{"id":"75ef5867-405f-41c1-b1ce-4e3db7f22cf9","arxiv_id":"2604.14649","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A weighted-residual-process test with smoothed bootstrap avoids the degeneracy of classical ICM tests in regression models with a diverging number of predictors.","lead":"The paper proposes a new goodness-of-fit test for regression models when the number of predictors grows with the sample size, built from weighted residual processes and a smoothed bootstrap. It claims the test keeps its nominal size and can detect local model misspecifications that shrink at the n^{-1/2} rate, where classical ICM tests degenerate.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test requires full error-covariate independence, not just mean independence; under heteroskedastic nulls it has wrong size, so it does not test the stated conditional-mean hypothesis.","rationale":"The reader's weakest assumption correctly identified the independence assumption as load-bearing. However, the consequence is more severe than the reader's conditional-accept suggests. The paper's stated goal is to test whether the regression mean belongs to a parametric family, and it explicitly says the error distribution is unrestricted. Yet the test statistic's centering requires full independence between ε and X. Under the null with only mean independence, the test is inconsistent against a broad class of correctly specified models with heteroskedastic errors. This is not a minor technical gap that can be fixed by clarifying assumptions; it changes what the test actually measures. The test can be repurposed as a test of error-covariate independence, but as a specification test for the conditional mean it fails. The theoretical results (Theorems 1–3) may be correct conditional on the independence assumption, but the paper's central claim—that this is a valid model check for the regression mean in diverging-p settings—is not established. The internal contradiction between the stated null and the added independence assumption is explicit in the text, making the issue central rather than peripheral. Therefore the verdict should be REJECT, unless the authors substantially reframe the contribution and hypotheses.","tokens_in":18440,"tokens_out":6799,"duration_ms":60718,"concrete_test":"Simulate from a heteroskedastic null: n=200 or 400, p=10, X~N(0, I_p), Y = β^T X + ε with ε|X ~ N(0, (1 + 0.5 X_1)^2) (so E(ε|X)=0 and the linear mean is correct). Apply the proposed WICM test with a weight function for linear alternatives and smooth residual bootstrap, α=0.05, over 1000 replications. If the rejection rate is near 0.05, the independence assumption is not essential; if it is substantially larger (e.g., >0.5 at n=400), the test has incorrect size for conditional-mean nulls and the concern lands. An analytic check: compute the probability limit of n^{-1}∫|U_n(t)|²φ(t)dt under this null; a nonzero limit indicates the test rejects this null asymptotically.","verdict_should_be":"REJECT","load_bearing_attack":"Under H0 in (1.2), only the conditional mean is specified; the error distribution is unrestricted. The proposed statistic is built on the identity E{g0(X) exp(it e)} = E{g0(X)}E{exp(it ε)} = 0, which holds only if e = ε is independent of X, not merely mean-independent. The paper assumes independence immediately after (1.1), contradicting its claim that ε remains unrestricted. If ε|X has conditional variance depending on X, then even when E(ε|X)=0 and the mean is correctly specified, U_n(t) has nonzero mean: for small t, E{g0(X)[cos(tε)+sin(tε)]} ≈ -(t²/2)E{g0(X)ε²} ≠ 0. Hence WICM_n diverges and the test rejects a correctly specified conditional mean with probability tending to 1. The test is actually one of independence between ε and X, not of the parametric form of the conditional mean. Theorems 1–3 condition on an independence assumption that is not part of the stated null H0; making the assumption explicit does not resolve the mismatch because the advertised hypothesis concerns only the conditional mean.