Noise ceiling estimation

Measurement noise in an fMRI experiment includes the physiological and neural noise in voxel activation patterns, fMRI measurement noise, and individual differences between subjects—even a perfect model would not result in a correlation of 1 with the voxel RDMs from each subject. Therefore an estimate of the noise ceiling is necessary to indicate how much variance in brain data—given the noise level—is expected to be explained by an ideal ‘true’ model.

We calculated the upper bound of the noise ceiling by finding the average correlation of each individual single-subject voxel RDM (Eqs 21 and 22) with the group mean, where the group mean serves as a surrogate for the perfect model. Because the individual distance structure is also averaged into this group mean, this value slightly overestimates the true ceiling. As a lower bound, each individual RDM was also correlated with the group mean in which this individual was removed.

We also tested whether a model’s noise ceiling was significantly greater than zero. We first Fisher transformed individual Spearman’s rank-order correlation values and then performed a one-sided t-test for the mean being greater than zero. The 5% significance threshold for the t-value was corrected for multiple comparisons (number of regions). For more information about the noise ceiling see [36].

In sum, we only considered a model evidence to be significant if it satisfied three criteria: (1) the model correlation with data was significantly greater across participants than the lower bound of the noise ceiling, (2) the lower bound of the noise ceiling was significantly greater than zero across participants, and (3) the average correlation for the item-mixture model (null-hypothesis) did not reach the noise ceiling.