WorksEconomics, Econometrics and Finance
Network-Adaptive Instrumental Variables
Abstract
Three mechanisms generate endogeneity in regressions under interference. We address them by constructing network-adaptive instrumental variables (NIV) for a spatial autoregressive (SAR) model with endogenous covariates and unknown network weights. Under network-conditional exclusion and information equivalence, the resulting instruments yield valid moments without knowing the true network. We establish joint identification of the structural mean coefficients and the network parameters, providing two sufficient routes. Then under additional covariance restrictions, we identify the disturbance spillover coefficient and the innovation variance. We develop NIV-GMM estimation, establish consistency and asymptotic normality, and give feasible inference and a test for mean network spillovers. Monte Carlo simulations examine their finite-sample performance.