|
Abstract:
|
Recently there has been growing interest in the human brain's functional architecture, i.e. how brain regions interact in networks. The association measure between regions is Pearson correlation between the fMRI time series of regions of interest, z-transformed to facilitate modelling using the normal distribution. A challenge for many network modeling techniques lies in deciding which possible connections to retain, and which to discard. Different thresholding rules have been suggested, all somewhat arbitrary. We propose a Bayesian hierarchical mixture model for resting-state brain connectivity, allowing inferences without mid-analysis thresholding. We use a two-component mixture, where the non-connected component has a normal distribution, while the connected component is lognormal. This model implies that only positive correlations may represent connections while non-connected regions may have negative or positive correlations. This contrasts with previous related work, which consider normal distribution mixtures. The hierarchical structure allows us to provide population- and individual-level inferences and explore covariate effects on functional connectivity.
|
