Dear Susanne,
the observed effect is due to filtering design (and data) with a weight matrix computed from error covariance estimation. Especially in a flexible factorial design where multiple contrast images per subject are entered, careful and statistically sound assumptions about error covariances need to be made. In your case, it seems that you allowed lots of freedom to error covariance (i.e. "Variance: Unequal" and "Independence: No" for almost all factors). This allows to fit almost all subtle patterns in error covariance.
It may reveal real issues such as two subjects (5? and 11?) having a much lower error covariance than the others. In that case, results would be dominated by these two subjects with little or no contribution of any other subject. Since you have got this information, it would be a good idea to check model fit and effect sizes for the two highlighted subjects on first level and compare to some of the other subjects.
More likely however, the error covariance model is too complicated to be fitted successfully. Did you notice unusual processing times during non-sphericity estimation while fitting the model? Usually, non-sphericity estimation should converge rather soon. If it does not and the check above does not give any reason to assume an issue with your data, you should revisit your error covariance specifications. There are some assumptions that will hold in many cases (you should check them for your specific experiment and document them when publishing results!). Assuming your subjects are drawn randomly from groups of homogenous populations, one may assume for the group factor "Variance: Unequal" and "Independence: yes" and for the subject factor "Variance: Equal" and "Independence: yes". Setting "Independence: no" for any of those factors would assume error variance of one group/subject to co-vary with error variance of the other groups/subjects. Also, "Variance: unequal" for the subject factor would imply that your groups are not drawn from homogenous populations. This might be the case if you have reasons to assume that there are per-subject effect variations that are not accounted for in your group factor. Error variance dependencies between your experimental conditions should only be modeled by your condition factor. Depending on your experiment, there may be other considerations, but you should try to model expected error covariance based on reasonable assumptions about your experiment.
Hope this helps,
Volkmar
|