Daer Joanna,
> We have two conditions, block design. We are interested in changes
> over time. We ran our task as four separate scans/sessions (ie. we
> started and stopped the scanner four times) only in order to give the
> subjects a little break.
> If we specify four sessions, then parametric modulation with time
> happens separately for each session - but we want to look at changes
> over the entire task, not within each session.
>
> Say we parametrically modulate only cond2. So the design matrix is
> like this
> cond1 cond2 cond2PM
>
> One option is to change the file names and treat them as though they
> were all collected in one session (post realignment).
>
> If we do this
> 1) will our power be hurt because we have a lot of variance due to session
> but no session column in the design matrix?
Possibly, but you can easily include user-specified covariates to model
the session effect.
> 2) In addition, might we get false positives due to linear scanner
> drift?
No. Linear scanner drift is a main effect of time. You are testing for
a time x condition interaction.
> 3) Could we decrease error variance by using a regressor with all 1s for
> files in the first session, 2s for files in the 2nd session, 3's for files
> in the 3rd session and 4's for files in the 4th session?
> 3) Would a better solution to decrease error variance attributable to
> session be to add four conditions named sess1 sess2 sess3 sess4 and give
> onset vectors for them corresponding to every file in in sess1, 2 etc
> respectively?
This would be better but you do not have to model session effects as trials,
you can simply enter 4 sessions with 1s for all scans in each session.
> 4) are there any other problems with treating four sessions as though they
> were one?
There are assumptions about sphericity but these are the same for any
deisgn.
I would suggest that you model all four seesions separately (as normal)
and look for 'within-session' time x condition interactions with the
appropriate modulation of trial-specific regressors (as normal) and
(ii) look for 'between-session' adaptation using a contrast that tested
for decreasing main effects of condition over sessions. This has the
advantage of decomposing time x condition effects into within and
between session components. This is a more realistic model if you had
4 distinct sessions. You could then do a conjunction of the two? (which
would be quite powerful).
I hope this helps - Karl
|