Thanks a lot for clarification. So, as this is not the reason, is there another step which fsl_glm is doing internally and by which it possibly scales the standard deviation of the residuals?
My approach simply was res*res^T/trace(P) with P being the residual forming matrix (computed from the time course matrx in dr stage 2). Actually, trace(P) ~ (number of time steps - number of components). With that I get z-scores with perfect correlation to the one from fsl, but always with a smaller scale in amplitude. Is there something else I am missing? Do you correct for autoregressive effects in the noise?
Cheers
|