Hi Loreen,
check out this page: http://mumford.fmripower.org/mean_centering/
The problem is probably that the mean effect in the group (the [1 0]
contrast) is at covariate = 0, when you do not center- it's more
interesting to look at the mean effect at the mean covariate-value
(which you get when subtracting the mean). So centering is the more
"correct" thing to do.
Best, Brian
Den 09-01-2015 kl. 11:00 skrev Loreen:
> Dear SPMers,
>
> I would greatly appreciate any comments you might be able to offer on the issue of centring continuous (nuisance) covariates in a one sample t-test (2nd level).
>
> We are testing whether activation patterns contained in the contrast images derived from 1st level analysis are significantly different from 0 across the group (one sample t-test), whilst controlling for effects of age and gender.
>
> Before introducing the two covariates, the activation patterns are in line with de/activation patterns suggested in the literature. When I subsequently introduce the two covariates without centring age, pretty much all of the significant (after FEW correction) clusters disappear. When I test the association between the individual covariates and mean activity however, no significant associations emerge (i.e. intensity change is not associated with age or gender). If I centre the continuous age variable, a few clusters are still taken out from the effects of gender and age (which is what we want to see), but overall the remaining patterns are more in line with what I would have expected and has been suggested by previous research.
>
> Would you be able to offer some insight into why centring of covariates in a one-sample t-test makes such a massive difference? Also, not wanting to cherry-pick which results to continue with, can you comment on the validity of the two test results (i.e. the different patterns obtained with and without centring age)?
>
> Thank you very much for your help!
>
> Kind regards,
> Loreen
|