Inter-Subject Analysis: A Partial Gaussian Graphical Model Approach.

Area of Science:

• Neuroscience
• Statistical Modeling
• Graph Theory

Background:

• Inter-subject analysis (ISA) explores dependencies between subjects, treating intra-subject data as noise.
• ISA is crucial for understanding functional brain connectivity under natural stimuli.
• Existing methods face challenges with dense intra-subject dependencies and sparse inter-subject structures.

Purpose of the Study:

• To develop a novel statistical modeling framework for inter-subject analysis (ISA).
• To address the challenge of estimating and inferring partial Gaussian graphical models with sparse inter-subject structures.
• To provide a robust method for analyzing functional connectivity in neuroscience.

Main Methods:

• Proposed a modeling framework based on Gaussian graphical models for ISA.
• Developed an estimation strategy using an alternative parameter to handle non-sparse intra-subject dependencies.
• Introduced an "untangle and chord" procedure for de-biasing estimators, valid without assuming sparsity on the inverse Hessian.

Main Results:

• The proposed estimation method achieves asymptotic consistency even with dense intra-subject dependencies.
• The "untangle and chord" inference procedure is general and theoretically robust.
• Numerical experiments on simulated and real brain imaging data validated the effectiveness of the proposed methods.

Conclusions:

• The developed framework offers a powerful approach for inter-subject analysis in neuroscience.
• The statistical methods provide robust estimation and inference for Gaussian graphical models with specific sparsity assumptions.
• This work has broader implications for statistical inference in various fields beyond neuroscience.