The Interplay Between Sample Size and Reproducibility of Results in fMRI Studies

• Activity Number: 547
• Type: Contributed
• Date/Time: Wednesday, August 3, 2016 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Han Bossier* and Ruth Seurinck and Sanne Roels and Simone Kuehn and Jean-Baptiste Poline and Beatrijs Moerkerke
• Companies: Ghent University and Ghent University and Ghent University and MPI for Human Development and University of California at Berkeley and Ghent University
• Keywords: Neuroimaging ; fMRI ; reproducibility ; large-scale databases

Abstract

In the study of brain functions using fMRI, reproducibility may be limited as fMRI data are characterized by large amounts of noise while studies typically have small sample sizes. The emergence of large-scale projects such as the Human Connectome Project (N=500, Van Essen et al., 2013) or the IMAGEN consortium (N = 1500, Schumann et al., 2010) enable researchers to further deepen the study on reproducibility of fMRI results. In this paper, we look at the similarity between brain images X and Y with increasing sample sizes (N). Using up to 1400 subjects from the IMAGEN database doing a math versus language task, we study (1) the similarity of thresholded images (activated regions) using the percent overlap of activation (Maitra, 2010) and (2) unconditional similarity using the Pearson product-moment correlation coefficient. We start by sampling 10 subjects for group X and 10 other subjects for group Y. We then add 10 subjects and continue until N = 700. For thresholded images, results show moderate levels of overlap (60% overlap in activation at N=200). When comparing brain images without thresholding, we observe reasonable similarity (r=0.83 at N=60).