Specifically, are the components calculated for each run and session separately, or is their some concatenation at the session level?
Since the confounds file is produced for each run, I suppose there is no concatenation. The slightly odd thing about this approach is that done this way the components are not “matched” across runs - a_comp_cor_XX in run N and a_comp_cor_XX in run M could reflect different noise sources. In this case, removal of the top 5 (for example) comps in run 1 and run 2 actually removes different noise sources. I feel that estimating the components across sessions rather then runs may be a bit better fit for multi session experiments (like my own).
Any insight appreciated.