James,
Thanks for bringing up the issue. I just haven't yet got chance to cover every detail regarding 3dMEMA.
The conventional meta analysis assumes that the estimated between-subject variability converges to its 'true' value, and the Z-statistic testing hinges around such an assumption. However, if the number of subjects is relatively moderate or small, for example, less than 10, the convergence assumption might be questionable. In recent couple of years, a t-statistic has been proposed to handle such a problem, and it's supposed to be a more accurate test with a moderate or small sample size of subjects. Therefore it's expected to see similar results between Z- and t-statistic, but t tends to be on the conservative side, and a couple of examples I've tested so far did verify this.
> I have run the same R script and only changed the option for t or z, and I get
> significantly different patterns in the results of each (specifically, the active
> voxels in the z-stat condition are a small subset of those in the t-stat condition
> at the same p-value threshold).
That seems to be the opposite of what I'd expect. If it's indeed the case as you're describing here, do you mind uploading all the input data plus the command lines to AFNI so that I can check whether there is any quirk in the program?
[
afni.nimh.nih.gov]
Thanks,
Gang