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new specification test for parametric regression models when the number of predictors p diverges with the sample size n. The test is based on a weighted residual process U_n(t) = n^{-1/2} Σ (g(X_i)-gbar)(cos(t e_i)+sin(t e_i)) for a real-valued weight function g, integrated over t to form WICM_n. The authors claim a nondegenerate Gaussian limit under the null (Theorem 1), consistency against fixed alternatives and nontrivial power at the n^{-1/2} local rate (Theorem 2), and validity of a smooth residual bootstrap (Theorem 3 and Corollary 1), under the rate condition p^3 log n / n → 0. They also propose data-driven choices of g for directional and nonparametric alternatives (Section 4.3). Simulations and a real-data application are reported.","tokens_in":18822,"tokens_out":5937,"duration_ms":51786,"significance":"If valid, this would be a useful addition to the diverging-p model-checking literature, avoiding the degeneracy of classical ICM statistics and the failure of the wild bootstrap. The asymptotic rate condition p^3 log n / n → 0 is weaker than earlier conditions in Tan and Zhu (2019) and Tan et al. (2025), and the bootstrap justification is an important contribution. However, the paper's central claim—that the test checks the conditional mean (1.2)—is compromised by the reliance on full independence of errors and covariates, and the theoretical results do not cover the estimated weight functions actually used in implementation. These are load-bearing issues that undermine the validity of the test as advertised.","major_comments":[{"comment":"The test is constructed on the identity E{g0(X) exp(ite)} = E{g0(X)}E{exp(ite)} = 0 under H0, which requires ε to be independent of X. The null (1.2) only specifies the conditional mean; the paper even states that the error distribution 'remains unrestricted.' If ε is heteroskedastic but E(ε|X)=0, then for small t, E{g0(X)(cos(tε)+sin(tε))} ≈ -(t²/2)E{g0(X)ε²} ≠ 0, so WICM_n diverges and the test rejects a correctly specified mean with probability tending to 1. Theorems 1–3 condition on independence, which is not part of the stated H0. The test is actually one of independence between ε and X, not of the parametric form of the conditional mean.","section":"Section 1 (after Eq. (1.1)) and Section 3 (Eq. (3.1))"},{"comment":"All theoretical results treat g as a fixed, nonrandom weight function (Assumption 1). In practice, g is estimated from the same data using (4.4) or (4.5). The theorems do not cover the estimated g. Consequently, the size and power simulations and the real-data analysis rely on a procedure without asymptotic justification. This is particularly problematic because the local-power argument in Section 4.3 depends on g being aligned with S; with an estimated g, the expansion (4.1) and the power claims may fail. The Fourier truncation level l and the structural dimension s are additional user-chosen quantities not covered by the theory.","section":"Section 4.3, Eqs. (4.4)–(4.5)"},{"comment":"The numerical studies are conducted in regimes where the rate condition p^3 log n / n → 0 is far from satisfied. For example, p=10, n=100 gives p^3 log n / n ≈ 46; the real-data example with p=68, n=1059 gives ≈ 2078. The paper does not discuss this discrepancy. It is unclear whether the reported control of size and power reflects the asymptotic theory or some other mechanism. The real-data application is thus not supported by the paper's own theoretical conditions.","section":"Section 5, Tables 1–4"}],"minor_comments":[{"comment":"There is a notation inconsistency: the introduction uses d and p for predictor and parameter dimensions, but Section 2 uses p for the covariate dimension and later p becomes the parameter dimension. Please standardize.","section":"Throughout"},{"comment":"The assumption states ∫ t^4 φ(t)dt < ∞, but φ was not introduced in Assumption 8; it should be l(t). Also, 'phi' in the bootstrap step should be consistent with φ(t).","section":"Section 4, Assumption 8"},{"comment":"The sentence 'This choice, however, is not practically useful...' correctly notes the degeneracy of the optimal weight g ∝ m(X,β0)−m(X). But the replacement using the projection onto the score space is ad hoc and no property is proved for it. Please clarify or cite a justification.","section":"Section 4.3, p. 22"},{"comment":"The reported p-value 'approximately equal to 0' is vague; provide a numerical upper bound. Also, the claim that the scatter plots 'suggest' nonlinearity is informal; the test result itself is enough.","section":"Section 5.2"}],"recommendation":"reject","confidential_remarks":"The paper addresses an important problem, but the independence assumption and the unestimated-g gap are not merely presentation issues. The test does not check the conditional mean as stated, and the implemented procedure lacks theoretical support. If the authors were to reframe the null as a joint hypothesis of conditional mean and error–covariate independence, and provide a full treatment of estimated g, a substantial revision might be viable. However, as it stands, the central claims are not supported, and the simulations run outside the theory's regime. