Addressing Missing Outcome Data in Randomized Experiments: A Design-Based Approach

• Activity Number: 510
• Type: Invited
• Date/Time: Wednesday, August 7, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Donald P. Green*+ and Holger L. Kern and Peter M. Aronow
• Companies: Columbia University and University of South Carolina and Yale University
• Keywords: experiments ; causal inference ; missing data ; Manski bounds

Abstract

Missing outcome data plague many randomized experiments. Common solutions such as selection models and multiple imputation rest on untestable assumptions that are not al- ways credible. In contrast, we propose a design-based approach for dealing with missing outcome data that makes minimal assumptions but still yields informative bounds on the average treatment effect. Our approach is based on a combination of the double sampling design (Neyman 1923; Hansen and Hurwitz 1946) and non-parametric worst case bounds (Birnbaum and Sirken 1950a,b; Manski 1990). We also derive the ex ante optimal exper- imental design when experimenters expect some fraction of outcome data to be missing. We illustrate the value of our approach through Monte Carlo simulations and the analysis of data from a randomized field experiment in Uganda.