Dear Pierre-Yves,
Please excuse my late reply! There is a paper by Russell Poldrack on ROI analyses (https://dx.doi.org/10.1093/scan/nsm006 ), but it treats the topic rather generally as far as I recall. Providing guidelines would be quite difficult anyway, as it is going to depend on your particular hypotheses.
One could be interested in the average response/estimate of a certain region = measuring. Alternatively, one could be interested in localising voxels/clusters with a certain behavior (e. g. A > B) within an a-priori region = mapping, for which one could rely on a small volume correction. Note that the term "ROI analysis" in e. g. MarsBaR or REX always refers to the first approach, but "ROI analysis" can also refer to SVCs. In any case, ROI analyses focusing on a composite score on the one hand and SVC analyses on the other respond to different questions. SVCs tend to be misinterpreted; a sig. finding within e.g. a hippocampus label does not imply that "the hippocampus" (as a whole) shows a certain behavior. In my opinion, when going with SVC, one should at least try to discuss the location of the sig. voxels/clusters in some way (otherwise one would have been trying to localise something wihout being interested in its location, which is quite odd) = is there any evidence from other studies that this could reflect a certain cytoarchitectonic, anatomical, functional (sub)region?
Another aspect is how to define the a-priori regions. This is really going to depend on the aims, hypotheses of the study. In case you are interested in regions whose extent or location differs considerably between subjects and/or which are defined by some functional property you would probably want to go with a localiser condition (possibly presented in a separate run), a condition which is not used otherwise for contrasts. This would be typical for studies dealing with visual subregions like V1, V2, ... or e. g. face- or object-selective regions. This is not as trivial as it might sound at first, as, depending on localiser stimuli employed, you might or might not affect and possibly bias the ROI definition, see e. g. Berman et al. (2010, http://dx.doi.org/10.1016/j.neuroimage.2009.12.024) for fusiform face area.
ROI analyses based on available anatomical labels and/or parcellation schemes from connectivity / resting state studies and/or coordinates from meta-analyses are easy to reproduce, which is certainly an advantage. Whether they are meaningful as well is going to depend on the study though, maybe you are interested in smaller subregions, maybe there is no meta-analysis available or it refers to a broader topic. However, you might still split those labels based on findings from other studies, or you might just go with spheres placed along a gradient from e. g. posterior to anterior. Several studies in cingulate subregions have done so. In the end, as long as ROI definition is reasonable with regards to content (and statistically unbiased) then any approach should be alright.
In any case, you should make sure about the approach in advance. I have the impression that biased ROI definitions, circular analyses, ... result from unplanned ROI analyses. People might have expected a sig. finding on whole brain, now that they don't detect any they might start to consider a ROI analysis, and as there are no localiser conditions they stick with findings from other contrasts, which can easily introduce bias.
Best regards
Helmut
|