Specification Testing for Dyadic Regression Models

arXiv:2607.26366 · econ.EM, stat.ME · Submitted 2026-07-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Specification Testing for Dyadic Regression Models".

Warren: Here's the gist of the paper. This paper develops specification tests for linear conditional-mean models with undirected dyadic data, establishing a uniform projection theorem that reduces the dyadic process to its latent first-order node projections and proposing a corrected Gaussian bootstrap to handle variance duplication in different dependence regimes.

Ruth: Model Setup and Hypothesis The study focuses on linear conditional-mean models with undirected dyadic data, where observations may be dependent whenever their dyads share a node.

Warren: The null hypothesis tests whether the conditional mean of the dyadic outcome is linear in the included regressors, stating that E

Y12 X12: = X'12β∗ almost surely.

Ruth: The population residual-marked moment is defined as ∆(x) = E

ε121q(X⪯ x): .

Paper summary: Iris: So we've talked about the setup, and now let's zero in on what the paper actually claims about the model structure. The main thesis of this work is establishing specification tests for linear conditional mean models with undirected dyadic data, which is a setting they say hasn't been fully addressed before one <ref:2607.26366#pg2>.

Warren: They lay out their central claim: they develop a uniform projection theorem that reduces the complicated dyadic process to its latent first-order node projections when there’s shared-node dependence one <ref:2607.26366#pg2>. This is the mechanism they use to handle that complexity.

Ruth: From an economic standpoint, this means they are showing a way to simplify the model structure from a complex pairwise relationship down to something manageable at the level of individual nodes, which is crucial for inference in network applications one <ref:2607.26366#pg2>.

Iris: And then building on that, they show that under shared-node dependence, the raw first-order node-multiplier bootstrap is valid, but when dyads are independent—meaning those pairs aren't sharing much—it double counts the variation specific to each pair one <ref:2607.26366#pg2>.

Warren: That’s where the nuance is. They then introduce this exact covariance decomposition which leads them to propose a corrected Gaussian bootstrap that handles that issue by removing the extra same-dyad contribution one <ref:2607.26366#pg2>.

Ruth: The paper also confirms that their implemented Kolmogorov-Smirnov and Cramér-von Mises tests are consistent against fixed alternatives captured by an evaluation grid, and they characterize the power against local alternatives where the effect is changing over time or space one <ref:2607.26366#pg2>.

Iris: So, to summarize their methodology: they use a projection theorem to reduce complexity, identify a failure mode in bootstrapping when dyads are independent, and propose a corrected method that handles both dependence regimes one <ref:2607.26366#pg2>.

Warren: The continuum moment characterization they do is omnibus across the whole setup, which gives them broad coverage over the different types of data structures involved one <ref:2607.26366#pg2>.

Conclusion: Ruth: Looking at the full scope of "Specification Testing for Dyadic Regression Models," it feels like this paper is really about making sure that when we use these powerful conditional mean models on network data, our tests actually tell us what they need to tell us one <ref:2607.26366#pg2>.

Warren: It’s about moving beyond just saying a linear model fits well, to actually providing a valid way to test if the true underlying relationship is even linear, especially given the messy nature of shared-node dependence one <ref:2607.26366#pg2>.

Iris: The authors have really made their point that accounting for shared-node dependence isn't optional; it’s central to getting valid inference in these dyadic settings one <ref:2607.26366#pg2>.

Ruth: And the practical implication, as shown in their empirical results, is that you need these corrected procedures if you want a reliable diagnostic tool when dealing with real network data, because the simpler methods can be misleading one <ref:2607.26366#pg2>.

Warren: So ultimately, the paper provides a robust framework—the corrected KS test and bootstrap—that seems to balance size accuracy with power across different dependence regimes one <ref:2607.26366#pg2>.

