> Let us consider that a single voxel shows activation during this comparison. [...] And what about deactivations?

"Activation" typically refers to a positive BOLD response, "deactivation" to a negative BOLD response. Note that the direction of the response is going to depend on the baseline. If you present small visual stimuli with a blank screen in between then you will probably detect a positive BOLD response in visual cortices for the stimuli. Now present large images with high contrast instead of the blank screen. When modeling the small stimuli you might detect a negative BOLD response because the visual input is lower relative to the rest of the experiment. This example is artificial of course, but e.g. in fast event-related designs the baseline can well be shifted relative to "default" paradigms with a certain amount of fixation periods.

> Let's consider the 2nd contrast: FF vs. NF [0 1 -1]. Let us consider that a single voxel shows activation during this comparison. Does this
> measure whether the BOLD signal for FF is greater than the BOLD signal for NF? If this is true, can we say that, this voxel is more activated
> during FF than during NF? 

Basically this voxel would be associated with a positive contrast estimate, meaning the beta estimate for FF is larger than the one for NF. But both the two might be negative, one might be positive and the other negative, ... thus it is incomplete and possibly misleading to simply state "the voxel was activated more strongly during FF than during NF" (or to a lesser extent during NF relative to FF), because this somewhat implies positive beta estimates (positive BOLD response) for both the two. To make sure whether this is really the case you would have to look at the beta estimates for the two conditions, in your case beta_0002 for FF, as it's the 2nd condition, and beta_0003 for NF, as it's the 3rd. E.g. if both the two were negative it would be better to state "this voxel was less deactivated during FF relative to NF".

Another issue is whether these voxels really show any *significant* BOLD response at all. Assume FF to be associated with positive beta estimates non-sig. different from zero and NF to be associated with negative beta estimates non-sig. different from zero. You could state "that voxel is activated more strongly during FF relative to NF", but in a strict sense there is no activation at all (in general people are not that strict though). 

If you want to test whether there are any sig. activations/deactivations for a certain condition you would have to set up contrasts like [0 1 0] (positive activations) and [0 -1 0] (negative activations aka deactivations).

> I am having a lot of trouble finding the difference between a perceived decrease in activation vs. deactivation.

If you want to be precise, then use the terms
- activation or deactivation if the conditions/simple contrasts like [0 1 0] are associated with significant positive or negative beta estimates
- increased/higher/larger or decreased/lower/reduced activations for differential contrasts like [0 1 -1] if both are associated with sig. positive beta estimates
- increased/higher/larger or decreased/lower/reduced deactivations for differential contrasts like [0 1 -1] if both are associated with sig. negative beta estimates
- for everything else check your tables and figures and where necessary provide corresponding information, e.g. "region xyz was associated with a BOLD response more negative for B relative to A. Closer inspection revealed that A was non-sig. different from implicit baseline for the corresponding cluster, while B was associated with sig. negative estimates, indicating a deactivation"

As a take-home message, for differential contrasts like [0 1 -1] it is never bad to look at the activation levels for the conditions under investigation. This especially holds for more complex contrasts. E.g. for (FF vs. NF) in patients vs. controls, that is the interaction Face x Group, you should probably also provide some plots for sig. clusters in case the activation levels cannot be derived directly. For interpretation it is definitely interesting to know whether 1) both controls and patients show a positive emotional effect (FF - NF), larger in controls relative to patients, 2) controls show a positive emotional effect and patients show none, 3) controls show a positive effect and patients show a negative, 4) ...

> How can one contrast measure both? 

It's the same contrast, just a different interpretation. [0 1 -1] tests whether FF is larger than NF, which is identical to whether NF is smaller than FF.

> Is the above (marked in red) a typo? Did you mean to say -1*2+1*2+1*2=2? That would make a lot more sense to me. 

This is indeed a typo.

> However, what do you mean when you say that the "contrast value will match the individual conditions better" with the latter contrast? Should I change my contrast to the latter value? 

Think of the beta estimates as amplitudes, contrast estimates as differences in amplitude. With [-1 1/2 1/2] the contrast estimates correspond to the average amplitude of the 2nd and 3rd condition relative to the 1st condition. With [-2 1 1] the contrast estimates correspond to the sum of the amplitudes of the 2nd and 3rd condition relative to twice the one of the 1st condition. When plotting the latter the difference would be twice the underlying difference in amplitude. If you label your figure correctly this is okay, but often people forget, so it's better to scale the contrast vector correctly right from the beginning.

Best,

Helmut