I’m checking if people have noticed that the fixelcfestats output only provides a one-sided response (positive effects) when the design and contrast matrices are set up in the following manner.
Design Matrix (Global intercept and group coding with the control group being the reference)
intercept|group|age
1 0 # ←- Control01
1 0 # ←- Control02
1 0 # ←- Control03
etc
1 1 # ←- Patient01
1 1 # ←- Patient02
etc
For the contrast matrix, if I put 1 for Age column (say col 3) –> I get positive effects only
0 0 1
For the contrast matrix, if I put -1 for Age column (say col 3) –> I get negative effects only
0 0 -1
I double-checked using R/ModelArray FDR statistics (where both positive and negative effects are generated when I set up the general linear model).
It is more conventional to use [-1, 1] values in a dummy variable for a binary category rather than [0, 1], as it keeps things zero-centred, thereby reducing collinearity with the global intercept and improving problem conditioning.
Performing a one-sided test and not automatically testing the opposite hypothesis is a very deliberate design decision, for multiple reasons:
One can choose to construct a single two-sided test rather than a pair of one-sided tests by using an F-test that contains a single row.
(This is quite strange in the context of neuroimaging since statistical enhancement will reinforce connected effects of opposite signs, but it’s nevertheless technically possible)
If one wishes to perform strong familywise error control, then the more hypotheses you test the lower your statistical power will be, and so you do not want to perform tests that are not of interest.
With a large number of elements tested, a complex statistical enhancement algorithm, and extensive non-parametric null distribution generation, the computations involved in additional hypothesis tests are nontrivial and should therefore be based on explicit user request.