REVIEW 4 major objections 5 minor 14 references
A Nonparametric Test for Cross-Unit Spillovers
T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash
Pith's one-line read A simple LM-style statistic, computed under the null of no spillovers, detects nonlinear cross-unit spillovers and is asymptotically standard normal.
desk verdict A useful-looking c-channel spillover test with an unverified outcome-channel instrument story and a p-choice that contradicts its own theorem. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the series-expanded gradient vector d̂ = −(2/n) U' P_Z (y − X β̂), where U stacks basis functions ψ_i(w'y) and ψ_i(w'c_j) (Hermite polynomials in the applications) and Z is an instrument matrix. The statistic S is a cluster-robust quadratic form in this gradient, centered and scaled by q = p(l+1) restrictions. Because the gradient is evaluated under the null, no nonparametric estimation of the spillover functions f or g_j is needed; the basis functions act as a growing-dimensional sieve that makes any departure from the null detectable in the projected moment conditions.
What would settle it
Run two simulations: (1) under the null of no spillovers but with an instrument that is correlated with a cluster-level shock, check whether the rejection rate of S stays near 5%; (2) under a known nonlinear spillover f(s)=s² with valid instruments, check whether power rises toward 1 with n. The paper reports only a size simulation with valid instruments, so neither of these checks is documented.
Extended reading notes
Core claim
Under the approximate null H0A of no spillovers (all series coefficients µ equal zero), the statistic S = (n d̂' Ĥ^{-1} d̂ − q)/√(2q), built from the gradient of a series-approximated IV objective evaluated at the restricted linear estimate, converges in distribution to N(0,1) as n and p grow with p^3/n → 0, and S is consistent against alternatives where any series coefficient is nonzero. The test simultaneously captures attribute-channel (c), outcome-channel (y), and joint (cy) spillovers, and requires only estimation of the null model y_i = x_i'β + ε_i.
Load-bearing premise
For the outcome-channel tests, the null distribution of S relies on valid instruments for the endogenous peer-outcome term w_i'y: the orthogonality condition E(ε_i z_j)=0 must hold for every instrument z_j, and the four empirical applications never state which instruments are used for the y and cy tests, so those rejections are credible only if such instruments exist.
Editorial extensions
If this is right
- Applied researchers can test for nonlinear cross-unit spillovers using only the residuals from a linear IV/OLS regression under the null; no nonparametric estimation is required.
- The test covers block-based and link-based interaction structures, extends to multiple weight matrices, and, under high-level conditions, tolerates noisy or incomplete network data.
- The composite rule-of-thumb 's test' rejects if any of the c, y, or cy component tests rejects, which guards against masking when one channel is strong, the other weak, or when the series terms are highly collinear.
- In four empirical settings (golf tournaments, university roommates, school deskmates, academic researchers), the test rejects the null of no spillovers in many specifications where linear peer-effect regressions do not reject.
- For small samples, using χ²_q critical values rather than the limiting normal can improve size control, as shown in the paper's simulations.
Reading between the lines
- Not stated in the paper: a rejection by S only indicates that some nonlinear dependence exists; it does not identify which function or mechanism generates it, so applied work should pair the test with an estimator of the spillover function.
- Because the composite s-test unions three hypotheses, its overall size may be larger than nominal; a researcher could apply a multiple-testing correction when using the rule of thumb.
- The paper's simulations cover only size, not power; a natural next step is a power experiment with a known nonlinear spillover function (e.g., a quadratic) to verify the consistency claim in finite samples.
- In the empirical y and cy tests, the instrument set is not reported; a careful replication should state the instruments used for the endogenous peer-outcome term, because the null distribution relies on their validity.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a nonparametric LM-type specification test for cross-unit spillovers. Spillovers may enter through peers' attributes (c), peers' outcomes (y), or both (cy), represented by unknown functions approximated by a divergent series expansion. Under the approximate null of no spillovers, the test statistic S in (2.11) is claimed to be asymptotically standard normal, and the test is claimed to be consistent. Estimation is required only under the null, which is an attractive feature. The paper also defines a composite 's-test' rejection rule based on the c, y, and cy tests, gives practical guidance on choosing the number of basis functions, and illustrates the method on four empirical applications.
