Between-movie variability severely limits generalizability of "naturalistic" neuroimaging
Movie-watching fMRI has become increasingly popular in neuroscience. Movie-fMRI data are commonly analyzed using inter-subject correlation (ISC), which quantifies the similarity of neural time series across individuals. Differences in ISC during movie viewing have been associated with psychological traits and clinical diagnoses. However, most studies investigating group differences in ISC or ISC-behavior associations have drawn conclusions from a single movie. Because ISC is inherently stimulus-driven, effects observed for one movie may not generalize to another. Yet, the extent to which ISC patterns and ISC-behavior associations depend on the specific movie being viewed has received limited systematic attention. Here, we analyzed three independent datasets comprising 318 subjects and 36 movies in total to quantify between-movie variability in ISC and assess its consequences for ISC-behavior associations. Across datasets, ISC varied between movies throughout the cortex. This variability was spatially heterogeneous: regions with stronger ISC showed greater between-movie variability. Movie-specific inter-subject representational similarity analysis revealed distinct spatial distributions of ISC-behavior associations, with limited overlap between movies. This pattern was observed for two distinct behavioral constructs. These findings suggest that ISC-behavior associations can be strongly movie-specific.