Ulrich Hounyo, Jiahao Lin, Xiaojun Song

University at Albany-State University of New York · Fuyao University of Science and Technology · Peking University

econ.EM, stat.ME

Submitted: 2026-07-29

Updated: 2026-11-09

Project page: https://pdhoff.github.io/amen/reference

License: http://creativecommons.org/publicdomain/zero/1.0/

Importance score: 99/100

The gist: This paper develops specification tests for linear conditional-mean models with undirected dyadic data…

Key concepts

Linear Conditional-Mean Model
This model assumes the expected value of an outcome (Y12) is a linear function of regressors (X12), conditional on the dyad itself. The null hypothesis tests if this relationship is strictly linear, meaning E[Y12 | X12] = X'12β* almost surely.
Uniform Projection Theorem
This theorem shows that for empirical processes indexed by a class of functions (VC-type), the complex second-order component of the dyadic data process becomes negligible uniformly across all possible indices at the node level, simplifying the analysis to focus on latent node projections.
Corrected Gaussian Bootstrap
This bootstrap method addresses issues where variance is duplicated, especially when dyads are independent. It corrects a naive multiplier bootstrap by removing the extra contribution from same-dyad variation, making it valid and accurate under various dependence structures.

Terminology

Summary

This paper develops specification tests for linear conditional-mean models with undirected dyadic data, establishing a uniform projection theorem that reduces the dyadic process to its latent first-order node projections and proposing a corrected Gaussian bootstrap to handle variance duplication in different dependence regimes. p. 0

Where the paper says it

Specification Testing for Dyadic Regression Models Ulrich Hounyo∗ Jiahao Lin† Xiaojun Song‡ August 25, 2026 Abstract This paper develops specification tests for linear conditional-mean models with undirected dyadic data, a setting that, to the best of our knowledge, has not previously been addressed in the literature. We establish a uniform projection theorem that reduces the dyadic process to its latent first-order node projections under shared-node dependence.

Model Setup and Hypothesis

The study focuses on linear conditional-mean models with undirected dyadic data, where observations may be dependent whenever their dyads share a node. The null hypothesis tests whether the conditional mean of the dyadic outcome is linear in the included regressors, stating that E[Y12 X12] = X'12β∗ almost surely. The population residual-marked moment is defined as ∆(x) = E[ε121q(X⪯ x)]. p. 6

Where the paper says it

Define M(x) = E[X121{X12 ⪯ x}] and qx(X) = 1{X ⪯ x} − M(x) ′Q −1X. (3) The population residual-marked moment is ∆(x) = E[ε121{X12 ⪯ x}].

Lemma 1 shows that the null hypothesis holds if and only if ∆(x) = 0 for every x ∈ R d. p. 6
Where the paper says it

(4) Since E[X12ε12] = 0, the following equality holds regardless of whether the null hypothesis is satisfied: ∆(x) = E[rx(Z12)]. (5) Lemma 1 (Omnibus characterization).

Asymptotic Theory and Process Characterization

The authors establish a uniform projection theorem for dyadic empirical processes indexed by a VC-type class, showing that the second-order component is negligible uniformly over the index set at the node-level rate. This reduces the leading dyadic process uniformly to an ordinary empirical process over the latent nodes. p. 2

Where the paper says it

After separating these components, we control the former as a Rademacher process and the latter as a decoupled second-order Rademacher chaos. The result reduces the leading dyadic process uniformly to an ordinary empirical process over the latent nodes and provides the functional weak convergence required for both the KS and CvM statistics.

The asymptotic behavior of the estimated residual-marked process Rbn is characterized by two distinct variance regimes: under nondegenerate shared-node dependence, it converges at the √n rate governed by first-order node projections, and under independent dyads, it converges at the faster √Nn rate governed by dyad-level marks. p. 0
Where the paper says it

We establish a uniform projection theorem that reduces the dyadic process to its latent first-order node projections under shared-node dependence. We then show that a raw first-order node-multiplier bootstrap is valid when this node component is nondegenerate but double-counts dyad-specific variation when dyads are independent.

Bootstrap Procedures and Validity

The paper develops two bootstrap procedures: a raw first-order node-multiplier bootstrap and a covariance-corrected Gaussian bootstrap. The raw node-multiplier bootstrap is valid under nondegenerate shared-node dependence but double-counts dyad-specific variation when dyads are independent. p. 0

Where the paper says it

We establish a uniform projection theorem that reduces the dyadic process to its latent first-order node projections under shared-node dependence. We then show that a raw first-order node-multiplier bootstrap is valid when this node component is nondegenerate but double-counts dyad-specific variation when dyads are independent.

