Denoised Conformal Alignment Boosts CATE Selection Reliability

In personalized medicine and policy, practitioners often deploy models only on a subset of individuals—those with the most reliable treatment effect predictions. But standard conformal guarantees can break when selection is based on predicted CATE. The new Denoised Conformal Alignment method tackles this by controlling the false discovery rate (FDR) among selected individuals, ensuring that errors stay below a tolerance. It uses doubly robust proxy errors (derived from pseudo-outcomes) but addresses a key flaw: under heteroskedasticity, naive proxy errors lose power because variance drowns out the signal. The solution subtracts an estimated conditional variance component to denoise the proxies, then applies conformal calibration followed by Benjamini-Hochberg multiple testing correction. The authors prove validity hinges on stability of label consistency, not perfect variance estimation.

Results across synthetic and semi-synthetic datasets show the method consistently outperforms baselines. It achieves higher selection power—meaning more true positives among chosen individuals—while strictly controlling FDR at the desired level. This holds even in high-heteroskedasticity regimes where naive methods fail. The paper provides rigorous theoretical guarantees and practical calibration procedures. For data scientists working with causal inference or A/B testing, Denoised Conformal Alignment offers a principled way to trust your model's decisions, ensuring you only act on predictions when the error risk is acceptably low.