Bayesian Variable Selection with Applications to Neuroimaging Data

In this dissertation, we discuss Bayesian modeling approaches for identifying brain regions that respond to certain stimulus and use them to classify subjects. We specifically deal with multi-subject electroencephalography (EEG) data where the responses are binary, and the covariates are matrices, with measurements taken for each subject at different locations across multiple time points. EEG data has a complex structure with both spatial and temporal attributes to it. We use a divide and conquer strategy to build multiple local models, that is, one model at each time point separately both, to avoid the curse of dimensionality and to achieve computational feasibility. Within each local model, we use Bayesian variable selection approaches to identify the locations which respond to a stimulus. We use continuous spike and slab prior, which has inherent variable selection properties. We initially demonstrate the local Bayesian modeling approach which is computationally inexpensive, where the estimation for each local modeling could be conducted in parallel. We use MCMC sampling procedures for parameter estimation. We also discuss a two-stage variable selection approach based on thresholding using the complexity parameter built within the model. A prediction strategy is built utilizing the temporal structure between local models. The spatial correlation is incorporated within the local Bayesian modeling to improve the inference. The temporal characteristic of the data is incorporated through the prior structure by learning from the local models estimated at previous time points. Variable selection is done via clustering of the locations based on their activation time. We then use a weighted prediction strategy to pool information from the local spatial models to make a final prediction. Since the EEG data has both spatial and temporal correlations acting simultaneously, we enrich our local Bayesian modeling by incorporating both correlations through a Kronecker product of the spatial and temporal correlation structures. We develop a highly scalable estimation approach to deal with the ultra-huge number of parameters in the model. We demonstrate the efficiency of estimation using the scalable algorithm by performing simulation studies. We also study the performance of these models through a case study on multi-subject EEG data