 |
 |
HiOn 20 Mar 2012, at 08:22, Kirstie Whitaker wrote:Hi FSL folk,
In our analyses of the effects of cognitive training on white matter we're running everyone through TBSS and using randomise to investigate our various questions. We also have some regions of interest that we'd like to analyze. My first question is 1) can we analyze this data using standard statistical tests (regressions, ttest, anovas and the like) in a standard statistics package, or should we be using permutation tests (such as randomise) as we do in the voxelwise analyses?
We do recommend permutation testing for TBSS - there's a good chance that parametric stats will not be valid.Given that I suspect the answer to the first question is - you should use permutation tests - then our next step was to write our roi values into a .nii.gz file (using a NiPy tool akin to fslascii2img) such that our data was 4D: 1 x 1 x number ROIs x number subjects. We then ran this 4D data set as the input to randomise and used the same design files we had used in our TBSS analyses.
My second question is then: does this make sense? Is there anything we're missing in creating these files? Obviously we aren't using the --T2 option as we would have done in TBSS (because that wouldn't make sense!) but we are using the -x option for voxelwise corrections for multiple comparisons. Which leads me to question 3) what is that correction? How would we describe it in our methods section?
Yes - this is exactly right - it's a valid "voxelwise" (in this case ROI-wise) test with multiple comparison correction giving control of family-wise error - and in general will be less conservative than Bonferroni correction. Make sure you use the *corrp* p-value output.FINALLY, we've run these analyses and some of our results look very sensible, but sometimes the values we input (eg for L1 and L23) are too small and Randomise says the mask is empty. I looked through the list and saw that randomise doesn't like "timeseries" which look constant so it doesn't test them. If I were to multiply all of our L23 values by, say, 1000 would this a) get rid of the problem (hint: I tried it and it does allow randomise to run) and b) cause any additional problems?
Yes - this is a perfectly good fix for that problem.Cheers.
Sorry for the rather long email, I hope this makes sense!
Thank you
Kx
--
It's that time again! I'm riding from San Francisco to LA with AIDS/Lifecycle for the fourth time in June 2012.
Please donate anything you can to help me reach my goal: http://www.tofighthiv.org/goto/kirstie
-----------------------
Kirstie Whitaker
Doctoral Candidate
Bunge Laboratory: Building Blocks of Cognition
Helen Wills Neuroscience Institute
University of California at Berkeley
134 Barker Hall, MC 3190
Berkeley, CA, 94707
tel: 510 684 2456
web: bungelab.berkeley.edu
---------------------------------------------------------------------------
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
---------------------------------------------------------------------------