Hello there,
I have a pretty basic question about running group-based ICA analysis with dual regression that I was hoping someone could help me with.
I would like to identify networks that are over/under expressed across 3 groups of subjects (1 control and two patient groups). When running MELODIC using multi-session temporal concatenation ICA across the entire sample (N = 63), with automatic dimensionality estimation, the output included a total of 29 components. Of these, approximately 9 seem to be biological.
I would like to run dual regression using this full dataset, and estimate group differences via the appropriate specification in the design matrix. However, I am concerned that differences across the 3 groups in the spatiotemporal structure of the data might be masking some components, given the relatively small number of non-artifactual components. I subsequently conducted an ICA separately for the 3 groups, and observed a much larger number of components in the control participants (N = 56) versus the meth dependent (N = 34) and psychotic individuals (N = 36), despite the fact that the imaging parameters are the same for all groups, and the sample sizes are comparable (+- 20 subjects per group).
I am therefore thinking that perhaps it would be a good idea to denoise the data on an individual-subjects level prior to concatenating all the data for the dual regression, as this should at least control for differences in the number of artifact components between the groups. My questions are, (a) whether this is necessary, and (b) if there is a standard way of identifying additional networks that are present in one group but not the others. According to my limited understanding, dual regression would only identify differences in networks that are present across the data from all subjects entered as input into MELODIC.
I would appreciate any help you can provide me with.
Thanks,
Jonathan
|