A0534
Title: Bayesian image mediation analysis
Authors: Jian Kang - University of Michigan (United States) [presenting]
Abstract: Neuroimaging data presents unique challenges for mediation analysis due to its high dimensionality, complex spatial correlations, sparse activation patterns, and relatively low signal-to-noise ratio. A new Bayesian image mediation analysis (BIMA) method that employs a spatially varying coefficient structural equation model is proposed to address these challenges. A soft-thresholded Gaussian process (STGP) is used for prior specifications of the spatially varying coefficients, which enables large prior support for sparse and piece-wise smooth functions. The spatially varying mediation effects of the exposure on the outcome mediated through imaging mediators under the potential outcome framework are defined. Posterior consistency is established for spatially varying mediation effects and selection consistency on important regions that contribute to the mediation effects. An efficient posterior computation algorithm for BIMA is developed and scalable for large-scale imaging data analysis. As demonstrated through simulations, BIMA improves estimation accuracy and computational efficiency for high-dimensional mediation analysis compared to existing methods. Additionally, BIMA is applied to analyze behavioural and fMRI data in the Adolescent Brain Cognitive Development (ABCD) study and infer the mediation effects of parental education level on children's general cognitive ability mediated through working memory brain activities.