If I had to guess it probably is due to using cluster stats with a fairly
conservative threshold (Z>2.3, p<0.01). Basically this requires that you
have fairly extensive clusters -- typically on the order of 1000 or more
voxels for a whole brain analysis with resels sizes of approx 50
voxels. So you may find that in each group there is a certain amount of
activation -- possibly even with very high Z-scores -- but it does quite
reach your cluster threshold. But if you combine them, then the clusters
merges and voila -- instant significant results.
Why not put the p value to 1.0 which would effectively mean you were doing
voxel stats? If you're use to SPM99, then a z threshold of 3.1 corresponds
to an uncorrected p=0.001 which is the default in SPM. That way you'll see
any voxels that pass that threshold and get a better feel for what's going
on in your data.
Joseph T. Devlin, Ph. D.
FMRIB Centre, Dept. of Clinical Neurology
University of Oxford
John Radcliffe Hospital
Headley Way, Headington
Oxford OX3 9DU
Phone: 01865 222 738
Email: [log in to unmask]