Abstract
Most fMRI studies average results across subjects, but there is substantial individual variability in anatomical structure and BOLD responses. Therefore, averaging is usually performed by transforming each subject's anatomical volume to a standard template, and then averaging functional data from all subjects within this common anatomical space. However, anatomical normalization is under-determined, so this process is likely to introduce some error into the averaged data set. How much tuning information is lost when we average across subjects in fMRI experiments? To investigate this issue we compared averaged and individual results, using a computational modeling approach used previously in our laboratory (Kay et al., Nature 2008, v.452, 352-355). The data consisted of fMRI BOLD activity recorded from the visual cortex of three subjects who viewed a large set of monochromatic natural images. We first estimated voxel-based receptive fields for each subject and calculated the correlation between observed and predicted BOLD responses. We then averaged the fMRI data across subjects (using a leave-one-out procedure to avoid over-fitting), estimated voxel-based receptive field models on the averaged data, and calculated the correlation between observed and predicted BOLD responses. We found that the predictions of models based on individual data were more highly correlated with the observed data than were the predictions of models based on averaged data. In summary, our data suggest that averaging across subjects reduces the information that can be recovered from fMRI data.