{"id":"0c5a1702-e754-4ade-ab78-fa465f4f8478","arxiv_id":"2412.14778","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A sieve-based, heteroskedasticity-robust LM test for linearity of the spatial lag in spatial autoregressive models is derived with standard normal asymptotics and applied to Finnish municipality tax data.","lead":"This paper proposes a statistical test for whether spatial interaction effects are linear, estimating only the simple linear model under the null hypothesis. A generalist might care because rejecting linearity changes conclusions about peer effects, tax competition, and how shocks spread through networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's null distribution is the load-bearing claim, but its proof is imported from Gupta (2018) without verification, and the simulations violate p^3/n=o(1).","rationale":"The reader's weakest assumption is exactly the missing proof of Theorem 2. Our independent read of Appendix A confirms that the proof of Theorem 2 invokes Gupta (2018) Remark 3 with no details. We also checked the quadratic-form structure: after standardizing by σ_i, Q is an idempotent quadratic form in independent variables, so the CLT is plausibly true, but it must be proved. The Monte Carlo design p=floor(n^{1/3}) violates p^3/n=o(1), so it cannot substitute for the proof. We therefore agree with the CONDITIONAL verdict: the paper should be accepted only after a complete proof of Theorem 2 (or a precise reference with verified conditions) and, ideally, simulations inside the stated rate condition.","tokens_in":28840,"tokens_out":9742,"duration_ms":64512,"concrete_test":"Independently derive the CLT for Q=nd'H^{-1}d. Write Q=η'C'Cη where η_i=ǫ_i/σ_i and C=(J'WΣW'J)^{-1/2}J'WΣ^{1/2} with W=M^{-1/2}M_N^MM^{-1/2}Z'; verify C C'=I_r with r=p+k+1. Then apply a quadratic-form CLT (e.g., de Jong 1987 or Hall 1984) for independent η_i with Eη_i=0 and Eη_i²=1: show (Q-r)/√(2r)→N(0,1) by checking max_i (C'C)_{ii}→0 and the Lindeberg condition under Assumptions 1 and 5. If max_i (C'C)_{ii} does not vanish or Var(Q)/r does not converge to 2, Theorem 2 fails; if it does, the proof gap is fillable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 2: under H0, Assumptions 1-6, ν>5/2, p^3/n=o(1), T→N(0,1). The proof in Appendix A consists of two sentences: 'By Theorem 1, it suffices to show... Under the null hypothesis, the above CLT is for a SAR model with an increasing number of instruments. By Remark 3 of Gupta (2018), the claim follows by a trivial modification of the arguments in the proof of Theorem 3.3 therein to allow for heteroskedastic innovations.' No statement of Gupta's Remark 3 or Theorem 3.3 is given, and no check is provided that its assumptions are satisfied by the present spatial, heteroskedastic, growing-instruments setting. In particular, the quadratic form nd'H^{-1}d must be shown to satisfy a CLT with mean p and variance 2p after studentization; the required conditions on the coefficient matrix and on the standardized innovations are never stated. Theorem 1 only justifies replacing \\hat d and \\hat H by population quantities; it does not deliver the distributional result. Thus the asymptotic null distribution of T is unproved in this paper. The Monte Carlo does not fill this gap: it sets p=floor(n^{1/3}), so p^3/n=O(1) rather than o(1), placing the experiments outside the theorem's stated conditions. The usefulness of the test depends entirely on Theorem 2; absent a full proof, the central claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a test for the null hypothesis that the spatial interaction function in a spatial autoregressive model is linear. The test is constructed from a sieve approximation to the nonlinear component and follows the Lagrange Multiplier principle, so that only the linear SAR model is estimated. The test statistic T in (3.9) is a standardized quadratic form in the 2SLS gradient. The paper claims that T is asymptotically standard normal under the null (Theorem 2), that the test is consistent (Theorem 3), and that it detects local alternatives of order p^{1/4}/sqrt(n) (Theorem 4). Monte Carlo evidence and an empirical application to Finnish municipal tax competition are also provided.","tokens_in":29188,"tokens_out":15236,"duration_ms":104651,"significance":"If the asymptotic results are correct, the paper offers a computationally simple and economically motivated specification