The issue of random intercept and slope: With voxel-wise analysis, it's hard (and not worthwhile) to fret about the possibility that one model is better than the other at some voxels and vice versa for the other way around. So, forget about model tuning and comparison, just adopt the model with both random intercept and random slope, and be done with it.
> Not really, I have run a t-test and they don't differ (p=0.4).
Forget about the p-value obsession. Think about the crucial question: do you have some prior knowledge as to whether one condition is expected to have a higher covariate value than the other condition? If so, it may make sense to perform centering for the covariate within each condition. If not, do global centering.
