Learning Gaussian Graphical Models from Correlated Data.

Area of Science:

• Statistics
• Network Analysis
• Genomics

Background:

• Gaussian Graphical Models (GGMs) represent complex variable relationships using partial correlations.
• Standard GGM inference assumes independent observations, which is often violated in clustered or longitudinal data.
• Ignoring within-subject correlation can lead to inflated Type I errors, misrepresenting network structures.

Purpose of the Study:

• To develop and validate a cluster-based bootstrap algorithm for inferring GGMs from correlated data.
• To address the limitations of traditional GGM methods when applied to non-independent observations.
• To accurately model complex biological networks from family-based genetic data.

Main Methods:

• A novel cluster-based bootstrap algorithm was proposed for GGM inference.
• Extensive simulations using correlated data from family-based studies were conducted.
• The proposed method was applied to learn the GGM of 47 Polygenic Risk Scores from the Long Life Family Study.

Main Results:

• The cluster-based bootstrap method effectively controlled Type I error rates.
• The proposed algorithm maintained statistical power compared to alternative methods.
• The method accurately identified complex relationships in Polygenic Risk Scores without inflating Type I error.

Conclusions:

• The cluster-based bootstrap algorithm provides a robust approach for GGM inference with correlated data.
• This method is suitable for analyzing complex biological networks in family-based and longitudinal studies.
• The proposed approach offers a reliable alternative to conventional methods that ignore within-cluster correlation.