Hi Adnan,
I think it is possible that if certain statistical assumptions are not met in the real world, you may see "artifactual" differences between conditions where the amplitude (peak-to-peak, max minus min etc) of your fMRI regressors are different between two conditions which you are directly comparing. In your case, for your short duration condition (I assume event-related design?) your predicted regressors after convolving with HRF will have smaller-amplitude fluctuations than your long-duration condition, since the height of the HRF/brain response increases as the event duration increases. In the invidividual-subject regression, if in fact the brain response (or perhaps task-correlated noise fluctuations in the BOLD signal) is not increased with duration exactly to the extent the model predicts, but instead is more similar than the model predicts, the regression will need to multiply the low-amplitude short-condition regressor by a bigger beta value (pe, regression slope) in order to "match" the fluctuations in the BOLD signal. You would then find that the amplitude of your parameter estimates is systematically larger in the short-duration condition than the long-duration condition. If this effect is in fact present you would expect it to run counter to/attenuate your predicted effects which seem to involve greater activation in "reflective processing regions" for the long duration condition.
Discussion of this scaling issue in a different context can be found in Lebreton & Palminteri, Revisiting the assessment of inter-individual differences in fMRI activations-behavior relationships,
https://www.biorxiv.org/content/biorxiv/early/2016/04/11/036772.full.pdf) - see 4th paragraph in the Discussion for a mention of the event duration/reaction time issue.
It is possible that in your context this scaling issue is safe to ignore.
Would be curious to hear what the experts think of this.
Good luck,
Dan