Statistical Methodologies for Group Comparisons of Brain Connectivity using Multimodal Neuroimaging Data

Recent advances in mental health research have led to novel treatments that target specific brain networks involved in the pathophysiology of psychiatric disorders. This has made it increasingly important for researchers to characterize the networks that are associated with these disorders. Although neuroscientists in independent studies have identified abnormalities in structural and functional brain connectivities associated with these diseases, the relationship between the two remains elusive. As complex psychiatric disorders (like depression) are not likely to be simply related to isolated brain regions of interest, we assessed structural (measured using high angular resolution diffusion imaging) and functional (measured using resting-state functional magnetic resonance imaging) connectivity in a model of geriatric depression using whole-brain graph theory-based connectivity analyses. To appropriately account for both the between-modality and within-subject correlations of the connectivity measures, we fitted bivariate linear mixed-effect models with random subject intercepts and heteroscedastic errors to neuroimaging data containing measures for structural and functional connectivities. The models were estimated using an expectation-maximization like approach by iterating between the empirical Bayes estimates of the random effects and maximum marginal likelihood estimates of the fixed and covariance parameters. By comparing study and control subjects using the model estimates, we performed unimodal, bimodal, and correlation analyses, and detected brain region links (connectivities) that demonstrated significant differences in single and both connectivity modalities, as well as in connectivity coherence between the two groups of subjects. We observed consistent results in similar analyses using Bayesian mixed-effect models, as well as in approaches that utilized machine learning techniques to determine the importance of each connectivity in group classification. Our approach represents a first attempt in simultaneous modeling of the two connectivity modalities in a population study while considering both the between-modality and within-subject correlations of the measures.