At the moment I am trying to interpret my beta weights from an fMRI study of motor preparation and execution. Overall our experiment had 4 different conditions that each consisted of a preparatory phase and an execution phase with either the left or the right hand. Thus I builded my GLM with overall 16 regressors of interest (2 hands x 2 phases x 4 conditions).

Unfortunately, it was not possible to completely delineate the regressors for preparation and execution and they correlated to some extent (one pair with r=77 the other pairs around r=40-50).

Afterwards I entered my regressors in a flexible factorial ANOVA. Hereby, I skipped the contrast images for the condition in which preparation and execution was highly correlated, because the design matrix looked not like expected for an ANOVA.

When I run the 2(right hand/left hand)x 6 (preparation cond 1, preparation cond 2, prepartaion cond 3, execution cond 1, execution cond 2, execution cond 3) and plot the beta weights for different areas depending on condition I get somehow mirrored beta weights for the preparation and execution. That means when the voxel exhibited a positive beta during preparation it was negative during execution and the other way round.

For me this looks like a statistical problem. So far, however, I do not have an idea why it occurs. Even though my regressors were correlated, I did not think that this may be the problem, since this should just decrease my statistical power (shared explained variance enters into error term) and my results should be valid.

I would be really glad if some one could help.