Testing for Conditional Mean Independence with Covariates through Martingale Difference Divergence

Ze Jin, Xiaohan Yan, David S. MattesonUAI 2018

Introduction

In regression analysis, determining whether a set of variables $X$ contributes additional information to the prediction of $Y$ beyond what is already provided by covariates $Z$ is a fundamental challenge. This is formally known as testing for conditional mean independence: $E(Y|X, Z) = E(Y|Z)$. While various tests for conditional independence exist, many are either computationally intensive or lack power in high-dimensional settings. This paper introduces LinMDD, a novel multivariate test that leverages the Martingale Difference Divergence (MDD) framework to address these limitations, particularly when the relationship between $Y$ and $Z$ can be approximated linearly.

Methodology

The LinMDD approach simplifies the testing problem through a transformation-based strategy:

  1. Linear Decomposition: It assumes the conditional expectation $E(Y|Z)$ is a linear function of $Z$ ($E(Y|Z) = BZ$).
  2. Residual Estimation: Residuals $V = Y - BZ$ are estimated using Ordinary Least Squares (OLS) or Penalized Least Squares (e.g., Lasso) to obtain $\hat{V}$.
  3. MDD Application: The test applies Martingale Difference Divergence to the pair $(\hat{V}, (X, Z))$. MDD is specifically designed to detect if the mean of one variable depends on another, making it sensitive to non-linear relationships between $V$ and $X$.
  4. Permutation Inference: P-values are calculated via a permutation approach, shuffling $X$ values to approximate the null distribution.

A key theoretical contribution is the proof that the estimation error from using $\hat{V}$ instead of the true $V$ is negligible ($O_p(n^{-3/2})$), ensuring the test's asymptotic validity.

Simulation Results

The authors evaluated LinMDD across several simulated models with varying degrees of linearity. In scenarios where the relationship between $Y$ and $Z$ was linear (Models 1 and 2), LinMDD consistently demonstrated higher statistical power compared to benchmarks like partial distance covariance (pdCov) and pMDD.

Figure 2: Power comparison for Model 2 (c=1/2), where LinMDD (red) exhibits superior power as sample size $n$ increases.

Furthermore, the test maintains appropriate size control, ensuring that the Type I error rate remains close to the nominal level under the null hypothesis.

Figure 5: Size (Type I error) comparison for Model 1 (c=0). All methods, including LinMDD, maintain a stable size around the 0.10 level.

Empirical Application: Fama-French Factors

The utility of LinMDD was validated using 53 years of financial data (1964–2016) for Boeing (BA) stock returns and the Fama-French five factors.

  • CAPM vs. 3-Factor Model: The test rejected the null hypothesis ($p=0.072$ at $\alpha=0.1$), indicating that "Size" (SMB) and "Value" (HML) factors provide significant information beyond the market risk factor.
  • Factor Redundancy: In the 5-factor model context, LinMDD failed to reject the null for the "Value" factor ($p=0.218$). This aligns with Fama and French’s (2015) finding that HML becomes redundant once Profitability (RMW) and Investment (CMA) factors are included.

Conclusion

LinMDD provides a powerful and computationally efficient tool for testing conditional mean independence. By incorporating a linear assumption for the known covariates $Z$, it achieves significant power gains over purely non-parametric methods while remaining robust to complex, non-linear dependencies on the variables of interest $X$. Future work may extend this framework to non-parametric covariate effects using splines or local regression.

Summary updated

BibTeX
@inproceedings{jin2018testing,
  author = {Ze Jin and Xiaohan Yan and David S. Matteson},
  title = {{Testing for Conditional Mean Independence with Covariates through Martingale Difference Divergence}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2018},
}
Figure