Quantifying uncertainty in network regressions
Social networks represent relationships between pairs of interconnected individuals and are widely used to understand complex social phenomena. In this work, we consider inference on regression coefficients in network regression models, where the presence/absence (or strength) of a connection between two individuals is modeled as a linear function of observable covariates and structured, network dependent, error. We leverage a joint exchangeability assumption, nearly ubiquitous in the statistics literature on networks but not previously considered in the estimating equations formulation for network regressions, to derive parsimonious estimators of the covariance within the network. We propose and evaluate testable conditions that guide practitioners in selecting an estimator. We also demonstrate our proposed estimators using data from a behavioral economics experiment.