Abstract
Despite its widespread use in neuroscience, the reliability of fMRI remains insufficiently understood. One powerful way to tap into aspects of fMRI reliability is via the inter-subject correlation (ISC) approach, which exposes different viewers to the same time-locked naturalistic stimulus and assesses the similarity of neural time series. Here we examined the correlations of fMRI time series from 24 participants who watched the same movie clips across three repetitions. This enabled us to examine inter-subject correlations, intra-subject correlations, and correlations between aggregated time series, which we link to the notions of inter-rater reliability, stability, and consistency. In primary visual cortex we found average pairwise inter-subject correlations of about r = 0.3, and intra-subject correlations of similar magnitude. Aggregation across subjects increased inter-subject (inter-group) correlations to r = 0.87, and additional intra-subject averaging before cross-subject aggregation yielded correlations of r = 0.93. Computing the same analyses for parietal (visuospatial network) and cingulate cortices (saliency network) revealed a gradient of decreasing ISC from primary visual to higher visual to post-perceptual regions. These latter regions also benefitted most from the increased reliability due to aggregation. We discuss theoretical and practical implications of this link between neural process similarity and psychometric conceptions of inter-rater reliability, stability, and internal consistency.