A simple test for the ignorability of non-compliance in experiments

This paper proposes a simple method for testing whether non-compliance in experiments is ignorable, i.e., not jointly related to the treatment and the outcome. The approach consists of (i) regressing the outcome variable on a constant, the treatment, the assignment indicator, and the treatment/assignment interaction and (ii) testing whether the coefficients on the latter two variables are jointly equal to zero. A brief simulation study illustrates the finite sample properties of the test.