Posted by B. Douglas Ward on March 22, 2001 at 14:05:41:
In Reply to: significance levels posted by Gary on March 22, 2001 at 11:47:08:
Gary:
Yes, the thresholding should only be applied to the final output of the
statistical analysis. The input data should be as "pure" as possible.
An example that I have used before is: Suppose that you are measuring the
heights of random samples from two groups of people, in order to test whether
the heights of the groups are significantly different. Suppose, for some
reason, the height measurement for one subject was not very accurate. You
definitely would NOT threshold this value, by entering that subject's height
as ZERO. Just so, you should not threshold the input data to 3dANOVA2.
Regarding the appropriate significance levels for the output: It is necessary
to make some correction for multiple inferences. However, due to the huge
number of voxels, the standard Bonferroni correction is extreme. One possible
alternative is to use a combination of individual voxel probability
thresholding and cluster size thresholding in order to achieve the desired
alpha level for the entire volume. You can read more about this in the
documentation for program AlphaSim (contained in files AlphaSim.ps and
AlphaSim_fig.ps).
The input to AlphaSim consists of the dimensions of the ORIG dataset. The
output of this Monte Carlo simulation allows you to determine the tradeoff
between voxel probability threshold and cluster size threshold in order to
achieve a given overall alpha. The output cluster size threshold is in number
of voxels. This must be converted to volume (by multiplying by the volume of
an ORIG voxel). Also, the (individual voxel) probability threshold must be
converted to a t-thr (F-thr, CC-thr) for use with an fitt (fift, fico) type
dataset. This can be done within afni itself, or by using program cdf.
You can then use these two numbers (vmul and thr) as input parameters for
3dmerge (or 3dclust) when forming clusters from the TLRC output from 3dANOVA2.
Doug Ward