Serial Autocorrelations in fMRI: Sources of Bias Within GLM Models

• Activity Number: 177
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Ariana Anderson*+
• Keywords: fMRI ; modeling ; GLM ; neuroimaging ; correlations ; HRF

Abstract

Discoveries using fMRI data often rely on statistical methods using the general linear model (GLM), which are considered theoretically optimal only after certain conditions have been met. The hemodynamic response function (HRF) is a mathematical model of how blood flow changes in response to a stimuli. The assumption of independence among residual errors in the GLM is known to be false given that the HRF varies by tissue type, location, and stimulus onset. Additional serial autocorrelations (ACLs) exist even during periods of rest that can be attributed to machine noise, motion, or inherent biological processes such as cardiac signal or respiration. We simulate that autocorrelated white noise, when unaccounted for in the GLM, leads to biased estimates of parameters, increasing the Type 1 and Type 2 errors. We demonstrate that if the HRF is misestimated by even a small temporal shift, the parameter bias is compounded even further in the presence of ACL noise. This study indicates that ACL noise and flawed HRFs have a significant impact on estimation of parameters and establishing statistical significance in fMRI studies.