A1386
Title: Semiparametric inference in panel data with interactive and high-dimensional confounding
Authors: Yukun Ma - University of Rochester (United States) [presenting]
Abstract: Factor-augmented double machine learning (FA-DML) is developed as a semiparametric framework for inference on low-dimensional causal parameters in panel data settings with interactive and high-dimensional confounding. Existing double/debiased machine learning methods for panel data typically rely on additive fixed effects and transformations such as demeaning or differencing, which fail when unobserved heterogeneity follows an interactive fixed effects structure. To address this limitation, orthogonalized machine learning is combined with principal-components-based factor estimation in a cross-fitted iterative procedure. In each fold, flexible learners first estimate the conditional expectations of the outcome and treatment given observed covariates; residualized outcomes are then used to recover latent factors and loadings, and these nuisance and factor estimates are iteratively updated until convergence. The resulting estimator delivers valid root-NT inference under joint large-N, large-T asymptotics, provided both the nuisance estimation error and factor estimation error are asymptotically negligible under stated rate conditions. Consistency of the variance estimator is also established. Monte Carlo simulations show that FA-DML attains near-nominal coverage in the presence of interactive confounding, whereas standard DML procedures that ignore latent factors exhibit severe undercoverage.
