I am using representational similarity analyses (RSA) to compare models to some fMRI data. As part of this analysis I have computed noise ceilings for each of my regions of interest (ROIs). What's weird is that, for one ROI, the correlation between my model and the aggregate participant data (RDMs averaged together) is higher than the noise ceiling.
I calculated the noise ceiling by splitting the participant data into two sets, averaging the RDMs from each set and the correlating the two sets to one another. The main model comparisons are done by correlating the model RDM to an averaged RDM from all of the participants.
Given that the noise ceiling is supposed to estimate the performance of a "hypothetical true model" and be the upper bound of your prediction accuracy, how should one interpret this?
Given that this is only one ROI, my interpretation is that this region is a little noisier than others and so the between-participant correlation isn't great, but when you aggregate many participants it averages out some of the noise in the data and allows for a decent model correlation. Is that reasonable?