Analysis of task fMRI studies is typically based on using ordinary least squares within a voxel- or vertex-wise linear regression framework known as the general linear model. This use produces estimates and standard errors of the regression coefficients representing amplitudes of task-induced activations. To produce valid statistical inferences, several key statistical assumptions must be met, including that of independent residuals. Since task fMRI residuals often exhibit temporal autocorrelation, it is common practice to perform “prewhitening” to mitigate that dependence. Prewhitening involves estimating the residual correlation structure and then applying a filter to induce residual temporal independence. While theoretically straightforward, a major challenge in prewhitening for fMRI data is accurately estimating the residual autocorrelation at each voxel or vertex of the brain. Assuming a global model for autocorrelation, which is the default in several standard fMRI software tools, may under- or over-whiten in certain areas and produce differential false positive control across the brain. The increasing popularity of multiband acquisitions with faster temporal resolution increases the challenge of effective prewhitening because more complex models are required to accurately capture the strength and structure of autocorrelation. These issues are becoming more critical now because of a trend toward subject-level analysis and inference. In group-average or group-difference analyses, the within-subject residual correlation structure is accounted for implicitly, so inadequate prewhitening is of little real consequence. For individual subject inference, however, accurate prewhitening is crucial to avoid inflated or spatially variable false positive rates.
In this paper, we first thoroughly examine the patterns, sources and strength of residual autocorrelation in multiband task fMRI data. Second, we evaluate the ability of different autoregressive (AR) model-based prewhitening strategies to effectively mitigate autocorrelation and control false positives. We consider two main factors: the choice of AR model order and the level of spatial regularization of AR model coefficients, ranging from local smoothing to global averaging. We also consider determining the AR model order optimally at every vertex, but we do not observe an additional benefit of this over the use of higher-order AR models (e.g. (AR(6)). To overcome the computational challenge associated with spatially variable prewhitening, we developed a computationally efficient R implementation using parallelization and fast C++ backend code. This implementation is included in the open source R package
We find that residual autocorrelation exhibits marked spatial variance across the cortex and is influenced by many factors including the task being performed, the specific acquisition protocol, mis-modeling of the hemodynamic response function, unmodeled noise due to subject head motion, and systematic individual differences. We also find that local regularization is much more effective than global averaging at mitigating autocorrelation. While increasing the AR model order is also helpful, it has a lesser effect than allowing AR coefficients to vary spatially. We find that prewhitening with an AR(6) model with local regularization is effective at reducing or even eliminating autocorrelation and controlling false positives.
Our analysis revealed dramatic spatial differences in autocorrelation across the cortex. This spatial topology is unique to each session, being influenced by the task being performed, the acquisition technique, various modeling choices, and individual differences. If not accounted for, these differences will result in differential false positive control and power across the cortex and across subjects.