Instrumental Variable Representation Learning under Confounded Covariates

Instrumental variable (IV) analysis is a crucial tool for causal inference across diverse domains—from genetics to chemistry—in the presence of unobserved confounders, but discovering true IVs from observed covariates is challenging. Recent approaches have focused on synthesizing representations that can serve as IVs, but under restrictive assumptions and settings. We propose CoCoIV to tackle a more challenging yet realistic problem of learning IV representations from observed covariates, potentially correlated with unobserved confounders. CoCoIV utilizes latent variable models to learn representations for both IVs and non-IVs from confounded covariates, guided by a dual prediction network with mutual information regularization, allowing both discrete and continuous treatments. Extensive experiments across various configurations of estimators and treatment types show the effectiveness and wide applicability of our framework.