Hi Charlie,

Thanks for the link. This was calculated like this:

The Cohen's is: d = (x1-x2)/s

The t-stat for 2 independent samples with unequal size is: t = (x1-x2)/s/sqrt(1/n1+1/n2)

The relationship is direct: the second can be written as: t = d/sqrt(1/n1+1/n2), which is the same as t = d/sqrt((n1+n2)/n1/n2), which is the same as t = d*sqrt(n1*n2/(n1+2)).

In your case, the value of this constant is sqrt(81/18) = 2.12, that is, d = t/2.12.

However, in my original post I said multiply, but that's for t. For the d, it's to divide. I'll email a follow up on that thread and the person that originally asked.

All the best,

Anderson

On 26 June 2014 23:37, Charles Leger <[log in to unmask]> wrote:

Thank you again for you help Anderson,

Here is the link to the post regarding TBSS effect size for a 2 group test. The search string was just <effect size TBSS> in the JISC Mail archive

https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1402&L=FSL&P=R8499&1=FSL&9=A&I=-3&J=on&X=3585CA6FE0D7B13C5F&Y=cslfalcon%40gmail.com&d=No+Match%3BMatch%3BMatches&z=4

All the best,

Charlie

On Thu, Jun 26, 2014 at 6:13 PM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Charlie,

On 26 June 2014 20:26, Charles Leger <[log in to unmask]> wrote:

Thank you Anderson,

With the global FA 1-sample test for experts completed, L-R proved to be significantly different from zero and positive, t (8) = 18.9, p = 6.379e-08 (p < .0001), d = 6.29 (as you mentioned, Cohens’ does not seem to be appropriate for paired left, right hemisphere difference values). A separate 1-sample t-test for controls had a similarly highly significant outcome.  FA certainly differs from zero in both hemispheres, the sum of left FA is greater than right FA (L: R ratio of ~ 1.08), the means differ (left M = 0.5005947, right M = .4664487), and the inordinately high significance level obtained would appear would appear consistent with these measures.

 In addition, as L-R proved to be significantly different from zero and positive an overall leftward FA asymmetry is also supported. Is this a valid interpretation?

Interpretation depends on a number of other things that aren't included in the design. You may want to discuss that with the PI of your study -- if not you, of course.

 

Moving to the randomise results, the 1-sample (single group with sign-flip) tbss_sym script (corrp_tstat1 images and cluster output) results clearly show where, in the main tracts, L>0 and R>0 are lateralized at an the exploratory significance level of .10 (many coordinates were significant at p < .05), and these coordinates differ for left and right, given a voxel can’t be most lateralized to the right and left simultaneously. If the randomise results (peak FA) are extracted with fslmeants, I assume it would be correct to annotate plots, e.g. bar or box, of left and right  as significantly different at the obtained significance level?  

Also, you may want to discuss this with the PI of your study whether these further plots are useful to characterise differences. In any case, L and R separately are surely both >0 near everywhere in the brain, so I'm surprised that you run that contrast to begin with, and that it's not highly significant. Maybe you mean something different?

 

 Overall and in respect to the 1-sample tests, the global FA sums of left and right significantly differ and the lateralization of left and right significantly differ. The between-group randomise tbss_sym results also showed a significant difference where experts had greater leftward FA, and controls higher rightward FA at the same coordinate. 

In respect to an effect size here, I don’t understand the variation of Cohen’s d noted previously from the FSL mail group: multiply the tstat1 value (I assume the corrp_stat1 value) by sqrt(n1*n2)/(n1+n2). In my case, with 9 experts and 9 controls this value would be .5, or 50% of the tstat value which does not seem correct. While I don’t use Cohen’s d, I recall the formula for 2 groups was the difference between the means divided by the pooled standard deviation. Any suggestions here?

Could you indicate which mailing list post are you talking about? Thanks!

All the best,

Anderson

 

On Mon, Jun 23, 2014 at 6:30 PM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Charlie,

The 1-sample t-test tests whether the mean is different than zero. In this case, if there's no difference between L and R, the L-R should be not significantly different than zero (any difference is due to chance alone). The test is to reject the possibility of being zero (at a certain significance level).

