Single-Network Finite-Sample Inference in Strategic Network Formation Models

2026-07-29

AI Summary

This study addresses the challenge of finite-sample inference for strategic network formation models when only a single network is observed. It proposes a novel approach that avoids imposing assumptions on network density, strategic dependence structure, or equilibrium selection mechanisms. The method constructs realization-specific sharp bounds via bounding-by-c techniques, incorporates identification constraints through group-level averages of exogenous covariates, and derives a test statistic—based on exogenous variables and error terms—that can be simulated exactly. By circumventing the need to solve or enumerate equilibrium networks, the procedure achieves computational efficiency while preserving finite-sample validity. Empirical results demonstrate that, across networks ranging from 300 to 9,500 nodes, the approach correctly identifies the sign of strategic coefficients at the 95% confidence level and provides strong evidence of positive interdependence in link formation.

Abstract

We develop a finite-sample valid inference procedure for strategic network formation models in which linking decisions depend on endogenous network statistics (say, the number of common friends). Only a single network is required to be observed, and we restrict neither its density, nor the dependence structure induced by strategic interaction, nor the equilibrium selection mechanism. We exploit a bounding-by-c technique to construct a set of sandwich inequalities that are valid realization by realization, with the middle term involving only the i.i.d. pairwise error. We then average the sandwich inequalities over cells of exogenous covariates, and obtain identifying restrictions under a nonstandard pathwise limit formulation. For inference, we construct test statistics whose finite-sample uncertainty can be controlled by statistics of the exogenous covariates and errors alone, whose conditional distributions are exactly simulable in both semiparametric and parametric settings. Our proposed inference procedure is also computationally tractable, with no need to solve, simulate, or enumerate equilibrium network structures. In simulations, our procedure easily scales to networks of size 10,000, and yields confidence sets that certifies the sign of the strategic coefficient. In two empirical applications (with network size about 300~9500), we find statistical evidence for positive link interdependence at 95% confidence level.

Problem

• strategic network formation
• finite-sample inference
• single-network observation
• endogenous network statistics
• link interdependence

Innovation

• finite-sample inference
• strategic network formation
• single-network observation
• bounding-by-c technique
• equilibrium-free inference