Hi Jeremy & David,
Just wanted to follow up on this and note that this question was recently covered by a recent thread:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1207&L=fsl&P=R19796&1=fsl&9=A&J=on&d=No+Match%3BMatch%3BMatches&z=4
While the degree of motion-related noise may have a small impact on the overall components produced by the tc-gica, it is likely to have a much greater effect on the results of the dual regression (stage 2); our paper describing this effect came out around a year ago. As Christian suggested in the thread referenced above, it probably makes the most sense to include motion regressors & possibly spike regressors in dual regression stage 2. Though we have only tested it extensively for seed and network analyses (see our paper in press at NeuroImage on this topic for this data), I suspect that higher order motion parameters (e.g. temporal derivatives, quadratic, quadratic of derivative a la Friston '96) and spike regression will provide a very substantial benefit over the standard 6 parameters alone. This can be easily implemented in FSL5 using the extend motion parameters feature and fsl_motion_outliers. As a final caveat, however, I would warn that even such an approach is unlikely to remove all artifact, and therefore for clinical studies where motion is related to the effect of interest (group, sx, etc) it may make sense to finally add MRD or some other summary metric of motion as a confound variable in the final DR stage 3.
Cheers,
t
PS. This all may be rendered moot if the soon-to-be-released (?) automated denoising of data using first-level ICA provides better control of motion artifact in the data than such regression-based approaches.
_______________
Hi Jeremy,
I would regress them out prior to running the gica (you can use unconfound for this) so that those confounds don't have a chance to affect the ICA results. However, I'm not entirely sure how possible it would be for the precise pattern of an RSN to be shaped by individual differences in a nuisance variable (e.g., head motion). It seems possible that such a difference could bias subsequent dual regression results in cases where you were examining a correlated factor (imagine a case where one group moved more than another group). I'm not sure if including subject-level nuisance regressor would completely mitigate that concern...
See fsl_motion_outliers to remove spikes.
Cheers,
David
On Dec 11, 2012, at 3:00 PM, Jeremy Elman wrote:
Hello all,
I was wondering what the best way to account for subject specific nuisance factors is when running a dual regression analysis. Is it preferable to regress out motion parameters first and running the group ICA on this clean data, or to include these subject specific motion regressors after the first stage of DR? I guess my main question is whether doing one versus the other affects the group ICA in a significant way.
I've also seen discussion that regressing out spikes may introduce some problems to the group ICA step. If I would like to regress out particular volumes due to motion or RF spikes, would this change the recommended process above?
Thank you for your help!
Jeremy
--------------------------------------------
David V. Smith, Ph.D.
Postdoctoral Fellow, Delgado Lab
Department of Psychology
Rutgers University
Newark, NJ 07102
--------------------------------------------
|