TL;DR
This paper introduces a novel, computationally efficient method for measuring conditional dependence in high-dimensional data, particularly useful for building graphical models. Their "projection-based conditional dependence" (P-DCov) test effectively identifies relationships even when the number of underlying factors is large and unknown, outperforming existing methods in various scenarios.
What Problem Does This Paper Address?
Understanding how variables depend on each other, especially when conditioned on other factors, is crucial in many fields like econometrics, finance, and bioinformatics. In high-dimensional settings (where the number of variables is large), traditional methods for assessing conditional independence become computationally prohibitive or rely on restrictive assumptions (like Gaussianity). Building dependency graphs, which visualize these relationships, is a key application. The challenge is to develop a method that is both statistically sound and computationally feasible for high-dimensional data, can handle non-Gaussian distributions, and can identify dependencies driven by a large number of underlying latent factors.
Methods
The core of the paper is a two-step conditional independence test.
- Projection: They first project the observed variables (x and y) onto a space orthogonal to the space spanned by the latent factors (f). This step effectively removes the influence of the factors, allowing for a cleaner assessment of the direct conditional dependence between x and y. They consider both linear and additive models for this projection.
- Distance Covariance: After obtaining the projected residuals, they use distance covariance (a measure that captures non-linear dependencies and doesn't assume Gaussianity) to assess the remaining dependence between the projected variables.
This approach is implemented as the "Projection-based Conditional Dependence" (P-DCov) test. For building graphical models, they propose an edge-by-edge testing procedure and incorporate False Discovery Rate (FDR) control for multiple testing.
Datasets
The paper uses simulated datasets for Monte Carlo experiments and a real-world financial dataset:
- Simulated Data: Used to evaluate the performance of the P-DCov test in various scenarios, including high-dimensional factor models, Gaussian graphical models, general graphical models (with t-distributions), and dependency graphs with external factors.
- Financial Data: Daily excess returns of 90 stocks from the S&P 100 index over one year (August 2004 - August 2005) were used to study stock dependencies conditional on the Fama-French three-factor model.
Key Insights and Ideas
- Decoupling Dependence: The projection step is key to isolating the conditional dependence between x and y by effectively "removing" the influence of the common factors.
- Distance Covariance Power: Using distance covariance is a significant advantage as it captures non-linear relationships and doesn't require the data to be Gaussian, making the method more robust.
- Handling High-Dimensional Factors: The method is designed to work even when the number of latent factors (K) is large and can grow with the sample size, which is a major limitation of many existing techniques.
- Graphical Model Construction: The P-DCov test provides a solid foundation for building conditional dependency graphs, even in complex, high-dimensional settings.
- FDR Control for Robustness: Incorporating FDR control in graph construction makes the results more reliable when performing many tests simultaneously.
Key Results
- Simulation Studies:
- The P-DCov test demonstrated good control over Type I errors in simulations.
- It achieved higher empirical power than the oracle test (which uses true residuals) when sample sizes were small, with the difference becoming negligible as sample size increased.
- In Gaussian graphical models, P-DCov methods (lasso.dcov, sam.dcov) performed comparably to penalized likelihood methods, showing no significant loss in accuracy despite using a more general dependence measure.
- For general graphical models (non-Gaussian), P-DCov methods outperformed likelihood-based methods in terms of AUC.
- When dealing with nonlinear factors, sparse additive model-based P-DCov methods showed substantial advantages.
- Real Data Analysis (Financial Stocks):
- When analyzing stock returns conditional on Fama-French factors, 15.46% of stock pairs were found to be conditionally dependent, suggesting the factors alone are insufficient to explain all dependencies.
- The P-DCov test identified more significant pairs than a Pearson correlation-based test, highlighting its power in detecting non-linear dependencies.
- Analysis of industry groups revealed interesting conditional dependencies, such as between "Conglomerates" and "Aerospace"/"Transportation," potentially due to supply chain relationships.
Limitations and Future Work
- Linear Factor Assumption: The current theoretical results primarily assume that the contribution of factors is sparse and linear.
- Future Work: The authors suggest extending the theory to sparse additive models for factor projection and to cases where the dimensions of x and y grow with n.
- R Package: An R package named
pgraphis available for implementing the proposed methodology.
Summary updated