test for a practically important question, complementing the existing literature on testing nonlinearity in spatial lags. The paper is clearly written, includes several structural models that justify the null hypothesis, and presents a credible empirical illustration. The main weakness is that the central null distribution is not proved in the paper: Theorem 2 is delegated to a 'trivial modification' of a result in the first author's earlier paper, and the Monte Carlo design violates the condition p^3/n=o(1). These are load-bearing issues that prevent acceptance in the current form.","major_comments":[{"comment":"The proof of Theorem 2 does not establish the asymptotic normality of the quadratic form. After Theorem 1, the proof says: 'By Remark 3 of Gupta (2018), the claim follows by a trivial modification of the arguments in the proof of Theorem 3.3 therein to allow for heteroskedastic innovations.' Neither Remark 3 nor Theorem 3.3 of Gupta (2018) is stated, and no verification is provided that the conditions of that result hold for the present SAR model with spatial dependence, a growing number of instruments (m ~ p), and heteroskedastic innovations. The required CLT for the studentized quadratic form nd'H^{-1}d, with H defined in (4.6), needs explicit conditions on the weight matrix, the instrument set, and the innovation distribution; these are not given. Since Theorem 2 provides the null distribution used in practice and is also invoked in the proofs of Theorems 3 and 4, this is a load-bearing gap.","section":"Section 4.1, Theorem 2 (proof in Appendix A)"},{"comment":"The Monte Carlo design sets p = floor(n^{1/3}). For the reported sample sizes n = 100, 200, 400, 700, 1000, 2000, this gives p = 4, 5, 7, 8, 10, 12 and p^3/n approximately 0.64, 0.63, 0.86, 0.73, 1.00, and 0.86, so p^3/n is O(1) and not o(1). Theorems 1 and 2 require p^3/n = o(1). Hence the simulations are outside the stated theoretical conditions, and the reported size results based on the normal approximation do not validate Theorem 2. The authors should either re-run the simulations with p = o(n^{1/3}) (for example p = floor(n^{1/4}) or p = floor(n^{0.3})), or extend the theory to allow p = n^{1/3}.","section":"Section 5 (Monte Carlo)"}],"minor_comments":[{"comment":"The notation 'n/p(ν+1/2) = o(1)' is ambiguous; it should be written as n / p^{ν+1/2} = o(1).","section":"Section 4, Assumption 5"},{"comment":"In the text after (A.17), the first two terms are said to be Op(p^2/n^{3/2}); based on the bounds \\|\\hat d - d\\| = Op(p^{3/2}/n) and \\|d\\| = Op(\\sqrt{p}/n), these terms are actually Op(p^2/n^2). The conclusion is unaffected under p^3/n = o(1), but the stated rate is a typo.","section":"Appendix A, proof of Theorem 1"},{"comment":"The note says 'For lattice, n = 100, 210, 400, 702, 992, 1980', but the table rows list n = 100, 200, 400, 700, 1000, 2000; the lattice sample sizes should be clearly displayed in the table itself to avoid confusion.","section":"Section 5, Tables 1 and 2"},{"comment":"In equation (2.2), the parentheses are likely misplaced: the expression inside the large parentheses should be \\tilde{k}_i + \\sum_{j\\ne i} \\tau_j / c(d_{ij}), not \\tilde{k}_i + \\sum_{j\\ne i} \\tau_j / c(d_{ij}) as currently appears to place \\tilde{k}_i inside the sum.","section":"Section 2, Example 1"},{"comment":"The table notes define significance stars, but no stars appear in the tables; either add the stars or remove the note.","section":"Section 6, Tables 5 and 6"}],"recommendation":"major_revision","confidential_remarks":"The central theoretical result is imported from the first author's earlier work (Gupta, 2018) with no proof or even statement of the cited result, and the Monte Carlo design violates the paper's own condition p^3/n = o(1). Both issues are fixable in a major revision, but they are substantial. I would not recommend acceptance in the present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real and useful extension: a Ramsey-style LM test for linearity of the spatial lag in SAR models, with diverging sieve dimension p and heteroskedasticity-robust 2SLS. It estimates only under the null, so it is cheap, and it is careful to distinguish itself from Hoshino (2022), which tests a different object. The structural examples and the Finnish tax competition application are well done. Second, the paper's central theorem is not actually proved. Theorem 2, which gives the N(0,1) null distribution, is dispatched in two sentences: 'By Remark 3 of Gupta (2018), the claim follows by a trivial modification...' There is no statement of Gupta's Remark 3 or Theorem 3.3, and no verification that their conditions hold in this spatial, heteroskedastic, growing-instrument setting. That is the load-bearing result; without it, the test has no justified null distribution. The passing references to Korolev (2019) and Robinson (2008) do not fill the gap. This needs a full proof, or at least a precise statement and verification of the imported CLT.