Dear Benjamin,

This recent paper by Chao-Gan Yan and colleagues explicitely looked at minimizing inter-site biases using a variety of ad-hoc normalization procedures:
http://www.sciencedirect.com/science/art...

You also have several examples of successful multisite rs-fmri studies, e.g. in autism research
http://www.ncbi.nlm.nih.gov/pmc/articles...
http://www.sciencedirect.com/science/art...

For direct comparison between task and resting-state fmri in multisite studies, I don't have a reference (the fBIRN consortium had a number of publications, but I think it was about task-based fMRI only). I don't think multisite acquisitions are necessarily a problem. In a scenario where you contrast two groups and both groups are represented at each site with a reasonable sample size, regressing out site effects in a linear model seems reasonable (although there can always be site x population effects or other issues, e.g. heteroscedasticity across sites). In the scenario suggested by Constantin, that is contrast a patient group from site A with a control group from site B, site bias is highly problematic and cannot be properly addressed.

Best,

Pierre