Causal Clustering: Design of Cluster Experiments Under Network Interference
Too Long; Didn't Read
This paper presents a novel approach to designing cluster experiments in network settings for estimating global treatment effects. The Causal Clustering algorithm is introduced, aiming to minimize the worst-case mean-squared error in treatment effect estimation by optimizing cluster design. The study explores the impact of clustering choices on bias and variance, providing conditions for selecting between cluster-level and individual-level randomization. Unique network data from Facebook users and existing field experiment data are utilized to illustrate the properties of the proposed method.