On Sat, Sep 11, 2010 at 7:03 PM, Xu Chen <[log in to unmask]> wrote:

Thanks for your explanation, Reza! That helps. I missed the nonparametric part. I am working on the cluster-wise inference using "randomise".

that is a nonparametric inference, however, if you take ONLY the T-stat image from randomise and do some-analysis-on-your-own, then you cannot say "using randomise". 

 

But seems to me , the threshold selection (the parameter after -c ) for the clusters seems a little bit "random". Do you have any recommendations on the strategy of  the threshold selection?

By random you mean arbitrary :)

This is the reason why TFCE is introduced to the field. 

In general, 3'ish (±.5) on your T-stat should be a safe threshold (though it's hard to say a single value for EVERY study).

 

Thanks

Jerry

On Sat, Sep 11, 2010 at 5:41 AM, Reza Salimi <[log in to unmask]> wrote:

Jerry,

On Sat, Sep 11, 2010 at 12:06 AM, Xu Chen <[log in to unmask]> wrote:

Hi, FSLers
  
       Just curious how  the uncorrected p in TFCE  get calculated?  I thought that the uncorrected-p  was simply computed based on the t stat value (in the file of *_tfce_tstat) with the consideration of the DOF. But I just noticed the inconsistency between the one I calculated based on the tstat and the one in the Uncorrected P file (*_tfce_p_tstat).  Seems I missed something there. 

randomise is a nonparametric inference, hence, no T-test with a DOF takes place here.

Anyway, as you said, we advocate the use of corrected P-value, however, there is an uncorrected P-value image in the result folder.

The uncorrected P-values are purely voxel-wise; calculated for each voxel using its individual distribution.

 

      I understand that ideally, the Corrected-P in TFCE should be used. However, if the resultant image in Corrected-P doesn't show any significant activations at a certain threshold (e.g. p=0.95) , which one should be used to report the  result, the tstat value (_tfce_tstat) or the Uncorrected -P (_tfce_p_tstat)?

In order to have the most reliable, i.e., statistically powerful inference, you better use a family-wise error measure, which is why corrected P-values of the TFCE are recommended.

Using the T-stat, you can only perform a voxel-wise parametric inference, which is shown to be less sensitive than cluster and TFCE.

And for the uncorrected TFCE P-values, you will have a result that is not corrected for multiple comparison, which makes it questionable.

hope it helps ...

 

Thanks

Jerry

--
Reza Salimi-Khorshidi,
DPhil Candidate, FMRIB Centre of the University of Oxford (Linacre College)
Associate Member, Oxford-Man Institute for Quantitative Finance, University of Oxford

Email: [log in to unmask]; Tel: +44 (0) 1865 222704;  Fax: +44 (0)1865 222717
Address: FMRIB Centre, John Radcliffe Hospital, Oxford OX3 9DU, UK

--
Reza Salimi-Khorshidi,
DPhil Candidate, FMRIB Centre of the University of Oxford (Linacre College)
Associate Member, Oxford-Man Institute for Quantitative Finance, University of Oxford

Email: [log in to unmask]; Tel: +44 (0) 1865 222704;  Fax: +44 (0)1865 222717
Address: FMRIB Centre, John Radcliffe Hospital, Oxford OX3 9DU, UK