Sources of Residual Autocorrelation in Multiband Task fMRI and Strategies for Effective Mitigation

In task fMRI analysis, OLS is typically used to estimate task-induced activation in the brain. Since task fMRI residuals often exhibit temporal autocorrelation, it is common practice to perform prewhitening prior to OLS to satisfy the assumption of residual independence, equivalent to GLS. While theoretically straightforward, a major challenge in prewhitening in fMRI is accurately estimating the residual autocorrelation at each location of the brain. Assuming a global autocorrelation model, as in several fMRI software programs, may under- or over-whiten particular regions and fail to achieve nominal false positive control across the brain. Faster multiband acquisitions require more sophisticated models to capture autocorrelation, making prewhitening more difficult. These issues are becoming more critical now because of a trend towards subject-level analysis, where prewhitening has a greater impact than in group-average analyses. In this article, we first thoroughly examine the sources of residual autocorrelation in multiband task fMRI. We find that residual autocorrelation varies spatially throughout the cortex and is affected by the task, the acquisition method, modeling choices, and individual differences. Second, we evaluate the ability of different AR-based prewhitening strategies to effectively mitigate autocorrelation and control false positives. We find that allowing the prewhitening filter to vary spatially is the most important factor for successful prewhitening, even more so than increasing AR model order. To overcome the computational challenge associated with spatially variable prewhitening, we developed a computationally efficient R implementation based on parallelization and fast C++ backend code. This implementation is included in the open source R package BayesfMRI.