Moderated t-tests for group-level fMRI analysis

In recent years, there has been significant criticism of functional magnetic resonance imaging (fMRI) studies with small sample sizes. The argument is that such studies have low statistical power, as well as reduced likelihood for statistically significant results to be true effects. The prevalence of these studies has led to a situation where a large number of published results are not replicable and likely false. Despite this growing body of evidence, small sample fMRI studies continue to be regularly performed; likely due to the high cost of scanning. In this report we investigate the use of a moderated t-statistic for performing group-level fMRI analysis to help alleviate problems related to small sample sizes. The proposed approach, implemented in the popular R-package LIMMA (linear models for microarray data), has found wide usage in the genomics literature for dealing with similar issues. Utilizing task-based fMRI data from the Human Connectome Project (HCP), we compare the performance of the moderated t-statistic with the standard t-statistic, as well as the pseudo t-statistic commonly used in non-parametric fMRI analysis. We find that the moderated t-test significantly outperforms both alternative approaches for studies with sample sizes less than 40 subjects. Further, we find that the results were consistent both when using voxel-based and cluster-based thresholding. We also introduce an R-package, LIMMI (linear models for medical images), that provides a quick and convenient way to apply the method to fMRI data.