Actually it is difficult for me to give a good answer to this one as I
do not think the global AR(1) is a good model anyway. But if the idea
is, that we only need to correct the residuals where we want to make
inference, then I would specify the regressors you want to draw
inference as covariates of interest and all others as covariates of no
interest.This will restrict the mask to the area where your effect of
interest is present. The trade of is this: If you think a global AR(1)
model is adequate of modeling the noise you would include as many voxels
as possible to get as precise an estimate as possible of this global
process. If on the other hand you do acknowledge that there is a
significant spatial structure in the non-whiteness of the noise then you
would like to use a more local estimate. Given that SPM2 does only use
ONE estimate of non-white noise the only way around this in SPM2, is to
restrict the mask in which the AR(1) process is estimated. In the future
I hope SPM will use regional AR(n) estimates :-)
Best
Torben
Cyril Pernet wrote:
> Dear Torben, thank you for all your comments
>
> I still have a question (two indeed) and maybe you can help me on this
> (other comments are also welcome).
>
> In the previous e-mails, we looked at the effect of setting motion
> parameters as variables of interest or not. My question concerns other
> variables, as for example the answer of subjects.
>
> Let say I have three conditions A, B, C with C the case in which
> subjects have to answer (I don't want to look at that .. it's just to
> keep awake my subjects). My idea was to set C as variable of no
> interest... Do you think it's a good idea or is it better to put all
> conditions as variables of interest?
>
> Similarly, if I have 3 conditions A B and C = rest, is it better to set
> C as variable of interest, of no interest or to leave this condition
> implicit. Mauro Pesenti told me that in this case the implicit condition
> should be better ...
> > when possible and if one is not interested in fix/rest per se, it is
> better to leave it implicit.
> > it has something to do with estimation of the model that is over
> determined if all conditions are declared,
> > i.e., the 3d being totally predicted if you know the other 2
>
>
> Thank you
> Best,
> Cyril
|