dependence","spatial dependence","Anderson-Rubin tests"],"falsifier":"Re-derive the omitted proof of Theorem 1 in this IV model; if it requires an additional condition not stated in the paper, such as a bound on error-dependence strength, the claim as stated would fail. Alternatively, simulate the feasible statistic under the null with strongly concentrated errors—one error carrying most of the norm—and check whether size is still near the nominal level.","tokens_in":20568,"feed_emoji":"📊","tokens_out":10023,"duration_ms":103543,"temperature":0.7,"pith_summary":"The paper aims to close a gap in instrumental-variable inference: tests that tolerate many weak instruments usually assume homoskedastic or independent errors, while tests that tolerate dependent errors usually require few instruments. It constructs a test statistic by self-normalizing the residual vector and projecting it on the eigenvectors of the instrument Gram matrix, then shows that this statistic converges to a standard normal distribution under the null when the number of instruments grows proportionally with the sample size. A plug-in version, in which the unknown spectral normalization is estimated from the data, inherits the same limit. The paper also derives a power theory: under local alternatives the statistic has a shifted normal limit, so the test can detect departures at a specific rate. If correct, practitioners can test the structural coefficient without estimating the error covariance matrix, even with many instruments and network or spatial dependence.","feed_headline":"IV test stays valid when instruments outnumber observations","feed_subtitle":"A self-normalized statistic stays standard normal under network or spatial error dependence.","key_machinery":"The central object is the self-normalized quadratic form \\(Q_N\\) built from the singular value decomposition of \\(Z/\\sqrt{N}\\). The instrument matrix is expanded as \\(Z/\\sqrt{N} = \\sum_{\\ell} \\sqrt{\\lambda_\\ell} q_\\ell w_\\ell'\\), and the statistic compares the projection of the self-normalized outcome \\(\\underline{Y}\\) onto the eigenvectors \\(q_\\ell\\) against the trace of the Gram matrix: \\(\\underline{Y}' S_N \\underline{Y} - \\operatorname{tr}(S_N)/N\\). Self-normalization by \\(\\|Y^*\\|\\) makes the statistic scale-free, while the spectral average over \\(\\lambda_\\ell\\) lets random matrix theory control the high-dimensional fluctuations without estimating the \\(N\\times N\\) error covariance matrix. The proof engine is a random-matrix concentration lemma for quadratic forms in iid entries, together with a plug-in estimator of \\(\\operatorname{tr}(\\$Sigma^{2}$)\\). The power theorem additionally decomposes the statistic into a null part and deterministic drift terms involving the first-stage coefficients and the correlation between the structural and first-stage errors.","core_discovery":"Under Assumptions 1–3, with instruments generated as \\(z_i = \\$Sigma^{{1/2}}$ f_i\\) for iid entries \\(f_i\\) and \\(K/N\\) converging to a positive constant, the oracle statistic \\[Q_N = \\sqrt{\\frac{$N^{2}$}{2\\operatorname{tr}(\\$Sigma^{2}$)}}\\left(\\underline{Y}' S_N \\underline{Y} - \\frac{1}{N}\\operatorname{tr}(S_N)\\right)\\] satisfies \\(Q_N \\xrightarrow{d} N(0,1)\\) under \\(H_0:\\$\\beta$=\\beta_0\\), where \\(\\underline{Y}\\) is the self-normalized residual vector and \\(S_N = Z'Z/N\\). Replacing \\(\\operatorname{tr}(\\$Sigma^{2}$)\\) with the ratio-consistent estimator \\(\\widehat{\\operatorname{tr}(\\$Sigma^{2}$)} = \\frac{1}{N(N-1)}\\sum_{i\\ne j}(z_i'z_j)^2\\) gives the feasible statistic \\(\\widehat{Q}_N\\) with the same standard normal limit. Under local alternatives defined by \\($h^{2}$ = N\\$\\Delta$^2/(2\\operatorname{tr}(\\$Sigma^{2}$))^{2/5}\\), the statistic is asymptotically normal with a non-central drift, yielding a nondegenerate power function. The paper’s stated contribution is that this is the first inference procedure that is simultaneously dimension-robust, weak-identification-robust, and robust to general error dependence of network or spatial type.","pith_inferences":["By inverting the test, one could construct identification-robust confidence sets for \\(\\beta\\) that retain the same dimension- and dependence-robustness; the paper itself only presents hypothesis tests.","The dependence of the