Dear Experts,
     I have data from a mixed design with 2 within subject factors (A, B
each with 2 levels) and a between subject factor (Group, 2 levels).
Unfortunately, it ended up that the number of subjects is different in the 2
groups. My understanding is that when the variance components are estimated
using ReML, it should work irrespective of whether the design is balanced or
not (right?). However I have the following questions:

If I specifiy A,B and Group as fixed effects (with subject as the only
random effect):

(1) How does SPM know how to calculate the F ratio for the main effect of
Group? Should I specify what the appropriate error term is?
(2) Given that all factors have only 2 levels, do I still need a sphericity
correction? Also, is there a way to specifiy the model differently to
circumvent dealing with sphericity assumptions?
(3) I have in the past read especially in SPLUS forums about issues with
determining the 'denominator' degree of freedom when using ReML.. Any
references on how to understand that would be great..

Any comments and references are most appreciated.