|
Abstract:
|
Recent years have seen renewed interest in design based approaches to analyzing randomized controlled trials. Design based inference, or randomization inference, relies on the known treatment assignment mechanism of the experiment to derive tests and estimators, rather than specifying parametric forms for outcomes. In many situations, these tests coincide with permutation approaches to IID data. In both cases, the null hypothesis allows constructing exact tests that do not require asymptotic justifications. In this paper, we extend design based inference to the study of networks. Existing parametric methods make assumptions that are incompatible with design based principles or obscure the role of treatment effects. Using only the randomization mechanism or sampling plan, we provide exact tests of tests of no treatment effects or equivalence between groups. We demonstrate our methods with simulation studies and applications in social science and genomics.
|