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does real work on a real problem. The classical ICM test degenerates when p diverges, the wild bootstrap stops working, and this paper constructs a weighted residual process that avoids the curse of dimensionality and gives a nondegenerate null limit under p^3 log(n)/n -> 0. That rate is weaker than previous conditions in Tan et al. (2025), and the smooth residual bootstrap is a sensible way to get critical values. The simulations are consistent with the theory, and the real-data example is a nice sanity check. If the paper is read as a test of independence between the error and the covariates, its internal logic is coherent.\n\nThe soft spots are not small. The key identity in Section 3 is E{g0(X) exp(ite)} = E{g0(X)}E{exp(itε)} = 0, which requires e = ε to be independent of X. The paper assumes this right after (1.1), but its abstract and discussion claim the hypothesis concerns only the conditional mean and that the error distribution is unrestricted. That is a false advertising problem. Under H0 with E(ε|X)=0 but Var(ε|X) depending on X, the statistic has nonzero mean and the test rejects a correctly specified mean with probability tending to one. This is not a technical quibble; the test is not testing the stated hypothesis. The simulations only use independent normal errors and X, so they never expose this.\n\nThere are two more issues in proportion. The proofs are all deferred to a supplementary file that is not part of the submission, so the main theorems cannot be verified from the manuscript. And the data-driven weight functions in (4.4) and (4.5) are outside the asymptotic theory: Theorems 1–3 treat g as fixed, while Section 4.3 selects g to maximize E{g0 S} against a guessed alternative. Section 4.3 even admits the oracle choice degenerates under the null, then substitutes a projection heuristic without asymptotics for the estimated g. For a methodology paper this is a real gap, though not fatal if the fixed-g case is the core claim.\n\nWho is this for? People working on high-dimensional specification tests will want to read it; it is a useful extension of Tan et al. (2025) and a good example of how residual process tests can be reshaped for diverging dimensions. A serious referee could help the authors fix the hypothesis mismatch, either by repositioning the paper as a test of independence or by redesigning the statistic so that only mean independence is needed. I would not desk-reject it; the construction is thoughtful and the problem is important. But as it stands, the conditional-mean claim is wrong, and that has to be confronted before publication.","headline":"Worth a referee, but the test tests error–covariate independence, not the conditional-mean null it advertises.","tokens_in":19223,"tokens_out":3986,"would_cite":false,"duration_ms":34277,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G10","62G08","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A weighted residual process test checks the parametric regression mean even when the number of predictors grows with the sample size, avoiding the collapse of classical integrated conditional moment (ICM) tests.","keywords":["model specification test","diverging number of predictors","weighted residual process","integrated conditional moment test","smooth residual bootstrap","high-dimensional regression","local alternatives"],"falsifier":"Simulate a correctly specified linear mean with heteroskedastic noise, for example Y = Xβ + (1 + |X_1|)ε with ε standard normal independent of X, at n = 400 and p = 10, and run the proposed test at the 5% level. If the rejection rate substantially exceeds 5%, the independence-of-error assumption is doing load-bearing work.","tokens_in":18362,"feed_emoji":"📊","tokens_out":5378,"duration_ms":48110,"temperature":0.7,"pith_summary":"This paper proposes a goodness-of-fit test for regression mean functions that keeps working when the number of predictors p grows with the sample size. Classical integrated conditional moment (ICM) statistics degenerate to fixed constants in this regime, and the wild bootstrap fails to control size. The paper replaces the high-dimensional covariate weight exp(it^T X) with a scalar weight function g(X) applied to residual-based trigonometric processes. The resulting statistic has a nondegenerate Gaussian limit under the null whenever p^3 log n / n → 0, diverges under fixed alternatives, and detects local alternatives at the parametric 1/√n rate. A smooth residual bootstrap recovers the null distribution, and simulations show nominal level and higher power than existing methods.","feed_headline":"Catches wrong regression shapes as predictors grow","feed_subtitle":"New weighted residual statistic keeps nominal size, finds misspecification at the 1/√n rate, and works with a residual bootstrap.","key_machinery":"The central object is the weighted residual empirical process Û_n(t) = n^{-1/2} Σ_{i=1}^n (g(X_i) - ḡ){cos(tê_i) + sin(tê_i)}. The