Dear Amy,
I performed a similar analysis concatenating 4 runs recently. I found the results were very similar to the normal approach (not concatenating, but modeling the 4 runs separately), although some t-values went up in the concatenated analysis.
You do in fact need the extra regressors for 3 of your four runs, to model a constant. This regressor is a zero for all scans not in that run, and a 1 for each scan from that run. You need this regressor because the "baseline" value for each run may be different. That is, the mean value (in scanner units) for one run may be 100, but in the next run may be 120, which will create an effect that looks like a blocked design. Failing to account for this (which is what the regressors do) will potentially cause serious problems with the model estimations.
But that is not enough. You should also create regressors to account for the temporal drift in each run. There is a good posting about this somewhere in the mailing list (I don't have the link but if you want you can search for it).
Each run needs a drift regressor. If you have four regressots, and 'nscans' per run, you code to create the regressors in matlab is:
% regressors for sessions, taking into account differing baslines
sess_regress1 = [ones(1,nscans) zeros(1,nscans)]';
sess_regress2 = [zeros(1,nscans) ones(1,nscans) zeros(1,nscans) zeros(1,nscans)]';
sess_regress3 = [zeros(1,nscans) zeros(1,nscans) ones(1,nscans) zeros(1,nscans)]';
%each block requires a separate linear regressor
lineregress1 = [(linspace(-1,1,nscans)) zeros(1,nscans)]';
lineregress2 = [ zeros(1,nscans) (linspace(-1,1,nscans))]';
lineregress3 = [zeros(1,nscans) zeros(1,nscans) (linspace(-1,1,nscans)) zeros(1,nscans)]';
lineregress4 = [ zeros(1,nscans) zeros(1,nscans) zeros(1,nscans) (linspace(-1,1,nscans))]';
I hope this helps. It worked for me!
Colin Hawco
Post doctoral Researcher
Centre for Therapeutic Brain Intervention
Centre for Addiction and Mental Health (CAMH)
University of Toronto
"Disorder Expands"
----------------------------------------------------------------------
Date: Fri, 9 Nov 2012 05:06:03 +0000
From: Amy <[log in to unmask]>
Subject: concatenating runs
Hello,
I'm running an fMRI analysis with four different runs. I'm concatenating the four runs into one long run in my design matrix (I realize this is not recommended). I created four different regressors to model each of the four runs (using dummy coded 1s and 0s).
Because these four regressors add up to the constant term (which is automatically created in SPM), they are not estimable. The boxes are grey below these regressors when reviewing the design matrix.
I don't care about specific session effects, but I want to make sure that this is okay (i.e. my betas for the events of interest are valid).
Do I need to remove the constant term?
Thanks for any advice!
Amy
|