When randomized experiments encounter noncompliance, estimates of average causal effects typically apply only to the subset of subjects known as compliers (those who take treatment if and only if assigned to the treatment group). In the presence of treatment effect heterogeneity, however, the complier average causal effect (CACE) does not provide insight into the average causal effects for other subgroups. I propose a new experimental design and associated estimator for combating noncompliance. The share of subjects to whom causal effect estimates pertain can be increased by resampling among those who fail to take first-round treatment into second-round treatment and control groups. Simulations show that the design outperforms alternatives under many circumstances.
Working paper here.
Working paper appendix here.