Speaker: Dr. Gang Chen (NIMH (National Institute of Mental Health), Bethesda MD, USA)

Title: “Do we have to deal with multiple comparisons in neuroimaging“

Host: Prof. Dr. Emrah Düzel

Abstract:
The conventional analytical approach in neuroimaging, massively univariate analysis, models
each and every spatial element (voxel or matrix element) separately with two assumptions:
1) the voxels or elements are unrelated with each other, and 2) no prior information is
available about their effects. However, we do know that the spatial elements are related to
some extent and share similar effect, and we do have some extent of prior knowledge about
the relevant effects (e.g., BOLD response usually less than 3%). The inefficiency resulting
from the first unrealistic assumption is only partially recouped through the correction for
multiple comparisons, but the information waste from the second assumption
fundamentally leads to the over-penalizing step of correction for multiple comparisons. In
addition, dichotomous decisions through thresholding under null hypothesis significance
testing are controversial in general and equally problematic in neuroimaging. For instance,
the popular practice of only reporting “statistically significant” results in neuroimaging not
only wastes data information, but also distorts the full results as well as perpetuates the
reproducibility crisis for several reasons. For instance, the difference between a “significant”
result and a “non-significant” one is not necessarily significant.
We believe that the heavy penalty lies in the low efficiency of the univariate GLM
methodology. With the assumption that the effects associated with brain regions follow a
Gaussian distribution, we build only one model that incorporates all relevant regions in
which the information across regions is shared and regularized, resolving the issue of
multiple comparisons that typically plagues the conventional statistical analysis in
neuroimaging. In addition to higher modeling efficiency, the methodology provides a
principled way to make statistical inferences, and allows us to emphasize the notion of full
results reporting through “highlighting,” instead of through the common practice of “hiding,”
thus minimizing loss of information while enhancing reproducibility. The modeling strategy
can be applied to three types of neuroimaging data: 1) type FMRI group analysis, 2) matrices
of either correlation coefficients or DTI properties (mean diffusivity, fractional anisotropy,
radial diffusivity and axial diffusivity), and 3) inter-subject correlation analysis for naturalistic
scanning. The related work is elaborated in two recent manuscripts:
https://www.biorxiv.org/content/early/2018/02/20/238998 and
https://www.biorxiv.org/content/early/2018/11/01/459545