A0851
Title: Nuisance parameter tuning for inference in observational studies
Authors: Rajarshi Mukherjee - Harvard T.H. Chan School of Public Health (United States) [presenting]
Abstract: The purpose is to discuss the issue of nuisance parameter tuning for estimating quantities in observational studies, such as the average treatment effect and measures of conditional dependence. Typical methods of estimating such quantities of interest rely on estimating nuisance functions often through the lens of nonparametric and/or high-dimensional machine learning methods. Whereas many popular ideas pertain to tuning these nuisance function estimation from a prediction perspective and subsequently perform downstream bias correction for valid inference of low dimensional summaries of interest in the observational studies of interest, cases are explored to show that there exists a delicate interplay between nuisance function estimation strategies, type of estimators that uses these nuisance functions in its pipeline of estimation of the final object of interest, and sample splitting strategies that are now popular to allow flexible methods of nuisance function estimation without jeopardizing the standard errors of estimators of the downstream objects of interest. The above is explored through the lens of specific functionals that arise in the context of causal inference, and both are studied in nonparametric and high-dimensional regimes.
