Primary Submission Category: Graphical Models
Recursive proximal identification when common causes or mediators are unobserved
Authors: Helen Guo, Ilya Shpitser,
Presenting Author: Beatrix Wen*
Nonparametric identifiability of the causal effect in the presence of hidden variables is characterized by the ID algorithm. In cases where the causal effect is not identified nonparametrically, prior work has augmented the ID algorithm via proximal causal learning methods, allowing some unobserved confounders to be handled via proxies, and others via the standard machinery of the ID algorithm based on the fixing operator.
We present a novel generalization of the ID algorithm that allows identification if either confounders or mediators are unobserved, provided informative proxies for these unobserved variables are available. Our generalization is based on a novel reformulation of the ID algorithm via a fixing operator that resembles sequential applications of either the backdoor or the front-door criterion.