The exact covariance decomposition motivates the corrected Gaussian bootstrap, which removes the extra same-dyad contribution and is essential under independent dyads. p. 2
Where the paper says it

This decomposition leads to a corrected Gaussian bootstrap that retains the shared-node component while removing the extra same-dyad contribution. We prove that the correction is asymptotically irrelevant under nondegenerate shared-node dependence and essential under independent dyads.

Empirical Results and Application

Simulations show that the corrected Kolmogorov-Smirnov test provides the most stable size control while retaining substantial local power across different designs. An application to a law-firm network rejects additive linear and quadratic specifications but does not reject a richer specification containing an economically relevant interaction, suggesting that professional connections cannot be adequately described by an additive linear conditional mean. p. 0

Where the paper says it

Simulations show that the corrected Kolmogorov-Smirnov test provides the most stable size control while retaining substantial local power. An application to the Lazega law-firm network rejects additive linear and quadratic specifications but does not reject a richer specification containing an economically relevant interaction.

The comparison of bootstrap methods demonstrates the empirical importance of accounting for dyadic dependence, as the naive dyad-level test overstates sampling variation when dependence is weak. p. 4
Where the paper says it

(2024), and Lee et al. Dyadic sampling differs from these settings because shared-node dependence generates competing node-level and dyad-level variance components, whose relative importance determines the appropriate normalization, the functional limit of the test process, and the validity of the bootstrap approximation.

Conclusion and Significance

The corrected KS test delivers the best finite-sample performance across the designs considered while retaining substantial local power against both quadratic and interaction alternatives. The results illustrate both the importance of accounting for shared-node dependence and the usefulness of the proposed tests for detecting economically meaningful nonlinearities in dyadic conditional-mean models. p. 3

Where the paper says it

The Monte Carlo experiments examine the finite-sample size and local power of the proposed tests across different sample sizes and strengths of shared-node dependence. The results illustrate the practical importance of accounting for dyadic dependence and show how the relative performance of the procedures changes as the node-level component becomes weaker.

The corrected procedures are asymptotically equivalent under nondegenerate shared-node dependence, while the raw node multiplier counts each dyad through both endpoints and produces twice the correct limiting covariance under independent dyads. p. 0
Where the paper says it

We then show that a raw first-order node-multiplier bootstrap is valid when this node component is nondegenerate but double-counts dyad-specific variation when dyads are independent. An exact covariance decomposition motivates a corrected Gaussian bootstrap, whose validity is established separately under shared-node dependence and independent dyads.

The rejection probabilities of the valid tests are at least α, providing a reliable diagnostic tool for assessing linear model misspecification.

Technical Foundations

The appendix develops the projection, node-recovery, conditional multiplier, and measurability arguments used in the main text. Key results include Lemma 1 (Exact projection identity) showing that the second-order component is op(1) for each fixed f, and Lemma 9 (Uniform recovery of node projections) which shows that the noise term is op(1). p. 2

Where the paper says it

First, we establish a uniform projection theorem for dyadic empirical processes indexed by a VC-type class. The theorem shows that the second-order component is negligible uniformly over the index set at the node-level rate.

The proof of Theorem 1 confirms that the corrected tests satisfy asymptotic validity under both shared-node dependence and independent dyads. p. 0
Where the paper says it

An exact covariance decomposition motivates a corrected Gaussian bootstrap, whose validity is established separately under shared-node dependence and independent dyads. The continuum moment characterization is omnibus, while the implemented Kolmogorov-Smirnov and Cramér-von Mises tests are consistent against fixed alternatives captured by the evaluation grid.

Bootstrap Comparison and Convergence

The analysis shows that the raw node multiplier process converges to GR in scenario (i) and GD in scenario (ii). The covariance estimators converge to KG and ΓG, respectively, under the respective scenarios. p. 4

Where the paper says it

Thus a process Zn = {Zn(t) : t ∈ T} is viewed as an ℓ ∞(T)-valued random element, and Zn ⇝ Z in ℓ ∞(T) denotes weak convergence of the entire process. We use →p, op(1), and Op(1) for convergence in probability, convergence to zero in probability, and boundedness in probability, respectively; an ≍ bn means that an/bn is bounded above and away from zero.

