Model specification and the reliability of fMRI results: implications for longitudinal neuroimaging studies in psychiatry

Functional Magnetic Resonance Imagine (fMRI) is an important assessment tool in longitudinal studies of mental illness and its treatment. Understanding the psychometric properties of fMRI-based metrics, and the factors that influence them, will be critical for properly interpreting the results of these efforts. The current study examined whether the choice among alternative model specifications affects estimates of test-retest reliability in key emotion processing regions across a 6-month interval. Subjects (N = 46) performed an emotional-faces paradigm during fMRI in which neutral faces dynamically morphed into one of four emotional faces. Median voxelwise intraclass correlation coefficients (mvICCs) were calculated to examine stability over time in regions showing task-related activity as well as in bilateral amygdala. Four modeling choices were evaluated: a default model that used the canonical hemodynamic response function (HRF), a flexible HRF model that included additional basis functions, a modified CompCor (mCompCor) model that added corrections for physiological noise in the global signal, and a final model that combined the flexible HRF and mCompCor models. Model residuals were examined to determine the degree to which each pipeline met modeling assumptions. Results indicated that the choice of modeling approaches impacts both the degree to which model assumptions are met and estimates of test-retest reliability. ICC estimates in the visual cortex increased from poor (mvICC = 0.31) in the default pipeline to fair (mvICC = 0.45) in the full alternative pipeline - an increase of 45%. In nearly all tests, the models with the fewest assumption violations generated the highest ICC estimates. Implications for longitudinal treatment studies that utilize fMRI are discussed.