re: First question. Does your design have gaps ("rest") between the trials, or is it one very long trial (e.g., a learning experiment)? If you just have one very long trial and want to see trends in activitation during that trial, you are pretty much out of luck. The "noise" contains long duration trends, too, so you can't reliably find steady activation that drifts down over a long period of time.

However, if you have a "normal" type of experiment (block design with short-ish blocks, or event-related design) that has interspersed "baseline" condition intervals, then you should be able to detect activation trends. There are a couple ways to imagine doing this.

Method A would be to divide the stimulus periods into 2 classes -- early and late -- and treat each class as a separate stimulus. Then a GLT with a [1 -1] contrast between these would bring out significant differences between early and late responses, while the GLT with the [1 1] contrast would bring out the sum of the responses -- you would probably want to first restrict yourself to look at voxels whose sum GLT is significant, then see which of these have a significant difference GLT.

The disadvantage of Method A is that you have to divide the responses at some point that probably isn't a priori obvious. Method B is to assume that the habituation takes place at some given rate (linear with stimulus index?). In that case, you would want to create a second regressor that incorporates this trend. If we call the habituation trend function w(i) [where i=stimulus index], then a reasonable regression model for the i-th response would be
ri(t) = a*h(t-si) + b*w(i)*h(t-si)
where
ri(t) = response at time t to i-th stimulus
h(t) = ideal response to a single stimulus at time 0 (from waver)
si = i-th stimulus time
a,b = response amplitudes to be estimated from data by 3dDeconvolve

The second regressor (stim_file) input to 3dDeconvolve would then have to be the sum over i of w(i)*h(t-si). How to produce such a thing? At the moment, waver is not set up to generate such a thing (e.g., with the -tstim option to provide the si's and a nonexistent -wstim option to provide the w(i)'s). This could be added, at the cost of intense suffering.

re: Second question. I think that putting regressor with no lags that has 1 at the TR immediately at-or-following each eyeblink is a reasonably way to try to remove the effect. You could use -stim_base to put this regressor into the baseline model, so it wouldn't count in the overall F statistic map.

bob cox