📈 economics

Specification Testing for Dyadic Regression Models

This paper develops and validates corrected Gaussian bootstrap-based omnibus specification tests for linear conditional-mean models with undirected dyadic data, demonstrating their consistency and superior size control in simulations and a real-world application to law-firm networks.

Original authors: Ulrich Hounyo, Jiahao Lin, Xiaojun Song

Published 2026-07-30

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Ulrich Hounyo, Jiahao Lin, Xiaojun Song

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand a massive, bustling city. You want to know why people hang out together. Is it because they live on the same street? Because they work at the same office? Or maybe just because they both love the same obscure band? In the world of data science, this is called studying "dyadic data." It's just a fancy way of looking at pairs of things—like two friends, two countries trading goods, or two companies buying from each other. The tricky part is that these pairs aren't independent. If Alice and Bob are friends, and Bob and Charlie are friends, Alice and Charlie might become friends too, not because of anything they did directly, but because they share Bob. This "shared connection" creates a ripple effect that makes standard math tools break.

For decades, statisticians have tried to build simple, straight-line models to predict these relationships. Think of it like drawing a ruler-straight line through a cloud of dots to see the trend. It's easy to use and easy to explain. But what if the real world isn't a straight line? What if the connection between Alice and Charlie depends on a complex mix of their shared hobbies, their jobs, and their ages, twisting and turning in ways a ruler can't capture? If you force a straight line onto a squiggly reality, your predictions will be wrong, and you might miss the most interesting parts of the story. The big question is: How do you know if your straight line is actually lying to you?

This paper is like a new, super-accurate lie detector test for those straight-line models. The authors, Ulrich Hounyo, Jiahao Lin, and Xiaojun Song, have developed a way to check if a simple linear model is doing a good job or if it's missing something important in these connected networks. They realized that the old ways of checking for errors were like trying to weigh a feather while standing on a trampoline; the bouncy, connected nature of the data made the scales jump around, giving false readings.

The team discovered that the data has two types of "noise." The first is the "neighbor noise"—the fact that everyone is connected to a central hub (like a person connected to many friends). The second is the "unique noise"—the specific, random quirks of just one pair. When the neighbor noise is strong, the math works one way. But when the connections are weak and everyone is basically independent, the math flips, and the unique noise takes over. The authors found that a popular method used by many researchers (called the "raw node-multiplier bootstrap") accidentally counts the unique noise twice when the connections are weak, making the test too cautious and missing real problems.

To fix this, they invented a "corrected Gaussian bootstrap." Imagine you are counting the number of people in a room. If you count everyone twice because they are holding hands in pairs, you get the wrong number. Their new method is like a smart counter that knows when to count the pairs once and when to count them twice, ensuring the math is perfect whether the group is tightly knit or loosely connected. They tested this new tool with thousands of computer simulations, creating fake networks with different levels of connection. The results showed that their corrected test is the most stable and reliable, catching errors that other methods miss without crying wolf when everything is actually fine.

Finally, they took their new test to a real-world dataset: a network of 71 lawyers in a law firm. They asked, "Can we predict who talks to whom just by adding up their years of experience, age differences, and office locations?" The answer was a resounding "No." The simple straight-line model failed. Even adding a curve to account for how much their experience levels differed wasn't enough. However, when they added a special "interaction" term—checking if two lawyers shared both an office and a practice area—the model suddenly worked perfectly. It turned out that lawyers only really connect when they share a specific combination of traits, not just one or the other. This paper proves that in the messy, connected world of human relationships, simple addition often fails, and you need a smarter, more flexible way to look at the data to find the truth.

Technical Summary: Specification Testing for Dyadic Regression Models

Problem Statement
Dyadic data, characterized by observations on pairs of nodes (e.g., international trade, social networks), present a unique inferential challenge: observations are not independent. Two dyads may be dependent if they share a common node (shared-node dependence). While existing methods for dyadic regression often assume a maintained specification for the conditional mean (typically linear), the functional form of this mean is rarely known in practice. Omitting relevant nonlinearities or interactions can lead to misleading conclusions. The primary problem addressed is the development of omnibus specification tests for linear conditional-mean models that remain valid under the complex dependence structures inherent in dyadic data, specifically accounting for the interplay between node-level and dyad-level variance components.

