Abstract

Pattern similarity analysis is widely used to assess response reliability in repeated-stimulus fMRI designs, but pairwise similarity estimates may be confounded by temporal structure, including slow drift, autocorrelation, hemodynamic smoothing, and fixed repetition order. This study presents a simulation-supported workflow for detecting and reducing such confounding. A fixed design with 40 events, corresponding to ten stimuli repeated four times, was simulated with and without stable stimulus-specific structure. Naive within-versus-between similarity was compared with a time-neighborhood-matched contrast and a distance-adjusted regression estimator. In 500 null simulations, temporal structure alone produced a biased naive contrast (mean ± SD = −0.1309 ± 0.0021), whereas the time-neighborhood-matched contrast was near zero (−0.0013 ± 0.0018). The time-neighborhood-matched contrast increased monotonically with stimulus-effect strength and remained robust after introducing hemodynamic convolution, spatial dependence, structured nuisance components, and variation in autocorrelation, voxel count, and drift specification. In an OpenNeuro music-fMRI case study, the naive contrast was strongly negative, whereas temporal matching yielded a small positive residual contrast. Cross-validated selection of functionally responsive voxels increased this residual effect, indicating partial dilution in the whole-EPI analysis, while the naive contrast remained negative under all mask specifications. The workflow provides a practical diagnostic framework for repeated-stimulus fMRI pattern similarity analyses.

IPC Classification

Keywords