Jointly Estimating Individual and Group Networks from fMRI Data

In fMRI research, graphical models are used to uncover complex patterns of relationships between brain regions. Connectivity-based fMRI studies typically analyze nested data; raw observations, e.g., BOLD responses, are nested within participants, which are nested within populations, e.g., healthy controls. Often studies ignore the nested structure and analyze participants either individually or in aggregate. This overlooks the distinction between within-participant and between-participant variance, risking Simpson's paradox, where group-level effects are opposite to individual-level effects. To address Simpson's paradox, we present an approach for jointly estimating both individual-level and group-level networks using a Bayesian multilevel model. Our approach flexibly specifies nearly any individual-level or group-level model. We used a Gaussian graphical model for the individual-level networks and a Curie-Weiss model for the group-level network. Simulations show that our method outperforms individual or aggregate analysis in edge retrieval as measured by AUC. We then apply the multilevel approach to a resting-state fMRI dataset of 724 healthy participants, examining both commonalities and individual differences. In the group-level network, we recover the seven previously found resting-state networks but also observe considerable heterogeneity in the individual-level networks. Finally, we discuss the necessity of a multilevel approach, additional challenges, and possible future extensions.