Verena,
A couple of possibilities:
(1) The PPI model is inaccurate because your design contained more than 2 conditions (including fixation/baseline); --> solution: use the gPPI toolbox to get more accurate estimates
(2) It looks like you used an interleaved acquisition, if you did, then this could be an effect of the acquisition interacting with subject motion; --> solution: identify in the raw data any timepoints that have a striping artifact and remove each with a separate regressor in the design matrix. Each regressor will be 0s, except for the timepoint to remove, which should have a value of 1. If you have 10 bad points, then you will want 10 regressors.
(3) It could be a problem with the acquisition, do you see rings around the brain or ghosting in the raw images?; --> solution: I'm not aware of a solution.
(4) A combination of the above issues.
In summary, I'd make sure to use the gPPI toolbox to get a more accurate estimate of the PPI values and always check the raw data for artifacts (as well as check at each stage of the processing).
Hope this helps.