[color=#000000]Hi Shady,[/color]
[color=#000000]That is an interesting question. First, in practical terms, you actually may, if you want, select the subjects that you want to include in your group-ICA analyses in CONN by, after pressing the 'Done' button, unchecking the 'all subjects' checkbox there and selecting the desired subjects (or if using batch commands simply by seting the batch.subjects field). The non-included subjects will in this case be labeled as missing data and automatically disregarded from any second-level analysis of ICA results. [/color]
[color=#000000]Then, regarding whether one "should" exclude entire subjects from ICA analyses, that is really still an open question. In general I typically recommend using scrubbing to remove only portions of "bad" datasets while savaging the remaining good data, and only in very extreme cases, where the total number of reamining "good" scans might be below a minimally-acceptable level (e.g. less than a few minutes) to then remove the entire-subject dataset. ICA, as well as standard second-level GLM analyses, are generally considered rather robust methods in the presence of noise and heteroscedasticity in subject-level connectivity measures, and the sensitivity of these analyses more directly relates to the number of subjects available than to the accuracy/noise of the subject-level measures. This means that, in general, if you want to be more conservative I would recommend removing "more portions" of each subject's data (e.g. using a more conservative thresholds in ART) rather than removing "more subjects" (e.g. removing subjects based on a more conservative max-motion or min-valid-data threshold), but of course recommendations will vary. [/color]
[color=#000000]Hope this helps[/color]
[color=#000000]Alfonso[/color]
[color=#000000]
[/color][i]Originally posted by Shady El Damaty:[/i][quote]Hi fellow CONN-artists,
I just realized today that group ICA concatenates all of the subjects in the data set when computing independent components, including those that may be excluded at the second level due to a high number of outlier scans.
Is it important to exclude subjects that have a large number of outlier scans (say >20%) from the group ICA analysis? I can see how having more subjects in general (irregardless of time points) may help fill in the distribution of variances and facilitate better estimation of ICA sources. However, I'm not sure how having subjects with drastically less scans will impact the number of components one can estimate..
Anyone have thoughts on this subject?[/quote]