Longitudinal design for resting-state data with 3 groups

Dear Neurostars,

I would appreciate your input on how to set up the analyses of the data collected for a resting-state study.

I have 3 groups: patients, their high-risk relatives, and healthy controls (80-90 subjects in each of the groups)

About half of these sample have been rescanned with the same resting state sequence (same scanner) with a follow-up time ranging between 6 months to 2.5 years.

Aims of the study: Identify functional RS connectivity abnormalities that the relatives have in common with the patients (vs HC), and abnormalities that are unique to the relatives.

What I did so far:

  1. Single-subject ICA (melodic FSL), and denoising with trained FIX

  2. Made the group ICA with 20 dimensions based on the 3 groups at baseline only

  3. Identified 15 on the networks showing good correlation with the Smith template.

  4. Run dual-regression on ALL scans (baseline & follow-up) based on these 15 networks

  5. Compared the baseline data of relatives and controls across 4 hypothesized networks and identified some clusters showing significant group differences.

  6. Extracted the parameter estimates from these clusters from the second stage of the dual regression from ALL scans and imported them in SPSS for longitudinal analysis using mixed models (that can handle both missing datapoints, variable follow-up time and family relation).

  7. Longitudinal voxel-wise (either relatives vs controls, or patients vs controls) including only participants that had both scans

My questions:

i) Was it fine to leave out the follow-up scans for the group ICA (point 2 above)?

ii) I have been criticized for step 6 as being “double-dipping”. This is the main issue I would need help with. There are no similar studies out there identifying other clusters that I could use to extract parameter estimate from in my data.

In addition, the problem with point 7 is that it leaves out a lot of baseline data, does not account for follow-up time or family relation, and it doesn’t yield any significant findings. I think PALM may be used instead of randomise to account for family relation, but this is my least of the concerns.

Kind regards