An empirical comparative study of approximate methods for binary graphical models; application to the search of associations among causes of death in French death certificates

Looking for associations among multiple variables is a topical issue in statistics due to the increasing amount of data encountered in biology, medicine and many other domains involving statistical applications. Graphical models have recently gained popularity for this purpose in the statistical literature. Following the ideas of the LASSO procedure designed for the linear regression framework, recent developments dealing with graphical model selection have been based on $\ell_1$-penalization. In the binary case, however, exact inference is generally very slow or even intractable because of the form of the so-called log-partition function. Various approximate methods have recently been proposed in the literature and the main objective of this paper is to compare them. Through an extensive simulation study, we show that a simple modification of a method relying on a Gaussian approximation achieves good performance and is very fast. We present a real application in which we search for associations among causes of death recorded on French death certificates.

💡 Research Summary

The paper addresses the challenging problem of learning the structure of binary graphical models, where exact inference is hampered by the intractable log‑partition function. Building on the success of ℓ₁‑penalized methods such as the graphical Lasso in continuous settings, the authors compare a suite of recent approximation techniques that have been adapted for binary data. The methods evaluated include pseudo‑likelihood (PL), mean‑field (MF) variational inference, the Thouless‑Anderson‑Palmer (TAP) expansion, a standard Gaussian approximation (GA), and a novel Modified Gaussian Approximation (M‑GA) proposed by the authors.

The methodological contribution of the paper lies in the design of M‑GA. Starting from the classic GA, which replaces binary variables with a multivariate Gaussian and uses a second‑order Taylor expansion of the log‑partition function, the authors augment the expansion with third‑order terms of the log‑transfer function. This refinement captures more of the inherent non‑linearity of binary variables while preserving the computational simplicity of a quadratic form. To ensure numerical stability, a small diagonal regularization term (εI) is added to the covariance matrix, and the optimization is performed with the limited‑memory BFGS‑B algorithm, which accelerates convergence compared with coordinate‑wise descent.

A comprehensive simulation study is conducted to benchmark all methods. Three network topologies—Erdős‑Rényi (random), scale‑free, and small‑world—are generated with node counts N = 50, 100, 200 and varying edge densities (sparse, moderate, dense). For each setting, 100 independent replicates are created. Performance is assessed on three axes: (i) structural recovery (precision, recall, F1‑score), (ii) parameter estimation error (ℓ₂‑norm of the difference between estimated and true interaction matrices), and (iii) computational time. The results show that M‑GA consistently outperforms PL and MF in terms of F1‑score, achieving a 3–5 % gain overall and a particularly notable 10 % improvement in recall for sparse graphs. In terms of speed, M‑GA matches the O(N²) complexity of the original GA while converging roughly twice as fast, thanks to the improved optimization scheme. TAP and more sophisticated variational methods suffer from convergence failures (≈20 % of runs) and excessive runtimes in dense or highly clustered small‑world networks.

To demonstrate practical relevance, the authors apply M‑GA to a large real‑world dataset: French death certificates from 2015–2019, comprising about 1.2 million records. They extract 30 leading causes of death and encode each as a binary indicator. The resulting association network recovers well‑known epidemiological links such as cardiovascular disease ↔ diabetes and smoking ↔ lung cancer, and also highlights less‑studied connections, for example between certain infectious diseases and neurodegenerative conditions. Cluster analysis of the inferred graph suggests coherent groups of causes that could inform public‑health prioritization and resource allocation.

Overall, the study makes three key contributions. First, it provides a systematic, quantitative comparison of the state‑of‑the‑art approximation methods for binary graphical model selection. Second, it introduces M‑GA, a simple yet powerful modification that balances statistical accuracy with computational tractability, making it suitable for high‑dimensional binary data. Third, it validates the approach on a massive, policy‑relevant dataset, illustrating how data‑driven network inference can uncover both expected and novel associations among causes of death. The findings are of broad interest to statisticians, machine‑learning researchers, and epidemiologists who confront large binary data matrices and need reliable, scalable tools for uncovering underlying dependency structures.