Nonseparable Dyadic Regression

Abstract

This paper studies a nonseparable model for dyadic outcomes, such as bilateral trade flows, in which the outcome depends on both agents' characteristics and on a scalar unobservable through an unknown function increasing in that unobservable. I establish identification of a normalized structural function and the error distribution, propose kernel plug-in estimators, and derive a two-regime central limit theory under dyadic dependence in which a shared-agent variance component generically dominates. An agent-level bootstrap is proved consistent in that regime. Simulations show independence-based intervals undercover severely while the bootstrap substantially improves coverage. In bilateral trade data, conditional dispersion falls by more than half between small and large exporters.