Significance. If the central claims hold, the paper provides a practically valuable diagnostic: a researcher can test for potentially nonlinear spillovers while estimating only a linear regression under the null, with cluster-robust standard errors. The series/LM construction is standard, and the paper is careful to provide cluster-robust variance formulas, Monte Carlo evidence, and extensions to heterogeneous and noisy interaction structures. The empirical applications are relevant and show that the test can sometimes detect dependence missed by linear specifications. These strengths are undercut by several load-bearing issues: the recommended tuning parameter violates the rate condition required by the theorems; the composite s-test has no stated or simulated overall size; the outcome-channel tests require an instrument set that is never described in the empirical work; and the core null CLT is imported from self-cited work without statement. These issues are fixable within the scope of the manuscript, so I regard the contribution as promising but not yet fully supported.
major comments (4)
- [Section 4, Eqs. (4.1)-(4.3); Theorem 1] The recommended tuning choices are p_c = [n^{1/3}/l], p_y = [n^{1/3}], and p_cy = [n^{1/3}/(l+1)]. For fixed l, each is Theta(n^{1/3}), so p^3/n = Theta(1), not o(1). Theorems 1 and 2 and the proof of Theorem C2 explicitly require p^3/n = o(1). Thus the implementation recommended in Section 4 is outside the range covered by the asymptotic theory. The statement that p^3(l+1)/n -> 0 is 'equivalent' to p^3/n -> 0 because l is fixed is incorrect at the recommended rate.
- [Definition 1 and Section 4] The s-test rejection rule rejects when at least one of the c, y, or cy tests rejects. No size analysis is supplied for this union rule. Even if each component test has asymptotic size alpha, the overall null rejection probability is not alpha and may be substantially larger, especially since the three tests are correlated. The Monte Carlo study in Appendix A reports size only for the cytest; there is no simulation of the s-test's overall size. Because all four empirical applications report 'stest result', the reported rejection frequencies have no stated null calibration.
- [Section 2, after Eq. (2.7); Tables 1-4] For the y and cy tests, valid instruments Z are required so that E(Z'epsilon)=0. The paper says OLS is a special case with Z=U, but when U contains psi_i(w_i'y), this fails under the null: y_i = x_i'beta + epsilon_i implies w_i'y depends on w_i'epsilon, so Z'epsilon is not mean-zero and Assumption 3's uncorrelatedness condition does not hold. The four applications report many rejections through the y and cy channels but never state what instruments Z are used for those tests. Without that information, the outcome-channel rejections in Tables 1-4 cannot be verified and could be driven by endogeneity rather than spillovers.
- [Proof of Theorem 1, Section C.2] The final step of the proof of Theorem 1 invokes Theorem A.1 of Gupta, Qu, Srisuma, and Zhang (2025) without stating it, and Lemma D1 is imported from Gupta et al. (2025) without proof. That CLT for growing-dimension quadratic forms is the main engine of the standard-normal result. Because the cited theorems are from the authors' own prior work and are not reproduced, the derivation is not self-contained and the conditions are not transparent to the reader. State the required result or provide a proof.
minor comments (5)
- [Section 4, Eq. (4.1)] The notation 'p = max{2, [n^{1/3}]}' is slightly confusing because the lower bound already ensures p>1, and the sentence 'always want p>1' is redundant with the max.
- [Appendix A, footnote 11] There is a duplicated sentence fragment: 'components in the same cluster. components in the same cluster.'
- [Section 5.4, Table 4 note] The note says standard errors are clustered at the department level, while the reproduced original specification in Appendix E.4 says clustering is at the university-year level. Clarify which clustering is used in the stest.
- [Appendix E.2, footnote 15] The explanation of how w_i^{race} is obtained from W^{race} is terse and would benefit from a more explicit description, especially since the full-sample specification is not an affine function of the peer-race variable used elsewhere.
- [General] The paper would be easier to follow if the instrument matrix Z were explicitly defined in one place for the three tests, including the exact number of instruments m and how the instruments differ between the c test and the y/cy tests.
Circularity Check
Central null distribution is imported from same-author prior theorems (Theorem A.1 and Lemma D1), so the derivation chain is not self-contained.
-
self citation load bearing
[Appendix C.2, Proof of Theorem 1, final step after (C.25)]
"Then (C.25) follows by Theorem A.1 of Gupta et al. (2025)."
Theorem 1's conclusion S→N(0,1) is the paper's central result. The proof reduces the final quadratic-form CLT to Theorem A.1 of a same-author arXiv paper (Gupta, Qu, Srisuma, and Zhang 2025), which is neither stated nor proved in this manuscript and is not machine-checked here. The asymptotic normality of the test therefore rests on a self-citation chain rather than on a self-contained derivation; if that external theorem were unavailable or false, Theorem 1 would not be established.
-
self citation load bearing
[Appendix D, Lemma D1]
"Lemma D1. Let p^2/n→0 as n→∞ and suppose that Assumptions 2-4 hold with ν>3/2. Then, as n→∞, || Phi|| = O_p(p/√n) ... Proof of Lemma D1: This is Lemma 1 in Gupta et al. (2025)."
