3dMEMA output interpretation

AFNI Message Board
1

Hello,
I am running a two group analysis with 3dMEMA (including the -unequal_variance -HKtest -model_outliers -residual_Z flags).

There are some subjects that seem like outliers in the dataset but I was unclear how best to use the output of 3dMEMA to identify them. Would you be able to help clarify if there are appropriate thresholds to examine output images for issues?

-For the main output image, are there appropriate threshold to examine the tau^2 and their QE images? Would particular extreme values indicate issues of note in the analysis?
-Also, I am seeing voxels with zero values in the tau^2 images that have non-zero values in the corresponding QE image, is that an error?

-For the ICC output, I get one sub-brik per subject, with values all between 0 and 1. Most voxels are a 1. This indicates the % of variance accounted for in each subject’s map? Is there a threshold at which these values are problematic and could used to indicate subject- or voxel-wise issues?

-For the resZ output, I again get one sub-brik per subject with Z scores indicate how much each subject deviates from the group mean, correct? Again, are there recommended thresholds or methods to identify if a subject is a particular outlier here?

Thank you!

2

are there appropriate threshold to examine the tau^2 and their QE images?

The QE is a ch-square statistic, so you should be able to assess the cross-subjects variability by using tau^2 and their QE as overlay and threshold on the AFNI GUI.

I am seeing voxels with zero values in the tau^2 images that have non-zero values in the
corresponding QE image, is that an error?

Sometimes the model renders negative tau^2 values, which are truncated to 0. Most likely those are the cases you see nonzero QE values.

For the ICC output, I get one sub-brik per subject, with values all between 0 and 1. Most voxels are a 1.
This indicates the % of variance accounted for in each subject’s map? Is there a threshold at which these
values are problematic and could used to indicate subject- or voxel-wise issues?

No such a hard threshold value available for such percentage.

For the resZ output, I again get one sub-brik per subject with Z scores indicate how much each subject
deviates from the group mean, correct? Again, are there recommended thresholds or methods to identify
if a subject is a particular outlier here?

Since they are Z-statistic, you can use the AFNI GUI to make some statistical inferences.