Unconfoundedness with Network Interference

This paper studies nonparametric estimation of treatment and spillover effects using observational data from a single large network. We consider a model of network interference that allows for peer influence in selection into treatment or outcomes but requires influence to decay with network distance. In this setting, the network and covariates of all units can be potential sources of confounding, in contrast to existing work that assumes confounding is limited to a known, low-dimensional function of these objects. To estimate the first-stage nuisance functions of the doubly robust estimator, we propose to use graph neural networks, which are designed to approximate functions of graph-structured inputs. Under our model of interference, we derive primitive conditions for a network analog of approximate sparsity, which provides justification for the use of shallow architectures.