This lemma supplies the convergence rates for Phi and J used in Theorem C1's bound on Phi, which is a premise of Theorem C2 and ultimately Theorem 1. The proof is literally a pointer to the authors' own previous arXiv paper rather than a derivation, and no independent verification is provided. Thus a load-bearing step in the proof of the test's null distribution is outsourced to a same-author result not established in the current manuscript.
full rationale
The test is not circular in the fitting sense: no spillover coefficients are estimated, no parameter is fitted to force rejections, and the empirical illustrations compare against independent original studies. The simulation study is a genuine size study under the null. However, the derivation chain for the chief asymptotic claim is not self-contained. The proof of Theorem 1 explicitly relies on Theorem A.1 of Gupta, Qu, Srisuma, and Zhang (2025), and Lemma D1 is asserted to be Lemma 1 of Gupta et al. (2025); both are same-author arXiv papers and are neither proved nor independently verified in this manuscript. These are load-bearing self-citations for the asymptotic standard normality result, not merely incidental references. The unstated instrument sets in the empirical y/cy applications are a validity and endogeneity concern, and the s-test's union rejection rule has no calibrated size, but these are not circularity steps under the definitions used here. I therefore score the circularity at 6: the central null distribution is partly forced by a same-author citation chain, even though much of the test construction and intermediate asymptotic approximation is original and independent.
Assumptions & free parameters
free parameters (2)
- p (number of basis functions) =
Varies by application: p_c=5,7,9,...; p_y=5,7,26,...; p_cy=2,3,4,...
- Basis family (Hermite polynomials)
assumptions (6)
- domain assumption Assumption 1: errors have zero mean, bounded 8+tau moments, and clusters are independent.
- domain assumption Assumption 3: spectral bounds on Sigma, M, L; approximation error decays as p^{-nu} with nu > 5/2; instruments are valid (E(epsilon_i z_j)=0).
- domain assumption Assumption 4: xi + kappa = O(n), bounding cross-sectional dependence of products of instruments and regressors.
- standard math Theorem A.1 of Gupta et al. (2025): quadratic-form CLT for independent cluster increments.
- standard math Lemma D1 = Lemma 1 of Gupta et al. (2025) for convergence of M-hat and J-hat.
- domain assumption Linear process representation epsilon = B eta in Theorem 1.
invented entities (1)
-
Latent social-space variables z_l (Appendix B.2)
Cite this review
Pith. "Pith review of A Nonparametric Test for Cross-Unit Spillovers." pith.science (2026). https://pith.science/paper/OFAK6SME
@misc{pith2026260800136,
author = {Pith},
title = {Pith review of: A Nonparametric Test for Cross-Unit Spillovers},
year = {2026},
howpublished = {\url{https://pith.science/paper/OFAK6SME}},
note = {Machine review of arXiv:2608.00136}
}
read the original abstract
Cross-unit dependence is pervasive in empirical applications and complicates econometric inference, especially when spillovers operate in nonlinear ways. We propose a novel nonparametric test for cross-unit spillovers that may operate through peers' attributes, peers' outcomes, or both. The test is straightforward to implement, as it requires only estimation under the null hypothesis of no cross-unit spillovers, and is shown to have a convenient asymptotic standard normal distribution. It is also versatile, accommodating data generated by a wide range of interaction structures. We present four empirical illustrations showing that the proposed test can yield substantively different conclusions about the presence of cross-unit spillovers than existing approaches.
Reference graph
Works this paper leans on
-
[1]
Acemoglu, D., A. E. ¨Ozdaglar, and A. Tahbaz-Salehi (2016). Networks, shocks, and systemic risk. In Y. Bramoull´ e, A. Galeotti, and B. Rogers (Eds.),The Oxford Handbook of the Economics of Networks, pp. 569–610. Oxford University Press. Arduini, T., E. Patacchini, and E. Rainone (2020). Treatment effects with heterogeneous externalities.Journal of Busine...
arXiv 2016
-
[2]
, ϵ′ (G))′, whereϵ (g) ∈R ng
For the cluster-correlation specification paired with cluster-robust standard errors, stack ϵ= (ϵ ′ (1), . . . , ϵ′ (G))′, whereϵ (g) ∈R ng . Then var(ϵ) = Σ =diag[Σ 1, . . . ,ΣG],Σ g = 1 2 Ing + 1 2 1ng 1′ ng ,(A.1) where1 ng is then g ×1 vector of ones. Hencevar(ϵ i) = 1,cov(ϵ i, ϵj) = 1/2 for distinct units in the same cluster, andcov(ϵ i, ϵj) = 0 acro...
