Methods

We use two different datasets. The first is a boxcar visual experiment with a reversing checkerboard boxcar stimulus (30 seconds on, 30 seconds off). The second is a single-event pain experiment, for which the stimulus was a thermal noxious stimuli of 3 seconds duration administered to the dorsum of the volunteer's left hand using an electrical resistor to generate heat with varying ISI (between $30$ seconds and $50$ seconds). For both experiments echo planar images (EPI) were acquired using a 3 Tesla system with TR=3 seconds, time to echo (TE) = 30ms, in-plane resolution 4mm and slice thickness 7mm. The first 4 scans were removed and the data was motion corrected using MCFLIRT (Jenkinson et al., 2002) and high-pass filtered as described in Woolrich et al. (2001). The data is not spatially smoothed. We use an HRF resolution of $\rho=6$ (i.e. 0.5 seconds). We then fit the model with an autoregressive order of $P=4$ and use a basis set of the top 3 eigenHRFs shown in figure 3. We use two different models, one with no HRF constraints ($m=0$ and $C=I$) and one with HRF constraints ( $m=\tilde{m}$ and $C=\tilde{C}$ as given in equation 17). We use an f-contrast to pick out the linear combinations of the basis functions. The resulting f-statistics are f-to-z converted to produce spatial maps of pseudo-z-statistics. The fully adaptive mixture modelling described in section 5 is then used to provide probabilities of a voxel being activated.