All the best,
Anderson

On 23 June 2014 23:23, Charles Leger <[log in to unmask]> wrote:

Thank you Anderson, my responses are below: 

Yes, looks fine, but:

1) These values (218, 182, etc), are for images with voxels of 1mm in standard space. If you are using 2mm, this needs to be changed.

-> Yes, I used 1mm standard space

2) The 0 9 isn't really necessary. It could be 0 1 instead (as after collapsing the 4th dimension with -Tmean, there will be just 1 volume anyway).

-> change from 0 9 to  0 1 

3) Most important of all: open the L-R_bin_allLminRmask and see how it looks. If it is a mask that covers just one of the hemispheres, yes, it's correct. If not, then something will need to be fixed.

-> Yes for  the orginal 4D file and flipped masks covered one hemisphere only, and as expected each differed 

then yes, subtract within subject, and run a 1-sample t-test.

-> I have subtracted within subject, and will run a 1-sample t-test, after checking for relative normality

I assume that for the 1 sample t-test mu=0? 

On Mon, Jun 23, 2014 at 1:06 PM, Anderson M. Winkler <[log in to unmask]> wrote:

On 23 June 2014 17:41, Charles Leger <[log in to unmask]> wrote:

OK,

To arrive at the all FA on the left, I took the mean of the all_FA_skeletonised_left_minus_right.nii.gz file and used this as the mask (binarized but not thresholded) for the all_FA_symmetrised_skeletonised.nii.gz in fslmeants. Is this correct?

fslmaths all_FA_skeletonised_left_minus_right.nii.gz -Tmean -bin -roi 91 91 0 218 0 182 0 9 L-R_bin_allLminRmask
fslmeants -i all_FA_symmetrised_skeletonised.nii.gz -m L-R_bin_allLminRmask -o L-R_allfa_ss_bin.txt

Yes, looks fine, but:

1) These values (218, 182, etc), are for images with voxels of 1mm in standard space. If you are using 2mm, this needs to be changed.
2) The 0 9 isn't really necessary. It could be 0 1 instead (as after collapsing the 4th dimension with -Tmean, there will be just 1 volume anyway).
3) Most important of all: open the L-R_bin_allLminRmask and see how it looks. If it is a mask that covers just one of the hemispheres, yes, it's correct. If not, then something will need to be fixed.

 

0.4739182588 
0.5086818159 
0.5008845372 
0.5587585085 
0.4754576326 
0.4959170728 
0.5060840501 
0.4846208321 
0.5010291775

If this is correct, then I follow the same procedure for the inverted version of the all_FA_skeletonised_left_minus_right.nii.gz file, subtract L-R and do a 1-sample dependent t-test...

You can flip LR and repeat the same fslmaths command, or change the parameters of the option -roi (instead of 91 91, use 0 91), then yes, subtract within subject, and run a 1-sample t-test.

All the best,

Anderson

On Mon, Jun 23, 2014 at 7:01 AM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Charlie,

Yes, fslmaths, the -roi option. You can take the average of 4D file (-Tmean), binarise it, then use -roi option. This will be the mask then for fslmeants.

All the best,

Anderson

On 23 June 2014 00:13, Charles Leger <[log in to unmask]> wrote:

Thanks Anderson for your quick response.

Re

"You can run a global test if you calculate the sum of all FA in one hemisphere, then all FA for another hemisphere This can be done with fslmeants with a mask for each hemi (you may need to create one, and the fslroi command that you used some weeks ago should give you an idea on how to do it; you can also use fslroi to "delete" one hemisphere, then use fslstats to calculate the global FA). With the results (two values per subject), subtract one from another (say, L-R) and run a 1-sample t-test in any statistical or spreadsheet software. Ideally, we should use permutation, but the central limit theorem should guarantee some normality, so that parametric tests can probably be used safely."

