Designs for estimating the treatment effect in networks with interference (Q2196188)

The authors construct and analyze new designs for estimating treatment effects from randomized experiments in networks with interference. The nodes of a network are considered as experimental units and the estimand of interest is the direct treatment effect. Typically for networks, the events that any two units receive treatment are not independent. The authors study the classical Neymanian estimator (\(Ne\)) that takes the difference between the mean of the outcome for treated nodes and the control nodes. For the \(Ne\) they focus on the design of treatment assignments. Conceptually, their main contribution is the idea of considering a treatment assignment as a ``quasi-coloring'' of a graph. A treatment assignment is a perfect quasi-coloring (\(pqc\)) if, for every treated node \(v\), there is a nontreated node \(v'\) that has the same number of treated and nontreated neighbors as that of \(v\). The notion of \(pqc\) is inspired by the notion of covariate balance in the context of matching in observational studies. For a \(pqc\), the \(Ne\) is unbiased. If \(pqc\) is not available, \(Ne\) becomes biased. The authors develop a new restricted randomization design that reduces bias and variance and provides easily implementable algorithms to construct such designs. Further, they derive upper bounds on the bias and variance of \(Ne\) and prove the asymptotic consistency of the \(Ne\) in both, dense and sparse asymptotic regimes. Finally, the authors demonstrate the efficacy of the proposed randomization scheme in a series of simulations.