NeuroImage (Feb 2020)
Fighting or embracing multiplicity in neuroimaging? neighborhood leverage versus global calibration
Abstract
Neuroimaging faces the daunting challenge of multiple testing – an instance of multiplicity – that is associated with two other issues to some extent: low inference efficiency and poor reproducibility. Typically, the same statistical model is applied to each spatial unit independently in the approach of massively univariate modeling. In dealing with multiplicity, the general strategy employed in the field is the same regardless of the specifics: trust the local “unbiased” effect estimates while adjusting the extent of statistical evidence at the global level. However, in this approach, modeling efficiency is compromised because each spatial unit (e.g., voxel, region, matrix element) is treated as an isolated and independent entity during massively univariate modeling. In addition, the required step of multiple testing “correction” by taking into consideration spatial relatedness, or neighborhood leverage, can only partly recoup statistical efficiency, resulting in potentially excessive penalization as well as arbitrariness due to thresholding procedures. Moreover, the assigned statistical evidence at the global level heavily relies on the data space (whole brain or a small volume). The present paper reviews how Stein’s paradox (1956) motivates a Bayesian multilevel (BML) approach that, rather than fighting multiplicity, embraces it to our advantage through a global calibration process among spatial units. Global calibration is accomplished via a Gaussian distribution for the cross-region effects whose properties are not a priori specified, but a posteriori determined by the data at hand through the BML model. Our framework therefore incorporates multiplicity as integral to the modeling structure, not a separate correction step. By turning multiplicity into a strength, we aim to achieve five goals: 1) improve the model efficiency with a higher predictive accuracy, 2) control the errors of incorrect magnitude and incorrect sign, 3) validate each model relative to competing candidates, 4) reduce the reliance and sensitivity on the choice of data space, and 5) encourage full results reporting. Our modeling proposal reverberates with recent proposals to eliminate the dichotomization of statistical evidence (“significant” vs. “non-significant”), to improve the interpretability of study findings, as well as to promote reporting the full gamut of results (not only “significant” ones), thereby enhancing research transparency and reproducibility.