detection rate on \\((2\\operatorname{tr}(\\Sigma^2))^{1/5}\\) suggests that instrument spectra matter for power: instruments whose eigenvectors spread the signal across many directions may make smaller effects detectable, a design implication the paper does not draw.","If the delocalization condition \\(\\sum_i |u_i|^3 = o_P(1)\\) fails—for example, when one error dominates the sample—the normal approximation may break down; a practical safeguard would estimate this quantity from residuals and flag cases where it is not small.","A bootstrap variant, in the spirit of dimension-agnostic bootstrap tests but combined with self-normalization, could extend the approach to fixed \\(K\\), which the authors note is outside the present theory."],"forward_implications":["Applied IV studies with many instruments—judge designs, shift-share instruments, and Fama-MacBeth regressions—could test the causal coefficient using standard normal critical values even when the number of instruments is proportional to or larger than the sample size.","The test remains valid under weak identification, so low first-stage F statistics do not invalidate it.","Because the limiting distribution does not require knowledge of the error covariance matrix, network and spatial dependence can be accommodated without cluster-robust or spatial-HAC corrections.","The local power result provides a concrete detection rate: departures of size \\(\\Delta\\) satisfying \\(h^2 = N\\Delta^2/(2\\operatorname{tr}(\\Sigma^2))^{2/5} > 0\\) are detectable, with power increasing in \\(h\\).","The feasible statistic inherits the standard normal limit, so the method is operational with no user-chosen tuning parameter beyond the instrument set."],"supporting_citations":[{"why":"The proofs of Theorem 1 and Corollary 1 are stated to be identical to the theorem and corollary of this companion paper, which supplies the entire asymptotic normality argument.","marker":"Feng et al. (2024)"},{"why":"It supplies the ratio-consistent estimator \\(\\widehat{\\operatorname{tr}(\\Sigma^2)}\\) used to turn the oracle statistic into the feasible statistic.","marker":"Li and Chen (2012)"},{"why":"It supplies the concentration lemma for quadratic forms of iid vectors that the power-proof in the appendix relies on.","marker":"Bai and Silverstein (1998)"}],"fun_headline_variants":["IV test survives when instruments exceed sample size","Weak-IV-robust test for high-dimensional instruments","Self-normalized statistic tames network-dependent IV errors","More instruments than data? IV test still works","Test for IV with many instruments and general dependence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on an unpublished companion theorem whose proof is not reproduced here applying unchanged to this IV setting, plus the delocalization condition \\(\\sum_i |u_i|^3 = o_P(1)\\), which is assumed rather than derived from primitive network or spatial dependence conditions.","fun_headline_variants_meta":{"raw":{"variants":["IV test survives when instruments exceed sample size","Weak-IV-robust test for high-dimensional instruments","Self-normalized statistic tames network-dependent IV errors","More instruments than data? IV test still works","Test for IV with many instruments and general dependence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001435,"raw_usage":{"total_tokens":5769,"prompt_tokens":912,"completion_tokens":4857,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":4785}},"tokens_in":528,"tokens_out":4857,"duration_ms":32263,"temperature":1.0,"reasoning_tokens":4785,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:30:10.666410+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-derive the omitted proof of Theorem 1 in this IV model; if it requires an additional condition not stated in the paper, such as a bound on error-dependence strength, the claim as stated would fail. Alternatively, simulate the feasible statistic under the null with strongly concentrated errors—one error carrying most of the norm—and check whether size is still near the nominal level.","supporting_citations":[{"cited_title":"Jaidee, G","cited_arxiv_id":null,"evidence_quote":"The proofs of Theorem 1 and Corollary 1 are stated to be identical to the theorem and corollary of this companion paper, which supplies the entire asymptotic normality argument."}],"review_version":1}