Dear CONN-artists,

Is it reasonable to assume that second level group comparisons of reconstructed ICA maps between groups of unequal sizes may be biased by the estimation of the group level ICA components?

In other words - if 100 subjects (20 patient 80 control) are used to estimate 10 components followed by second level modeling comparing a particular component between subject groups, will observed differences be biased by the number of subjects in each group?

Is there a way (besides splitting your groups) to ensure unequal sample sizes at group ICA estimation does not effect  second level comparisons?

I've done a bit of research and think I have answer to this question :: second level effects are not biased by the estimation of group level ICA components since individual subject maps and time courses are reconstructed independently of each other.

Back reconstruction of individual subject time courses and spatial maps from the group level estimate is done by partitioning the mixing matrix that was estimated during the ICA procedure.  GICA1-3 back reconstruction techniques generally only involve the parameters of the mixing matrix that are specific to that subset of the concatenated temporal series that corresponds to an individual subject.

See original Calhoun 2001 or Erhardt 2011 for a more detailed evaluation of back reconstruction methods.