Modeling intra-subject correlation among repeated scans in positron emission tomography (PET) neuroimaging data

Many in vivo positron emission tomography (PET) neuroimaging studies record correlates of regional cerebral blood flow (rCBF) in a series of scans for each individual, usually under different experimental conditions. Typical methods for statistical analysis involve fitting voxel-specific general linear models (GLM) that assume spherical normal errors, implying that all voxel-specific rCBF measurements are independent and arise from identical normal probability distributions. While the spherical GLM provides a unified and computationally efficient approach to estimation, the likely correlations among an individual's repeated scans and heteroscedasticity between conditions prompt the use of extended statistical methodology. We outline a more general method to analyze PET data using random effects and correlated errors to model unequal variances across conditions as well as covariances (correlations) among the repeated scans for each individual. We introduce correlation maps to display intra-subject correlations between an individual's rCBF measurements from different scans. We illustrate the application of our model using data from a study of social anxiety and highlight analytical advantages over the spherical GLM.