TFCE, effect size and degrees of freedom

[Ana Paula Salazar]

Dear SwE experts,

I am working with longitudinal data and one of the reviewers asked us to show the TFCE values, effect size, and degrees of freedom.
Is there a way to find/extract these values (TFCE, effect sizes, and degrees of freedom) using the SwE toolbox?


Thanks in advance,

Ana Salazar

[Thomas Nichols]

Dear Ana,

I am working with longitudinal data and one of the reviewers asked us to show the TFCE values, effect size, and degrees of freedom.

Is there a way to find/extract these values (TFCE, effect sizes, and degrees of freedom) using the SwE toolbox?

All the key values should be written out as NIFTI/CIFTI files.  See the man pages http://www.nisox.org/Software/SwE/man under "Post-processing & display of results".

The contrast values are in swe_vox_beta_c{c#}.nii; the T-statistic values are in swe_vox_Tstat_c{c#}.nii; if you take the first divided by the 2nd you'll get the standard error of the contrast.  The (error) degrees of freedom are in swe_vox_edf_c{c#}.nii.

The TFCE values are really hard to interpret and involve arbitrary discretisation choices... I wouldn't ever report TFCE values.

Is this enough to go on?

-Tom

 

Thanks in advance,

Ana Salazar

--
You received this message because you are subscribed to the Google Groups "SwE-Toolbox Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to swe-toolbox...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/swe-toolbox/07190257-903c-4807-83ba-8c7110256243n%40googlegroups.com.

[Ana Paula Salazar]

Dear Tom,

Thank you for your prompt reply; it was very helpful.

Ana Salazar

[Kathleen Hupfeld]

Hi Dr. Nichols, 

I had several follow-up questions on this. 

1) I agree the TFCE values don't make sense to report. Going off of your previous response + Christian Gaser's response here regarding his TFCE toolbox, reporting the T stat would be ignoring cluster size contribution? --so, if we're reporting TFCE results, it wouldn't really make sense to report the T statistic values, right?

"It's not possible to estimate effect size from the TFCE statistic. You could estimate the effect size from the T statistics. However, this would ignore the contribution from cluster size to the TFCE statistics and would only consider the height of the effect."

2) Thanks for further clarifying the NIFTI image outputs. However, I think I'm tripping up on terminology here, but what specifically would be the "effect size" in a longitudinal SwE analysis? Perhaps a silly question I'm getting confused on.. but can you calc (a meaningful) "effect size" using the SE of the contrast? 

3) Re: degrees of freedom, I don't think I've fully wrapped my head around this... but as DFs could vary across voxels (SwE manual excerpt below), in a results table would you rec extracting the DFs from swe_vox_edf_c{c#}.nii at the peak MNI coordinates we're reporting to appease the reviewer? i.e., is this 'meaningful' vs. would you instead rec explaining to a reviewer why this doesn't really make sense to report? :)  

"First, as the degrees of freedom with the SwE method may vary across voxels, for each contrast, the equivalent Z-score or 1-degree-of-freedom chi-squared image are displayed instead of the traditional t-score or F-score image (note that, for each parametric contrast, the t- or F-score image is saved in the working directory alongside the degrees of freedom image)." 

Thanks, as always, for your help. 

Best wishes, 
Kathleen 

On Wednesday, April 21, 2021 at 9:34:52 AM UTC-4 ten.p...@gmail.com wrote:

[Thomas Nichols]

Dear Kathleen,

1) I agree the TFCE values don't make sense to report. Going off of your previous response + Christian Gaser's response here regarding his TFCE toolbox, reporting the T stat would be ignoring cluster size contribution? --so, if we're reporting TFCE results, it wouldn't really make sense to report the T statistic values, right?

"It's not possible to estimate effect size from the TFCE statistic. You could estimate the effect size from the T statistics. However, this would ignore the contribution from cluster size to the TFCE statistics and would only consider the height of the effect."

There's no good answer here; best you can do is report the size of the cluster and the peak T stat... but, indeed, it's complete representation.

