Using the results of analysis A, for each paradigm, we tested whether the session variability was Gaussian. At each voxel in standard space we took the (first-level) parameter estimates (effect sizes) from the relevant voxel in each of the 33 relevant first-level analyses, (i.e., the same data that was fed into the group-level ME analysis). The variance of these is the ME variance. For each set of 33 first-level parameter estimates, we ran the Lilliefors modification of the Kolmogorov-Smirnov test [17] for non-Gaussianity, with a significance threshold of 0.05. We would therefore expect, in null data, rejection of the Gaussianity null hypothesis at this 5% rate by random chance.
We calculated the fraction of voxels failing the normality test both across the whole brain, and within the FE-derived masks described above. In both cases, and for all three paradigms, the fraction of failed tests was less than 7.5% (range 4.5-7.3%), which is very close to the expected 5% rate of null-hypothesis rejections if in fact all the data is normal. This provides strong quantitative evidence for the normality of the session variability in this data. Qualitatively, the voxels where the null hypothesis was rejected were scattered randomly through the images, not clumped, again suggesting that they were rejected by pure random chance rather than because of some true underlying non-Gaussian process.