Primary Submission Category: Proximal causal inference
Proximal causal inference for continuous point-exposure treatments
Authors: Antonio Olivas-Martinez, Andrea Rotnitzky,
Presenting Author: Antonio Olivas-Martinez*
The recently introduced proximal causal inference framework has revolutionized the identification and doubly-robust inference of causal effects, particularly in scenarios where confounding arises from an unobserved variable, yet observed proxy variables for the confounder are available. In such scenarios, estimating the parameter of interest requires estimating nuisance functions which are solutions to integral equations. While various machine learning methods exist for estimating these nuisance functions for binary treatments, limited attention has been given to the case of continuous point-exposure treatments. In this work, we extend the proximal causal inference framework to encompass continuous point-exposure treatments, employing a minimax optimization approach for estimating the nuisance functions. Our work fills a critical gap in the literature by demonstrating how finite sample properties of the estimator vary with the level of correlation between the proxy and unobserved confounders. To illustrate the method’s applicability, we apply it to the context of COVID-19 vaccines, focusing on identifying correlates of protection for use in immunobridging trials.