All Times EDT
Virtual
Graphical Modeling of Multi-Study Data with Multi-Study Factor Analysis (309570)
Raji Balasubramanian, Department of Biostatistics and Epidemiology, University of Massachusetts AmherstRoberta De Vito, Department of Biostatistics and Data Science Initiative, Brown University
*Katherine Hoff Shutta, Department of Biostatistics and Epidemiology, University of Massachusetts Amherst
Keywords: factor analysis, multi-study factor analysis, graphical modeling, networks
Multi-study factor analysis (MSFA) is a method for performing factor analysis on measurements obtained across multiple studies. MSFA identifies a set of latent factors shared among studies and a set of latent factors that are specific to each study. The MSFA model formulation admits a decomposition of the covariance matrix of the variables into the sum of covariance due to shared factors, covariance due to study-specific factors, and covariance due to noise. Gaussian graphical models (GGMs) are networks of nodes corresponding to variables and weighted edges corresponding to partial correlations between variables. GGM estimation provides scope for network-level analysis of conditional dependencies present in a set of variables. In this work, we leverage the MSFA framework to estimate shared and study-specific GGMs. We extend the MSFA model formulation by decomposing the noise into global and study-specific components; this leads to an expression for shared and study-specific inverse covariance matrices that can be translated into global and study-specific GGMs. We demonstrate via simulation that our method has the potential to recover important study-specific network properties.