Hi all,
Searching on the archives it seems that the recommended way to perform a second-level conjunction of two contrasts (e.g. first-level contrasts A and B) obtained from a single group of N subjects would be to define a two-sample t-test, enter the A contrast volumes (N files) as 'group 1 scans' and the B contrast volumes (N files) as 'group 2 scans' (making sure to select 'unequal variances', and possibly 'non-independence' between the two levels of this factor), and then define the two separate contrasts ([1 0] and [0 1]), right-click both contrasts and select a conjunction-null hypothesis. I am nevertheless confused by the difference in degrees of freedom between each of the individual contrasts in this model (e.g. [1,0] with degrees of freedom 2N-2) and the degrees of freedom resulting when considering that contrast in isolation (e.g. performing a one-sample t-test on the A contrast volumes, with degrees of freedom N-1). This discrepancy seems to stem from the repeated-measures nature of the data that would seem to be disregarded in the two-sample t-test design. Compared to this discrepant dof values, the effect size (contrast values) and T- statistics would seem to be identical between the ones resulting from the two-sample t-test model and those resulting from the individual one-sample t-tests. Does this difference in degrees of freedom mean that we are obtaining slightly liberal results when performing the conjunction in this way? (compared for example to using a 'manual' conjunction, e.g. computing using imcalc the intersection of the supra-threshold voxels for the two separate one-sample t-tests). And if so, is this something that could/should be 'corrected' for when selecting the voxel-level threshold (e.g. instead of using a given alpha uncorrected voxel-level threshold, one might choose instead 1-spm_Tcdf(spm_invTcdf(1-alpha,n-1),2*n-2) ) or some other method?
Thanks in advance for any help!
|