Experimental Design under Network Interference
share
This paper discusses the problem of the design of experiments under network interference. We allow for a possibly fully connected network and a general class of estimands, which encompasses average treatment and average spillover effects, as well as estimands obtained from interactions of the two. We discuss a near-optimal design mechanism, where the experimenter optimizes over participants and treatment assignments to minimize the variance of the estimators of interest, using a first-wave experiment for estimation of the variance. We guarantee valid asymptotic inference on causal effects using either parametric or non-parametric estimators under the proposed experimental design, allowing for local dependence of potential outcomes, arbitrary dependence of the treatment assignment indicators, and spillovers across units. We showcase asymptotic optimality and finite-sample upper bounds on the regret of the proposed design mechanism. Simulations illustrate the advantage of the method over state-of-art methodologies.
READ FULL TEXT