On Fri, May 13, 2011 at 7:03 PM, Matthew Webster <[log in to unmask]> wrote:

Hello Michael,
I've copied the text from the 4.1.8 randomise documentation below:

The pre-FSL4.1.8 version of randomise had a bug in the -e option that could generate incorrect permutation of scans over subject blocks. Incorrectly permuting scans over subject blocks does two things: Under permutation, it will randomly induce big positive or negative effects (inflating the variability in the numerator of the t statistics over permutations), but it will ALSO increase the residual standard deviation for each fit (inflating the denominator of the t statistic on each permutation). Hence it isn't clear which will dominate, whether the permutation distribution of T values will be artifactually expanded (wrongly decreasing significance), or artifactually contracted (wrongly inflating significance).

With some simulations, and with some re-analysis of real data, it appears that the effect of the bug is to always to wrongly deflate significance. Thus, it is anticipated that results with the corrected code will have improved significance.

Hope this helps,

Matthew

> Hi Matthew,
> Can you please elaborate as to what the issue with the -e option in
> randomise is? In particular, are results using the -e option
> potentially wrong, and if so, under what conditions?
>
> thanks,
> -MH
>
> On Thu, 2011-05-12 at 09:52 +0100, Matthew Webster wrote:
>> Hello Raphael,
>> This may be due to a current issue with the -e option in randomise - we will hopefully be releasing FSL 4.1.8. with a patched version of randomise in the next few days, I would strongly recommend rerunning the analysis with this new version.
>>
>> Many Regards
>>
>> Matthew
>>
>>> Dear FSL-experts,
>>>
>>> I`m very surprised by the difference between two tbss analyses i ran. In fact, I forget to include the DTI images of one subject into a longitudinal analysis with 3 points in time and 2 groups. Thus I reran the analysis with altogether 13 subjects in each group and surprisingly, results differed extremly. Testing for a main effect of time and an interaction of groups with all 26 subjects included in the model all the skeleton shows up when looking at tfce-corrected fstat-images at p < .05 (> .95 in fsl-terminology). However, the same results look way more plausibel (two cluster at the expected location) when I only consider 25 subjects, that is to say with the subject that I initally forgot to incorporate in the analysis. I visually inspected the results of the preprocssing steps before subjecting data to tbss analysis as well as the results of the tbss analysis (e.g. the all_FA_skeletonized image, etc.), results look fine.
>>>
>>> I attached pictures of the design matrices and contrasts as well as the exchangeability block (EB) or to be precise ".grp" file/s. As can be seen from the design files, I only added one subject and of course adjusted the EB file accorindgly. The first 26 columns of my design matrix code subject factors, the last four columns the experimental group and control group at T1 and T2. The columns are respectively labeled in the design matrix. The last four columns look somewhat scrabbled which is due to the odering of subjects in the all_FA file which makes the membership of subjects to groups most likely only obvious to me.
>>>
>>> Randomise was used with the following commands:
>>>
>>> randomise -i all_FA_skeletonized.nii.gz -m -o ... -m mean_FA_skeleton_mask.nii.gz -d ... .mat -t ... .con -f ... .fts -e ... .grp --fonly -n 5000 --T2 -V.1
>>>
>>> I would vey much appreciate any help on this as I´m a little bit desperate on how to explain this differences and on how to proceed.
>>>
>>> Thank you very much in advance.
>>>
>>> Raphael
>>>
>>>
>>> <13vs12.zip><13vs13.zip>
>