7. Keynote Address
Slides from this talk can be downloaded here.The video recording of this talk can be viewed here.
Tom Nichols
Professor of Neuroimaging Statistics, Nuffield Department of Population Health, University of Oxford
Title: Neuroimaging Statistics for Population Scale Data
Abstract:
While for years brain imaging studies have been limited to 2-digit sample sizes, with the advent of projects like the UK Biobank and Adolescent Brain Cognitive Development (ABCD) have made 5-digit sample sizes a reality. I will present two case studies of work that has been motivated or facilitated by such large scale projects. First, I'll discuss the problem of valid single-subject inference for fMRI connectomes. Large portions of neuroimaging research today depends on connectivity matrices derived from Pearson's correlation between 100's brain regions converted to Z-scores via Fisher's transformation. It is generally not appreciated that the variance of Fisher's-transformed correlation depends not only on the distinct autocorrelation within each time series but also the lagged cross-correlation. I will describe a practical solution (developed with Soroosh Afyouni) to obtain unbiased estimates of the variance Fisher's-transformed correlations, and review validation conducted on large-scale datasets that demonstrates the dramatic impact of incorrectly ignoring cross-correlations. Second, I pick up the issue of cluster inference for population scale data. With Armin Schwartzman's group we have been developing methods that account for spatial uncertainty in clusters. For thresholded maps of either the sample mean or Cohen's d, we produce inner and outer confidence sets for observed clusters, controlling an image-wide confidence level. When 5-digit sample-sizes make the entire brain significant, these methods will be essential for precisely characterising the spatial uncertainty in the localisation of effects. I will review the theory of these methods and demonstrate them on real task fMRI data.