Hi Mark,
You need one extra EV for each subject that is to be excluded. In this example, there would be 2 EVs coding group membership, and 3 EVs coding the excluded subjects, one per subject. The overall number of EVs would be 5. Since here you have 9 subjects total, this design would have 4 degrees of freedom.
Suppose you had excluded by hand the volumes from the 4D file. Then there would be 6 subjects. The 2 EVs coding group membership would stay, and this other design would likewise have 4 degrees of freedom.
Using a single EV for all subjects isn't correct in that the regression coefficient for that 3rd EV would yield an average value for the "excluded" subjects (that wouldn't actually have been excluded, but rather constituted a 3rd group), and the residuals for these subjects would not be equal to zero, thus affecting the estimate of the variance. Plus, the degrees of freedom would be 9-3 = 6, which is incorrect as can be seen from the case in which you had excluded the subjects by hand (that has df = 4).
All the best,
Anderson