Also, orthogonalizing your EV3 wrt EV1, and EV4 wrt EV2 is equivalent to
demeaning HPA separately within group, rather than across all subjects,
which is in general not an analysis you want to run unless you are
trying to remove nuisance variance in a rather particular (and non-
typical) way. The more standard approach would be to simply demean the
HPA regressor across ALL subjects, and then split it into separate EVs
if you want to model the possibility of a group*HPA interaction. See
recent post/comment on this from Tom Nichols.
Best,
-MH
On Thu, 2011-02-24 at 20:12 +0000, Eugene Duff wrote:
> Hi Hongyu,
>
> On 24 February 2011 19:49, Hongyu Yang <[log in to unmask]> wrote:
> Hello FSL list,
>
> I have a problem about orthogonalisation. I need to do
> correlation analysis for two groups (controls and patients). I
> created 4 EVs, EV1 for group 1 mean, EV2 for group 2 mean, EV3
> for group 1's HPA response, EV4 for group 2's HPA response.
> Based on the example of FSL webiste, EV3 was orthogonalised
> wrt the EV1 (group 1 mean EV), and EV4 was orthogonalised wrt
> the EV2 (group 2 mean EV), and then in Contrasts
> 0 0 -1 0 means the negative correlation of group 1; 0 0 0 -1
> means the negative correlation of group 2. Is this correct?
>
> I am a little confused about one thing. If wrt the group mean
> means exclude the effect of group mean for each group, then
> what will happen if I only make two EVs, EV3 and EV4 for HPA
> response, do not make EVs for group mean, do not
> orthogonalise them wrt the group means, will that influence my
> correlation results significantly? For I think demean the
> group mean maybe exclude all the areas activated in group
> mean, if some areas were activated in group mean, and also
> showed some correlation to HPA response, I will miss those
> areas. Is my concern reasonable?
>
>
>
> I'm not exactly sure I follow you, but if you exclude group mean
> regressors and do not demean, then your HPA regressors will fit to the
> mean, even if there is no correlation in the data with the variations
> in HPA. Thus you could not interpret significant regions reflecting
> correlation with HPA response. Demeaning does not exclude those areas
> showing a mean effect, it just allows the constant regressor fit to
> the study/group mean in those regions, and the HPA regressor to
> account for the variability around the mean.
>
>
>
>
> Eugene
>
> Thanks,
> Hongyu
>
>
>
>
|