Primary Submission Category: Causal Inference in Networks
Theoretical insights and new algorithms for model-assisted design of experiments under network correlated outcomes
Authors: Amir Ghasemian, Minzhengxiong Zhang, Edoardo Airoldi,
Presenting Author: Minzhengxiong Zhang*
Randomized trials are methodologically justified in order to achieve an unbiased estimate in causal inference.
However, an estimator even unbiased can be inefficient if it has a large variance. The standard difference-in-means estimator in a traditional randomized design setting such as bernoulli or complete randomization may output large variance estimates and renders the inference pointless.
This situation is worse on networks, where the systematic relation among the units, through the network topology of the treated and controlled units, can increase the variance of the inferences through mechanisms such as Homophily and interference.
Therefore, for a precise inference, we need to restrict the randomizations through an efficient design to reduce the variance of the estimators.
In this paper, utilizing a model-assisted design paradigm for a network setting in the presence of linear homophily, we (i) investigate a set of constraints needed for optimal randomizations that will satisfy some desired properties such as unbiasedness and minimum variance, (ii) propose mechanisms to satisfy these constraints, (iii) derive the marginal mean square error equations for the proposed mechanisms, (iv) develop efficient tools to satisfy these mechanisms, and (v) illustrate the efficiency of these suggested solutions through a set of simulations.