Reproduce “Spectral Representation for Causal Estimation with Hidden Confounders” 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

Haotian Sun, Antoine Moulin, Tongzheng Ren, Arthur Gretton, Bo Dai

The authors propose SpecIV, which learns compact, low-rank representations of conditional relationships before estimating causal functions with a saddle-point objective. Neural feature maps learn density-ratio factors from paired and mismatched observations, while a dual function tests residuals against instruments or proxies, avoiding double-sampling bias. On synthetic Demand Design data, SpecIV reports OOS MSE 1.67 versus DFIV’s 2.40, in units of 10^4; this result relies on low-rank assumptions and does not establish real-world validity.

Key insights

  • The low-rank factorization of conditional densities provides a principled way to construct realizable and closed function classes for conditional moment estimation.
  • Adaptive feature learning, driven by contrastive objectives, enhances the model's flexibility and ability to handle complex, high-dimensional causal relationships.
  • Saddle-point optimization effectively transforms the estimation problem into one that is robust to biases common in two-stage methods and compatible with neural network training.

Summary by alphaXiv.

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