{"id":"fc6321f5-ec0a-44f5-8041-0742f686fb49","arxiv_id":"1908.02552","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A fully modified GLS estimator for multivariate cointegrating polynomial regressions, using a banded modified Cholesky inverse-covariance estimator, is derived and shown to improve finite-sample estimation and inference.","lead":"Econometricians propose a generalized least squares estimator for multi-equation nonlinear cointegration models, built on a banded estimate of the inverse error covariance matrix and bias corrections. Simulations show gains in accuracy and test performance, but the illustrative CO2-income application is rejected by the paper's own specification tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Assumption 2 is derived in the proof of Theorem 3(c), not merely asserted, and the FM-GLS limit (3.13) is internally consistent.","rationale":"The reader's central worry is that Assumption 2 is an unproved primitive. In fact, the manuscript proves it in the course of Theorem 3(c), and the proof is standard: OLS in cointegrating polynomial regressions has the usual mixed convergence rates, and the scaled design matrix has a well-behaved limit. I checked the surrounding argument for internal consistency: the score term in Theorem 3(c) converges to \\int J\\Omega_{uu}^{-1}dB_u + B_{\\epsilon\\eta}; the subtracted v-term converges to \\int J\\Omega_{uu}^{-1}\\Omega_{uv}\\Omega_{vv}^{-1}dB_v + B_{vu}; the estimator \\hat B^+ is explicitly constructed to cancel B_{\\epsilon\\eta} - B_{vu}; hence the zero-mean Gaussian mixture limit in (3.13) follows. The Wald statistic in Theorem 4 uses the correct sandwich form and the conditional chi-square argument is valid. Remaining concerns are finite-sample: the empirical application proceeds after all three cointegration tests reject the quadratic specification, and the calibration simulation shows confidence intervals under-cover. These are honestly disclosed in the paper and do not undermine the asymptotic theorem. Therefore the conditional verdict can stand as is; the specific Assumption 2 objection does not land.","tokens_in":52123,"tokens_out":36219,"duration_ms":358359,"concrete_test":"The one check worth running is an independent verification of the one-line proof of Assumption 2: analytically or by simulation, confirm that \\|G_T Z'Z G_T\\| = O_p(1) for the SUCPR design with deterministic trends and integer powers up to s_i, and that first-stage OLS satisfies \\|G_T^{-1}(\\hat\\beta_{OLS}-\\beta)\\| = O_p(1). If either fails, the residual bound and Theorem 3(c) would need repair; if both hold, the reader's Assumption 2 concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I could not find a load-bearing flaw in the central theoretical claim. The reader's flagged weakest assumption, Assumption 2, is not left unproved: the proof of Theorem 3(c) derives it from first-stage OLS rates, writing \\|\\hat u-u\\|_2 \\le \\|G_T^{-1}(\\hat\\beta_{OLS}-\\beta)\\| \\|G_T Z'Z G_T\\|^{1/2} = O_p(1), where \\|G_T^{-1}(\\hat\\beta_{OLS}-\\beta)\\| = O_p(1) is the standard OLS convergence in cointegrating polynomial regressions and \\|G_T Z'Z G_T\\| = O_p(1) follows because the fixed-degree polynomial regressors have scaled sample moments converging to integrals of r^k or B^k. The proof of Theorem 3(c) then uses only these rates and the consistency results in Theorem 2 and Supplement Theorem S2; the bias terms cancel as B^+ = B_{\\epsilon\\eta} - B_{vu}, and \\hat B^+ is constructed to match this. This gives exactly (3.13), and Theorem 4's Wald statistic follows from the conditional mixed normal limit. The paper's own stated limitations (rejection of cointegration in the empirical model, confidence intervals below nominal coverage) concern finite-sample and application validity, not the asymptotic central claim. No internally inconsistent step was identified.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Fully Modified GLS estimator for seemingly unrelated cointegrating polynomial regressions with deterministic trends and integer powers of I(1) regressors. The inverse autocovariance matrix of the errors is estimated directly through a modified Cholesky block decomposition based on first-stage OLS residuals, and the estimator applies a second-order bias correction so that the scaled estimation error converges to a zero-mean mixed Gaussian distribution. The paper also develops multivariate KPSS-type cointegration tests, studies their size and power by simulation, and applies the methods to a six-country Environmental Kuznets Curve model. The central theoretical result is Theorem 3(c), equation (3.13), giving the limiting distribution of FM-GLS, and Theorem 4, which derives a chi-square Wald statistic.","tokens_in":52475,"tokens_out":18670,"duration_ms":205392,"significance":"If the asymptotic results are correct, the paper makes a useful contribution by providing a feasible fully modified estimator for multivariate cointegrating polynomial regressions that avoids leads-and-lags augmentation and supports standard chi-square inference. The supplement contains the auxiliary lemmas underlying the consistency of the banded inverse autocovariance