 |
 |
Hi Donal, On 9 December 2015 at 12:02, Hill, Donal <[log in to unmask]> wrote: > Hi Anderson, > > Many thanks for your input! On your point regarding the apples and oranges > comparison - do you know of any way to construct an apples to apples > comparison here? In order to choose a method that is most sensitive, there > must be a metric whereby one can compare TFCE with regular threshold based > permutation and indeed parametric cluster formation. > In a sense, ultimately all methods are trying to do the same: infer about activity or significance, while dealing with the spatial spread of the signal. In this sense they could be compared, and in this was in fact how TFCE was compared with cluster in the first paper. You'd be re-doing what was done some years ago. > > I am most interested in knowing the sensitivity threshold of all methods > i.e. at what critical point can an effect be seen. I think this is still > possible, even if you are comparing oranges and apples. If the oranges > sensitivity threshold is better than the apples one (when analysing the > same data), you still learn something even if a direct comparison isn’t > possible…I think! > You can do that: simulate known signal and make an ROC curve. All the best, Anderson > > Cheers, > Donal > > > On 9 Dec 2015, at 10:39, Anderson M. Winkler <[log in to unmask]> > wrote: > > Hi Donal, > > A few comments comments on this: > > - It is not appropriate to compute a p-value for a cluster after TFCE has > been applied. Yes, of course an image with TFCE results will have > contiguous voxels that can be called "clusters", but computing a p-value > for these would require a threshold, which would defeat the core purpose of > TFCE. > > - There is no formula to derive p-values for TFCE, hence results from the > random field theory don't extend trivially. There is no point in estimating > smoothness, etc, as there is nowhere to plug these values into. > > - Comparing the peak value of TFCE with the value of a cluster (even after > conversion to p-value) isn't meaningful -- it's an apples and oranges > problem: they aren't the same thing. > > - Yes, the smallest p-value attainable with a permutation test is 1/J, > where J is the number of permutations. > > All the best, > > Anderson > > > > On 8 December 2015 at 13:35, Donal Hill <[log in to unmask]> wrote: > >> Hi again Matthew, >> >> I think I know why I only get p-values in units of 0.01 actually. I just >> ran 100 permutations in TFCE to test, so I guess the p-value "granularity" >> is 1/100 i.e. 0.01. I'll crank up the number of permutations to get more >> accuracy on the p-values. >> >> Cheers, >> Donal >> >> > >