Group Modeling
Thomas Nichols
University of Michigan
MRI Engineering Group
Optimal fMRI multi-subject analysis requires modeling multiple sources of variation. The two principal sources are intrasubject variability, in the form of scan-to-scan (possibly autocorrelated) noise, and intersubject variability, reflecting individual differences in response magnitude to the same stimulus.
In this talk, we review the history of hierarchical modeling in functional neuroimaging and highlight a widely used shortcut: the summary statistic approach (Holmes and Friston, 1998). We describe how early PET modeling often neglected the repeated-measures issue and suffered little consequence, whereas fMRI exposed this issue dramatically.
We examine the motivation for the summary statistic approach and discuss its strengths and weaknesses. Although this method is unparalleled in its ease of application, its simpler implementations have several limitations—some well known (e.g., that only a single image per subject can be considered) and some less widely recognized (e.g., the implicit assumption of homogeneous intrasubject variance).
We then review standard mixed-effects modeling approaches for such data, followed by two widely used fMRI methods implemented in FSL3 and SPM2.
