Optimal Nonparametric Inference on Network Effects with Dependent Edges

Wenqin Du, Yuan Zhang, Wen ZhouarXiv 2024 · DOI · arXiv

Abstract

Testing network effects in weighted directed networks is a foundational problem in econometrics, sociology, and psychology. Yet, the prevalent edge dependency poses a significant methodological challenge. Most existing methods are model-based and come with stringent assumptions, limiting their applicability. In response, we introduce a novel, fully nonparametric framework that requires only minimal regularity assumptions. While inspired by recent developments in $U$-statistic literature, our approach notably broadens their scopes. Specifically, we identified and carefully addressed the challenge of indeterminate degeneracy in the test statistics — a problem that aforementioned tools do not handle. We established Berry-Esseen type bound for the accuracy of type-I error rate control. Using original analysis, we also proved the minimax optimality of our test's power. Simulations underscore the superiority of our method in computation speed, accuracy, and numerical robustness compared to competing methods. We also applied our method to the U.S. faculty hiring network data and discovered intriguing findings.

Introduction

In network science, quantifying "network effects" is essential for understanding the underlying mechanisms of social and economic interactions. These effects include phenomena such as reciprocity (mutual interaction between two nodes) and same-sender bias (the tendency of a node to initiate many interactions). Traditional inferential methods, such as the Social Relations Model (SRM), typically rely on heavy parametric assumptions, often assuming Gaussian random effects. However, real-world network data frequently violate these assumptions, and the inherent dependency between edges complicates standard statistical analysis.

Defining Network Effects

The study focuses on four primary network effects: reciprocity, same-sender, same-receiver, and sender-receiver. These are formally defined as covariances between specific edge weights, as illustrated in the following table:

Definitions and illustrations of the four primary network effects: reciprocity, same-sender, same-receiver, and sender-receiver.

A major technical hurdle in testing these effects is degeneracy. In many scenarios, the first-order variance of a test statistic vanishes under the null hypothesis, causing standard normal approximations to fail. Existing tools often require expensive subsampling or ignore the issue entirely, leading to poor power or invalid Type-I error control.

The NET Framework and Indeterminate Degeneracy

We introduce NET, a novel nonparametric framework designed to handle these challenges. A key innovation of NET is its ability to address indeterminate degeneracy. Unlike previous methods that assume a fixed degeneracy status, NET recognizes that the degree of degeneracy can depend on the underlying network structure and the specific null hypothesis being tested.

Summary of the testing framework for different null hypotheses, accounting for varying levels of degeneracy.

As shown in the table above, the framework employs different test statistics and decision rules depending on whether the statistic is degenerate or non-degenerate. For effects like "same-sender" ($\eta_3$) or "same-receiver" ($\eta_4$), the null hypothesis ($H_0: \eta = 0$) actually implies first-order degeneracy, making standard tests inherently unreliable. NET incorporates a diagnostic test to detect this degeneracy and adjust the downstream inference accordingly.

Computational Efficiency and Scalability

To manage the computational complexity of calculating $U$-statistics on large networks—which can involve millions of node combinations—NET utilizes a $U$-statistic reduction technique. By using a subsampled version of the statistic controlled by a parameter $\lambda \in [1, 2)$, the method restores asymptotic normality and significantly speeds up computation. This allows NET to scale to networks with over $10^5$ nodes, whereas traditional likelihood-based methods often fail at much smaller scales.

The parameter $\lambda$ offers a trade-off: values closer to 1 prioritize Type-I error control (accuracy), while values closer to 2 enhance the test's power (sensitivity).

Empirical Applications: U.S. Faculty Hiring

We applied the NET framework to U.S. faculty hiring data across Computer Science, Business, and History. The results revealed several key insights:

  • Universal Same-Sender Effect: All fields exhibited a strong same-sender effect, indicating a rigid prestige hierarchy where a few elite institutions dominate faculty production.
  • Field-Specific Reciprocity: Reciprocity was significant in Computer Science and History but absent in Business Schools.
  • History's Unique Pattern: History departments showed a pronounced sender-receiver effect, suggesting that institutions recruiting from elite counterparts are more successful at placing their own graduates in influential positions.

Conclusion

The NET framework provides a robust, model-free approach to network inference that overcomes the limitations of parametric models and the technical challenges of degeneracy. While currently focused on 3-node and 4-node motifs, future work will explore extensions to higher-order motifs and dynamic network structures.

Summary updated

Abstract

Testing network effects in weighted directed networks is a foundational problem in econometrics, sociology, and psychology. Yet, the prevalent edge dependency poses a significant methodological challenge. Most existing methods are model-based and come with stringent assumptions, limiting their applicability. In response, we introduce a novel, fully nonparametric framework that requires only minimal regularity assumptions. While inspired by recent developments in $U$-statistic literature (arXiv:1712.00771, arXiv:2004.06615), our approach notably broadens their scopes. Specifically, we identified and carefully addressed the challenge of indeterminate degeneracy in the test statistics $-$ a problem that aforementioned tools do not handle. We established Berry-Esseen type bound for the accuracy of type-I error rate control. Using original analysis, we also proved the minimax optimality of our test's power. Simulations underscore the superiority of our method in computation speed, accuracy, and numerical robustness compared to competing methods. We also applied our method to the U.S. faculty hiring network data and discovered intriguing findings.
BibTeX
@misc{du2024optimal,
  author = {Wenqin Du and Yuan Zhang and Wen Zhou},
  title = {{Optimal Nonparametric Inference on Network Effects with Dependent Edges}},
  year = {2024},
  eprint = {2401.03072},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2401.03072},
}
Figure