Dear Volkmar, Torben and Manfred,
We specify the same settings for variance in both models but we did used different images for estimating the two different models. However see the details before about how we estimated the two sets of images. (GRACIAS!)
We are analyzing a complex set of data that has the following design structure,
Pre: Participants see their own face and a stranger's face across 3 emotions (happy, neutral and sad)
Intervention: Participants attempt to change their Bilateral Amygdala and Hippocampus (ROI) activity via a neurofeedback procedure:
Condition 1: See the Self face and increase ROI.
Condition 2: See the Other face and count backwards.
Post: Participants see their own face and a stranger's face across 3 emotions (happy, neutral and sad).
We are interested in analyzing the data Pre versus Post, and we used a Full Factorial Design with 3 factors: Time, Face and Emotion.
The columns in the attached matrixes are as follows 1 Face a= self, b= other, 2 Emotion a=happy, b=neutral, c=sad.
Column 1: Pre, 1a, 2a
Column 2: Pre, 1a, 2b
Column 3: Pre, 1a, 2c
Column 4: Pre. 1b, 2a
Column 5: Pre, 1b, 2b
Column 6: Pre, 1b, 2c
Column 7: Post, 1a, 2a
Column 8: Post, 1a, 2b
Column 9: Post, 1a, 2c
Column 10: Post. 1b, 2a
Column 11: Post, 1b, 2b
Column 12: Post, 1b, 2c
The first level analysis contrasts that we used to create these second level full factorial batches are coming from the same source in terms of pre-processing files.
The only difference is that we calculated the first level contrast twice (using the same pre-processing data and initial fmri factorial specification ) and for one of the sets of first level analysis batches (after using the same initial fmri factorial design specification) we calculated more follow up t contrasts. That was all.
However when we run two parallel second level analysis with those two sets of first level contrasts we are getting very different results because SPM 12 seems to be choosing to model the data differently in terms of covariance structure:
Batch 1 in the figure we attach is the batch using images derived from 1st level batches where we calculated less t contrasts. SPM created a covariance matrix that assumes higher collinearity between column 1 an column 7 (i.e. Pre Happy Self Face, and Post Happy Self Face). It gives a value of cosine = -0.75.
Batch 2 in the figure we attach is the batch using images derived from 1st level batches where we calculated more t contrasts. Using the same initial Batch 1 but loading the different images, it seems that SPM 12 created a different covariance structure. Cosine = -0.08 (more orthogonal compared to Batch 1)
In addition to that, we had also looked into the SPM.mat file for both of the second level batches. We took a screenshot of the SPM.xX.W values for both batches. If you would like, we could send you the SPM.mat batches for you to look at.
Our questions are:
1. Why is SPM 12 making such different decisions with regards to covariance structures?
2. Given our design which model is more appropriate Batch 1 or Batch 2?
3. We cannot run a flexible factorial design without loosing one of our conditions (in order to model participants). Is that still ok?
One of our colleges is adamant about running a flexible factorial but that means that we will lose either emotion or self in the design, We think that the best model is Batch 1 because it assumes more collinearity between the Pre and the Post conditions, what is your opinion?
Warmly yours,
Karina
<design matrix information SPM_xX_W.jpg><matrix and covariate structures.jpg>