Testing Conditional Independence with Deep Neural Network Based Binary Expansion Testing (DeepBET)

Yang Yang, Kai Zhang, Ping-Shou ZhongAISTATS 2025 · OpenReview

Introduction

Conditional Independence (CI) testing—determining whether $X \perp \perp Y | Z$—is a cornerstone of statistical analysis and causal discovery. However, as the dimensionality of the conditioning set $Z$ increases, traditional methods often succumb to the "curse of dimensionality," leading to either inflated Type-I errors or a lack of power to detect non-linear relationships.

Presented at AISTATS 2025, DeepBET is a novel framework that addresses these issues by integrating Deep Neural Networks (DNNs) with Binary Expansion Testing (BET). This combination allows for a test that is computationally efficient, theoretically grounded, and uniquely interpretable.

Methodology

The DeepBET framework follows a three-stage pipeline designed to handle high-dimensional data while ensuring statistical validity.

The DeepBET pipeline: residualization via DNNs followed by Binary Expansion Testing (BET) to assess conditional dependence.
  1. Residualization via DNNs: The data is split into training and testing sets. DNNs are trained on the first split to estimate the conditional means $E[X|Z]$ and $E[Y|Z]$, effectively capturing the influence of high-dimensional confounders.
  2. Binary Expansion Testing (BET): Residuals are calculated on the second split. BET transforms these continuous residuals into binary "bits." By testing for correlations between these bits, the framework can detect complex, non-linear dependencies (e.g., U-shaped patterns) that standard correlation measures miss.
  3. Multiple Data Splitting: To stabilize the test and maximize power, DeepBET employs a multi-split aggregation procedure. This ensures that the results are robust to the specificities of any single random data split.
Empirical power comparison between single-split and multiple-split procedures, demonstrating the significant power gains achieved through aggregation in DeepBET.

Theoretical Breakthrough: Root-n Consistency

A major challenge in using DNNs for statistical testing is their slow convergence in high dimensions. DeepBET overcomes this through a theoretical breakthrough: the authors prove that the test statistic achieves a $\sqrt{n}$ convergence rate. The binary nature of the BET step acts as a filter, protecting the final inference from the "pollution" of high-dimensional estimation errors. This makes DeepBET as theoretically robust as linear tests while maintaining the flexibility of deep learning.

Key Results and Clinical Insights

DeepBET was rigorously tested against modern alternatives like GCM and DGCIT. In high-dimensional simulations ($d_Z = 500$), DeepBET significantly outperformed competitors in terms of power while strictly maintaining Type-I error control at 0.05. The method is also highly efficient; a single replication takes approximately 1.5 seconds, compared to several minutes for sampling-based methods like CCIT.

Beyond synthetic data, DeepBET was applied to a real-world clinical dataset for Dry Eye Disease (DED). The method's "BET plots" provided visual interpretability, revealing a non-linear relationship where patients with extreme corneal epithelial thickness (either very thin or very thick) exhibited lower tear production. This level of nuance is often lost in traditional linear or "black-box" high-dimensional models.

Conclusion

DeepBET provides a scalable and interpretable solution for conditional independence testing in the era of high-dimensional data. By bridging the gap between deep learning and classical statistical testing, it offers a robust tool for researchers across various domains, from clinical medicine to automated causal discovery.

Summary updated

Abstract

This paper focuses on testing conditional independence between two random variables ($X$ and $Y$) given a set of high-dimensional confounding variables ($Z$). The high dimensionality of these confounding variables presents a challenge, often resulting in inflated type-I errors or insufficient power in many existing tests. To address this issue, we leverage the power of Deep Neural Networks (DNNs) to handle complex, high-dimensional data while mitigating the curse of dimensionality. We propose a novel test procedure, DeepBET. First, a DNN is used on part of the data to estimate the conditional means of $X$ and $Y$ given $Z$. Then, binary expansion testing (BET) are applied to the predicted errors from the remaining data. Additionally, we implement a multiple-split procedure to further enhance the power of the test. DeepBET is computationally efficient and robust to the tuning parameters in DNNs. Interestingly, the DeepBET statistic converges at a root-$n$ rate despite the nonparametric and high-dimensional nature of the confounding effects. Our numerical results demonstrate that the proposed method controls type-I error under various scenarios and enhances both power and interpretability for conditional dependence when present, making it a robust alternative for testing conditional independence in high-dimensional settings. When applied to dry eye disease data, DeepBET reveals meaningful nonlinear relationships between the epithelial thickness and the tear production in the central region of eyes, given other regions.
BibTeX
@inproceedings{yang2025testing,
  author = {Yang Yang and Kai Zhang and Ping-Shou Zhong},
  title = {{Testing Conditional Independence with Deep Neural Network Based Binary Expansion Testing (DeepBET)}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2025},
  url = {https://openreview.net/forum?id=5oZC5NeWsI},
}
Figure