CFE 2020: Start Registration
View Submission - CMStatistics
Title: The impact of autocorrelation in fMRI task and rest analysis Authors:  Soroosh Afyouni - University of Oxford (United Kingdom) [presenting]
Thomas Nichols - University of Oxford (United Kingdom)
Abstract: Time series obtained using fMRI are notoriously autocorrelated. Although the source of the autocorrelation is not completely known, the dependency between the observations induced by autocorrelation violates the assumption behind the majority of conventional statistical methods used in rest and task fMRI analysis. We show that in resting-state functional connectivity, the variance of Pearson correlations, the most widely used measure of connectivity on subject level, is excessively biased due to autocorrelation. We propose a novel variance estimator for sample correlation coefficients which accounts for such dependencies. Further, we show that the existing methods for accounting autocorrelation in noise for subject-level task fMRI come short in modern rapidly sampled fMRI scans. Using a combination of spectral methods, we propose a whitening technique which successfully flattens the spectrum of the noise across various task paradigms in fMRI acquisitions above 1Hz.