Bounding the Effects of Social Experiments: Accounting for Attrition in Administrative Data

Social experiments frequently exploit data from administrative records. However, most administrative data systems are state-specific, designed to track earnings or benefit payments among residents within a single state. Once an experimental participant moves out of state, his earnings and benefits in his state of origin consist entirely of zeros, giving rise to a form of attrition. In the presence of such attrition, the average treatment effect of the experiment is no longer point-identified, despite random assignment. I propose a method to estimate such attrition and, for binary outcomes such as employment, to construct bounds on the average treatment effect. Results from a welfare-reform experiment considered to have sizeable effects appear quite ambiguous after accounting for attrition. The results have important implications for planning social experiments.