Bayesian meta-analysis of fMRI image data

ABSTRACT

We composed an R-based script for Image-based Bayesian random-effect meta-analysis of previous fMRI studies. It meta-analyzes second-level test results of the studies and calculates Bayes Factors indicating whether the effect in each voxel is significantly different from zero. We compared results from Bayesian and classical meta-analyses by examining the overlap between the result from each method and that created by NeuroSynth as the target. As an example, we analyzed previous fMRI studies focusing on working memory extracted from NeuroSynth. The result from our Bayesian method showed a greater overlap than the classical method. In addition, Bayes Factors proved a better way to examine whether the evidence supported hypotheses than p-values. Given these, Bayesian meta-analysis provides neuroscientists with an alternative meta-analysis method for fMRI studies given the improved overlap with the NeuroSynth result and the practical and epistemological value of Bayes Factors that can directly test the presence of an effect.

KEYWORDS:

• fMRI
• meta-analysis
• p-value
• Bayesian inference
• Bayes factors
• Bayesian random-effect meta-analysis

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplementary material

Supplemental data for this article can be accessed here.

Notes

¹ Although it could be argued that this assertion is false because the decision-theoretic approach, or Neyman-Pearson school of frequentism, allows one to accept the null, it requires one to specify the utility, or loss, function to be incorporated in the decision process, which might be hard to do due to the uncertainty of relative costs in practice. Also, because the decision-theoretic approach is, lamentably, usually intermingled with the Fisherian approach that does not allow one to accept the null, and it is rarely the case that the pure decision-theoretic approach is used in practice, we do not consider such a possibility seriously.