Domain Adaptation under Hidden Confounding

Reference codebase implementation of the Generative Invariance (GI) estimator from
"Domain Adaptation under Hidden Confounding" by Carlos García Meixide and David Ríos Insua.

---

Repository contents

This repository contains the GI estimator and the scripts to reproduce the main simulation studies and the real-data experiment discussed in the paper.

Files

• main.R
  Implements the Generative Invariance (GI) estimator.

• aux_gi.R
  Helper functions for efficient environment-combination selection aimed at optimizing asymptotic variance.

• generate_6_2.R
  Data-generating process used in the simulation setup of Section 6.2.

• sect6_4.R Reproduces the comparisons in Section 6.4 of the Generative Invariance paper. Compares GI, Causal Dantzig, and Instrumental Variables (IV) in terms of predictive accuracy in an unseen domain.

• sect6_6.R
  Reproduces the analysis of the SPRINT trial in Section 6.6.
  Requires access to the dataset available upon request at: https://biolincc.nhlbi.nih.gov/login/?next=/requests/type/sprint/

Method summary

Generative Invariance (GI) leverages cross-environment structure to enable domain adaptation under hidden confounding. The estimator produces predictors that transfer effectively to unseen domains.

For full details and theoretical results, see the paper.