Panel data applications often use instrumental variables (IV) to address endogeneity, but when instrument validity requires conditioning on high-dimensional covariates, flexible adjustment for confounding is essential and standard estimators like two-stage least squares (2SLS) break down. This paper proposes a novel Double Machine Learning (DML) estimator for static panel data with instrumental variables which accommodates unobserved individual heterogeneity, endogenous treatment assignment, and flexible nuisance components. Our estimator extends the DML framework to the static panel setting using first-difference transformation and Neyman orthogonal score functions to ensure consistent estimation and valid inference. We also derive robust tests for weak instruments, including panel analogues of the first-stage F-statistic and the Anderson-Rubin (AR) Wald test, tools not previously available in DML settings. We apply the method to three prominent studies on immigration and political preferences using shift-share instruments, finding a strong causal effect in one case and weak instrument concerns that cast doubt on their causal conclusions in the other two. Monte Carlo simulations confirm that the panel IV DML estimator outperforms 2SLS in both bias and root mean squared errors when the instrument is strong, and provides more conservative inference when instruments are weak. The method broadens the empirical toolkit for researchers working with endogenous treatments in panel data, especially when covariate complexity precludes standard approaches
Presented by:
Annalivia Polselli
Date & time:
October 8, 2025 12:30 pm - October 8, 2025 1:30 pm
Venue:
2N2.4.16
Internal seminars home