Average treatment effects in the presence of unknown interference (Q2039779)

It is noted that causal inference usually assumes the absence of interference among the units of the population. However, in certain situations ,especially in social and medical fields, interference is inevitable. In such cases, it is possible to allow moderate amounts of interference in an unknown and arbitrary way. A sample of units is selected and a random subset is assigned to a particular treatment. Here the parameter of interest is the average effect of the assignments under spill over effects. The authors investigate the expected average treatment effect (EATE). They obtain consistent estimates of EATE as well as their variance estimates using Horvitz-Thompson and Ha'jek estimators. Furthermore, they show that EATE for a given experiment is informative of the effect under the designs that are close to the implemented design. Thus external validity is established with a proper definition of measure of closeness, which is an important practical aspect. There is a simulation study in one of the three supplements provided at the end of the paper.