estimator, and the proof of Theorem 3(c) derives the previously flagged Assumption 2 from standard OLS rates rather than simply assuming it. The simulations compare feasible and infeasible estimators, and the paper is unusually candid about finite-sample failures, including the large MSE for high persistence at T=100 and the below-nominal coverage of empirical confidence intervals. The theoretical core is internally consistent: the bias terms cancel by construction in the definition of the FM-GLS correction, and the claimed mixed normal limit follows from standard cointegrating polynomial regression asymptotics.","major_comments":[],"minor_comments":[{"comment":"Assumption 2 is stated as a mild condition satisfied by OLS residuals, but the verification is deferred to the proof of Theorem 3(c). Since Theorem 2 already relies on Assumption 2, the verification should be moved to Section 3.2 or explicitly referenced there so that the feasibility of the GLS estimator does not depend on a later proof.","section":"Section 3.2, Assumption 2"},{"comment":"The displayed bound for the OLS residuals reads ||hat u - u||_2 <= ||G_T^{-1}(hat beta_OLS - beta)|| ||G_T Z'Z G_T||, but the correct intermediate factor is ||Z G_T|| = ||G_T Z'Z G_T||^{1/2}, not the full norm ||G_T Z'Z G_T||. The conclusion O_p(1) is unaffected because ||G_T Z'Z G_T|| = O_p(1), but the inequality as written should be corrected.","section":"Proof of Theorem 3(c)"},{"comment":"Assumption 3 is printed as \"1/q_T + q_T^3/T T -> 0\"; the intended condition is 1/q_T + q_T^3/T -> 0, and Theorem 5(b) similarly appears to omit a division sign in the requirement q_T/b_T + b_T/T -> 0. These typos should be fixed.","section":"Assumption 3 and Theorem 5(b)"},{"comment":"All three cointegration tests reject the null of cointegration at the 5% level, yet the paper proceeds to estimate and interpret the EKC model. The authors disclose this, but the framing should make explicit that the empirical application is illustrative and cannot serve as evidence in favor of the quadratic EKC specification, given the formal rejection.","section":"Section 5, Table 5"},{"comment":"The reference list contains two entries, Ing, Chiou, and Guo (2016a) and (2016b), with the same title and journal information. The duplication should be reconciled or the two distinct contributions should be clearly distinguished.","section":"References"},{"comment":"The title uses \"Efficient Estimation,\" but the paper does not establish a formal efficiency lower bound or compare the limit distribution to a semiparametric efficiency bound. Consider qualifying the claim, for example by referring to efficiency in the GLS sense, to avoid overstating the theoretical result.","section":"Title and Section 1"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is within the scope of econometric theory and methods. The self-citation to Beutner et al. (2019) is used only as a tool for the modified Cholesky block decomposition and does not appear to be a citation-padding concern. I did not find a load-bearing technical error in the central asymptotic claims; the issues identified are presentation and framing problems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe bottom line: this is a solid, incremental paper. The FM-GLS estimator is a genuine addition to the seemingly unrelated cointegrating polynomial regression toolkit, and the asymptotic machinery checks out. The infeasible GLS limit coincides with MSUR, as the authors note, so the real novelty is the feasible estimator built on banded inverse autocovariance estimation plus the bias corrections. That part is well executed and the mixed normal limit (3.13) is derived from primitive OLS rates, not assumed. I do not see a load-bearing flaw in the central theory; Assumption 2 is actually proved in the proof of Theorem 3(c), so that concern disappears on close reading.\n\nWhat they do well: the BIAM construction is computationally clean and avoids leads-and-lags; the Wald test in Theorem 4 is standard and works; the Monte Carlo is honest and relatively comprehensive. It includes the known failure case (high persistence, high endogeneity, T=100) and reports it openly. That kind of transparency counts.\n\nThe soft spots are in the empirical application and, to a lesser extent, the manuscript's completeness. All three cointegration tests reject the maintained quadratic EKC specification at 5%, and the authors proceed anyway. They justify this by comparability with Wagner et al. (2020), but it does mean the reported EKC coefficients come from a model the paper itself rejects. The confidence intervals also undercover significantly—89% for FM-GLS at nominal 95%, worse for FM-SOLS and FM-SUR—and the authors are upfront about it, even calibrating the scaling factors. Still, the applied reader should treat the EKC numbers as illustrative rather than decisive. The proofs depend on supplementary lemmas S1-S4 that are not fully in the preprint, and no code or replication data are shipped. These are addressable requirements for a journal version, not deep flaws.\n\nThe citation pattern is fine. The self-citation to Beutner et al. (2019) is for the covariance parametrization tool, not to inflate novelty.