Hi Jesper,
It's great that you are weighing in on our paper. We think it is an important issue that potentially affects not only TBSS but, as we mention in the paper, any voxelwise analysis where one is using the same images for both spatial normalization and statistical testing.
You bring up a couple of potentially interesting issues but could you please elaborate a bit more as I don't quite follow how those issues relate to the arguments presented in our paper? Essentially, our argument is very simple:
* Following Eq. 2, spatial normalization to a template using the SSD similarity metric causes the average voxelwise variance over the region of interest to decrease (which does not necessarily reflect increased anatomical alignment accuracy).
* Since the t-statistic can be approximated as the difference in means over the variance, the t-statistic will increase as the variance is decreased.
Could you discuss where the issues you raise apply to these two points?
Also, I'm not sure you are aware but your statement "Since TBSS uses non-parametric testing..." is not quite accurate. In the second column, second paragraph of page 1494 of the original 2006 TBSS paper (http://www.ncbi.nlm.nih.gov/pubmed/16624579), Steve discusses both parametric and non-parametric statistical testing within the TBSS paradigm.
Thanks,
Nick
|