next up previous
Next: Small Scale Variation Up: Spatio-temporal Noise Modelling Previous: Scale of Variation

Large Scale Variation

In general, the large-scale variation represents deterministic trends across one or more of the 4 dimensions of the observed data space. Previous work (16,43) has shown large scale temporal noise processes. We do not incorporate these processes into the model, but instead remove the worst of them by using high-pass filtering as a preprocessing step (i.e. in equation 2 we set $ l_{it}=0$). In this work the high pass filtering cut-off is chosen as the one which removes as much of the temporal low frequencies as possible without degrading the signal. There is increasing evidence that there is large scale spatial structure in FMRI data possibly attributable to neuronal networks which are not locked to an imposed task (34). When looking for task related activation these large scale spatial networks are effectively ``noise/confounds''. Ideally, a spatio-temporal model would attempt to model these confounds out. However, in this work we do not attempt to do so. This is an important area for future work.