\n\nThe Monte Carlo has a related problem: it sets p = floor(n^{1/3}), so p^3/n = O(1), not o(1), which violates the condition used in Theorems 1 and 2. The simulations show good size with chi-squared-based critical values, but that does not rescue the theory. A referee should ask either for a design inside p^3/n = o(1) or for a discussion of why the condition can be relaxed in practice.\n\nWhat is genuinely good here: the test is novel, computationally straightforward, and the paper is honest about what Hoshino (2022) does and does not test. The sieve LM construction is sensible, and the power results, while based on lattice simulations, look reasonable. The paper's own Theorem 1 and the appendix do real work to show T is approximated by a quadratic form in the innovations; what is missing is the final CLT for that quadratic form, and that is exactly the imported step.\n\nMy overall take: this is a promising paper with a real gap. The idea deserves referee time, but the current version is not self-contained on its main distribution theorem. I would send it to review with a strong request to complete the proof of Theorem 2 and fix the simulation rate mismatch. If the CLT can be verified under these spatio-heteroskedastic assumptions, this will be a useful contribution to spatial econometrics.","headline":"A genuinely new Ramsey-style test for linearity of the spatial lag, but the null distribution is inherited from Gupta (2018) rather than proved, and the simulations sit outside the stated p^3/n = o(1) condition.","tokens_in":29683,"tokens_out":5430,"would_cite":true,"duration_ms":43180,"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":"A single score statistic, computed from the linear spatial model alone, tests whether the spatial interaction function is truly linear.","keywords":["Spatial autoregression","Series expansion","Specification testing","Nonparametric","Spatial interaction function","Ramsey RESET","Lagrange multiplier test","Heteroskedasticity robustness"],"falsifier":"Simulate the null model with spatially dependent heteroskedastic errors under the paper's Assumptions 1-6, choose the sieve dimension as $p \\approx n^{1/4}$ so that $p^3/n \\to 0$, and check whether the empirical rejection rate of $T$ at the 5% normal critical value converges to 5% over many Monte Carlo replications; a clear size distortion would show that the borrowed CLT does not govern this statistic.","tokens_in":28681,"feed_emoji":"🧮","tokens_out":13332,"duration_ms":93171,"temperature":0.7,"pith_summary":"The paper proposes a test of the null hypothesis that the spatial interaction function $f$ in a spatial autoregressive model $y_i = \\lambda w_i' y + f(w_i' y) + x_i' \\beta + \\varepsilon_i$ is identically zero, so that the familiar linear SAR specification is correct. The test expands $f$ in a sieve of $p$ basis functions and builds a Lagrange-multiplier statistic from the gradient of the 2SLS objective function evaluated under the null, which means only the linear SAR model must be estimated. The statistic is centred and rescaled as $T = (n \\hat{d}'\\hat{H}^{-1}\\hat{d} - p)/\\sqrt{2p}$, and the paper claims it is asymptotically standard normal under the null when the sieve dimension satisfies $p^3/n \\to 0$, consistent against fixed nonlinear alternatives, and able to detect local alternatives of order $p^{1/4}/\\sqrt{n}$. Monte Carlo experiments show good size and power across several network designs, and an application to Finnish municipal property tax data corroborates the earlier conclusion of no tax competition when municipality fixed effects are included, while rejecting linearity in the same model without fixed effects. The practical payoff is a specification test for the most common empirical model of social and spatial interaction that avoids estimating any nonparametric function.","feed_headline":"New