On Fri, Mar 09, 2012 at 12:49:34AM +0000, In Kang wrote:
> As I knew that ICA maximizes the non-gaussianity of the components.
> When I use ICA to extract components from brain image to compare two or
> more independent component, can I use t-test which assumed the normality
> for assessing the significance of activated/ deactivated voxels in SPM?
It is a fairly reasonnable assumption to model the null as a Gaussian,
centered on 0 and with unit variance (as long as the variance of your ICA
maps is 1). See:
Varoquaux et al, ICA-based sparse feature recovery from fMRI datasets,
ISBI 2010
http://hal.inria.fr/hal-00489506/en
For statistical arguments in favor of this assumption,
Gaël
|