\n\nBottom line: the paper deserves a serious referee. The theory is a modest but sound step forward, the simulations support the claims, and the empirical caveats are at least acknowledged. I would send it to a good econometrics journal and ask the authors to make the supplement available and to strengthen the empirical discussion, possibly by presenting the specification tests as evidence against the quadratic model rather than continuing as if they had passed.","headline":"A credible and useful FM-GLS for cointegrating polynomial regressions; the asymptotic core holds up, but the empirical application and under-covering intervals keep it from being a clean yes.","tokens_in":52934,"tokens_out":2385,"would_cite":true,"duration_ms":27647,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","62P20","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"FM-GLS makes nonlinear cointegration inference chi-square.","keywords":["Cointegrating Polynomial Regression","Seemingly Unrelated Cointegrating Polynomial Regression","Fully Modified Estimation","Generalized Least Squares","Modified Cholesky Block Decomposition","Banded Inverse Autocovariance Matrix","Cointegration Testing","Environmental Kuznets Curve"],"falsifier":"Run the paper's Setting B at the high-persistence, high-endogeneity corner (eigenvalues in [0.8,0.95], theta=0.5) with T=100 and record the Monte Carlo distribution of the residual-distance measure and of the normalized FM-GLS estimator; if the distance is not bounded or the empirical distribution fails to converge to the claimed zero-mean Gaussian mixture as T grows, Assumption 2 or the limit in Theorem 3(c) is false.","tokens_in":51918,"feed_emoji":"📊","tokens_out":10234,"duration_ms":102199,"temperature":0.7,"pith_summary":"This paper develops a feasible fully modified generalized least squares (FM-GLS) estimator for seemingly unrelated cointegrating polynomial regressions, where cointegrating relations contain deterministic trends, stochastic trends, and integer powers of integrated regressors. The paper's central claim is that, after correcting second-order biases, the normalized FM-GLS estimator converges to a zero-mean Gaussian mixture distribution, so Wald tests are asymptotically chi-square under the null. This matters because existing fully modified estimators in this setting require leads-and-lags augmentation or kernel-based long-run covariance estimation, and they tend to understate parameter uncertainty in applications. The paper also constructs multivariate KPSS-type cointegration tests based on filtered residuals. In the Environmental Kuznets Curve application, FM-GLS supports the inverted-U hypothesis but produces wider confidence intervals than FM-SUR or FM-SOLS, with closer-to-nominal coverage in a calibrated simulation.","feed_headline":"FM-GLS yields chi-square tests for nonlinear cointegration","feed_subtitle":"Avoids leads-and-lags augmentation and gives more honest uncertainty in the Environmental Kuznets Curve.","key_machinery":"The central object is the Modified Cholesky Block Decomposition (MCBD) of the inverse autocovariance matrix, $\\Sigma_u^{-1}=M_u'S_u^{-1}M_u$, built from linear minimum MSE predictors of the innovations; banding the sub-diagonal blocks yields the Banded Inverse Autocovariance Matrix (BIAM), which is estimated directly from first-stage OLS residuals rather than by inverting a large covariance matrix. The MCBD lets the GLS estimator be read as weighted least squares applied to filtered data, and the same estimated VAR coefficients feed the long-run covariance estimators and the second-order bias correction that removes the endogeneity and serial-correlation terms from the limiting distribution. Rate conditions on the banding parameter, inherited from fitting VARs of increasing order, make the estimation error asymptotically negligible.","core_discovery":"The paper's central discovery is a feasible estimator whose normalized error converges to a zero-mean Gaussian mixture. In notation, under the paper's assumptions, $G_T^{-1}(\\hat\\beta_{\\text{FGLS}}^+ - \\beta)$ converges weakly to $\\left(\\int_0^1 J(r)\\Omega_{uu}^{-1}J(r)'dr\\right)^{-1} \\int_0^1 J(r)\\Omega_{uu}^{-1}dB_{u.v}(r)$, a distribution that is normal with zero mean conditional on the driving Brownian motion. Because the conditional variance is consistently estimable, the Wald statistic in Theorem 4 is asymptotically $\\chi^2_k$. This is the first fully modified estimator for the SUCPR framework that obtains the mixed normal limit without leads-and-lags augmentation, using instead direct estimation of the inverse autocovariance matrix and an explicit second-order bias correction. The paper further claims the same machinery delivers multivariate KPSS-type tests whose null limit is $\\int_0^1 \\|W(r)\\|^2 dr$, free of nuisance parameters.","pith_inferences":["A testable extension the paper leaves implicit is to build the BIAM from a first-stage residual that is itself bias-corrected rather than OLS; the high-persistence, high-endogeneity corner of the simulations is where such a variant would show whether Assumption 2 or the second-order correction is the limiting