Abstract: We propose a new algorithm for estimating treatment effects in contexts where the exogenous variation comes from aggregate time-series shocks. Our estimator combines data-driven unit-level weights with a time-series model. We use the unit weights to control for unobserved aggregate confounders and use the time-series model to extract the quasi-random variation from the observed shock. We examine our algorithm's performance in a simulation based on Nakamura and Steinsson [2014]. We provide statistical guarantees for our estimator in a practically relevant regime, where both cross-sectional and time-series dimensions are large, and we show how to use our method to conduct inference.
Link to work
Presentation slides