The activation data set (iii) was analysed using standard GLM techniques as implemented in FEAT [Smith et al., 2001]. Final statistic maps were used to define activation 'masks' by thresholding at and clustering with , subtracting this threshold and re-scaling so that the final mask range was within the range , where any value signifies `level' of activation. These masks were transformed into the space of the resting data set using FLIRT [Jenkinson et al., 2002].
Next, activation was linearly added into the resting data (i) using artificial timecourses, modulated spatially by the activation masks described above. The timecourses were created by taking simple box-car designs (matching the paradigm of the activation data (iii) described above) and convolving with a standard gamma-based HRF kernel function (std.dev.=3s, mean lag=6s). Various overall levels of activation were added to create various test data sets, with the maximum resulting activation signal of 0.5%,1%, 3% and 5% times the mean baseline signal intensity. The average activation level within the clusters was of the peak activation level. In the real activation data, the highest activation was % peak to peak. Note that this is more realistic than the artificial data presented in [Lange et al., 1999] where all activated voxels have identical scores. The above procedure was carried out for auditory and visual 'activation' using a separate spatial activation mask and activation timecourses.