An integrative Bayesian approach to matrix‐based analysis in neuroimaging

Abstract

Understanding the correlation structure associated with brain regions is a central goal in neuroscience, as it informs about interregional relationships and network organization. Correlation structure can be conveniently captured in a matrix that indicates the relationships among brain regions, which could involve electroencephalogram sensors, electrophysiology recordings, calcium imaging data, or functional magnetic resonance imaging (FMRI) data—We call this type of analysis matrix‐based analysis, or MBA. Although different methods have been developed to summarize such matrices across subjects, including univariate general linear models (GLMs), the available modeling strategies tend to disregard the interrelationships among the regions, leading to “inefficient” statistical inference. Here, we develop a Bayesian multilevel (BML) modeling framework that simultaneously integrates the analyses of all regions, region pairs (RPs), and subjects. In this approach, the intricate relationships across regions as well as across RPs are quantitatively characterized. The adoption of the Bayesian framework allows us to achieve three goals: (a) dissolve the multiple testing issue typically associated with seeking evidence for the effect of each RP under the conventional univariate GLM; (b) make inferences on effects that would be treated as “random” under the conventional linear mixed‐effects framework; and (c) estimate the effect of each brain region in a manner that indexes their relative “importance”. We demonstrate the BML methodology with an FMRI dataset involving a cognitive‐emotional task and compare it to the conventional GLM approach in terms of model efficiency, performance, and inferences. The associated program MBA is available as part of the AFNI suite for general use.

Thetaburst TMS to the posterior superior temporal sulcus decreases resting-state fMRI connectivity across the face processing network

“…Since the ROIs are apriori selected from a network where we expect to see disruption Pitcher et al, 2017) -with the hand motor regions as a control, the statistical changes of interest will focus on the ROI analyses. A matrix based analysis (MBA) through Bayesian multilevel modeling was used to identify pairs of ROIs where a decrease in correlation magnitude was larger than expected along with a measure of statistical evidence (Chen et al, 2019). The advantage of this approach is that, instead of adopting a univariate GLM with the assumption that each ROI pair is an independent entity that shares no commonality or similarity with its peers, the contrast magnitude estimates and uncertainties of all ROI pairs are assessed as part of a single integrative model.…”

A comprehensive macaque fMRI pipeline and hierarchical atlas

“…These Bayesian methods are presently available in AFNI (e.g., programs RBA and MBA) for ROI-based analyses. They offer additional advantages, including containing built-in model validation, allowing for full results reporting, and not requiring arbitrary thresholding (Chen et al, 2020(Chen et al, , 2019a. Having a hierarchical atlas such as the CHARM allows a great deal of flexibility for the researcher to perform such analyses on an appropriate scale for their study and experimental design.…”

Untangling the Relatedness among Correlations, Part III: Inter-Subject Correlation Analysis through Bayesian Multilevel Modeling for Naturalistic Scanning

“…(2) from, our previous work for ROI-based group analysis for neuroimaging data (Chen et al, 2019a) as well as the BML approach for matrix-based analysis (Chen et al, 2019b),…”

“…Recently we applied the BML modeling approach to matrix-based analyses (Chen et al, 2019b) when the input data are either functional (e.g. inter-region correlation) or structural (e.g., white matter properties among gray matter regions) attribute matrix from each subject.…”

“…To address those limitations, here we propose a Bayesian multilevel (BML) framework that integrates all the spatial elements (i.e., regions of interest) into one model. Such a framework has been applied to typical task-related FMRI experiments (Chen et al, 2019a;Xiao et al, 2019) as well as matrix-based data analysis (Chen et al, 2019b;Yin et al, 2019). A dataset of naturalist scanning is utilized to illustrate the modeling approach and to showcase the modeling capability, flexibility and advantages in reporting results.…”