Efficient Difference-in-Differences Estimation with High-Dimensional Common Trend Confounding

Understand

This study considers various semiparametric difference-in-differences models under different assumptions on the relation between the treatment group identifier, time and covariates for cross-sectional and panel data.

• The variance lower bound is shown to be sensitive to the model assumptions imposed implying a robustness-efficiency trade-off.
• The obtained efficient influence functions lead to estimators that are rate double robust and have desirable asymptotic properties under weak first stage convergence conditions.
• This enables to use sophisticated machine-learning algorithms that can cope with settings where common trend confounding is high-dimensional.