Methodology
The paper develops a testing framework based on a residual-marked empirical process indexed by lower orthants of the regressor distribution. The methodology proceeds through several theoretical and computational steps:

Uniform Projection Theorem: The authors establish a uniform projection theorem for dyadic empirical processes indexed by a VC-type class. This theorem demonstrates that the dyadic process can be uniformly reduced to its latent first-order node projections under shared-node dependence. Crucially, it shows that the second-order remainder (containing dyad-specific variation and joint node variation) is negligible at the node-level rate ( $\sqrt{n}$ ).
Variance Regimes and Limits: The analysis identifies two distinct variance regimes:
- Nondegenerate Shared-Node Dependence: The limiting process is governed by node projections, converging at rate $\sqrt{n}$ .
- Independent Dyads: The node projection vanishes, and the process is governed by dyad-specific variation, converging at the faster rate $\sqrt{N_n}$ (where $N_n = n(n-1)/2$ ).
Bootstrap Procedures:
- Raw Node-Multiplier Bootstrap: A standard multiplier bootstrap is constructed using estimated node projections. The paper proves this is valid under nondegenerate shared-node dependence. However, under independent dyads, this procedure "double-counts" the same-dyad variation (since each dyad enters two node averages), resulting in an asymptotic covariance that is twice the correct value.
- Covariance-Corrected Gaussian Bootstrap: To address the failure of the raw bootstrap in the independent-dyad regime, the authors derive an exact finite-sample covariance decomposition. This separates the same-dyad contribution from the shared-node contribution. A corrected Gaussian bootstrap is proposed that retains the shared-node component while removing the extra same-dyad contribution. This corrected procedure is shown to be valid in both regimes.
Test Statistics: The framework utilizes Kolmogorov-Smirnov (KS) and Cramér-von Mises (CvM) statistics derived from the residual-marked process. These are implemented on a fixed finite grid to avoid bandwidth selection.

Key Contributions

Uniform Projection Theory: The paper provides a nontrivial uniform projection theorem that controls the dyad-specific empirical process component and the node-pair Rademacher chaos uniformly. This allows for the functional weak convergence of the process to a Gaussian limit under both variance regimes.
Identification of Bootstrap Failure: The authors rigorously identify that the standard node-multiplier bootstrap fails under independent dyads due to a factor-of-two variance inflation.
Corrected Bootstrap Construction: An exact covariance decomposition motivates a corrected Gaussian bootstrap that adapts to the underlying variance regime without requiring prior knowledge of whether dyads are independent or dependent.
Local Power Analysis: The paper establishes that the tests are consistent against fixed alternatives and possess nontrivial power against local alternatives at the appropriate rate for each regime. The analysis accounts for the re-estimation of linear coefficients and identifies the detectable component of local departures.

Results

Theoretical Validity: The corrected grid tests are proven to be asymptotically valid in both the nondegenerate shared-node regime and the independent-dyad regime. The raw node-multiplier test is valid only in the former.
Simulation Findings: Monte Carlo experiments indicate that the corrected Kolmogorov-Smirnov test provides the most stable size control across different dependence strengths while retaining substantial local power. The raw node-multiplier test is severely conservative when dyadic dependence is weak or absent. The naive approach (treating dyads as independent) leads to severe overrejection as dependence increases.
Empirical Application: Applied to the Lazega law-firm network, the tests reject additive linear and quadratic specifications. However, after including an interaction term between shared office and shared practice area, the tests find no remaining misspecification. This suggests that professional connections are driven by complementarities among characteristics rather than simple additive homophily.

Significance and Claims
The paper claims to provide a robust diagnostic tool for dyadic regression that does not rely on knowing the specific dependence structure (independent vs. shared-node) in advance. By explicitly handling the boundary between these regimes, the proposed corrected Gaussian bootstrap avoids the pitfalls of both the standard node-multiplier (which fails under independence) and the naive dyad-level bootstrap (which fails under dependence).

The authors explicitly note the boundary of their results: the theory covers partial degeneracy and the totally first-order degenerate case of mutually independent dyads. They do not claim validity for every totally degenerate exchangeable array, noting that in more general cases, the second-order component may become leading and possess a non-Gaussian limit. The paper positions its contribution as filling a gap in the literature regarding specification testing for exchangeable arrays, extending beyond existing high-dimensional Gaussian approximations and nonparametric methods to specifically address the competing variance components of dyadic data.

More like this