Dear Christine,

I would not worry about this.

There are often several ways of setting up equivalent designs.

It is the overall contrast of interest that matters, so if you see consistent, sensible results for that then this should be fine.  You are definitely looking at any changes related to variation in dB level across your group.  The absence of individual t-contrasts might just mean that the differences to the arbitrarily chosen reference level are not strong enough to reach significance on their own, but either taken all together, or when combined to look at differences between other pairs of levels (which is what the F-test will be doing), these differences are then strong enough to reach significance in the F-test.  I would not worry about interpreting the individual t-contrasts. Also, stick to Anderson's advice on the best way to set things up for randomise, as it is important to make sure that the permutations are done correctly.

All the best,

Mark

 

On 29 Oct 2013, at 11:08, Christine Wyss <[log in to unmask]> wrote:

Dear Anderson, 

I am a bit confused, because I modeled the design as described in the randomise FslWiki (respeated measure ANOVA) and as far as I can assess this differs from the "factor effects model". To specify, I would like to figure out which area is involved in intensity variation (A=40dB, B=50dB, C=60dB, D=70dB, E=80dB) within the group. Does my model hold for that?

With the adapted model I get  significant results in the F-statistic that are exactly the same as in the old model. But none of the t-maps do reach significance. How could that be and how can this result be interpreted? Does an alternative approach exist?

Thank you for your help.

Christine

Am 24.10.2013 um 15:40 schrieb Anderson M. Winkler:

Hi Christine,

The same p-values in the fstat files or the same actual F-statistic? But regardless, good you found the same result, and as for the t-stat, I'd stay with the 2nd. For reference, this is the "factor effects model", described in the manual in the link below, with the first n-4 columns taking the place of the intercept.
http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#ANOVA:_1-factor_4-levels__.281-way_between-subjects_ANOVA.29

All the best,

Anderson

On 23 October 2013 16:05, Christine Wyss <[log in to unmask]> wrote:

Hi Anderson, I appreciate very much your help.

I changed the matrix as per your suggestion and expected a different fstat result compared to the old version (containing all 0s).  But the fstats results look the same, only the tstats (A-E) differs. I am relatively new to FSL/randomise and I am not sure if i interpreted your answer correctly. Did you mean that this change would have an effect on the fstats or the tstats?

Best regards, 

Christine

Am 11.10.2013 um 21:02 schrieb Anderson M. Winkler:

Hi Christine,

One more change: the cells in the last 4 columns and last 16 rows should all contain -1, otherwise contrasts won't produce the equivalent to a comparison between means of all 5 conditions, as I guess is what you'd like to do.

All the best,
Anderson

On 11 October 2013 14:30, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Christine,

Design and contrasts look fine to me. However, the .grp file you sent contains all 1s. Use instead one index per subject, so that the permutations happen within subject only and across the 5 conditions.

All the best,

Anderson

On 10 October 2013 14:56, Christine Wyss <[log in to unmask]> wrote:

Dear FSL experts,
I would like to compare 5 conditions within 1 group (n=16). I designed a 1x5 ANOVA model as in the attachment and modified the .grp file regarding the structure of the data. The contrasts are A-E, B-E, C-E, D-E. The randomise function would look like:

randomise -i merged_files.nii -o ANOVA_results -d 16x5.mat -t 16x5.con -f 16x5.fts -e 16x5_neu.grp -T -n 14400 -v 5

I would be very grateful if any of you could tell me if the design and contrasts are right.

Best,
Christine