In a previous email for calculating the LI, one step involved the use of fslmaths -roi on the LI quotient

fslmaths all_FA_symmetrised_skeletonised_LI.nii.gz -roi 91 91 0 218 0 182 0 9 allfaSSLI_L-Ronehem {retains just one hemisphere, so stats apply to only one side}

Here, I am attempting to arrive at the text fiel for FA in one hemishpere but it doesn't seem correct...

fslmaths all_FA_symmetrised_skeletonised.nii.gz -roi 91 91 0 218 0 182 0 9 L-R_onehem_roi
fslmaths L-R_onehem_roi -thr 0.90 -bin L-Rroi_th
fslmeants -i all_FA_skeletonised_left_minus_right.nii.gz -m L-Rroi_th L-RFA.txt

On Sat, Jun 21, 2014 at 5:06 AM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Charlie,

Sure, please see below:

 

However, in respect to the 1-sample test is it possible to test if total or aggregate left side FA differs from right side FA. Normally, I would run a paired t-test but doing so on the fslmeants extracted previously tested data presents, as you have warned,  a circularity violation. The tested data extracted with fslmeatns shows a L>R higher mean FA than R>L but a significant difference between the means can't be claimed without a t-test. 

You can run a global test if you calculate the sum of all FA in one hemisphere, then all FA for another hemisphere This can be done with fslmeants with a mask for each hemi (you may need to create one, and the fslroi command that you used some weeks ago should give you an idea on how to do it; you can also use fslroi to "delete" one hemisphere, then use fslstats to calculate the global FA). With the results (two values per subject), subtract one from another (say, L-R) and run a 1-sample t-test in any statistical or spreadsheet software. Ideally, we should use permutation, but the central limit theorem should guarantee some normality, so that parametric tests can probably be used safely.

 

Also, can effect size be extracted from the tbss_sym data  1-sample and between-group analyses? The difference between means  using fslmeants extracted peak values could server as non-standardized effect-size estimates but Cohen's d or r would be preferable.  I recall seeing a another researcher's correspondence inquiring about effect size and the response was, in part, that Cohen's d could be arrived at my multiplying the t-stat by the sqrt(n1*n2)/(n1+n2) but I don't know if this would apply to the tbss_sym data.

This formula looks like for a 2-sample t-test, and it's fine to use (as is Cohen's) for the group comparisons. However, for an effect size comparing both sides of the brain, the laterality index is probably a better choice. I'm not sure Cohen's work for paired data, as for the latero-lateral comparisons.

All the best,

Anderson

On Tue, Jun 17, 2014 at 8:47 AM, Charles Leger <[log in to unmask]> wrote:

Thanks very much Anderson, you have made the entire process very clear. 

All the best,

Charlie

On Tue, Jun 17, 2014 at 4:01 AM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Charlie,

On 16 June 2014 16:28, Charles Leger <[log in to unmask]> wrote:

Hi Anderson,

A point of clarification...

The 1-sample tbss_sym FA corrp_tstat1 results have shown tracts or areas where L>R significantly differ,

Not exactly. The 1-sample t-test is testing where the difference L-R is significantly different than zero and positive.

 

and all such fslmeants extracted values are positive;

This is a very likely output for fslmeants, given that the peak of significance is the one with the average farthest from zero, but still a coincidence. It would have been ok to have found a few negatives there.

 

ditto for the R>L extracted fslmeants values.

Same applies here. A few negatives would have been ok, still having the average the farthest from zero (towards the positive side, as the contrast is +1, not -1).

 

In addition, L>R area and structure coordinates differ from those the for R>L.

That's fine. A voxel cannot be the most lateralised to L and at the same time the most lateralised to the R.

 

The 1-sample test, with the  -1 sign-flip option, then is testing for the absolute asymmetry of L>R then, optionally for R>L on the inverted image.

Yes.

 

By contrast, the between group tbss_sym test (without the -1 sign-flip option), is, as you have clearly articulated, testing for bilateral differences in FA of structures between the groups

Yes

 

which accounts for the mixture of positive and negatives in fslmeants output.  Correct?

No, the last bit isn't right. There could have been a few negatives in the 1-sample t-test, although this is certainly less likely than in the between group comparison, in which the largest difference interval is more likely to include the zero, with a more even distribution of positives and negatives.

 

All the best,

Anderson

On Sat, Jun 14, 2014 at 2:05 PM, charlie <[log in to unmask]> wrote:

L>R as well as R>L was tested in the  tbss_sym 1-sample test of FA asymmetry. If the peak FA data is extracted with fslmeants (with use of a thresholded and binarized corrp_tstat1 mask), I assume that although the 1-sample option tests if the mean > 0,  a two-sample t-test (e.g. Welch, Wilcox)  could be run to test left vs right side FA? This is somewhat redundant but supplements the randomise and cluster report output.