|
Abstract:
|
Multivariate data with several different types of data collected on the same sampling units are ubiquitous in modern science and arise in many settings. In neuroimaging studies different imaging modalities, such as structural/functional magnetic resonance imaging, diffusion tensor imaging and others, provide different views of the human brain and are often accompanied by demographic or medical data. Such collections motivate simultaneous exploration to gain insights into relationships between data sets. We examine two Joint and Individual Variation Explained (JIVE) approaches (one iterative, the other angle-based) via simulation and case studies. Each method uncovers latent signals within data blocks, which can be further decomposed into subject scores and variable loadings to provide information about potential multivariate associations, aid in discriminant analyses or tease out features unique to a type of data. Results: the iterative method is better able to detect the correct number of components, but the angle-based approach provides better estimation of the latent signals. Both methods were also applied to datasets from the Philadelphia Neurodevelopment Cohort Study.
|
