These two things are not directly related it is not one or the other. RFX allows you to take into account the variance between individuals when averaging over the participants which is needed if you want to conclude on whole populations with the same characteristic (e.g. every person with the same disease, vs. all controls) as oppose to just your sample (ie the 2 x16 people you scanned).

Monte Carlo Simulation Cluster Thresholding is one way to correct for multiple comparisons (wrt to voxels) when determining your threshold for statistical significance.

Many thanks for clarifying this, Oury, and sorry for the confusion. I just noticed my mistake having mixed up the RFX-GLM with Gaussian Random Fields (GRF) for multiple comparison correction. I have been trying to get my head around why apparently for both methods the smoothness of the data matters.

Do I understand correctly that in GRF, the estimated smoothness of the data (intrinsic + induced by proprocessing) is used to compute the corrected p-Value (as part of the RESEL term R)? Similarly, in Monte-Carlo Methods the smoothness of the real data is estimated to smoothe simulated data that has been generated under the null-hypothesis - these are then thresholded with some p-value to estimate Cluster sizes?   

My main question now is IF and why the degree of data smootheness also matters when computing the RFX-GLM? My guess would be that it is for this reason: RFX only difference to FFX is that it takes the pooled (within + between subject) variability into account. I remember that one reason why we smoothe data in the preprocessing is to get rid of some (arbitrary) individual variability, i.e. we want to make the images between subjects a bit more similar to another. So a higher degree of smoothness also lends more power to the RFX-GLM by decreasing the within, and especially the between subject variance (which is usually the bigger one anyway)?   

Amongst other things smoothing your data reduces the number of voxels you are looking at and hence the degrees of freedom (the reason you need input the amount of smoothing at every step). In my experience unless you are using tiny voxels or comparing hugely different values of smoothing, it won't have a very big effect on your results ie the global story (including when you use RFX averaging). There are many free slides online that address these issues at least in the context of SPM or FSL. Even if you use something else the principles are the same.....