Population Inference for Node Level Differences in Multi-subject Functional Connectivity

Using Gaussian graphical models as the basis for functional connectivity, we propose new models and test statistics to detect whether subject covariates predict differences in network metrics in a population of subjects. Our approach emphasizes the need to account for errors in estimating subject level networks when conducting inference at the population level. Using simulations, we show that failure to do so reduces statistical power in detecting covariate effects for realistic graph structures. We illustrate the benefits of our procedure for clinical neuroimaging using a resting-state fMRI study of neurofibromatosis-I.