Unobserved confounding presents a major threat to the validity of causal inference from observational studies. In this paper, we introduce a novel framework that leverages the information in multiple parallel outcomes for identification and estimation of causal effects. Under a conditional independence structure among multiple parallel outcomes, we achieve nonparametric identification with at least three parallel outcomes. We further show that under a set of linear structural equation models, causal inference is possible with two parallel outcomes. We develop accompanying estimating procedures and evaluate their finite sample performance through simulation studies and a data application studying the causal effect of the tau protein level on various types of behavioral deficits.
Download