Dear Joelle,
I agree with everything Donald has already stated. If you have ten trials you should try to account for ten trials, even if you're only interested in the first and the last one. Within the "fMRI model specification" module specify ten conditions, the first condition reflecting the first trial onset, the second the second and so on. This will result in a design matrix Cond1 Cond2 Cond3 Cond4 ... To look at the differences between first and last trial (~ condition) you would have to specify a t-constrast [1 0 0 0 0 0 0 0 0 -1 ...] to test for Cond1 > Cond10 and [-1 0 0 0 0 0 0 0 0 1 ...] for Cond10 > Cond1.
However, as previously discussed your paradigm is special due to 1) the very long blocks ("trials") 2) the relatively short interval between successive blocks. The default high-pass filter removes all the low frequencies. In your case the low frequencies can be expected to include not just noise (that's why we usually want to filter) but also relevant signal. Thus it is likely that you can't detect much with a single condition reflecting all ten onsets. The same issue holds if you turn to ten separate conditions. For a comparison between something at the beginning and something at the end of the session it would especially be important to account for signal drifts. Thus you could of course adjust or disable the HPF, but this way you would also introduce lots of noise. Therefore I would not try to go on with these long blocks, as they are very disadvantageous from a theoretical point of view.
If you really have to rely on these long blocks, and if they really can only be modeled as a single trial (which would be surprising, usually there are many trials/stimuli/events within a long block), then you should alter the paradigm. For each of the trials, you could go with an extra fMRI run at the scanner. The run should start with several tens of seconds rest and also end with several tens of seconds rest (so much more rest compared to the current paradigm). Due to the rest periods it should be easier to properly estimate BOLD response during the trial. Due to the separate sessions HPF would be less of an issue, as the HPF is applied separately for each of the sessions. As long as activation during rest remains the same across fMRI runs then you could contrast the ten different trials on single-subject level. But this is just a consideration, I have no idea about the usual settings for experiments like yours. It's just a fact that the current design is definitely suboptimal when it comes to fMRI analysis.
Best
Helmut
