 |
 |
Dear Philippe, > > In the second level of a random analysis with PET data, I have three > conditions A, B, and C: > > Conditions B and C are well matched except for cognitive components of > interest (CCI). However, condition A differs unfortunately from > conditions B and C for unspecifics aspects (stimulus presentation > mode). I posit that the CCI involved in condition A is also involved in > condition B, but not in condition C. Hence, subtraction [ A - B ] will > remove the CCI and highlight only brain structures related to > unspecific effects, while subtraction [ A - C ] will highlight both CCI > and unspecific effects. > > Since I want to segregate the CCI from unspecifics effects in the > context of condition A (otherwise B - C was sufficient), I perform the > subtraction [ A - C ] masked exclusive by [A - B ] at p = 0.05. > > I wonder about the validity and the interpretative limitations of this > procedure, given that I cannot assume the independence between the two > subtractions. In one sense, it looks like an interaction, but I > strongly suspect that it do not have the real taste of an interaction Your exclusive masking approach is sensible but anecdotal. The reason it is anecdotal is that exclusive masking is being used to reject voxels where the null hypothesis (A - B) can be rejected. However, you cannot accept the null hypothesis (no unspecific effects) at the the remaining voxels. To properly examine those components of the responses to the CCI that depend on the context prevalent in A (i.e. the interaction) you would need a fourth condition that was like A but did not include the CCI. All you can do with your design is exmine a subset of simple main effects (which may be sufficient to address some questions). I hope this helps - Karl %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%