2025
-
[3]
Students completed baseline (pre-intervention) and endline (post- intervention) surveys of academic performance and personality traits
In the mixed-seating (MS) and mixed-seating-with-reward (MSR) classes, students are randomly paired as deskmates within height groups. Students completed baseline (pre-intervention) and endline (post- intervention) surveys of academic performance and personality traits. For each endline outcome, the authors estimate yend i =α+βy base i +γ w′ iybase +x ′ i...
2023
-
[4]
Hence, the second term in (C.15) iso p(√p/n) as long asp 3/n=o(1), concluding the proof
We conclude that ∆ ˆH H =O p p/√n .(C.23) By Assumption 4, the last term in (C.15) is thusO p(p2/n3/2), given∥d∥=O p( p p/n) and 33 ˆd =O p( p p/n). Hence, the second term in (C.15) iso p(√p/n) as long asp 3/n=o(1), concluding the proof. C.2 Proofs of main theorems Proof of Theorem 1:We haveEϵ igϵjg ′ = 0 forg̸=g ′ andEϵ igϵjg = Png r=1 birbjr . Thusσ ijg...
2025
-
[5]
By embedding the observed graph into a latent social space, we interpret it as a noisy measurement of an underlying latent structure
#) ,cytest.(B.5) B.2 Embedded graphs We now consider a setting in which the observed social matrix is generated by an embedded graph model rather than treated as exogenous. By embedding the observed graph into a latent social space, we interpret it as a noisy measurement of an underlying latent structure. This reframing is particularly valuable in applied...
2020
-
[7]
The last term dominates the former five ones and thus, underH 1A, for allη >0,P |S| −1 ≤η/n √p →1 asn→ ∞and hence consistency ofSfollows
By standard norm inequalities, the second term in (C.45) isO p(√n), the third isO p(n/√p), the fourth isO p(p3/2), the fifth isO p(p√n) and the sixth isO p(n√p). The last term dominates the former five ones and thus, underH 1A, for allη >0,P |S| −1 ≤η/n √p →1 asn→ ∞and hence consistency ofSfollows. D Auxiliary lemmas Lemma D1.Letp 2/n→0asn→ ∞and suppose t...
2025
-
[10]
For each academic outcome, the authors estimate yik =α+λw ′ iRace+x ′ iβ+µw ′ ix+δ k +ε ik,(E.2) wherey ik is the outcome of studentiin residencek, measured at the end of the first academic year;Raceis a vector of race dummies of all students in the sample;x i is a set of individual baseline controls (gender, UCT admission score, household wealth, monthly...
2009
-
[13]
(2022), which serve as the basis for the tests in Table
Table E4 reproduces the OLS peer-effect estimates of Table 3 in Bosquet et al. (2022), which serve as the basis for the tests in Table
2022
Show all 14 references
-
[14]
(2022) (Table E4)
contains 42,861 author–year observations, compared with the 42,521 reported by Bosquet et al. (2022) (Table E4). The difference arises because Bosquet et al. (2022) compute peer averages using the full university–year roster before dropping observations with missing age. To sa...
2022
-
[190]
Jochmans, K. (2023). Peer effects and endogenous social interactions.Journal of Econo- metrics 235, 1203–1214. Kelejian, H. H. and I. R. Prucha (2007). HAC estimation in a spatial framework.Journal of Econometrics 140, 131–154. Lalive, R. and M. A. Cattaneo (2009). Social inte...
2023
-
[992]
For disturbances generated viaN(0,1) shocks, there appears to be some benefit from using theχ 2 q critical values even at (n, p) = (1000,5) for the circulant and lattice cases
The results are reported in Table A1. For disturbances generated viaN(0,1) shocks, there appears to be some benefit from using theχ 2 q critical values even at (n, p) = (1000,5) for the circulant and lattice cases. For thet 10 case, this remains true. In the remaining four des...
2017
-
[2009]
(2009), who study peer effects in pro- fessional golf tournaments
Table E1 reproduces the estimates in Guryan et al. (2009), who study peer effects in pro- fessional golf tournaments. Column (i) reports the authors’ baseline regression from yi,tr =α+βAbility i +γw ′ i,trAbility+δ tc +ε i,tr,(E.1) wherey i,tr is playeri’s score in roundrof to...
2009
-
[2022]
(2022), which serve as the basis for the tests in Table
Table E2 reproduces the academic performance estimates of Table 4 in Corno et al. (2022), which serve as the basis for the tests in Table
2022
-
[2023]
(2023), which serve as the basis for the tests in Table
Table E3 reproduces the deskmate-level peer-effect estimates of Table 5 in Wu et al. (2023), which serve as the basis for the tests in Table
2023
Reviewed August 4, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.