We now describe the experiment and original analysis carried out in [18]. A healthy 23-year-old right-handed male was scanned on 33 separate days (over 2 months) with as many factors as possible held constant. On each day, three block-design paradigms were run (all using rest=24.6s, activation=24.6s block lengths): visual (8Hz reversing black-white chequerboard, 36 time points after deleting the first two), motor (finger tapping, right index finger at 1.5Hz, 78 time points) and cognitive (0.66Hz random number generating vs counting, 78 time points), with the paradigm order randomised. The data was collected on a Siemens Vision at 2T; TR=4.1s, 64x64x48 3x3x3mm voxels. A single T1-weighted 1.5x1x1mm structural scan was taken.
Original analysis was carried out with SPM99 (www.fil.ion.ucl.ac.uk/spm). All 99 sessions were realigned (motion-corrected) to the same target (the first scan of the first session of the first day) and then a mean over all 99 sessions was created. This was used to find normalisation (to a T2-weighted target in MNI space [8]) parameters for all 99 sessions (using 12-parameter affine followed by 7x8x7 basis-function nonlinear registration). Sinc interpolation on final output was used.
Sessions containing ``obvious movement artefacts'' were identified by eye and removed from consideration (3 motor, 2 visual, 3 cognitive). Cross-session analysis was carried out for voxels in standard space which were present in all sessions. Spatial filtering with a Gaussian kernel of FWHM 6mm was applied. Each volume of each session was intensity-normalised (rescaled) so that all had the same mean intensity.
Voxel time-series analysis was carried out using general linear modelling (GLM). The data was first pre-coloured by temporally smoothing the data with a Gaussian of 6s FWHM. Slow drifts in the data were removed by including drift terms in the model - a set of cosine basis functions effectively removing signals of period longer than 96s.
For presentation of within-session results, voxel-wise thresholding
(
) was used, correcting for multiple comparisons using
Gaussian random field-theory (GRF) [11].
Both fixed- and ``random''-effects analyses were carried out to examine the effects of using different variance components, and an extra-sum-of-squares (ESS) F-test was performed across all sessions of each paradigm to assess the presence of significant inter-session variance.