In this paper we focus on the issues surrounding temporal autocorrelations in FMRI time series. These include understanding the nature of the autocorrelation, the effects of temporal filtering, the effect of different experimental designs and ways of performing efficient and accurate statistical tests. In particular, the aim is to deduce an autocorrelation estimation technique which gives acceptably low bias when prewhitening.
We start with a brief overview of previous work in the area. We then set up a familiar GLM framework to define the different strategies for dealing with autocorrelations in FMRI. Four different approaches to temporal autocorrelation estimation are considered and a qualitative data analysis is used to examine the way in which the estimated autocorrelation varies spatially, the effects of temporal filtering, and the effect of different design types. We then introduce nonlinear spatial smoothing of the autocorrelation as a means to improving the estimation further. Finally, quantitative assessment of the bias (calibration) for the different autocorrelation estimation techniques is performed, by computing null distributions from null/rest data and comparing them with the expected theoretical distributions.