test statistic is WICM_n = ∫|Û_n(t)|²φ(t)dt with an even integrable weight φ; with standard normal φ it collapses to a pairwise sum of covariate-centered weights times exp(-(ê_i-ê_j)²/2). This one-dimensional trigonometric construction separates the covariates from the residual differences, avoiding the interpoint distance concentration that makes ICM statistics degenerate in high dimensions. The smooth residual bootstrap adds smoothed noise v_n z_{i,j} to centered resampled residuals, refits the model, and recomputes WICM; Theorem 3 shows the bootstrap and the","core_discovery":"The paper's central claim is that parametric regression misspecification can be tested in diverging-p settings by studying a one-dimensional weighted residual process. Under the null, the residual equals the error ε, which is assumed independent of X, so the centered process has zero mean; under alternatives, the mean shift m(X) - m(X, β~0) makes the expectation nonzero. The main theorem shows that, under regularity conditions and p^3 log n / n → 0, the statistic WICM_n converges weakly to the nondegenerate integral ∫|U∞(t)|²φ(t)dt of a Gaussian process, and that the same limit is reproduced by a smooth residual bootstrap. Against fixed alternatives the statistic grows linearly in n; against","pith_inferences":["The null theory needs ε independent of X, not just E(ε|X)=0; under heteroskedastic errors with a correctly specified mean, E{g0(X) exp(itε)} need not vanish, so the test may over-reject. The paper states this independence assumption explicitly, but practitioners should verify it.","The asymptotics fix g while the implementation estimates g (directional via a working alternative family, or nonparametric via a Fourier/dimension-reduction basis); a formal proof that estimated g preserves the null limit would close a gap between theory and code.","Because the test targets the conditional mean only, it will not detect variance misspecification or other distributional departures; a different residual transformation would be needed for those.","The weight-choice analysis implies a practical extension: if g can be chosen close to the projected departure m(X, β~0) - m(X), the test becomes nearly adaptive, but guarding against its null degeneracy is exactly the paper's concern."],"forward_implications":["Under the null, WICM_n has a nondegenerate Gaussian limit instead of collapsing to a constant, so the test can hold nominal size as the predictor dimension grows.","Under any fixed misspecification, the test rejects with probability tending to 1, and under local alternatives within 1/√n of the null it retains nontrivial power.","The smooth residual bootstrap is asymptotically valid for size and power, replacing the wild bootstrap that fails in the diverging-p regime.","The statistic is computationally light: a closed-form pairwise kernel on residuals, with no d-dimensional numerical integration.","The dimension condition p^3 log n / n → 0 is weaker than previously available conditions for such empirical-process tests."],"fun_headline_variants":["Regression shape test that survives many predictors","Weighted residual process rescues ICM test in high dimensions","Smooth residual bootstrap works for diverging predictor count","Detect wrong regression form at parametric rate in HD","New model check for regressions with many predictors"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The construction assumes the error ε is independent of X under the null; if only E(ε|X) = 0 holds, the weighted characteristic moment E{g0(X) exp(itε)} need not vanish, and a correctly specified mean could be rejected.","fun_headline_variants_meta":{"raw":{"variants":["Regression shape test that survives many predictors","Weighted residual process rescues ICM test in high dimensions","Smooth residual bootstrap works for diverging predictor count","Detect wrong regression form at parametric rate in HD","New model check for regressions with many predictors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000369,"raw_usage":{"total_tokens":1821,"prompt_tokens":754,"completion_tokens":1067,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":994}},"tokens_in":498,"tokens_out":1067,"duration_ms":10252,"temperature":1.0,"reasoning_tokens":994,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T16:11:10.696330+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a correctly specified linear mean with heteroskedastic noise, for example Y = Xβ + (1 + |X_1|)ε with ε standard normal independent of X, at n = 400 and p = 10, and run the proposed test at the 5% level. If the rejection rate substantially exceeds 5%, the independence-of-error assumption is doing load-bearing work.","supporting_citations":[],"review_version":2}