Dear FSL support forum,
For the purpose of addressing my current concerns, let’s assume the I have a very simply paradigm involving only the following conditions:
Left Hand (LH) - movements of the left hand
Left Foot (LF) - movements of the left foot
Rest - baseline (fixation)
I have 4 runs per subject, and what I would like to determine at the single subject level (averaged across runs - i.e. second-level analyses) are those areas showing LH > LF AND LH > rest.
To this end, I have compared results derived from the following two approaches:
A. “Pre-threshold masking”
Step A-1: Calculate the results of the contrast: LH > rest (at z =2.3, cluster-size corrected at p <.05); second-level (gfeat directory).
A-2 use this output: thresholded z-stat map - to constrain the contrast: LH > LF
i.e. Step A-3: Calculate the results of the contrast LH > LF (at z=2.3, cluster-size corrected at p<.05) --- Constrained by the results of A-1: --- using Pre-thresholding masking --- selecting the threshold_zstat (that reflects LH > rest).
--- I COMPARED THESE RESULTS WITH APPROACH B ---
B. Contrast masking (at the second-level)
Step B-1: Set data input to: “3D cope images from FEAT directories”
B-2: Select the COPEs from the gfeat that reflect: (1) LH > rest and (2) LH > LF
B-3: Specify these as EVs, and then as contrasts C1: LH > rest and C2: LH > LF, in the model
B-4: Use post-stats, contrast masking: Mask C2 with C1 (with the option "mask using z>0" NOT selected)
Again, I applied the default threshold settings: z 2.3, cluster-size correction at p <.05.
The maps are virtually identical.
However, the results of method B - using contrast masking - are not corrected for multiple comparisons using cluster-size thresholding methods.
That is, the resultant maps from methods A and B overlap entirely, accept that the thresholded z-maps from Method A are cluster-size corrected whereas the thresholded z-maps from Method B are not.
That no correction is being applied to the resultant maps from B is clear:
i. Some of the active clusters comprising the thresholded maps from Method B are only 1/2 voxels - these voxels are absent from the thresholded results from Method A.
ii. When these maps are compared with corresponding uncorrected z-maps, they are identical.
Q-1: Does the contrast masking option not allow for implementation of cluster-based correction of the resultant z-maps (in the case of conjunction-derived outcomes: e.g. C2: LH > LF masked by C1: LH > rest)?
Q-2: In general, is there a consensus about whether one of these methods - A versus B - is more appropriate? I understand that B is a ‘true’ conjunction analysis, while method A is not, but is there anything invalid or worrisome about performing method A?
One of the features I find potentially advantageous about using method A is that the one may apply different statistical thresholds to either contrast. For example, one might reasonably argue that a contrast of a stimulus Condition versus Rest merits application of more stringent statistical thresholds - greater than z =2.3 (which, in general, is a very liberal minimum statistical threshold to use for fMRI data, in my experience). With Method A above, one could create the map of Condition > Rest at a higher z-min, (e.g. z = 3.09, p < .001, uncorrected) then apply this map to constrain more selective contrasts of interest, set to more liberal thresholds.
Of course, Method A also reduces the number of multiple comparisons - making step two of map thresholding - correction for multiple comparisons - more reasonable. After all, in this situation, I have no hypothesis about areas/voxels showing greater activity for Hand versus Foot movements, but NOT for Hand above baseline (rest = fixation). Here, it seems sensible to constrain the number of voxels being considered to those that are activated by the task versus rest.
To all you FSL gurus - your input is greatly appreciated.
Many thanks in advance,