> Is there a way to quantify the change in cross-subject variability?
I would not worry too much about it because the impact is usually small. In case you want to toy with modeling, you could obtain standard error for each contrast (or contrast of contract in the case of interaction effect) for each subject when you do averaging across the voxels within an ROI. Then you can bring both the average effect estimate plus its standard error for group analysis using function rma() in R package metafor: [
www.metafor-project.org]
Gang