Chad Brown and Kyle Butts
Working Paper
How to conduct inference in time-series settings where a first-step machine-learning estimator is 'plugged into' a second-step estimator without sample splitting.
We begin by outlining the semiparametric framework. We then present general results that enable inference on a finite-dimensional parameter after first-stage estimation of nuisance functions in time-series settings. Following Chen, Syrgkanis, and Austern (2022), we restrict attention to cases where the moment conditions are linear in the target parameter of interest. This includes common cases such as partially linear models, partially linear IV models, and impulse response models.