One of my questions is how do we incorporate a trend into an event-related model. I wish to test for which voxels show habituation or a progressive decrease in signal within a run and across runs per trial type. Obviously, one would not simply add trial number as a regressor. For a block design model I believe one would simply put a “trend” regressor which would represent the same .1D file that loads the overall “stimulus” variance but would have the waver file multiplied by a trend (e.g. linearly decreasing or increasing, exponential or any other function). This regressor would have no lag and would be incorporated into 3dDeconvolve as another stim_file with the trend .1D file as stated. However, in the case of an event-related model, would we have a “trend regressor” that had for example a 6TR lag so there would be 7 regressor coefficients instead of the one coefficient as in the case of block-design? Also we wouldn’t have to run the .1D file through “waver” first and then multiply by the trend since we are using a “pure deconvolution” instead of a regular multiple regression as is the case in block design. However, if we do have 7 IRF parameters/coefficients resulting from the “trend”, how do we interpret them? I don’t believe one could just collapse all of them to one coefficient and retain the necessary information or can you? The question remains how does one include a trend in the “pure” deconvolution model? This was my first question.
My second question is about how to create a regressor which allows one to find the correlation between our EMG data (e.g. a dataset collected over the time of the scan) and the fMRI data by using the pure deconvolution model. Since we have a TR=2s and the duration of each eyeblink response caused by the same stimulus as presented for the fMRI lasts only ~ 400ms which is less than a TR, it would seem allright to take the .1D file for that stimulus and multiply each “1” in it by the magnitude of the eyeblink response for a given presented stimulus. This would be the same as a regular motion regressor since there one takes a value of the motion parameter at each TR in the time series. However, here there is no “response” at TRs which are not at stimulus presentation points so the value at all other TRs is zero. Also I assume that this regressor would have no lag and would be incorporated into 3dDeconvolve the same way as a motion regressor. Am I thinking about this the right way?

Anyway your insights and comments would be very helpful.
Thanks,
Linda Heidinger
Research Assistant
Department of Psychiatry
University of Chicago Hospitals