Dear Angela,
if one condition is empty, don't model it. It might be a pain for individually defining contrast vectors, but it is certainly better this way.
Concerning your 3x3, the problem is that there might be differences between the sessions (learning effects, fatigue, ...). If you average across sessions for some subjects, but not for others (or only for a subset of conditions), then some of the estimates are going to be biased. From a theoretical perspective it would be best then to exclude the whole session of these subjects (to avoid a bias due to averaging across sessions for e.g. A1B1, A1B2, ... but not for A3B3), and as you only have two sessions, probably reject the whole data set in affected subjects (to avoid a bias due to averaging across sessions for some subjects and using a single session for others).
Leaving this aside, it sounds as if you actually have a rather low number of trials per condition per session. This might be another problem. It probably doesn't make much sense to model a condition based on two or three trials.
Hope this helps a little,
Helmut
|