Abstract:
|
In fields such as marketing, social science, and public health, designed and natural experiments will frequently occur in a networked environment. In this setting, it is reasonable to assume that the treatment of one individual may impact nearby individuals in a network. We explore a series of assumptions on the set of potential outcomes which parsimoniously allow for this interference in a setting where the network is observed. Our focus is on estimation and we explore Bayesian estimates, optimal design-unbiased estimates, and minimax estimates of direct, interference, and total effects. We compare these estimates via simulations on random graph models and real-world graphs and consider the tradeoffs between them.
|