 |
 |
Modeling rest is an interesting topic, and I looked at it more closely for two independent fMRI data sets in the past. Might be of relevance for you. Note that this is descriptive only! I did some statistical testing as well, but it would take some time to get the things together again. One data set was about a cognitive task (visual and auditive stimuli presented simultaneously, 3 different task conditions + rest, task conditions were presented twice as often as rest), the other was about saccades (pro, anti, rest - all three conditions had the same frequency). Rest blocks had the same duration as task block (about 20 s each). For both data sets, the finding was the same: When rest was not modeled, activations for the conditions (like [1]) were relatively low in group analyses. Deactivations for the conditions [-1] were also rather small. When rest was explicitely modeled, the activations/deactivations, for example [1] or [-1], were quite similar compared to the other model. The beta values were quite similar as well. When computing the differential contrasts relative to rest [1 -1] or [-1 1] task-related activations (eye fields, insula) were much stronger, same for deactivations (frontal medial, PCC/precuneus). This had no impact on differential contrasts between different task conditions though. I think this can be explained as follows: If the regression analysis were based on boxcar functions only, then indeed one would overfit the data by adding another regressor for rest (everything which is not task condition A or task condition B is rest condition C, so the time course can already be explained by two conditions). But the boxcars are convolved with the HRF, and then the third condition does not simply follow from the other two time courses. Thus it might be useful to explicitely model the rest condition if you want to show some contrasts comparing one condition with the implicit baseline/rest (like "prosaccades > rest" in the data set with pro/anti/rest to show that the paradigm worked well). For differential contrasts between task conditions (which have their own regressor in any case) it seems to be irrelevant. But it would definitely be necessary to further validate this with statistical tests.