Partial-Identification of Networks from Short Panels
Abstract
Partial Identification of Network Structure from Short Panel Data
We study identification of social network structure from short panel data in the linear-in-means model. When T is fixed and small relative to N, the reduced-form parameter matrix cannot be consistently estimated, yet we show the network is partially identified: a subset of eigenvalues and eigenvectors of the adjacency matrix—those not orthogonal to the covariate column space—can be recovered. We sharply characterize the identified set as a function of (N, T) and covariate rank. We then show these spectral components enable economically relevant inference. For optimal treatment targeting under network interference, the recoverable eigenvectors define the feasible set of treatment rules, yielding welfare strictly above random assignment—even without observing the network. We also discuss implications for key player identification and eigenvector centrality. Simulations calibrated to realistic networks demonstrate that with T as small as 3 to 5, the dominant spectral components most relevant for targeting are recoverable with useful precision.
Keywords: Social interactions, partial identification, network econometrics, eigenvalue methods, optimal targeting, short panels
JEL codes: C31, C33, D85