During model specification (or estimation, whatever) SPM determines a global value in a two-step procedure (see spm_global) for each of the volumes. First it calculates the mean intensity across all voxels within a volume, call it x. Then it determines another mean or global value, call it y, based on only those voxels whose intensity is larger than one eighth of the mean intensity x. The idea is to exclude non-brain "background" voxels (which usually have very low intensities) and then average across the remaining voxels. The defaults.mask.thresh refers to this global value y, meaning with default settings, those voxels "survive" whose intensity is above 80% of the global value. The procedure is repeated for all volumes of a model. The final analysis is based on only those voxels that survived the thresholding in all volumes. I think SPM additionally discards voxels associated with constant values in all volumes.
Thus, defaults.mask.thres does not affect the preprocessing, but it will have an impact on the model.
There seems to be no particular reason why the default thresholds are exactly 1/8 and 80%, but usually they result in reasonable brain masks. If you go with a defaults.mask.thresh of 0.10 more voxels become part of the analysis, especially around the boundaries. However, usually we want to focus on "brain voxels" associated with a certain amount of intensity, and we assume meaningful effects to be present in those "brain voxels". Effects showing up in "non-brain" low intensity voxels only might be difficult to interpret. The default settings might be problematic if signal intensities are very inhomogenous (e.g. due to B1 inhomogeneitiy), but note that a less strict threshold will not just lead to inclusion of low-intensity brain regions but also non-brain regions.
Concerning thesholding and statistical issues with boundary voxels you might want to look at Ridgway et al. (2009, Neuroimage) and Ridgway et al. (2012, Neuroimage). Note that during first-level analyses a small amount of variance is added automatically to all voxels as described in the 2nd paper.
Best
Helmut
|