test spots nonlinearity in spatial spillovers","feed_subtitle":"Runs on a standard linear spatial model, yet catches curvature that misspecified tax-competition studies hide.","key_machinery":"The central object is the truncated series (sieve) expansion $f(z) = \\sum_{j=1}^{p} \\alpha_j \\psi_j(z) + r(z)$ of the unknown spatial interaction function, which turns the hypothesis $H_0: f = 0$ into the linear restriction $\\alpha_1 = \\cdots = \\alpha_p = 0$ on $p$ sieve coefficients. The test statistic is the centred and scaled score (Lagrange multiplier) statistic $T = (n \\hat{d}'\\hat{H}^{-1}\\hat{d} - p)/\\sqrt{2p}$, where $\\hat{d}$ is the gradient of the 2SLS objective evaluated at the null estimates and $\\hat{H}$ is its robust covariance; the centring by $p$ and scaling by $\\sqrt{2p}$ mirrors the standardization of a $\\chi^2_p$ variable. The asymptotic argument approximates $T$ by a quadratic form in the innovations with population-anchored weights (Theorem A1 in the appendix) and then invokes a central limit theorem for quadratic forms with a growing number of instruments to obtain the standard normal null distribution.","core_discovery":"Under $H_0: f \\equiv 0$ in the spatial autoregressive model with an unknown link function, the statistic $T = (n \\hat{d}'\\hat{H}^{-1}\\hat{d} - p)/\\sqrt{2p}$ is shown to be asymptotically standard normal, provided that the sieve approximation error decays at rate $\\nu > 5/2$ and that the sieve dimension grows no faster than $p^3/n \\to 0$ (Theorem 2). Here $\\hat{d}$ is the gradient of the IV objective under the null and $\\hat{H}$ is its heteroskedasticity-robust covariance estimator. The paper proves further that the test is consistent against any fixed nonlinear alternative (Theorem 3) and that it detects local alternatives of the form $\\alpha_j = p^{1/4} n^{-1/2} \\delta_j$ with a non-central normal limit (Theorem 4). The test requires estimation only under the linear null, and the paper recommends comparing the statistic against either standard normal critical values or the standardized $\\chi^2_p$ distribution for moderate $p$. The empirical section finds that the test fails to reject linearity for Finnish property tax rates with municipality and year fixed effects, matching the earlier finding of no tax competition, but strongly rejects linearity for the level specification without fixed effects, supporting the interpretation that earlier evidence of tax competition came from a misspecified model.","pith_inferences":["Inference: If the null approximation holds uniformly in the sieve dimension, the same statistic could be applied to test linearity of the link function in other network- and peer-effect models where $f$ acts on a network-weighted average of outcomes.","Inference: The Monte Carlo design $p = \\lfloor n^{1/3} \\rfloor$ makes $p^3/n = O(1)$, outside the theorem's condition; rerunning the size experiments with $p \\approx n^{1/4}$ would test whether the claimed normal approximation holds where the theory says it should.","Inference: The paper's empirical contrast — linearity not rejected with municipality fixed effects but rejected without them — suggests the test can serve as a diagnostic for omitted cross-sectional confounders, not only for functional form."],"forward_implications":["Estimating only the null linear SAR model, practitioners can now test for hidden nonlinearity in the spatial lag, avoiding the computational cost of nonparametric estimation of the link function.","The test is consistent: any fixed nonlinearity in the spatial interaction function will be detected asymptotically.","Local alternatives of order $p^{1/4}/\\sqrt{n}$ are detectable, a dampened rate typical of series-based specification tests and noticeably slower than the parametric $1/\\sqrt{n}$ rate.","Using standardized $\\chi^2_p$ critical values gives reliable size for small $p$ and moderate samples, complementing the asymptotic normal approximation."],"supporting_citations":[{"why":"Supplies the central limit theorem for quadratic forms with a growing number of instruments that the proof of Theorem 2 imports to obtain the null distribution of T.","marker":"Gupta (2018)"},{"why":"Establishes the series-based specification-testing approach and the standardization of chi-square-type statistics with p growing to infinity that the test statistic mimics.","marker":"Hong and White (1995)"},{"why":"Provides the limit behaviour