Family‐wise error rate control in Gaussian graphical model selection via distributionally robust optimization

Chau Tran, Pedro Cisneros‐Velarde, Sang‐Yun Oh, Alexander PetersenStat 2022 · DOI

TL;DR

This paper bridges the gap between a robust optimization framework (DRO) and graphical model selection, showing that a parameter in the DRO method directly controls the family-wise error rate (FWER) of detecting false edges. This provides a principled way to tune the graphical lasso, a popular method for inferring network structures, and demonstrates its effectiveness in simulations and real-world applications.

What Problem Does This Paper Address?

Estimating the structure of a Gaussian graphical model involves inferring which pairs of variables are conditionally independent given all others. This is often done by estimating the precision matrix (the inverse of the covariance matrix), where zero entries indicate conditional independence. The "graphical lasso" is a widely used method for this, but it requires a tuning parameter to control sparsity. Choosing this parameter, known as graphical model selection, is crucial for obtaining accurate network structures. Traditional methods like cross-validation and BIC can be computationally expensive and may not perform optimally in finite samples. This paper addresses the challenge of principled and efficient graphical model selection by connecting it to a distributionally robust optimization (DRO) framework.

Methods

The core of this paper lies in establishing a theoretical link between the confidence level parameter ($\alpha$) in the Distributionally Robust Optimization (DRO) formulation of graphical lasso (proposed by Cisneros-Velarde et al., 2020) and the asymptotic Family-Wise Error Rate (FWER) of detecting false edges. The DRO framework, which minimizes the worst-case expected risk over an uncertainty set of distributions, is shown to be equivalent to the graphical lasso objective function. The authors prove that by setting the confidence level $\alpha$ in the DRO method, one can control the asymptotic FWER of the resulting graphical lasso estimator. This theoretical connection allows for a principled way to select the regularization parameter for graphical lasso. They also leverage a bootstrap procedure (Algorithm 1) to approximate this parameter in finite samples.

Datasets

The paper evaluates its method using:

  • Simulated Data: Generated from multivariate zero-mean Gaussian distributions with a fixed true precision matrix. Datasets were generated for various sample sizes ($n \in \{50, 100, 200, 400, 800, 1600, 3200\}$) and for different significance levels ($\alpha$).
  • Real-World Datasets:
    • In silico dataset: From the DREAM5 transcriptional network inference challenge, containing $d=195$ transcription factors and $n=805$ arrays.
  • Escherichia coli (E. coli) dataset: From the DREAM5 challenge, with $d=334$ transcription factors and $n=805$ arrays.
  • Saccharomyces cerevisiae (S. cerevisiae) dataset: From the DREAM5 challenge, with $d=333$ transcription factors and $n=536$ arrays.

These real-world datasets were used to validate the inferred networks against experimentally validated interactions.

Key Insights and Ideas

  • Principled Tuning via FWER Control: The most significant insight is the direct theoretical link between the DRO confidence level ($\alpha$) and the FWER of false edge detection in graphical lasso. This transforms the abstract tuning parameter of graphical lasso into a concrete statistical guarantee.
  • DRO as a Robust Model Selection Tool: The paper demonstrates that the DRO framework, beyond its theoretical elegance, offers a practical solution for selecting the regularization parameter in graphical lasso, providing a robust alternative to traditional methods.
  • Finite Sample Performance: While the theoretical guarantees are asymptotic, the simulations show that the FWER control provided by the DRO-based method (RobSel) holds reasonably well even in finite sample settings.
  • Comparison with Hypothesis Testing: The paper draws parallels between the DRO-based approach and traditional hypothesis testing methods for graphical model selection, highlighting that both can achieve FWER control, but graphical lasso offers the advantage of performing model selection and parameter estimation simultaneously.

Key Results

  • Theoretical FWER Bound: Theorem 1 establishes that for a fixed $\alpha$, the asymptotic probability of estimating at least one false edge is strictly less than $\alpha$.
  • Simulation Performance: Figure 1 shows that RobSel generally exhibits lower FWER and True Positive Rate (TPR) compared to the Holm-based testing method at smaller $\alpha$ values, indicating a more conservative approach. As $\alpha$ increases, RobSel becomes less conservative, with TPR increasing while FWER remains bounded. The Jaccard index suggests that RobSel and Holm-based methods produce increasingly similar graphs as sample size increases.
  • Real Data Analysis: Table 1 demonstrates that RobSel is computationally faster than EBIC and cross-validation for the DREAM5 datasets, while achieving comparable or better validated edge proportions. Notably, RobSel consistently produces sparser graphs than EBIC and CV, even when tuned for a denser graph (using $\alpha=0.9$). Figure 2 visually confirms this sparsity and accuracy, showing that RobSel and Holm-based methods produce graphs with fewer false positive edges compared to cross-validation.

Limitations and Future Work

  • Asymptotic Guarantees: The theoretical FWER control is asymptotic. While simulations show good finite sample performance, further theoretical work might be needed to fully characterize the finite sample FWER behavior.
  • Computational Cost of RobSel: While faster than EBIC and CV in the presented real-data scenarios, the bootstrap procedure in Algorithm 1 for RobSel can still be computationally intensive, especially for very large datasets or when a high degree of accuracy is required.
  • Comparison with Other Methods: The paper primarily compares RobSel with the Holm-based testing method and mentions cross-validation and EBIC. Further comparisons with other state-of-the-art graphical model selection methods could provide a more comprehensive evaluation.
  • Future Work: The authors suggest that the theoretical connection between DRO and FWER control provides practical insights for practitioners. Future work could explore extending these findings to other graphical model settings (e.g., non-Gaussian distributions) or investigating the performance of RobSel in more complex network structures.

Summary updated

Abstract

Recently, a special case of precision matrix estimation based on a distributionally robust optimization (DRO) framework has been shown to be equivalent to the graphical lasso. From this formulation, a method for choosing the regularization term, that is, for graphical model selection, was proposed. In this work, we establish a theoretical connection between the confidence level of graphical model selection via the DRO formulation and the asymptotic family‐wise error rate of estimating false edges. Simulation experiments and real data analyses illustrate the utility of the asymptotic family‐wise error rate control behavior even in finite samples.
BibTeX
@inproceedings{tran2022familywise,
  author = {Chau Tran and Pedro Cisneros‐Velarde and Sang‐Yun Oh and Alexander Petersen},
  title = {{Family‐wise error rate control in Gaussian graphical model selection via distributionally robust optimization}},
  booktitle = {Stat},
  year = {2022},
  doi = {10.1002/sta4.477},
}
Figure