Primary Submission Category: Difference in Differences, Synthetic Control, Methods for Panel and Longitudinal Data
Difference-in-Differences with Non-Ignorable Attrition
Authors: Javier Viviens,
Presenting Author: Javier Viviens*
Unbalanced panels are frequently used in Difference-in-Differences (DiD) applications. In this paper, I employ principal stratification analysis to highlight the potential drawbacks of the DiD research design when the outcome is missing for some units. Specifically, the conventional ATT estimand may not be well defined, and the DiD estimand cannot be interpreted causally without additional assumptions. To address these issues, I develop an identification strategy to partially identify causal effects on the set of units for which the outcome is observed and well-defined under both treatment and control. I adapt Lee bounds to the DiD setting, replacing the unconfoundedness assumption in the original trimming strategy proposed by Lee (2009) with a principal parallel trend assumption. I also explore how to leverage multiple sources of attrition to relax the monotonicity assumption, thereby allowing the four latent strata to exist, which may be of independent interest. Alongside the identification results, I present estimators and their asymptotic distributions. I illustrate the relevance of the proposed methodology by analyzing a job training program in Colombia.