The Design and Analysis of Experiments under the K-Nearest-Neighbors Interference Model

Traditional methods for estimating causal effects from experimental data often assume that an intervention only affects the unit receiving the intervention and does not impact the behavior of any other unit in the experiment. However, experimental settings where this condition fails to hold are increasingly common. For example, in an experiment on a social media network, an individual receiving an intervention may engage with other users on the network, thereby impacting the other users' responses. When this occurs, we say that the experiment exhibits treatment interference. This project investigates a new model of treatment interference- the K-nearest neighbors interference model (KNNIM)-in which a unit's response may be affected by the intervention given to its 'K' closest connections. Notably, this model allows for interventions given to closer connections of a unit to have a greater impact on that unit's response. This project also provides research training opportunities for graduate students. The project will derive estimators for useful causal estimands, in particular, nearest neighbor treatment effects, which quantify the amount of influence that neighboring units have on a unit's response under KNNIM and relaxations of this model. Tests for determining whether these relaxed models are plausible will also be developed. Furthermore, the project will derive effective experimental designs for improved estimation of and inference of treatment effects under KNNIM. Finally, borrowing approaches for detecting communities in networks, this project will develop methods for simultaneously determining the correct KNNIM interference structure-i.e., the correct value of 'K'- and estimating treatment effects under repeated experimentation on the same set of units. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.