Dear Joaquim,
I have extracted data from an ROI following a whole-brain SDM analysis, in order to do forest plots. I have two questions:
- the SDM output gives an Estimate and a Variance for each study, as well as a mean Estimate and Variance across all studies. However, I calculated the average Estimate across all studies, but that is systematically higher than the mean Estimate output by SDM. Is that normal?
- also, I'd like to plot confidence intervals (as is common for forest plots), is it valid to calculate it from the Variance using the formula ± Z * √(Variance/n)?
Thank you very much for your help.
Best,
Guillaume

Dear Guillaume,

the SDM output averages the values of the voxels of the ROI using the arithmetic mean, which is useful for example for creating a plot.

However, note that the average effect size of the voxels of the ROI is indeed different than the effect size of the ROI. I guess formulas are too much for this forum, but you can empirically check it with the following code in R:

voxel1 = rnorm(10)
effectsize_voxel1 = mean(voxel1) / sd(voxel1)
voxel2 = rnorm(10)
effectsize_voxel2 = mean(voxel2) / sd(voxel2)
voxel3 = rnorm(10)
effectsize_voxel3 = mean(voxel3) / sd(voxel3)
roi = (voxel1 + voxel2 + voxel3) / 3
effectsize_roi = mean(roi) / sd(roi)
(effectsize_voxel1 + effectsize_voxel2 + effectsize_voxel3) / 3
effectsize_roi

Dear Joaquim,
thank you very much for this clarification!
Do you think there's a straightforward way to derive confidence intervals from the Variance estimates (I'm not too sure about which sample size to use, since my input T-maps represent contrasts between 2 groups)?
Or do you think it is just fine to use forest plots with the Variance instead of confidence intervals?
Thanks for your enlightenment!
Best,
Guillaume

Dear Guillaume,

currently SDM is not parametric and thus there isn't a straightforward way to derive confidence intervals. Maybe one way is to use the normal distribution but base the intervals on the p-value (rather than on the variance), though it will be just an approximation.

Best,

Joaquim