Hi -
>
> /NumWaves 2--What does this mean?This is from an unpaired two-
> sample t-test.
>
This is the number of regressors. I guess you have one for each group?
> /NumPoints 24--OK
> /PPheights 1 1--peak to peak heights? What does this mean? This is
> from an
>
This is the max value of each of the regressors I think. It is not
used in tbss, but I think it is used in e.g. featquery
> unpaired two-sample t-test.
>
> And are these /comments input into the program, or are they strictly
> informational comments?
They are used by some programs but not others, but I think you will
need them for your design file to be read correctly.
>
> 2. There are several types of output images.
>
> <output>_vox_tstat
> --when would you accept these results as valid, i.e. is there ever
> a case
> when the lack of multiple comparisons corrections is acceptable?--
> see next
> question for related part of this.
>
Some people like to know what the uncorrected t-stats are. e.g. if
you have a strong prior hypothesis about the regions where you expect
change. Note that these t-stats assume gaussian variation in the
population. Steve tested this in the tbss paper, and it looked pretty
good for skeletonised data: Much better than for e.g. VBM style FA
analyses (see the paper).
> <output>_max_tstat
> --what is the extent of the multiple comparisons? is there a way to
> find out
> the number of comparisons being run and corrected for?
The multiple comparison correction uses something called the maximum
distribution. i.e. clusters or voxels are tested against null
distribution of the maximum cluster-size, or t-value in the brain for
each permutation.
This effectively protects against the number of "independent" tests
that you are doing.
Tom Nichols has papers describing this.
> our statistician
> suggested that it may be useful to know, even an approximate value, to
> inform thinking about the number of permutations entered into
> Randomise in
> the first place.
>
Well - not quite, because you are not looking for a p-value at each
voxel and then correcting. The correction is integral to the null
distribution formed by the randomisation. So the number of
permutations will not have the same effect that you might think.
> <output>_maxc_tstat
> --we don't get the cluster approach. is this also a way to increase
> power by
> decreasing sample size? what is empirical experience on FA image
> analysis of
> useful cluster size?
This is just the same as clustering in FMRI data, and therefore has
the same advantages and disadvantages.
You may gain statistical power by acknowledging that neighbouring
voxels are related (although you may not in some circumstances) but
you have to set arbitrary thresholds, so interpretation becomes less
clear!
I'm afraid I can't give you figures about cluster sizes, because it
depends enormously on your experimentm and your set-up. e.g. it
depends on the z-value you choose, on the smoothness of your data etc.
Hope this is helpful
T
>
> Thank you in advance for all your work on FSL, it is a tremendous
> program.
>
> Rob
|