Another question regarding MVPA:

I'm trying to figure out the optimal number of components.
I have 169 participants with 1 session of resting-state scan; but in my second-level analysis, only 120+ participants would be included due to the analysis design (longitudinal prediction that requires data at both timepoints).

I've tried out running second-level PCA with 4 components (cumulative percent GM variance was ~80%), 5 components (~84% variance), and 8 components (variance explained in the attached screenshot).

And when I do the analysis, bigger clusters are around the same peak coordinates, but smaller ones sometimes show up and sometimes disappear depending on the number of PCA components. 

Now, I'm a little bit confused whether I should increase the number of components at second-level PCA up to the point where more than 90% of variance can be captured, or 4-5 components that capture 80~85% variance would work fine.