Statistical Inference for Linear Mediation Models with High-dimensional Mediators and Application to Studying Stock Reaction to COVID-19 Pandemic

Xu Guo, Runze Li, Jingyuan Liu, Mudong Zeng

Abstract

Mediation analysis draws increasing attention in many scientific areas such as genomics, epidemiology and finance. In this paper, we propose new statistical inference procedures for high dimensional mediation models, in which both the outcome model and the mediator model are linear with high dimensional mediators. Traditional procedures for mediation analysis cannot be used to make statistical inference for high dimensional linear mediation models due to high-dimensionality of the mediators. We propose an estimation procedure for the indirect effects of the models via a partial penalized least squares method, and further establish its theoretical properties. We further develop a partial penalized Wald test on the indirect effects, and prove that the proposed test has a $\chi^2$ limiting null distribution. We also propose an $F$-type test for direct effects and show that the proposed test asymptotically follows a $\chi^2$-distribution under null hypothesis and a noncentral $\chi^2$-distribution under local alternatives. Monte Carlo simulations are conducted to examine the finite sample performance of the proposed tests and compare their performance with existing ones. We further apply the newly proposed statistical inference procedures to study stock reaction to COVID-19 pandemic via an empirical analysis of studying the mediation effects of financial metrics that bridge company's sector and stock return.

Introduction

Mediation analysis is a fundamental statistical tool used to uncover the mechanisms through which an exposure variable ($X$) influences an outcome ($Y$) via intermediate variables known as mediators ($M$). In the modern "Big Data" era, researchers frequently encounter scenarios where the number of potential mediators is vast—often exceeding the number of available observations ($p > n$). For instance, in financial econometrics, one might want to understand how a company’s industry sector ($X$) influences its stock return ($Y$) through hundreds of different financial metrics ($M$).

Traditional methods, such as Structural Equation Modeling (SEM), are ill-suited for these high-dimensional settings because they rely on inverting covariance matrices that become rank-deficient and unstable when $p > n$. While recent "debiased" Lasso techniques have attempted to address this, they often suffer from significant computational overhead and "variance inflation," which can obscure real effects. This paper introduces a novel High-Dimensional Mediation Model (HDMM) framework designed to provide fast, accurate, and theoretically grounded inference for these complex systems.

Methodology

The Partial Penalized Least Squares (PPLS) Approach

The core innovation of this research lies in the Partial Penalized Least Squares (PPLS) method. In high-dimensional statistics, researchers typically use penalties (like Lasso or SCAD) to enforce "sparsity," assuming that only a few variables have a non-zero effect. However, the authors take a more surgical approach: they apply the penalty only to the high-dimensional mediators while leaving the low-dimensional direct effect unpenalized.

By avoiding the penalization of the direct effect, the model bypasses the need for the computationally expensive "debiasing" steps required by other methods. This prevents variance inflation and ensures that the statistical power of the test remains high, making it easier to detect genuine relationships between variables.

The "Difference" Estimation Strategy

Rather than attempting to estimate a massive, complex matrix of all possible relationships between mediators and exposure, the authors utilize a clever mathematical identity: the indirect effect is the difference between the total effect and the direct effect. By estimating the total effect via standard least squares and the direct effect through PPLS, the framework simplifies a high-dimensional problem into a manageable calculation.

New Statistical Tests

To facilitate rigorous inference, the paper introduces two new tests:

  1. Wald Test: Used for the indirect effect.
  2. F-type Test: Used for the direct effect.

Both tests are shown to follow well-understood $\chi^2$ distributions, allowing researchers to calculate p-values and determine statistical significance with confidence.

Empirical Results

Simulation Performance

Through Monte Carlo simulations, the authors compared their PPLS method against existing state-of-the-art techniques. The results demonstrated that the PPLS-based Wald test achieved significantly higher statistical power and lower bias. Notably, the algorithm proved to be approximately 1,000 times faster than competing debiased Lasso methods, making it viable for extremely large datasets.

Case Study: COVID-19 and the Stock Market

The researchers applied their framework to analyze the stock returns of 490 S&P 500 companies during the initial COVID-19 crash (February to April 2020). They examined 550 financial metrics—including growth rates, liquidity ratios, and profitability markers—to see which ones mediated the relationship between a company's sector and its market performance.

Key Findings: * Sector Impact: The Healthcare and Technology sectors exhibited the strongest positive direct effects on returns during the pandemic. * Identified Mediators: Out of 550 potential variables, the model identified six critical mediators that explained the "how" of sector performance: * Profitability: Return on Assets (ROA) and Gross Margin. * Efficiency: Operating Income Growth. * Liquidity: Quick Ratio Growth. * Solvency & Collection: Debt to Assets and Receivables Turnover (both of which had a negative impact on returns).

The study found that both the direct effects of the sectors and the indirect effects through these specific operational metrics were highly significant ($p < 10^{-9}$).

Conclusion

This paper provides a robust and efficient framework for high-dimensional mediation analysis. By treating mediators as random variables and utilizing a partial penalization strategy, the authors have developed a tool that is both mathematically rigorous and computationally superior to existing methods. While the current model assumes linear relationships and relies on the sparsity of mediators, it sets a new standard for identifying the underlying mechanisms in complex, high-dimensional data across fields as diverse as finance, genomics, and social science.

Summary updated

Abstract

Mediation analysis draws increasing attention in many scientific areas such as genomics, epidemiology and finance. In this paper, we propose new statistical inference procedures for high dimensional mediation models, in which both the outcome model and the mediator model are linear with high dimensional mediators. Traditional procedures for mediation analysis cannot be used to make statistical inference for high dimensional linear mediation models due to high-dimensionality of the mediators. We propose an estimation procedure for the indirect effects of the models via a partial penalized least squares method, and further establish its theoretical properties. We further develop a partial penalized Wald test on the indirect effects, and prove that the proposed test has a $χ^2$ limiting null distribution. We also propose an $F$-type test for direct effects and show that the proposed test asymptotically follows a $χ^2$-distribution under null hypothesis and a noncentral $χ^2$-distribution under local alternatives. Monte Carlo simulations are conducted to examine the finite sample performance of the proposed tests and compare their performance with existing ones. We further apply the newly proposed statistical inference procedures to study stock reaction to COVID-19 pandemic via an empirical analysis of studying the mediation effects of financial metrics that bridge company's sector and stock return.
BibTeX
@article{guo2022statistical,
  author = {Xu Guo and Runze Li and Jingyuan Liu and Mudong Zeng},
  title = {{Statistical Inference for Linear Mediation Models with High-dimensional Mediators and Application to Studying Stock Reaction to COVID-19 Pandemic}},
  journal = {Journal of Econometrics},
  year = {2022},
  doi = {10.1016/j.jeconom.2022.03.001},
}
Figure