Hi Julia,

It may not make much difference at the end: adding them as nuisance absorbs known effects (i.e., the other ICs) that otherwise would go into the residuals of the GLM of the 1st and 2nd stages. However, these residuals aren't used for the statistical inference using randomise later. The effect on the parameter estimates for these stages (i.e., subject-specific time courses and spatial maps) may be minimal given that the ICs are, as the name implies, independent. That said, adding these shouldn't hurt either.

All the best,

Anderson

On 5 November 2015 at 14:57, Julia Schumacher <[log in to unmask]> wrote:

Dear FSL experts,

I have a few questions regarding Melodic and dual regression:

I am looking at resting state network differences between patients and healthy controls using Melodic with 70 components (estimating the maps from the control subjects only) and dual regression to observe group differences.

In other studies there seem to be two different approaches to what components are used as input for the dual regression:
1. Use all ICs (including those identified as noise) for dual regression
2. Delete noise components and only use identified network ICs for further analysis

Can someone please explain what exactly the difference is between the two approaches? I.e. what difference does it make statistically if the noise components are included? Which one would you recommend to use in my case?

Any comments are very much appreciated!
Thanks,
Julia