A common experimental research design is one in which individuals are randomly allocated into groups and then interact within the groups under different group-level treatment conditions. We present methods for design-based inference for such ``group interaction'' experiments. A key consideration is that group interaction implies potential interference: an individual's potential outcomes depend on the groups to which others are allocated. This yields dependencies that should be accounted for when making inferential claims. We show that when group-member interference is present, standard cluster-robust inference is consistent, with reference to a superpopulation regime with sparse sampling, in accounting for such dependencies for inference on marginalized causal effects that account for interference. When interference is not present, but groups are formed through individually randomized assignment, individual-level heteroskedasticity robust inference is consistent for inference on the usual average treatment effect.
Joint with Ye Wang and Jiawei Fu
