Primary Submission Category: Machine Learning and Causal Inference
Causal estimands and estimators for evaluating the impact of preventive interventions on long term outcomes
Authors: David Benkeser, Razieh Nabi, Elizabeth Rogawski-McQuade, Allison Codi, Mats Stensrud,
Presenting Author: Allison Codi*
Establishing the long-term effects of interventions aimed at preventing intermediate outcomes poses significant challenges. For example, vaccines designed to prevent diarrhea caused by Shigella bacteria in children may also positively impact long-term growth, as Shigella-induced diarrhea is a known cause of growth faltering. However, given the relatively low frequency of Shigella-related diarrhea, the vaccine’s marginal causal effect on growth may be too small to detect in a typical randomized controlled trial. Nevertheless, clinicians and policymakers are highly interested in demonstrating the broader benefits of vaccination on growth outcomes.
To address this challenge, we propose alternative causal estimands that enjoy improved power for detecting effects on long-term outcomes in realistic trial settings. Specifically, we introduce estimands based on principal stratification and interventional causal frameworks and demonstrate that both approaches yield the same identifying functional under different assumptions. Notably, the principal stratification approach relies on cross-world independence assumptions, whereas the interventional estimand does not.
We further derive nonparametric efficient, and doubly robust estimators for these estimands, leveraging machine learning techniques for nuisance parameter estimation. Through realistic simulations, we illustrate that these estimators can provide clinically meaningful inferences even within the constraints of practical Shi