Thanks for your reply, Matthew. That certainly helps put into perspective what was going on with the zero clusters. In my particular case, each voxel was associated with a value based on fisher-z transformed correlation coefficients from a searchlight. So many of the values going into the sign flipping test definitely were negative (and/or low).
I take it that the output from the newer versions is valid? That is, is there any reason why we *should* only introduce positive values (above some threshold) into Randomise? I suppose z-stats, t-stats, and countless others would be both positive and negative. Do such things require a transformation to play nice with the program and yield interpretable results?
Out of curiosity, do you happen to know how the threshold was determined in the old version (loosely speaking)? I suppose the intention had been to find non-brain (or CSF) voxels and mask them out?
Thanks, as always, for the time you put into answering questions (on top of everything else!),
Justin
|