Hi,
I agree with Ted and he's making a very important point that is worth re-iterating: in ICA you do not need to remove/explicitly model the global mean because it is implicitly modelled by the superposition of the different components that are modelled simultaneously. Rephrased this means that in seed analysis the removal of the global mean is a poor-man's approach to turning a univariate technique (single seed analysis) into a pseudo-multivariate approach. The global mean time course itself is a simplistic way if modelling the existence of all the other networks and other effects that are otherwise ignored in univariate seed analysis. Modelling the global mean is a fudge technique to capture in a single regressir what really ought the be expressed properly in a multivariate way. It's for this reason that we do not recommend it in ICA - it's hopefully already been dealt with by modelling all structure in the data simultaneously.
Hth
Christian
Sent from my iPhone
On 12.12.2012, at 08:55, Ted Satterthwaite <[log in to unmask]> wrote:
> Hi Jeremy,
>
> The global/wm/csf regression is somewhat controversial, I agree. However, in our experience, global signal regression seems to be clearly beneficial for control of motion artifact in traditional seed or network analyses, as motion causes large drops in BOLD signal across the brain parenchyma in a way that is effectively captured by the global timeourse. The benefits are less clear for ICA. Previously Christian and others have not recommended it prior to tc-gica (i.e., component generation). Furthermore, though I have not investigated it myself, I suspect that some of the advantages of global signal inclusion would be obviated by the fact that multiple component timecourses are included together as part of stage 2 GLM in dual regression, as the shared variance among these timecourses is likely to be similar to the global signal itself. But if you are curious, this is quite testable.
>
> cheers,
> t
|