factor.","Because the BIAM controls parameter proliferation by banding, the construction may extend to panels with cross-section dimension growing with T, but the fixed-n rate conditions in Assumptions 3 and 5 would need to be reworked for that regime.","The reported power loss of the prefiltered KPSS test suggests a size-power tradeoff that could be explored by choosing the filter order or mixing filtered and unfiltered residuals; this is not pursued in the paper."],"forward_implications":["Applied researchers can test linear hypotheses on individual coefficients of a seemingly unrelated cointegrating polynomial regression with conventional chi-square critical values, without leads-and-lags augmentation or simulation-based critical values.","FM-GLS should dominate the existing FM-SOLS and FM-SUR estimators in mean squared error, with the gains increasing in serial correlation and endogeneity; the simulations show the gains persist even for infeasible versions, pointing to the GLS weighting as the source.","The three multivariate KPSS-type tests give a nuisance-parameter-free null distribution, so the cointegration specification of a system can be checked without tabulating new critical values.","In the Environmental Kuznets Curve data, all three estimators support an inverted-U relation for all six countries, but confidence intervals differ sharply in width; the FM-GLS intervals are wider and their empirical coverage is closest to the nominal 95% level."],"supporting_citations":[{"why":"Defines the SUCPR framework and the FM-SOLS and FM-SUR estimators that FM-GLS is compared against; also supplies the six-country EKC dataset.","marker":"Wagner et al. (2020)"},{"why":"Introduces cointegrating polynomial regressions and the single-equation FM-OLS and KPSS-type test that this paper generalizes.","marker":"Wagner and Hong (2016)"},{"why":"Supplies the subsampling and Bonferroni testing strategy and the nonlinear-cointegration test setup used for the multivariate KPSS tests.","marker":"Choi and Saikkonen (2010)"},{"why":"Provides the fully modified correction idea that removes second-order bias and produces the zero-mean Gaussian mixture limit.","marker":"Phillips and Hansen (1990)"},{"why":"Introduces the modified Cholesky decomposition of the inverse covariance matrix that the MCBD generalizes.","marker":"Pourahmadi (1999)"},{"why":"Extends the modified Cholesky idea to a block decomposition for multivariate data, the MCBD used to estimate the inverse autocovariance matrix.","marker":"Kim and Zimmerman (2012)"},{"why":"Gives the rate conditions for fitting finite-order VARs to infinite-order processes, which underlie Assumption 3 on the banding parameter.","marker":"Lewis and Reinsel (1985)"},{"why":"Provides the univariate inverse-autocovariance estimation result that the paper's Theorem 2 proof generalizes to the multivariate setting.","marker":"Cheng et al. (2015)"},{"why":"Supplies the truncation and estimation bounds for inverse autocovariance matrices used in the consistency proof of the BIAM.","marker":"Ing et al. (2016b)"},{"why":"Justifies estimating long-run variances from autoregressive approximations, the alternative to kernel estimation used within the FM-GLS framework.","marker":"Berk (1974)"}],"fun_headline_variants":["FM-GLS makes nonlinear cointegration tests feasible","Fast fully modified GLS for polynomial cointegration","New estimator yields chi-square tests in cointegrated regressions","FM-GLS avoids leads-and-lags in nonlinear cointegration","Chi-square inference for Environmental Kuznets Curve via FM-GLS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that first-stage least-squares residuals remain within a bounded distance of the true innovations in overall size; the paper asserts this is mild and satisfied by least squares, but it is assumed rather than proven from primitive conditions, and the consistency of the banded inverse covariance estimate together with the FM-GLS normal-mixture limit both depend on it.","fun_headline_variants_meta":{"raw":{"variants":["FM-GLS makes nonlinear cointegration tests feasible","Fast fully modified GLS for polynomial cointegration","New estimator yields chi-square tests in cointegrated regressions","FM-GLS avoids leads-and-lags in nonlinear cointegration","Chi-square inference for Environmental Kuznets Curve via FM-GLS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2888,"prompt_tokens":925,"completion_tokens":1963,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":1883}},"tokens_in":541,"tokens_out":1963,"duration_ms":14858,"temperature":1.0,"reasoning_tokens":1883,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:40:33.950325+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the paper's Setting B at the high-persistence, high-endogeneity corner (eigenvalues in [0.8,0.95], theta=0.5) with T=100 and record the Monte Carlo distribution of the residual-distance measure and of the normalized FM-GLS estimator; if the distance is not bounded or the empirical distribution fails to converge to the claimed zero-mean Gaussian mixture as T grows, Assumption 2 or the limit in Theorem 3(c) is false.","supporting_citations":[],"review_version":1}