A1422
Title: Endogenous interference in randomized experiments
Authors: Mengsi Gao - University of Southern California (United States) [presenting]
Abstract: Identification and inference for treatment effects in randomized trials with social interactions are studied. Two key features introduce endogeneity: (1) latent variables that affect both networks and outcomes, and (2) treatment-induced network changes that mediate treatment effects. A potential-outcomes framework with post-treatment networks is developed, and direct and network-mediated effects are defined. For estimation, a shift-share IV strategy with asymptotic theory for sparse networks is proposed, but it fails in denser settings. A denoised, eigen-decomposition-based SSIV estimator is then developed that restores consistency in denser networks and improves convergence rate. An application illustrates both direct effects and network-mediated effects.
