We now describe the analysis approaches used for this paper. The two packages used for our investigations were SPM99b (Statistical Parametric Mapping, www.fil.ion.ucl.ac.uk/spm) and FSL (FMRIB's Software Library, www.fmrib.ox.ac.uk/fsl, version 1.3, June 2001). Both are freely available and widely used.
SPM includes a motioncorrection (``realignment'') tool, a tool for registration (``normalisation'') to standardspace, GLMbased timeseries statistics [23] and GRFbased inference [11]. SPM carries out standardspace registration before timeseries statistics. The SPM99b time series statistics correct for temporal smoothness by precolouring [10].
GLMbased analysis in FSL is carried out with FEAT (FMRI Expert Analysis Tool), which uses other FSL tools such as BET (Brain Extraction Tool [20]), an affine registration tool (FLIRT  FMRIB's Linear Image Registration Tool [14,13]), and a motioncorrection tool based on FLIRT (MCFLIRT [13]). FEAT carries out standardspace registration after timeseries statistics. FSL time series statistics correct for temporal smoothness by applying prewhitening, as described in [22].
6 different, complete analyses were carried out with various combinations of preprocessing and timeseries statistics options, in order to allow a variety of comparisons to be made. In tests A,C,G, FSL was used for preprocessing and registration whilst in tests D,E,F, SPM was used. For tests A,D,G, FEAT timeseries statistics was used whilst for C,D,F, SPM timeseries statistics was used.
In tests AE the various controlling parameters were kept as similar as possible, both to each other and to default settings in the relevant software packages. Tests A vs D and C vs E hold the statistics method constant whilst comparing spatial methods, therefore showing the relative merits of the ``spatial'' components (motion correction and registration). Tests A vs C and D vs E hold the spatial method constant whilst comparing statistical components, thus showing the relative merits of the statistical components (timeseries analysis). A vs E tests pureFSL against pureSPM. F and G test pureSPM and pureFSL respectively, with these analyses set up to match the specifications of the original analyses in [18] as closely as possible, including turning on intensity normalisation in both cases. For a summary, see Table 1.
(For B, ICAbased temporalmodelfree analysis was carried out; the modelfree results are not included in this paper, but will be presented elsewhere.)
Because the methods for highpass temporal filtering in FSL and SPM are intrinsically different, they cannot be set to act in exactly the same way (within AE and within F,G) by choosing the same cutoff period in each; instead, the cutoff choices were made so as to match as closely as possible the extent to which the relevant signal and noise frequencies were attenuated by the different methods. For the purposes of this paper, highpass temporal filtering is considered to be part of the temporal statistics, where it most naturally fits.
The nondefault ``Adjust for sampling errors'' motioncorrection option in SPM was not used.
8 sessions (of the 99) were excluded from the original analysis in [18], due to ``obvious movement artefacts''. These were however included in our analyses as we did not consider that there was sufficient objective reason to exclude them; the estimated motions for these sessions were not in general significant outliers relative to the average motion across sessions and any apparent (activation map) motion artefacts were not in general significantly different from the majority of the sessions. The quantitative results given in Sections 4.3 and 4.6 were in fact recalculated without these 8 sessions (i.e., reproducing the same dataset as used in [18]), but without any significant change in results, and therefore are not reported here.
