Primary Submission Category: Causal inference with continuous exposures
A new causal estimand for continuous exposures
Authors: Anand Hemmady, Marco Carone, Andrea Rotnitzky,
Presenting Author: Anand Hemmady*
To study the causal effect of a continuous point-exposure on an outcome of interest, investigators often contrast mean counterfactual outcomes under different exposure patterns. Most commonly, exposure patterns under which all participants are assigned the same exposure value are considered. While it results in intuitively interpretable estimands, such an approach has been critiqued for various reasons, including that it can lead to exposure patterns that would be impossible in the real world, which violates positivity. This has led to the development of alternative frameworks. In this work, motivated by recent work on modified treatment policies and stochastic interventions, we define a novel exposure pattern by specifying the cumulative odds ratio between the new and factual exposure patterns. The proposed exposure pattern has a desirable interpretation while avoiding violations of positivity. We derive an identification for the causal parameter resulting from the proposed exposure pattern and propose a debiased machine learning approach for inference. We also consider the identification and inference of the parameter under commonly encountered complications, including censoring and two-phase sampling, and use the resulting methodology to analyze data from COVID-19 vaccine efficacy trials.