[log in to unmask]" class="GroupWiseMessageBody">Dear David,


yes, global calculation / "Mean" takes into account all the voxels of the images. First the mean is determined, then the mean is divided by 8 and all those voxels with values above that threshold are used to get a global value for each of you images/subjects. It seems it does not matter for that calculation whether you specify any masks or implicit/explicit thresholds during model estimation or not. However, it DOES matter which data (masked/original) you use, see below. As the global values are taken into account during Global normalisation to scale the data different global values might indeed affect explained variance and the FWHM. As a note, the FWHM is calculated just based on the voxels that go into the analysis.

In your case, the global value of the masked data seems to be close to the average value of the remaining non-zero voxels. When looking at the images it seems about 80 % of the voxels have been zeroed due to the mask. Take an average value of the non-masked voxels of 2, then the average overall is about 2 * 20/100 = 0.4. The global threshold would be 0.4/8, so 1) all the masked zero voxels are rejected 2) only a few of the non-masked voxels (those with very low intensity) are not taken into account for the global value => average in-mask and global are very similar, and indeed the correlation is almost 1.

In contrast, the average value of the (unmasked) original data is about 1. The global threshold is thus 1/8. As the original data has lots of voxels with intensity above 1/8 (which fall outside the mask at the same time) the global value (taking into account these voxels) is much lower than the average value of the in-mask voxels. The correlation is "just" about .67.

Now, which of the models are "better"? No idea ;-) Actually this is very interesting to know, as it seems to be quite common to mask animal data at some point during the preprocessing, whereas it is less common (?) for human PET?


Interesting to hear some expert's opinion,

Helmut