Reproduce “Outcome-Aware Spectral Feature Learning for Instrumental Variable Regression” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Dimitri Meunier, Jakub Wornbard, Vladimir R. Kostic, Antoine Moulin, Alek Fröhlich, Karim Lounici, Massimiliano Pontil, Arthur Gretton

AugSpecIV extends spectral feature learning for instrumental-variable regression by adding outcome information while preserving the treatment–instrument operator structure. It learns compact features for downstream two-stage estimation. On a difficult dSprites ellipse-orientation task, it matches or slightly exceeds DFIV, while standard SpecIV performs poorly; a small positive δ reduces error by 20% on an easier heart-image task, subject to tuning and representation conditions.

Key insights

  • Spectral misalignment is a critical limitation where outcome-agnostic features fail to capture essential aspects of the causal function when it resides in low singular value regions of the instrument-treatment operator.
  • By augmenting the operator with outcome information, ASFL effectively realigns the dominant singular subspaces of the modified operator (Tδ) with the relevant components of the causal function.
  • The theoretical framework provides non-asymptotic, high-probability error bounds for the ASFL estimator, demonstrating its ability to control approximation and statistical errors through perturbation analysis.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.