Posted by B. Douglas Ward on July 27, 2001 at 12:04:22:
In Reply to: the questions in the anaphase of data analysis using AFNI posted by Shiyong Xu on July 27, 2001 at 02:47:48:
Hello Shiyong Xu:
1) You could apply cluster analysis to an individual subject's Orig data from
fim or 3dDeconvolve **IF** this is the final result. That is, the clustered
data must NOT be used as input to further voxel-by-voxel statistical analysis.
The effect of cluster size thresholding is to set certain voxels to zero. This
is NOT consistent with the assumptions underlying t-test or ANOVA. Therefore,
the data used as input to t-test or ANOVA must be as "raw" as possible.
To repeat: The output from the cluster analysis should be viewed as the
"end-of-the-line" as far as voxelwise statistical analysis is concerned.
2) The value of rmm used for the cluster connection radius depends on the
type of data that you are clustering. For clustering of Orig data, use
the Orig voxel dimensions; for clustering of Tlrc data, use the Tlrc voxel
dimensions. In any case, the choice of rmm is somewhat arbitrary.
3) The Orig voxel dimensions should be used as input to AlphaSim, even if
the clustering itself will eventually be performed on the Tlrc data. The
reason is that the assumptions underlying the Monte Carlo simulation used
by AlphaSim are not compatible with Tlrc-transformed data.
The AlphaSim output cluster threshold is in units of voxels. To perform
the clustering with 3dmerge or 3dclust, you must convert this to a volume
(i.e., the -vmul input). Therefore, whether clustering Orig data or Tlrc data,
you must multiply the cluster threshold from AlphaSim (in voxels) by the volume
of an Orig voxel (since AlphaSim uses Orig voxels).
4) Since AlphaSim works in Orig space, the AlphaSim -rmm input should be
consistent with the Orig voxel dimensions. If you later perform clustering on
the Tlrc data, the -rmm input for 3dmerge or 3dclust should be based on the
Tlrc voxel dimensions.
Now, this might mean using a larger value for -rmm as input to AlphaSim, and a
smaller value for -rmm as input to 3dmerge or 3dclust when doing the actual
clustering. However, this is a conservative approximation, since using the
smaller -rmm for actual clustering makes it harder to form clusters, resulting
in fewer false positives.
5) Once again, the answer depends on the type of data that will be clustered.
For clustering of Orig data, you could use program 3dFWHM to obtain a
(crude) estimate of the amount of spatial correlation of the Orig dataset
that you are using as input to 3dmerge or 3dclust. For clustering Tlrc data,
assuming that you have used the 3dmerge command "-1blur_fwhm 4" in converting
the Orig data to Tlrc space, you should use the "-fwhm 4" command as input
to AlphaSim.
Doug Ward