Joint work with Vít Illichmann. Previously circulated as Convolutional Peer Effects. Python package available here.

Note: this paper is a prototype. Work is in progress on a more general framework. This paper is mostly intended to showcase its computational approach.

Abstract. We study structural estimation on networks in the empirically common case of a single large observed graph. We propose an adversarial estimator that minimizes statistical distance between observed and simulated node-specific distributions of local network neighborhoods. The paper provides two theoretical results: population identification via a divergence characterization of the estimation objective, and consistency under growing-graph asymptotics with cross-observation dependence. A key contribution is computational. We provide a reproducible estimation workflow that integrates fixed-point simulation, efficient focal-neighborhood data construction, and alternating minimax training with stabilization tools suitable for large-scale runs. The workflow is model-agnostic in a broad class of network structural models and is straightforward to implement with modern software. In benchmark simulations, the procedure scales to large graphs and recovers structural parameters with high precision.

Download .pdf