Abstract
We propose a partially linear additive Gaussian graphical model (PLA-GGM) for the estimation of associations between random variables distorted by observed confounders. Model parameters are estimated using an $L_1$-regularized maximal pseudo-profile likelihood estimator (MaPPLE) for which we prove $\sqrt{n}$-sparsistency. Importantly, our approach avoids parametric constraints on the effects of confounders on the estimated graphical model structure. Empirically, the PLA-GGM is applied to both synthetic and real-world datasets, demonstrating superior performance compared to competing methods.
Introduction
In fields such as neuroscience and social network analysis, Gaussian Graphical Models (GGMs) are widely used to map dependencies between variables. However, real-world data is frequently contaminated by observed confounders—external factors that distort correlations. A prominent example is fMRI imaging, where subject head motion can create "phantom" connections, obscuring the true functional connectivity. Existing methods typically rely on rigid linear assumptions or incorporate confounders as additional nodes, often resulting in over-dense graphs that lack biological plausibility.
Methodology
The proposed Partially Linear Additive Gaussian Graphical Model (PLA-GGM) addresses these limitations by decomposing the precision matrix $\Omega$ into two components: a target structure $\Omega_0$ representing the true relationships, and a nuisance structure $R(g)$ induced by the confounder $g$.
Unlike traditional approaches, PLA-GGM does not impose a specific parametric form on $R(g)$. Instead, it assumes the effect is smooth and negligible when the confounder is small. To estimate $\Omega_0$, the authors introduce the Maximal Pseudo-Profile Likelihood Estimator (MaPPLE). This method utilizes kernel smoothing to "profile out" nuisance effects, focusing optimization on the underlying network. Theoretically, MaPPLE is shown to achieve $\sqrt{n}$-sparsistency, meaning it identifies the true sparse structure with high efficiency as sample size increases.
Empirical Evaluation
The performance of PLA-GGM was validated using both synthetic simulations and real-world neuroimaging data.
Synthetic Benchmarks
In simulations with varying dimensions ($p = 10$ to $100$), PLA-GGM consistently achieved higher Area Under the Curve (AUC) scores for structure recovery compared to baseline methods like TV-GGM and CON-GGM. The performance gap widened as the number of variables increased, highlighting the model's robustness in high-dimensional settings.
Real-World Application: Schizophrenia Connectivity
The model was applied to the COBRE dataset, consisting of fMRI scans from 72 patients with schizophrenia and 75 healthy controls, with motion as a confounder. PLA-GGM identified significant connectivity differences in the Occipital Pole and Central Opercular Cortex, findings that align with established psychiatric literature. Furthermore, the connectivity features derived from PLA-GGM provided superior diagnostic accuracy for schizophrenia compared to the standard linear regression approach (LR-GGM).
Discussion and Conclusion
PLA-GGM offers a flexible yet statistically efficient framework for network estimation under confounding. By anchoring on samples with low confounder levels, it effectively subtracts non-linear noise without the "statistical tax" usually associated with non-parametric methods. While the model requires the presence of some relatively "clean" samples, its ability to produce sparse, biologically relevant networks makes it a powerful tool for complex data analysis. Future work may extend this framework to non-Gaussian data types, such as discrete counts within the Univariate Exponential Family.
Summary updated