I have an experiment set up to test the activation difference to two types of long stimuli. Subjects see a type 1 movie of 12 seconds, or a type 2 movie of 12 seconds. Each movie is followed by a picture presentation and button response. The activation to picture and response is not of particular interest, but I have included it in the model in order to account for it. So... for a deconvolution, the maxlag for both types of movie events is long (I've been using 14) but for the picture event it is shorter, because it is a shorter event (up to 6 seconds, I've been using maxlag 10). I know my maxlags are long, but I'm trying to include return from undershoot, and the irf's I've generated, give a reasonable looking wave estimate using these long lookaheads. Additionally, I jittered the presentation of the picture/response (2, 4, or 6 secs), thinking that would make the response portion easier to pull apart from the preceding movie.
I'm not sure if using two different maxlags in a deconvolution is valid--I have to admit that I assumed it would be since the program allows it. It seems to have worked pretty well, except that for some of my subjects, the contrasts (area under the curve type1-type2) look awful, while the point by point activation (looking at the peak response brick/F-stat for each condition) look good. I haven't yet tried to do a regression with an extrinsic wave form. I'm wondering if there's an obvious feature of the design/analysis that would explain the problems with the contrast activation maps when the peak response maps look good (by good, reasonable locations of activation, and clean).
Thanks,
Arika