Outlier scores in the ‘Parzen_outlier_mvtroi’ table are p-values associated with the test ‘the observation is normal’. That means that one has a probability p of rejecting a normal subject when discarding an observation with associated p-value p. Thus, one will discard subjects with associated p-value < 0.01 if he tolerates 1% error in the outlier removal process. Yet, one should be aware of the multiple comparison problem and the different strategies to alleviate this (see [1])

The p-values computation strongly rely on precise assumptions about the data structure. In practice, these assumptions can not be checked and one should be careful when discarding outliers at a given level. For this reason, a common practice within the neuroimaging community is to fix a proportion of observations that one is ready to discard, and then remove that proportion of observations according to a ranking index. The p-values from the 'Parzen_outlier_mvtroi' table can be used as a ranking index (observations with smallest p-value to be discarded first).

Former outlier scores are deprecated.

[1] Brent R. Logan, Daniel B. Rowe, An evaluation of thresholding techniques in fMRI analysis, NeuroImage Volume 22, Issue 1, May 2004, Pages 95–108