 |
 |
Concerning thresholds: The most common thresholds on voxel level are .05 FWE corrected or .001 uncorrected, although you might also find .005 uncorrected or others in some instances. But see Woo et al. (2014, Neuroimage) for a recent paper on that issue. It is not recommended to use FDR as a voxel threshold anymore (see Chumbley & Friston, 2009, Neuroimage), and this option is disabled in more recent SPM versions. If you go with an uncorrected voxel threshold, you should use a cluster correction additionally, typically .05 FWE or .05 FDR. There are quite many papers which do NOT use a cluster correction, but rather an arbitrary (and probably liberal) extent threshold like k > 10 voxels. Often this is justified by citing Bennett et al. (2009, Soc Cogn Affect Neurosci) or related papers. The problem is that these parameters proved to be reliable in some specific instances, which do not necessarily transform to similar studies and likely do not hold anymore for current fMRI data. For example, data is resampled to a resolution of 2x2x2 mm^3 during normalisation nowadays, so the number of voxels is more than three times larger than when going with 3x3x3 mm^3. Note that if you use a corrected voxel threshold of .05 FWE, no (statistical) cluster threshold is required, as you've already corrected for multiple comparisons. However, it is common to use an additional extent threshold like k > 10 voxels to avoid many small one-voxel clusters.