All fmri analysis software I have used gives you the option to get estimates for condition contrasts during first-level (single subject) regression. Whenever I’ve set up single-subject regression, I’ve always included computation of every contrast someone could ever conceivably want (or sometimes all possible combinations of variables), out of fear that otherwise I may someday have to reanalyze the data to get those contrasts. However, this can eat up a lot of space if you have more than a few conditions.
However, another approach would be to just get estimates for each individual condition, and then use those to make contrasts as needed - by subtracting one condition from another with something like AFNI’s 3dcalc or fslmaths.
I’ve never been quite clear on whether these two methods should be mathematically equivalent, close enough that it shouldn’t matter, or different enough that one method is preferred.
Does anyone have any insight on this? Is it preferred to create contrasts during single-subject regression, or does it not matter?
I think that the only difficulty is to estimate first-level variance for a contrast, which we requires to take into account the covariance between the conditions: so you need more than per-condition effects and variance estimates to do that.
Now, if in a second-level analysis you disregard the first level variance, I don’t see any issue.
Is it preferred to create contrasts during single-subject regression, or does it not matter?
The conventional population-level analysis in neuroimaging solely relies on point estimates of an effect from the subject level and fully ignores the uncertainty information. Such an approach, as an approximation, may work roughly fine most of the time; however, if possible, it would be preferable to have both the point estimate of a contrast and the associated t-statistic (which contains the uncertainty information – standard error – in the denominator) available for more accurate population-level modeling (e.g., using 3dMEMA, RBA and TRR in AFNI).