Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

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.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A weighted-residual-process test with smoothed bootstrap avoids the degeneracy of classical ICM tests in regression models with a diverging number of predictors.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Worth a referee, but the test tests error–covariate independence, not the conditional-mean null it advertises. the 3 major comments →

arxiv 2604.14649 v2 pith:PSEJT6MQ submitted 2026-04-16 stat.ME math.STstat.TH

Model Checking for Regressions Based on Weighted Residual Processes with Diverging Number of Predictors

Yue Hu , Haiqi Li , Xintao Xia This is my paper
classification stat.ME math.STstat.TH MSC 62G1062G0862G20
keywords model specification testdiverging number of predictorsweighted residual processintegrated conditional moment testsmooth residual bootstraphigh-dimensional regressionlocal alternatives
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

Core claim

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

What carries the argument

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

Load-bearing premise

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.

What would settle it

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.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • 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.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

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.

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 (3)
  1. [Section 1 (after Eq. (1.1)) and Section 3 (Eq. (3.1))] 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.
  2. [Section 4.3, Eqs. (4.4)–(4.5)] 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.
  3. [Section 5, Tables 1–4] 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.
minor comments (4)
  1. [Throughout] 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.
  2. [Section 4, Assumption 8] 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).
  3. [Section 4.3, p. 22] 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.
  4. [Section 5.2] 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.

Circularity Check

0 steps flagged

No significant circularity: the derivation is self-contained given its modeling assumptions; the main concerns are a null-hypothesis mismatch and unsupported estimated-g asymptotics, not circular reductions.

full rationale

No circular step is exhibited. The test statistic WICM_n in (3.2) is an integral of the squared weighted residual process (3.1), and it is not defined as the fitted value of the quantity it claims to predict; no fitted parameter is renamed as a prediction and no equation reduces to its input by construction. The null moment identity E{g0(X) exp(ite)} = E{g0(X)}E{exp(itε)} = 0 used in Section 3 relies on the independence assumption stated after Eq. (1.1). This is a real correctness/validity concern: the paper's advertised null in (1.2) concerns only the conditional mean, while the construction effectively tests independence between the error and X, so under heteroskedastic errors with E(ε|X)=0 the test can reject a correctly specified mean. That is a hypothesis mismatch rather than circularity, because the independence assumption is an input to the derivation, not the conclusion being derived. Likewise, the estimation of the weight function g in Section 4.3, Eqs. (4.4)-(4.5), is not covered by Theorems 1-3, which treat g as fixed under Assumption 1; this is an omitted proof or unsupported extension, but it does not make the local-power claim equivalent to the fitted input. The citations to Tan et al. (2025), Dette et al. (2007), Neumeyer (2009), and others are external rather than self-citations, and the bootstrap approximation is a standard residual bootstrap. Accordingly, the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central results rest on standard high-dimensional M-estimation regularity, a strong independence assumption between errors and predictors, a fixed weight function whose data-driven plug-in is not covered by the theorems, an orthogonality condition on the local alternative direction, and smooth bootstrap conditions. No new physical or abstract entities are introduced.

free parameters (4)
  • bootstrap smoothing parameter v_n = 0.2 in simulations
    Set by hand to 0.2, following Dette et al. (2007); Assumption 8 only requires log n = o(n v_n^4). Not derived from data.
  • weight function phi(t) = standard normal density
    Chosen for closed-form evaluation of WICM_n; any even phi with bounded fourth moment is allowed, so this is a conventional tuning choice, not a fitted parameter.
  • Fourier truncation level l for nonparametric weight = unspecified
    In Eq. (4.5) the truncation level l is 'left to the researcher'; it controls the complexity of estimated g and is not covered by any theorem.
  • structural dimension s / central subspace B for WICM^(2) = estimated via CSE and MERE
    The orthonormal basis used to build g is estimated from the same data through sufficient dimension reduction; this estimation step is outside the stated asymptotics.
axioms (6)
  • domain assumption epsilon is independent of X
    Stated after Eq. (1.1). Needed for E{g0(X) exp(it e)} = E{g0(X)} E{exp(it epsilon)} under H0; if only E(epsilon|X)=0, the weighted characteristic moment need not vanish.
  • domain assumption Under H0, Y - m(X, beta_tilde_0) = epsilon and the model contains the true regression function
    Defines the null hypothesis in (1.2) and the population least-squares target in (2.1); all limiting arguments use this specification.
  • ad hoc to paper The weight function g(X) is fixed and satisfies an envelope condition; estimated g is ignored in the proofs
    Assumption 1 bounds g0(X), and Theorems 1-3 are stated for fixed g. Section 4.3 replaces g by data-dependent estimators (4.4)/(4.5), but no theorem covers this plug-in effect.
  • standard math Standard high-dimensional M-estimation regularity: Assumptions 1-7, p^3 log n / n -> 0, Gaussian-process tightness
    These are the standard conditions for the least-squares expansion and weak convergence of the residual process; the p^3 log n / n condition is quoted in Theorem 1.
  • domain assumption Local alternative perturbation S satisfies E{S(X) ḏm(X, beta_tilde_0)} = 0 and g is aligned with S
    Needed for the nonzero shift K^(2)(t) in Theorem 2. If the estimated g is orthogonal to the actual misspecification direction, the local power vanishes.
  • domain assumption Smooth residual bootstrap kernel l and smoothing parameter v_n allow uniform density estimation
    Assumption 8 imposes a positive, symmetric, twice-differentiable kernel and log n = o(n v_n^4); standard for smooth residual bootstrap validity.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Model Checking for Regressions Based on Weighted Residual Processes with Diverging Number of Predictors." pith.science (2026). https://pith.science/paper/PSEJT6MQ

