Michael Amlung Wrote:
-------------------------------------------------------
> We're using AlphaSim to estimate a cluster extent
> threshold for our group maps, and a reviewer for
> one of our articles suggested that we must used
> 3dFWHMx to estimate the true smoothness in X, Y,
> and Z rather than just putting the size of the
> blur we applied to the data (using the -FWHM
> option).
The reviewer is right. There's already some smoothness in your original data, so blurring it further will give you a smoothness that is higher than the size of your smoothing kernel. In other words, using the size of your smoothness kernel will underestimate the smoothness and therefore yield false-positive clusters.
> We applied a 3.5mm FWHM blur to the data
> using 3dmerge, but our actual values for
> smoothness in the group map is much larger (x =
> 7.2, y = 6.97, and z = 5.79). Obviously, when we
> run AlphaSim with these updated values, the output
> indicates that we must use a much larger cluster
> size (21 voxels) to achieve an alpha < .05.
>
> Our question is how should we handle this when we
> have multiple group maps that need to be corrected
> for multiple comparisons? For example, we have two
> conditions that are compared against an active
> baseline condition. So do we need to estimate the
> smoothness for each map individually, and then use
> the smallest cluster size that achieves alpha <
> .05 for all maps? We would prefer to apply the
> same cluster size across maps, rather than using
> unique cluster sizes for each map.
As explained in the documentation, you should *not* use the statistical maps to estimate the smoothness, as you are interested in the smoothness of the noise only.
The proper way is to run 3dDeconvolve (or better, 3dREMLfit, to account for temporal autocorrelation) for each participant, and save the residuals (with -errts). Run 3dFWHMx for each participant on the residuals and average the smoothness estimates, then feed this into AlphaSim or 3dClustSim.
Also make sure you use a proper group mask for both 3dFWHMx and Alphasim/3dClustSim, otherwise you are losing power because it is correcting for a larger volume.
best,
Nick