Abstract
To address the bias of the canonical two-way fixed effects estimator fordifference-in-differences under staggered adoptions, Wooldridge (2021) proposedthe extended two-way fixed effects estimator, which adds many parameters.However, this reduces efficiency. Restricting some of these parameters to beequal (for example, subsequent treatment effects within a cohort) helps, but adhoc restrictions may reintroduce bias. We propose a machine learning estimatorwith a single tuning parameter, fused extended two-way fixed effects (FETWFE),that enables automatic data-driven selection of these restrictions. We provethat under an appropriate sparsity assumption FETWFE identifies the correctrestrictions with probability tending to one, which improves efficiency. Wealso prove the consistency, oracle property, and asymptotic normality of FETWFEfor several classes of heterogeneous marginal treatment effect estimators undereither conditional or marginal parallel trends, and we prove the same resultsfor conditional average treatment effects under conditional parallel trends. Wedemonstrate FETWFE in simulation studies and an empirical application.