Dear Donald, 

Thank you for your reply.

I have some follow up questions that I placed inline after your replies. Could you please take a look at them?

Op 24 jan. 2014, om 13:29 heeft MCLAREN, Donald <[log in to unmask]> het volgende geschreven:

Please see inline responses below.

On Fri, Jan 24, 2014 at 10:19 AM, Vincent Koppelmans <[log in to unmask]> wrote:

Dear FSL experts,

I have longitudinal VBM data of 8 subjects. Each subject completed 7 measurements. I pre-processed using the data using the longitudinal VBM8 pipeline and now I would like to compare time point 1 and 5.

SPM offers the option to use a flexible factorial design as well as the paired t-test design. I compared both models in SPM using the following design/contrast. In addition I tried to analyze the VBM data in FSL using Randomise because of the advantages that non-parametric testing offers when analyzing small numbers of subjects.

The results are as follows:<Designs_small.jpg>

(a higher resolution version of this image can be download here: https://dl.dropboxusercontent.com/u/6747155/Designs.jpg)

Blue to Light-blue colors indicate an increase in GM volume (data are modulated); Red to Yellow colors indicate a decrease in GM volume.

I have a couple of issues that I would like some help with:

1) I tried to model a design in FSL that is similar to SPM's flexible factorial design. In the design matrix I therefore entered all time points and I used the design.grp to make sure within subject scans were treated as blocks by randomise. Is this a valid approach and is this design more or less comparable to SPM's flexible factorial design?

>>> No. There are two key differences between the SPM and FSL models: 

(1) In your FSL model, you've not coded each of the timepoints. You'll need 1 column for each timepoint. Then you can run randomise on the contrast of interest. 

(2) In FSL, the covariances are assumed to be 0; which is not the case in SPM. 

If you change the models to have the same number of columns, I suspect that you will get more similar results.

I previously created such a model in FSL, but it was rank deficient:

(see also: https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=fsl;dd87429.1311)

It was then suggested to create contrast files and use the contrast files in a higher level analysis.

How could I recreate a design in FSL Randomise that has the same number of  columns in the design matrix as the SPM flexible factorial design?

 

2) Is it correct to assume that the large differences in significant clusters between the paired t-test design and the flexible factorial/random effects design is because the flexible factorial/random effects design pool the error terms across all time points and thereby are better able to remove noise from the data?

I wouldn't say that they are "better able to remove noise". I would say that you are pooling the error term across all timepoints.

 

3) The output of FSL Randmise shows much larger significant areas than SPM's output. Could this be ascribed to the non-parametric nature of Randomise and/or TFCE? What would be other potential explanations?

Yes and No. Non-parametrics stats work for small samples as in your case. However, I suspect that your FWE was the voxel-wise FWE. If you used a cluster FWE with an voxel-wise p of .01 or .005, then I think you would get closer to the FSL results. 

Alright. But TFCE with voxel wise FWE is preferred over using cluster wise analysis with FWE, right?

Thanks for your help.

Best,

Vincent