 

2) Thanks for further clarifying the NIFTI image outputs. However, I think I'm tripping up on terminology here, but what specifically would be the "effect size" in a longitudinal SwE analysis? Perhaps a silly question I'm getting confused on.. but can you calc (a meaningful) "effect size" using the SE of the contrast? 

Without repeated measures, for a one-sample t-test, or a two-sample t-test, effect size is well defined via the corresponding Cohen's d statistics; for a correlation, r or R^2 is the appropriate measure of effect size.  In longitudinal and repeated measures models there is no single agreed measure of effect size.  The basic problem is that, for independent data, if you tell me the population standard deviation (along with sample sizes) you've just told me everything I need to know to make an effect size for any effect magnitude/change.  In repeated measures, for one particular effect (within subject, between subject) there is a complex function of the design and the repeated measures variance and covariance that determines the standard error... i.e. for repeated measures design, if you tell me a population standard deviation and an effect estimate, I still don't have enough to estimate power.

See e.g. Nakagawa & Schielzeth (2013) for how complicated it is to define R^2 outside of linear regression with independent errors.

BUT, there is one out:  If you have a one-sample setting, then *roughly* taking a T statistic value and divided by the sqrt of the number of subjects will give you a Cohen's d like effect size.  Aside from that, I fear there aren't any easy answers.

 

3) Re: degrees of freedom, I don't think I've fully wrapped my head around this... but as DFs could vary across voxels (SwE manual excerpt below), in a results table would you rec extracting the DFs from swe_vox_edf_c{c#}.nii at the peak MNI coordinates we're reporting to appease the reviewer? i.e., is this 'meaningful' vs. would you instead rec explaining to a reviewer why this doesn't really make sense to report? :)  

"First, as the degrees of freedom with the SwE method may vary across voxels, for each contrast, the equivalent Z-score or 1-degree-of-freedom chi-squared image are displayed instead of the traditional t-score or F-score image (note that, for each parametric contrast, the t- or F-score image is saved in the working directory alongside the degrees of freedom image)." 

If you report a T statistic at a peak, then it makes sense to be ready to report what the DF are for that T statistic. It makes sense.  But, if you're *not* reporting a T/F statistic, then, no, I don't see the purpose of reporting the DF.  (Maybe you could report median DF over the brain, or something, but that's really a general aspect of reporting the overall model.)

-Tom

Nakagawa, S., & Schielzeth, H. (2013). A general and simple method for obtaining R2 from generalized linear mixed-effects models. Methods in Ecology and Evolution, 4, 133–142. https://doi.org/10.1111/j.2041-210x.2012.00261.x

 

Thanks, as always, for your help. 

Best wishes, 
Kathleen 

On Wednesday, April 21, 2021 at 9:34:52 AM UTC-4 ten.p...@gmail.com wrote:

Dear Ana,

I am working with longitudinal data and one of the reviewers asked us to show the TFCE values, effect size, and degrees of freedom.

Is there a way to find/extract these values (TFCE, effect sizes, and degrees of freedom) using the SwE toolbox?

All the key values should be written out as NIFTI/CIFTI files.  See the man pages http://www.nisox.org/Software/SwE/man under "Post-processing & display of results".

The contrast values are in swe_vox_beta_c{c#}.nii; the T-statistic values are in swe_vox_Tstat_c{c#}.nii; if you take the first divided by the 2nd you'll get the standard error of the contrast.  The (error) degrees of freedom are in swe_vox_edf_c{c#}.nii.

The TFCE values are really hard to interpret and involve arbitrary discretisation choices... I wouldn't ever report TFCE values.

Is this enough to go on?

-Tom

 

Thanks in advance,

Ana Salazar

--
You received this message because you are subscribed to the Google Groups "SwE-Toolbox Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to swe-toolbox...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/swe-toolbox/07190257-903c-4807-83ba-8c7110256243n%40googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "SwE-Toolbox Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to swe-toolbox...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/swe-toolbox/c1e82d0d-e891-4e07-95a3-47adce72df97n%40googlegroups.com.

[Kathleen Hupfeld]

Awesome, thank you SO much for the detailed answers. This clarifies all of my questions. 

Best, 

Kathleen