Fig. 5.

Investigating the effect of patch extent. Data were simulated on the superficial (left) or deep (right) surfaces models using either a patch extent of FWHM = 10 mm (upper panels), or FWHM = 5 mm (lower panels). Three sources were modelled, at a per trial SNR of 0 dB. In addition, rotational coregistration errors of 0, 2 and 5° were simulated.

Panel A in the top row shows relative free energy differences for simulations based on the superficial surface and using a patch extent of FWHM = 10 mm. Looking at the leftmost set of bars (corresponding to reconstructions using FWHM 10 mm), we observe strong evidence in favour of the correct model (positive values) for reconstructions at true patch extent, both for 0 and 1, but not 5° of coregistration error. This pattern is destroyed when underestimating patch extent (FWHM = 5 mm), as illustrated by the rightmost set of bars where no clear difference between surfaces emerges.

However, looking at Panel B, which shows the comparison based on the same patch size (10 mm), but using the deep surface model, it is clear that even if we underestimate patch extent (5 mm), the strong evidence in favour of the true model is preserved. This suggests that when we underestimate patch extent, we are introducing a bias towards deeper surface models.

The bottom row shows relative free energy differences for simulations based on smaller patch extent (FWHM = 5 mm). Panels C and D correspond to the true surface models being superficial and deep, respectively. Here, in the case of superficial-surface-based data, when overestimating patch size, the evidence in favour of the superficial surface model is preserved (Panel C). On the other hand, as shown in Panel D, in the case of deep-surface-based simulations, overestimating patch size decreases the evidence in favour of the deep surface model. This indicates that by overestimating patch extent, we are introducing a bias towards superficial surface models.