The Proximal ID Algorithm
Ilya Shpitser, Zach Wood-Doughty, Eric J. Tchetgen Tchetgen; 24(188):1−46, 2023.
Abstract
Unobserved confounding poses a significant challenge in establishing valid causal conclusions from observational data. To address this challenge, two types of approaches have emerged: using external aids like instrumental variables or proxies, and employing the ID algorithm, which utilizes Markov restrictions on graphical causal models to identify causal relationships. In this paper, we propose a synthesis of these approaches to develop the most comprehensive identification algorithm known as the proximal ID algorithm. Our method not only enables nonparametric identification in all cases where the ID algorithm succeeds, but also allows for systematic utilization of proxies to adjust for unobserved confounders that would otherwise hinder identification. Furthermore, we present a set of estimation strategies for causal parameters identified by our method in a specific case. We validate our approach through simulation studies and a real-world data application.
[abs]