Published online by Cambridge University Press: 23 February 2017
Missing outcome data plague many randomized experiments. Common solutions rely on ignorability assumptions that may not be credible in all applications. We propose a method for confronting missing outcome data that makes fairly weak assumptions but can still yield informative bounds on the average treatment effect. Our approach is based on a combination of the double sampling design and nonparametric worst-case bounds. We derive a worst-case bounds estimator under double sampling and provide analytic expressions for variance estimators and confidence intervals. We also propose a method for covariate adjustment using poststratification and a sensitivity analysis for nonignorable missingness. Finally, we illustrate the utility of our approach using Monte Carlo simulations and a placebo-controlled randomized field experiment on the effects of persuasion on social attitudes with survey-based outcome measures.
Authors’ note: The authors thank Sebastian Bauhoff, Bill Berry, Chris Blattman, Jake Bowers, Matias Cattaneo, Kosuke Imai, Molly Offer-Westort, Rocio Titiunik, and participants of the 2013 Joint Statistical Meetings for very helpful comments on previous versions of this manuscript. The authors especially thank Peter Aronow for his contributions to previous versions of this paper. This research was approved by the Columbia University IRB (Protocol AAAP1312) and the empirical analyses were preregistered at egap.org (ID: 20150702AA). Easy-to-use software for the statistical programming language R that implements the methods described in this paper is available at github.com/acoppock/attrition. The replication materials for all analyses reported here are available at dx.doi.org/10.7910/DVN/AQB4MP.
Contributing Editor: Kosuke Imai