Bagging the Network

This paper studies parametric estimation and inference in a dyadic network formation model with nontransferable utilities, incorporating observed covariates and unobservable individual fixed effects. We address both theoretical and computational challenges of maximum likelihood estimation in this complex network model by proposing a new bootstrap aggregating (bagging) estimator, which is asymptotically normal, unbiased, and efficient. We extend the approach to estimating average partial effects and analyzing link function misspecification. Simulations demonstrate strong finite-sample performance. Two empirical applications to Nyakatoke risk-sharing networks and Indian microfinance data find insignificant roles of wealth differences in link formation and the strong influence of caste in Indian villages, respectively.