Identification of Peer Effects with Negative Controls for Unobserved Confounding

• In Event: Observational and Experimental Causal Inference

Abstract

Political scientists have been interested in causal peer effects to understand how people’s behaviors are affected by their network peers. However, it is well known that identification and estimation of causal peer effects are challenging in observational studies due to unmeasured confounding that exhibits network dependence, for example, homophily bias and contextual confounding. Negative control variables, also known as placebo variables, have a long-standing tradition to rule out non-causal explanations, although they have been used primarily for bias detection. In this article, we establish nonparametric identification of causal peer effects using double negative control variables in the presence of network-dependent unmeasured confounding. We then propose the generalized method of moments estimator for causal peer effects, and a network heteroskedasticity and autocorrelation consistent (HAC) variance estimator. We illustrate our methods with an application to campaign donations and lobbying.

Author

• Naoki Egami, Columbia University