Confidence intervals of treatment effects in panel data models with interactive fixed effects

We augment the factor-based estimation of treatment effects proposed by Bai and Ng (2021) with easy-to-implement and nonparametric confidence intervals of treatment effects on every treated unit at every post-treatment time. The construction of confidence intervals entails a residual-based bootstrap resampling procedure. This method does not rely on any parametric assumption on the distribution of idiosyncratic errors and it is robust to weak cross-sectional and time-series dependence among idiosyncratic errors. We prove the asymptotic validity of the proposed confidence intervals as the numbers of control units and pre-treatment times go to infinity. We also extend this method to cases where the common factors and covariates (if any) are unit root processes. Monte Carlo experiments show that the proposed confidence intervals are well-behaved in finite samples and outperform confidence intervals based on normal quantiles. Empirical applications with two classical datasets add informative confidence intervals to existing point estimates of treatment effects.