Conditional Randomization

Conditional randomization tests (CRTs) are statistical methods used to assess conditional independence between variables, offering a powerful tool for causal inference and feature selection, particularly when randomized controlled trials are infeasible. Current research focuses on improving CRT efficiency and power through techniques like nearest-neighbor sampling, novel loss functions for model training, and incorporating methods from differential privacy to handle sensitive data. These advancements enhance the reliability and applicability of CRTs across diverse fields, from causal discovery in complex systems to high-dimensional variable selection in machine learning.