A1648
Title: Machine learning and difference-in-differences: A practitioners dilemma
Authors: Paolo Libenzio Brignoli - ETH Zurich (Switzerland) [presenting]
Andrea Naghi - QMUL (United Kingdom)
Alessandro Varacca - Universita cattolica del sacro cuore (Italy)
Abstract: Recent econometric advances in difference-in-differences (DiD) have developed robust tools for staggered treatment adoption and expanded machine learning (ML) methods for selection on observables. Yet, an explicit methodological bridge connecting these advances is missing. A double/debiased machine learning staggered DiD estimator is introduced, showing that the statistical properties enabling ML in standard DiD naturally extend to staggered policy adoption. By formally connecting orthogonalized estimation in two-period settings with modern staggered designs, this approach enables valid, flexible estimation in complex, high-dimensional contexts. Simulation evidence demonstrates substantial improvements in empirically relevant settings with nonlinear relationships and high-dimensional sparse covariates. An application to evaluate the European Union Agri-Environmental Schemes (2014-2020) accounts for both enrollment timing and complex farm characteristics, finding that most farmers maintain previous practices after receiving environmental payments. This suggests significant windfall effects in current policy implementation, offering direct implications for future agricultural policy design.
