Novel Methods for Statistical Analysis of Covariance Structures

Neuroscientists increasingly understand brain development and pathology through relationships between complex measurements. These measurements include neuroimaging, genetic, and mobile health data that contain distinct but complementary information. However, the associations between these data and outcomes of interest can be difficult to detect and often require samples acquired across multiple study centers. This multi-site design can introduce bias in the form of site effects, which have been demonstrated to severely impact downstream analyses. Here, we develop methods for analysis of covariance structures in structural imaging, functional imaging, and mobile health studies. In structural imaging, we find that site effects in covariance can bias machine learning results and propose methodology for mitigating this bias. In functional imaging, we discover that site effects in subject-specific covariance structures can impact downstream network analyses and we develop several methods for addressing these effects. We additionally develop a multimodal regression framework that leverages the covariance among data modalities, which we apply in imaging and mobile health studies. We evaluate the performance and utility of our methodologies through simulations and applications to several notable multi-site and multimodal studies.