of chi-square-type tests when the number of conditional moments grows, the same centring-and-scaling device used here.","marker":"de Jong and Bierens (1994)"},{"why":"The RESET principle, testing neglected nonlinearity by adding polynomial terms under the null, is the conceptual template the statistic follows.","marker":"Ramsey (1969)"},{"why":"The spatial two-stage least squares estimator used to estimate the model under the null.","marker":"Kelejian and Prucha (1998)"},{"why":"The regularity conditions on the spatial weight matrix W that Assumptions 3 and 4 borrow to control spatial dependence.","marker":"Lee (2002)"},{"why":"Supplies the cross-sectional dependence bound formalized in Assumption 6 for the weak law of large numbers on the instrument matrices.","marker":"Lee and Robinson (2016)"},{"why":"Gives the sieve approximation error decay rates that Assumption 5 uses to set nu greater than 5/2.","marker":"Chen (2007)"},{"why":"The Finnish property tax dataset and the policy-based instrument used in the empirical application, and the prior finding of no tax competition that the test corroborates.","marker":"Lyytikäinen (2012)"}],"fun_headline_variants":["New test reveals nonlinear spatial spillovers","Spatial linearity test flags tax competition misspecification","Simple test catches nonlinear spatial interaction","Test roots out false tax competition in spatial models","Linear spatial model test exposes hidden nonlinearity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The asymptotic null distribution of $T$ is the load-bearing premise, and it is imported from a central limit theorem proved in earlier work in a different (non-spatial) setting, with the paper asserting rather than verifying that the required quadratic-form CLT holds under its spatial, heteroskedastic, growing-instruments assumptions.","fun_headline_variants_meta":{"raw":{"variants":["New test reveals nonlinear spatial spillovers","Spatial linearity test flags tax competition misspecification","Simple test catches nonlinear spatial interaction","Test roots out false tax competition in spatial models","Linear spatial model test exposes hidden nonlinearity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000851,"raw_usage":{"total_tokens":3690,"prompt_tokens":924,"completion_tokens":2766,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":2698}},"tokens_in":540,"tokens_out":2766,"duration_ms":18416,"temperature":1.0,"reasoning_tokens":2698,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:55:04.061871+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the null model with spatially dependent heteroskedastic errors under the paper's Assumptions 1-6, choose the sieve dimension as $p \\approx n^{1/4}$ so that $p^3/n \\to 0$, and check whether the empirical rejection rate of $T$ at the 5% normal critical value converges to 5% over many Monte Carlo replications; a clear size distortion would show that the borrowed CLT does not govern this statistic.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the central limit theorem for quadratic forms with a growing number of instruments that the proof of Theorem 2 imports to obtain the null distribution of T."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the series-based specification-testing approach and the standardization of chi-square-type statistics with p growing to infinity that the test statistic mimics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the limit behaviour of chi-square-type tests when the number of conditional moments grows, the same centring-and-scaling device used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The RESET principle, testing neglected nonlinearity by adding polynomial terms under the null, is the conceptual template the statistic follows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The spatial two-stage least squares estimator used to estimate the model under the null."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The regularity conditions on the spatial weight matrix W that Assumptions 3 and 4 borrow to control spatial dependence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the cross-sectional dependence bound formalized in Assumption 6 for the weak law of large numbers on the instrument matrices."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the sieve approximation error decay rates that Assumption 5 uses to set nu greater than 5/2."}],"review_version":1}