 |
 |
Dear Ling, The essential change with randomise v2.1 is a change in the way that confounds were handled. When there are no confounds, and model is a simple one-sample, two-sample test or a correlation, an exact permutation test is available. When you have confounds, there is no one 'right' permutation test, and there are many possible methods. Randomise v2.0 and before used a method due to Kennedy (1995; see randomise help page for references), and randomise v2.1 uses a method due to Freedman & Lane (1983). The reason for the change is that, while both the Kennedy & Freedman-Lane method are accurate for large n, for small n the Kennedy method is actually not so good and tends to falsely inflate significances. Hence, we changed to the more accurate method; unfortunately, you might see some decrease in significance, but this was only because the Kennedy method was too generous :) Sorry for any trouble! If you have have *no* confounds, then I can think of any reason why it would change. Please let me know if don't have any confounds. -Tom PS: The documentation for randomise is behind... it still references Kennedy... I'll update it asap. On Sun, Sep 14, 2008 at 8:36 AM, Wang, Ling <[log in to unmask]> wrote: > Dear All, > > Previously I ran a TBSS analysis with v1.1, and used randomise v2.0 to do > the statistical analysis. Now, I just wanted to try the new version of > randomise, v2.1, but on the identical data, design martix and contrasts, I > got different results, e.g., much less significant results in voxel level > in v2.1 than in v2.0. The command I used in randomise v2.1 was "randomise -i > all_FA_skeletonised -o tbss -m mean_FA_skeleton_mask -d design.mat -t > design.con -n 5000 --T2 -x -V". > > I read the "what's new" about TBSS v1.2, which suggests re-run the TBSS > analysis from the first step. Should I re-run the analysis on my data and > then do the statitical analysis with randomise v2.1, or any other > suggestions? Many thanks. > > Best wishes, > Ling > > -- > Ling Wang > Cognitive Neurology > Institute of Neuroscience and Biophysics - Medicine > Research Centre Juelich > 52425 Juelich, Germany > phone: +49-2461-612481 > Fax: +49-2461-612820 > ____________________________________________ Thomas Nichols, PhD Director, Modelling & Genetics GlaxoSmithKline Clinical Imaging Centre Senior Research Fellow Oxford University FMRIB Centre