Dimension constraints improve hypothesis testing for large-scale, graph-associated, brain-image data.

Area of Science:

• Statistical methodology
• Bioinformatics
• Neuroimaging analysis

Background:

• Large-scale hypothesis testing with complex, graph-associated data presents analytical challenges.
• Existing methods often overlook the structural information inherent in graph-associated datasets.
• Accurate identification of significant findings is crucial in fields like neuroimaging.

Purpose of the Study:

• To introduce an empirical Bayes mixture technique for scoring local false-discovery rates (FDRs) in graph-associated data.
• To enhance statistical power by leveraging the graph structure in hypothesis testing.
• To evaluate the performance of the proposed method against conventional approaches.

Main Methods:

• Development of the Graph-based Mixture Model (GraphMM) using empirical Bayes mixture techniques.
• Regularization of parameter contrasts between testing units to incorporate graph information.
• Simulation studies to assess FDR control and power across various settings.
• Application to magnetic resonance imaging (MRI) data from Alzheimer's disease studies.

Main Results:

• GraphMM demonstrated increased power in detecting effects when non-null cases formed connected subgraphs.
• Simulations confirmed that GraphMM generally controls the false-discovery rate (FDR), with potential loss of control under excessive regularization.
• The method yielded greater results on Alzheimer's disease MRI data compared to standard large-scale testing procedures.

Conclusions:

• The Graph-based Mixture Model (GraphMM) offers a powerful approach for large-scale testing on graph-associated data.
• Leveraging graph topology through regularization improves the detection of true effects, particularly in neuroimaging studies.
• GraphMM provides a valuable tool for analyzing complex biological data and advancing disease research.