 |
 |
Aga, What I did is: I demeaned the values in EV1 and also used D option fir > randomise (somehow I did it correct:). Yes, that's correct. > Then, I am not sure if I should model my mean or not? I did it both with > and > without the EV with ³1² only, results look roughly similar, what does it > change and what does it mean? > If you use the -D option, there is no mean in the data to model, hence including it doesn't hurt anything, but it is fitting nothing. > In group comparison, I dont assume the same variances. Randomise, as it is just using a vanilla GLM, assumes the variance is the same for all scans. In correlation within one group I may. Is it connected to this mean > modeling? No, totally separate issues. So is this single EV all I need (yes, my values are > demeaned and I used D option) or do I need to model my mean ? And then I > have 2 EVs, does it change anything with demeaning? As above, "-D" ==> No mean to model > Also, the third contrast "x(y)" is probably not meaningful... it's testing > if > > the average of the two regression coefficients are zero. > > > Hmm, I in fact thought this is the main contrast to look at:(. So is it > then > the first contrast that gives me info I am looking for (how TBSS values > correlate with x, controlling for y)? > Multiple linear regression automatically does this. E.g contrast [1 0] is assessing the evidence for x while controlling for y. -Tom ____________________________________________ Thomas Nichols, PhD Director, Modelling & Genetics GlaxoSmithKline Clinical Imaging Centre Senior Research Fellow Oxford University FMRIB Centre