@misc{pith2026260414649,
  author       = {Pith},
  title        = {Pith review of: Model Checking for Regressions Based on Weighted Residual Processes with Diverging Number of Predictors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PSEJT6MQ}},
  note         = {Machine review of arXiv:2604.14649}
}
Share X Bluesky LinkedIn Reddit HN
abstract

The integrated conditional moment (ICM) test is a classical and widely used method for assessing the adequacy of regression models. Although it performs well in fixed-dimension settings, its behavior changes dramatically when the predictor dimension diverges: in such regimes, the limiting null and alternative distributions of the ICM statistic degenerate to fixed constants. Moreover, when the number of predictors diverges, the commonly used wild bootstrap no longer approximates the null distribution of the ICM statistic well, leading to size distortion and substantial power loss. To address these challenges, we propose a new specification test based on weighted residual processes for evaluating the parametric form of the regression mean function in high-dimensional settings where the number of predictors increases with the sample size. We establish the asymptotic properties of the test statistic under the null hypothesis and under global and local alternatives. The proposed test maintains the nominal significance level and can detect local alternatives that deviate from the null hypothesis at the parametric rate $1/\sqrt{n}$. Furthermore, we propose a smooth residual bootstrap to approximate the limiting null distribution and establish its validity in high-dimensional settings. Two simulation studies and a real-data example are conducted to evaluate the finite-sample performance of the proposed test.

Figures

Figures reproduced from arXiv: 2604.14649 by Haiqi Li, Xintao Xia, Yue Hu.

Figure 1
Figure 1. Figure 1: (a) The scatter plot of Y against the fitted values βb⊤X, and (b) the scatter plot of the residuals against the fitted values βb⊤X for the Geographical Origin of Music data set. 6 Discussion Although widely used, the ICM test exhibits fundamentally different asymptotic behavior in high-dimensional settings, and the associated wild bootstrap is no longer valid. To address this issue, we propose a test based… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Testing for correct model specification in copula regression models

    math.ST 2026-07 conditional novelty 6.0

    A kernel-based test of the weighted L2 distance between the true regression function and the copula-regression approximation is consistent and asymptotically normal, with pivotal self-normalized confidence intervals f...

Reference graph

Works this paper leans on

4 extracted references · cited by 1 Pith paper

  1. [1]

    Bierens, H. J. (1982). Consistent model specification tests.Journal of Econometrics, 20(1):105–134. Bierens, H. J. (1990). A consistent conditional moment test of functional form.Economet- rica, 58(6):1443–1458. Bierens, H. J. and Ploberger, W. (1997). Asymptotic theory of integrated conditional moment tests.Econometrica, 65(5):1129–1151. Cook, R. D. (200...

    1982

  2. [3]

    Van Keilegom, I., Gonz´ alez Manteiga, W., and S´ anchez Sellero, C

    Cambridge university press. Van Keilegom, I., Gonz´ alez Manteiga, W., and S´ anchez Sellero, C. (2008). Goodness-of-fit tests in parametric regression based on the estimation of the error distribution.Test, 17:401–415. Zheng, J. X. (1996). A consistent test of functional form via nonparametric estimation techniques.Journal of Econometrics, 75(2):263–289....

    2008

  3. [961]

    and Lavergne, P

    Guerre, E. and Lavergne, P. (2005). Data-driven rate-optimal specification testing in re- gression models.The Annals of Statistics, 33(2):840–870. Guo, X., Wang, T., and Zhu, L. (2016). Model checking for parametric single-index models: a dimension reduction model-adaptive approach.Journal of the Royal Statistical Society Series B: Statistical Methodology...

    2005

  4. [1947]

    Hastie, T., Tibshirani, R., and Wainwright, M. (2015). Statistical learning with sparsity. Monographs on statistics and applied probability, 143(143):8. 33 Horowitz, J. L. and H¨ ardle, W. (1994). Testing a parametric model against a semipara- metric alternative.Econometric theory, 10(5):821–848. Khmaladze, E. V. and Koul, H. L. (2004). Martingale transfo...

    2015

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.