Hi,
Instead of using two EVs for the groups, just use one with 1s for group
1 and -1s for group two. So, if you replace the zeros of your EV1 with
-1 and delete your EV5, the contrasts would be:
testing group 1<group 2: -1 0 0 0
testing group 1>group 2: 1 0 0 0
good luck,
wolf
On 05/07/2013 06:21 PM, R.P. Hylkema wrote:
> Dear FSL experts,
>
> I compare 2 groups of Parkinson's Disease patiƫnts with a vertexanalysis. I want to correct for some variance in the results caused by gender, age and disease duration - these three variables are my covariates. The constructed designmatrix is as follows (some random data, just to give you an idea):
>
> group EV1 (group 1) EV2 (gender) EV 3 (disease duration) EV 4 (age) EV 5 (group 2)
> input 1 1 0 -0.3333 4.5 1 1
> input 2 1 0 0.6666 -1.5 -10 1
> input 3 1 0 -0.3333 6.5 4 1
> input 4 1 1 0.6666 -5.5 1 0
> input 5 1 1 -0.3333 -3.5 8 0
> input 6 1 1 -0.3333 0.5 -4 0
>
> The actual number of patiƫnts is 125.
> I demeaned my covariates as showed in the example design.
>
> contrasts are as following:
> testing group 1<group 2: -1 0 0 0 1
> testing group 1>group 2: 1 0 0 0 -1
>
> I got the following error when trying to save the design:
> ''Problem with processing the model: Warning: at least one EV is (close to) a linear combination with others. You probably need to alter you design. (design matrix is rank deficient - ratio of min:max eigenvalues in SVD of matrix is 7.29781e-17) Contrasts involving these combinations will be set to zero.''
>
> I tried running the analysis anyway, but it changed all the values of EV2 (gender) to zeros.
> What is causing this problem and how to resolve this matter? I read somebit about rank deficiency and someone stated it could be caused by multi-collinearity. But I tested for multi-collinearity in SPSS and it didn't showed anything significant.
>
> All the best,
> Ruben
|