Abstract
Multivariate Gaussian graphical models are defined in terms of Markov properties, i.e., conditional independences associated with the underlying graph. Thus, model selection can be performed by testing these conditional independences, which are equivalent to specified zeroes among certain (partial) correlation coefficients. For concentration graphs, covariance graphs, acyclic directed graphs, and chain graphs (both LWF and AMP), we apply Fisher’s z-transformation, Sidák’s correlation inequality, and Holm’s step-down procedure, to simultaneously test the multiple hypotheses obtained from the Markov properties. This leads to a simple method for model selection that controls the overall error rate for incorrect edge inclusion. In practice, we advocate partitioning the simultaneous p-values into three disjoint sets: a significant set S, an indeterminate set I, and a non-significant set N. Then our SIN model selection method selects two graphs, a graph whose edges correspond to the union of S and I, and a more conservative graph whose edges correspond to S only. Prior information about the presence and/or absence of particular edges can be incorporated readily.
Introduction
In multivariate statistics, graphical models are essential for visualizing complex dependencies between variables, such as gene regulation networks. However, "learning" the structure of these graphs from data is notoriously difficult. Existing approaches—like constraint-based algorithms or Bayesian score searches—often struggle with a fundamental trade-off: aggressive selection leads to "false edges" (noise), while conservative selection misses real connections.
Crucially, many standard methods lack a clear mechanism to control the Family-Wise Error Rate (FWER)—the probability of making even one mistake in the entire graph. This paper introduces a statistically sound framework for model selection with guaranteed error bounds, treating edge selection as a multiple hypothesis testing problem.
Methodology: The SIN Framework
The authors translate graph selection into simultaneous tests on (partial) correlation coefficients. The core of the approach involves:
- Markov Properties: Leveraging the fact that in Gaussian models, the absence of an edge corresponds to a specific (partial) correlation being zero.
- Fisher’s z-transformation: Converting sample correlations into approximately normally distributed test statistics to make the mathematics tractable.
- Multiple Testing Procedures: Comparing methods to handle the "multiplicity" problem. While marginal methods like Bonferroni or Holm’s step-down are used, the paper highlights the max-T procedure. This novel approach uses Monte Carlo simulations to estimate the joint distribution of test statistics, accounting for correlations between edge tests and thus reducing conservatism.
- Prior Knowledge: Incorporating "upper" and "lower" graphs to specify known edges, thereby increasing the statistical power of remaining tests.
Results and Insights
The method was evaluated using both simulated and real-world genomic data.
Simulated Benchmarking
Simulations across sample sizes ($n=25$ to $500$) for a 7-variable system demonstrated that the step-down max-T method provided the most exact error control. Unlike the Bonferroni correction, which becomes overly conservative as variable counts grow, max-T maintained high power.
Biological Discovery in Arabidopsis thaliana
Applying the method to microarray gene expression data ($n=118$, $p=13$ genes) from the MEP metabolic pathway, the authors reconstructed a Directed Acyclic Graph (DAG) that closely mirrored known biological pathways. By relaxing error criteria from FWER to GFWER (allowing up to $k$ mistakes), they recovered a denser, more informative network (24 edges vs. 19) while still maintaining high confidence in the overall structure.
The Duality of Graphs
The paper emphasizes that different graph types capture different nuances. Undirected graphs (concentration graphs) represent conditional independence, while bidirected graphs (covariance graphs) represent marginal independence. Utilizing both can reveal distinct layers of a system's architecture.
Limitations and Future Directions
While powerful, the current SIN approach assumes the sample size ($n$) is larger than the number of variables ($p$). In high-dimensional "Big Data" scenarios where $p \gg n$, additional sparsity assumptions or different estimators are required. Furthermore, for DAGs, the method currently requires a known "causal order" of variables, though using partial orderings (blocks of variables) is a promising area for future research. Finally, extending these rigorous controls to discrete or mixed data remains an open challenge.
Summary updated