Testing Attrition Bias in Field Experiments

We approach attrition in field experiments with baseline outcome data as an identification problem in a panel model. A systematic review of the literature indicates that there is no consensus on how to test for attrition bias. We establish identifying assumptions for treatment effects for both the respondent subpopulation and the study population. We then derive their sharp testable implications on the baseline outcome distribution and propose randomization procedures to test them. We demonstrate that the most commonly used test does not control size in general when internal validity holds. Simulations and applications illustrate the empirical relevance of our analysis.