outliers

[annib]

Hi Chris and team
Do you guys have any feeling for how outliers may influence the PCA.
For example how much will the presence of a partial volume effect in
inferior temporal lobes in 1 subject.

thanks
Anna

[Christian Habeck]

Hi Anna,

this can be checked from the workspace You can see whether here is an
overly influential subject whose causing the first PCA all by himself.

If this isn't in the demonstration, I could show you if you make the
workspace available to me.

Chris

--
Christian Habeck
ch...@columbia.edu
http://www.cumc.columbia.edu/dept/sergievsky/cnd/habeck.html

[annib]

Hi Chris
Sorry only got round to looking at this. So would I be looking at the
SSFs of the best fit composite or, the separate PCs in this or just
PC1?
thanks
anna

[Christian Habeck]

Hi Anna,

you would look at the SSF of PC1. If one subject is overly influential, it might create its own Principal Component. This means that SSF1 for this subject would be a lot larger than for any of the other subjects. It also means that if you take the voxel loadings of PC1 and correlate them with the brain map of this subject that went into the group analysis, you should get a very high correlation, like R^2=0.95 or so.  Basicually, PC1 is created by the outliers itself, and not by any meaningful across-group variance in the data.

Often I have seen this wen there is an error in the processing. If you look at PC1 or at the brain map of this problematic subject, is there a voxel with very high intensity? For instance several orders of magnitude higher than any other voxel? This would strongly hint at some co registration problem.