Hi,
If you are interested in differences in the relationship with the covariate in the different groups then you need to split up your covariates into different EVs (to allow each EV to fit a separate slope per group). So your EV4 would split into three, and the same for EV5. Therefore your new design would need 9 EVs. In addition, you need to demean the covariate values prior to putting them into the GLM. This needs to be done once for each covariate and *not* once for each EV. That is, take the values for covariate1 (as a single set of values), calculate the mean of this set, and subtract this mean from all the covariate1 values to get the demeaned values. Then split the demeaned values into the three groups and put these into the three EVs for covariate1. Do not demean each EV separately.
Once you’ve done this you’ll be able to look at differences in the slopes between groups (e.g. a contrast of 0 0 0 1 -1 0 0 0 0 would give the difference in the slopes for covariate1 between groups 1 and 2 - assuming that EVs 4-6 are the split versions of covariate1). If you want to then know if there is any change between any of the three groups then you can use an F-test: e.g. 0 0 0 1 -1 0 0 0 0 and 0 0 0 1 0 -1 0 0 0 would cover any difference for covariate1 slopes between any groups. Similarly, you can use an F-test to see if either covariate caused a difference, or you could use contrast masking to see where _both_ covariates caused a difference (although these become very underpowered when there is strong correlation between the covariates, whereas the F-tests are far less sensitive to that). More information about F-tests and contrast masking can be found in the wiki and the FSL Course material.
I hope this helps.
All the best,
Mark
> On 27 Nov 2014, at 15:23, Charlotte <[log in to unmask]> wrote:
>
> Hello.
>
> I would like some advice regarding a design matrix please.
>
> I have some longitudinal, repeated-measures data which I am running VBM on. In my study, I had three populations, and collected two behavioural measures from each of these. I am using the difference between acquisition, within participant, as the input for Randomise to avoid the issue of exchangeability with repeated measures.
>
> I am interested in looking at the relationship between Group and the behavioural measures. Amongst some of the analyses, one question of interest is to look at whether we can identify the basis of a (behavioural) correlation that exists between the two behavioural measures and, further, if this differs across the groups.
>
> I can see how to compute the simple contrasts for the mean effects, e.g if i had 6 subjects, 2 in each group (note, these are just made up values)
>
> design matrix:
>
> Group1 Group2 Group3 Cov1 Cov2
> 1 0 0 10 12
> 1 0 0 12 10
> 0 1 0 20 15
> 0 1 0 18 12
> 0 0 1 9 10
> 0 0 1 10 12
>
> contrasts:
>
> group1 1 0 0 0 0
> group2 0 1 0 0 0
> group3 0 0 1 0 0
> cov1 0 0 0 1 0
> cov2 0 0 0 0 1
>
> (plus reverse contrasts)
>
> The difficulty I'm having is how to identify which brain regions, if any, are correlated across the two measures and whether these differ according to groups.
>
> I'd be grateful of any advice.
>
> Thank you.
>
> Charlotte
|