Dear Lorena, Helen and all,

In addition to the mean signal, which can be modeled as an extra EV, the scanners may also differ in the variance of the signals. Some modalities (e.g. VBM) tend to be quite robust to these differences, but the same may not apply to all methods. These variance differences can be accounted for in randomise defining the design.grp file to contain one group per scanner, so that the permutations happen only between subjects acquired in the same device.

In other words, make sure the design is balanced, include the extra EV to account for mean differences between scanners and use a .grp file so that exchangeability is asserted within scanner. If the design is not balanced, you may consider regressing out the mean effect of the scanner first, then running randomise on the residuals, with all the remaining covariates included the design, plus the .grp file.

Finally, while this will account for mean and variance differences, other possible subtle scanner effects may remain present (e.g. different types of geometric distortion possibly affecting differently the registration, the possibility of biased study-specific template for unbalanced groups, etc).

All the best,

Anderson