Predictions from complex machine learning models are increasingly used as input data in subsequent statistical analyses, yet their errors can bias estimators and lead to invalid confidence intervals. Existing approaches attempt to remedy this issue but often impose strong assumptions or lack generality. As an alternative, we present the Predict-Then-Debias bootstrap, developed in Kluger et al. (2025). The method yields valid confidence intervals provided that a small complete sample from the population of interest is available. Its bootstrap construction applies to a broad class of estimators and can be modified to account for weighted, stratified, or clustered samples.
Kluger et al. (Fri,) studied this question.