Gang Chen wrote:
> I assume the manual you read is about 3dFDR, which is the old
> approach to obtaining FDR. With 3dFDR you do get a z-value,
> which you can use for thresholding. However, such a z-value is
> provided simply for the convenience of thresholding. In other
> words, each z-value eventually is associated with a q-value.
> With the new approach of 3drefit -addFDR, the q-value is
> provided without the intermediate z-value.
Yep, that was the one I read.
>
> > What I'm unclear on is how to come up with the critical q
> value for thresholding.
>
> It is pretty arbitrary: 0.05 or below is generally acceptable.
A ok that makes sense.
>
> > Also, when you say that the use of alphasim is moot because
> it assumes the
> > normality of the data. Is that because the random numbers are
> chosen from
> > a normal distribution or because of the Gaussian filtering? I
> should point out
> > that one of the stages of the VBM is smoothing with a
> Gaussian of specified
> > sigma, a number which I have been passing in to AlphaSim ( in
> the form of
> > FWHM, I've only just realised that AlphaSim can take a sigma
> arg). Would
> > that make a difference?
>
> My suggestion of FDR was based on the fact that you used
> non-parametric approach for group analysis, which does not
> require normality assumption about the beta values across
> subjects. Is the reason you run non-parametric analysis in the
> first place because of not enough number of subjects?
We run non-parametric tests because the FSL guys say it best not to assume normality of distribution of the grey maps. We've run permutation tests on it but it appears to be very conservative, so we wanted to try other methods to see how they compare. However, Mann-Whitney has been used in the literature on VBM.
Thanks for your help, enjoy the turkey tomorrow!