Hi Kirstie,
As Steve says, orth'ing behaviour with respect to age won't change the
contrast for behaviour, and there's no problem with correlated EVs in
recent versions of randomise.
If they are strongly correlated though, you might find that you have
very little power to detect significant associations with behaviour
when adjusting for age. In this situation, people occasionally want to
orthogonalise the nuisance (age) with respect to the interest
(behaviour), but there is usually very little justification for that.
A more valid alternative is simply to look at the simple regression
for behaviour alone, and present results for both the simple and
partial (adjusting for age) correlation, discussing the implication of
any differences between the results.
Another thing you could try would be to determine regions of
significant association with either of the covariates, using an
F-contrast that combines [1 0] and [0 1] t-contrasts, and then to
colour these significant voxels based on which of the partial
correlations for age [1 0] and for behaviour [0 1] were strongest.
You'd have to be careful not to imply *significant* differences
between the associations, but if you explained that the covariates
were jointly significant in region(s) X with behaviour explaining
(non-significantly) more of the variation in the data there than age,
then I think that would be valid, and might well be helpful. Of
course, if you do find significant age-adjusted associations with
behaviour, then that is much stronger and simpler to interpret, but if
not, I hope some of this helps!
Best,
Ged
2009/12/11 Kirstie Whitaker <[log in to unmask]>:
> Hi Jedi Masters!
>
> I have a TBSS analysis which is looking for regions of white matter which
> explain individual differences in behavior after the effects of age are
> removed. My design matrix has two columns: age and behaviour (contrast: 0
> 1). They're demeaned (and I use the -D flag in randomise) and my question
> is: Given that age and behavior are very strongly correlated, should I
> orthogonalise behaviour with respect to age in order to see these individual
> differences? Can I even do this design given the multicollinearity of
> behaviour and age? If not what should I do instead?
>
> Thank you so much for all your help!
>
> Kx
>
> --
> Kirstie Whitaker
> Doctoral Candidate
> Cognitive Control and Development Laboratory
> Helen Wills Neuroscience Institute
> University of California at Berkeley
> tel: 510 684 2456
> web: bungelab.berkeley.edu
> blog: http://bungelab.blogspot.com/
>
|