Dear everyone,
basically it doesn't make much sense to run a parametric test with such a small group (not that a non-parametric test is much better, it has less/no assumptions about the distribution, but you still should have a reasonable number of subjects of course). For the moment you might have highly sig. results, but this is going to change very probably. These glassbrain results look like noise more or less, single voxels or small clusters here and there spotted all over the brain (although it might look "smoother" with a different uncorrected voxel threshold like .001).
If you plan to acquire some more data and just want to make sure everything worked well so far I would suggest to look at the single-subject models. Use a more unspecific contrast like (task1 + task2)/2 vs. baseline, something like "is there a neural correlate of the visual stimulus material / of the button press", as specific effects like task1 vs. task2 might be highly variable or very small in individual subjects.
Aside from that, looking at the individual data sets can be quite useful, as you might also detect some relationship between "poor results" and large head motion or other issues related to data quality. Alternatively you could also run some fixed-effects group analysis based on those subjects. Although you should actually not test your data as long as the data acquistion has not yet been completed ;)
Best,
Helmut
|