# Conditional Randomization Rank Test

**Authors:** Yanjie Zhong, Todd Kuffner, Soumendra Lahiri  
**arXiv:** [2112.00258](https://arxiv.org/abs/2112.00258)  
**Date:** 2021-12-01  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2112.00258v1/

## Key points

- Authors Yanjie Zhong, Todd Kuffner, and Soumendra Lahiri introduce the Conditional Randomization Rank Test (CRRT) for assessing conditional independence between Y and X given Z.
- CRRT builds upon the Conditional Randomization Test (CRT) and leverages knowledge of the conditional distribution of X|Z.
- The approach is grounded in conditional sampling techniques, making it easy to implement and interpret.
- CRRT ensures precise control over type 1 errors and surpasses CRT in computational efficiency.
- It offers a more flexible framework with superior performance in power and computational efficiency compared to existing methods.
- Bounds on type 1 error probabilities are established using total variation norm and observed Kullback-Leibler divergence metrics.
- Extensive simulations validate the advantages of CRRT over alternative conditional sampling-based methods.

## Summary

In their paper titled "Conditional Randomization Rank Test," authors Yanjie Zhong, Todd Kuffner, and Soumendra Lahiri introduce a novel method known as the Conditional Randomization Rank Test (CRRT) for assessing the conditional independence between a response variable Y and a covariate variable X, given the remaining covariates Z. The CRRT builds upon the foundation laid by the Conditional Randomization Test (CRT) proposed in a previous study by [CFJL18] and leverages knowledge of the conditional distribution of X|Z. This approach is grounded in conditional sampling techniques, making it straightforward to implement and interpret. One key advantage of the CRRT is its ability to ensure precise control over type 1 errors while also surpassing the CRT in terms of computational efficiency. By offering a more flexible framework, the CRRT demonstrates superior performance compared to existing methods when considering both power and computational efficiency. The authors establish bounds on type 1 error probabilities using total variation norm and observed Kullback-Leibler divergence metrics. This is particularly useful when there are discrepancies in the specified conditional distribution of X|Z. To validate their theoretical findings, extensive simulations were conducted which confirmed that the CRRT exhibits significant advantages over alternative conditional sampling-based methods. Overall, this innovative approach presents a promising solution for testing conditional independence in statistical analysis contexts.