The corrected procedures are asymptotically equivalent under nondegenerate shared-node dependence and yield valid grid tests even when dyads are independent. p. 0
Where the paper says it

We then show that a raw first-order node-multiplier bootstrap is valid when this node component is nondegenerate but double-counts dyad-specific variation when dyads are independent. An exact covariance decomposition motivates a corrected Gaussian bootstrap, whose validity is established separately under shared-node dependence and independent dyads.

The local asymptotic power analysis confirms that the corrected KS test provides the most favorable balance between size accuracy and power. p. 0
Where the paper says it

We also characterize their power against rate-appropriate local alternatives. Simulations show that the corrected Kolmogorov-Smirnov test provides the most stable size control while retaining substantial local power.

Empirical Study

The empirical study on a law-firm network shows strong evidence against the additive linear specification, and the corrected bootstrap p-values reject the null hypothesis for both KS and CvM statistics at the 5% significance level. The results provide strong evidence against an additive linear specification. p. 3

Where the paper says it

The empirical analysis examines whether professional links among lawyers can be adequately represented by an additive linear conditional mean based on individual and pairwise characteristics. The results indicate that professional connections cannot be adequately described by adding the separate contributions of experience, demographic similarity, and organizational proximity.

The disagreement in the last row of Table 1 between the naive and corrected procedures illustrates why a dyadic-dependence-robust specification test is needed. p. 3
Where the paper says it

Instead, these characteristics appear to operate jointly. The application also demonstrates how the proposed procedures can be used as formal diagnostic tools before interpreting a linear dyadic regression and illustrates their ability to detect economically meaningful departures from the maintained specification.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Development of robust specification tests for linear conditional-mean models with undirected dyadic data allows AI systems to detect economically relevant nonlinearities or interactions that standard linear specifications omit, as demonstrated by rejecting additive and quadratic specifications in the Lazega law-firm network application.

  2. Implementation of a corrected Gaussian bootstrap procedure enables AI systems to maintain validity across two variance regimes—nondegenerate shared-node dependence and independent dyads—providing more stable size control than raw node-multiplier methods.

  3. The uniform projection theorem reduces the dyadic process to its latent first-order node projections under shared-node dependence, enabling AI systems to recover these unobserved node projections uniformly from incident-dyad averages.

  4. The identification of the failure of the raw node-multiplier bootstrap under first-order degeneracy and the proposal of a covariance correction allow AI systems to adjust sampling variation estimates, ensuring that critical values are not overly conservative when dyadic dependence is weak or absent.

  5. Application in network analysis allows AI systems to distinguish between additive linear specifications and richer models containing economically relevant interactions, such as the shared office by shared practice interaction, leading to more nuanced interpretation of professional connectivity.

Abstract

Of the 88 empirical papers on dyadic or other relational data that we reviewed in the top five economics journals, 55 (62.5%) use a linear specification. To assess whether such a specification describes the conditional mean, we develop Kolmogorov-Smirnov and Cramér-von Mises tests for undirected dyadic data that also accommodate smooth nonlinear specifications. To our knowledge, this paper is the first to develop omnibus specification tests for undirected dyadic regression models. A corrected first-order multiplier bootstrap is asymptotically exact in two standard Gaussian cases: the first-order nondegenerate case and independent dyads case. For a general class of such processes, we establish that a second-order multiplier bootstrap consistently approximates the limiting law uniformly over the indexing function class, even when the indexing functions are estimated. For the specification tests, we augment this bootstrap to account for two additional sources of non-Gaussianity: parameter estimation and the curvature of the regression function. We then construct an envelope test that is asymptotically exact in both Gaussian cases and controls size in the other degenerate cases. Simulations illustrate the size and power of the tests. In an application to housing sales, we reject a common mean price gap within distance bands but do not reject separate linear relationships by metro access.

Sources

Related papers