Dear Uta,
I had a similar question to yours recently, and so I'll answer this as I now
understand it. Unfortunately, it's not quite kosher to inflate your second-level
degrees of freedom (dfs) by dropping the distinction between sessions/subjects.
Your true second-level dfs should be something like 18 for the study you
describe (10+10-2 for the two groups). I assume that in your first (low-power)
analysis you averaged (in some sense) the four sessions to get a single contrast
image/subject. By doing this, you'd have hoped that you were reducing
intra-subject error and thus increasing the sensitivity of your analysis. This
is where you gain by having 4 sessions/subject, and not by modelling them at the
2nd-level.
Best,
Dave.
ps If it's any consolation, the significance of my results vanished in a similar
fashion when I performed the 'correct' 2nd level analysis.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|