Reproduce “On the Hardness of Conditional Independence Testing In Practice” 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

Zheng He, Roman Pogodin, Yazhe Li, Namrata Deka, Arthur Gretton, Danica J. Sutherland

KCI tests whether two variables remain conditionally independent after accounting for a continuous context by estimating conditional means, forming residuals, and using a kernel to localize dependence across contexts. Power-focused kernel selection can detect dependence hidden by pooled covariance, but poor conditional-mean estimates can inflate Type-I error. With true means, bounded kernels yield finite-sample validity and consistency against fixed alternatives.

Key insights

  • The primary practical challenge in kernel-based CI testing stems from errors in estimating conditional mean embeddings, rather than inherent issues with the kernel framework itself.
  • The choice of the conditioning kernel's lengthscale is critical for detecting localized conditional dependencies and achieving adequate test power, independently of its use for regression.
  • A fundamental trade-off exists: optimizing the conditioning kernel for power can substantially inflate Type-I error rates when conditional mean embeddings are poorly estimated, as it amplifies spurious correlations from regression noise.

Summary by alphaXiv.

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