Perusing the AFNI messageboards, I found the idea of using age-specific templates in samples that are not well matched for age.
However, how can one compare across age-specific templates in group analysis? Let’s say I want to determine whether a quantitative risk factor explains variation in a sample of 8 to 20 year olds. I used 3dttest++ for group analysis including age, sex, and the risk factor as covariates over scans all registered to TT-N27. Is this valid analysis? Or should I be more careful about age-related differences?
Thank you so much!
There’s really not a perfect solution. We know that aligning pediatric data to adult templates is not ideal. The basics are:
- At age 6, cerebral volume is about 95% of that of an adult (Lenroot & Giedd, 2006).
- However, grey matter follows an inverted U-shape, which peaks at different times for different cortical (and subcortical) regions (Lenroot & Giedd, 2006).
- In general, using a nonlinear warp can help with local deformations (Kochunov et al., 2002)
- Some estimates show that 8 year olds have 20% more deformation when going to an adult template vs. a child template (Yoon et al., 2009)
Using covariates is still a voxel-wise process in 3dttest++, so it’s not going to fix the problems with deformation distances and transform accuracies. I would be hesitant of how much to trust the alignment of small cortical and subcortical areas. You might consider the following:
- Wilke et al. (2008) have a Template-o-Matic that has the ability to generate a template from 4-19 years or so.
- Aligning the pediatric data to our Haskins Pediatric Template (nonlinear preferred), Aligning the adult data to TT_N27. Trying to make indirect comparisons between the samples.
- Create your own template based on the sample
- Make extended use of the phrase “exploratory” in your limitations section and do any of these three or your current path of using the TT_N27.
Thank you so much! This makes it a whole lot more clear!
Thank you again for your help; I am following your suggestions and they are helpful.
I wanted to ask you one thing regarding the Haskins atlas. In the first-pass, I performed group analysis on sample ages 8 to 21 all having non-linear registrations to the TT-N27 template.
There after, I identified significant clusters with 3dclust (placed in clust+tlrc). Then I ran whereami of these clusters using both CA_N27_ML and the Haskins_Pediatric_Nonlinear_1.0 atlases, that is,
whereami -atlas CA_N27_ML -omask clust+tlrc
3drefit -space HaskinsPeds clust+tlrc;
whereami -atlas Haskins_Pediatric_Nonlinear_1.0 -omask clust+tlrc
The idea here is, for the first-pass analysis, I would report ROIs overlapping with significant clusters under both the CA_N27_ML and Haskins_Pediatric_Nonlinear_1.0 atlases (separately).
For example, it could be somehting like:
CA_N27_M: Left Insula (73.9%), Left pars triangularis (9.9%)
Haskins Pediatrics: Left pars opercularis (53.6%), Left Insula (26.2%), Left lateral orbitofrontal (12.6%)
Is it kosher to use Haskins Pediatrics atlas for reporting results in group analysis of scans registered to TT-N27? More pointedly, is it fundamentally wrong from technical point of view?
The Haskins atlas is smaller in size than the TT_N27 or the MNI, reflecting the size of the brain differences in children compared to either 152 adults or Colin. The issue is really normalizing the pediatric data to the adult template or the adult data to a pediatric template.
You could try to make a nonlinear warp between the two spaces (Colin → Haskins) and then warp the stat map or cluster mask to Haskins and try that. But I’d still be wary.
Many papers have claimed “willful ignorance” by just putting everything to an adult template and calling it a day.
I’ll try to ponder it over the weekend and likely confer with some of the co-authors on the pediatric atlas.
Thank you for your suggestions. I followed your recommendations and found some interesting differences in group analysis over children with scans registered to Haskins atlas versus the same group analysis over scans registered to TT-N27. The locations of the significant clusters under the Haskins template align closer with what I would have expected from previous findings (which, to me, is good news!)
Is there a paper to cite for the Haskins Pediatrics template & atlas? I cannot seem to find, although the AFNI documentation says it is under preparation.
(Edited: Removed previous question; I found what the problem was.)
Peter is actively writing this up (ahem…), but if you’re in a rush, you can cite this other paper that used the Haskins Pediatric Atlas: