Estimating Spillovers Using Imprecisely Measured Networks

Working Paper
Published on 31 January 2020


In many experimental contexts, whether and how network interactions impact outcomes of both treated and untreated individuals are key concerns. Networks data is often assumed to perfectly represent the set of individuals who might be affected by these interactions. This paper considers the problem of estimating treatment effects when measured connections are, instead, a noisy representation of the true spillover pathways. Hardy et al. show that existing methods yield biased estimators in the presence of this mismeasurement. They develop a new method that uses a class of mixture models to model the underlying network and account for missing connections, and then discuss its estimation via the Expectation-Maximization algorithm. The authors check their method's performance by simulating experiments on network data from 43 villages in India (Banerjee et al., 2013). Finally, they use data from Cai et al. (2015) to show that estimates using their method are more robust to the choice of network measure than existing methods.


Morgan Hardy

New York University, Abu Dhabi

Rachel Heath

University of Washington

Wesley Lee

University of Washington

Tyler H. McCormick

University of Washington