Coverage guarantee for feasible bias-aware confidence intervals with estimated bias and variance

Establish a formal asymptotic coverage guarantee for the feasible bias-aware confidence interval constructed for the parameter \theta using the sigmoid-based debiased machine learning estimator, where both the worst-case bias and the variance are estimated from data rather than known, including conditions under which the interval attains nominal coverage.

Background

The paper proposes bias-aware confidence intervals following Armstrong and Kolesar (2020), adapting them to the smoothing-based debiased machine learning setting. The interval uses an estimated worst-case bias bound and an estimated variance tied to the behavior of the CATE function.

While the theoretical framework provides asymptotic distributions and suggests practical estimation of tuning parameters and moments, the authors explicitly note that a formal coverage guarantee for the implementable (feasible) procedure—using estimated bias and variance—has not yet been established.