We have two groups of subjects. Each subject from one group is paired with a subject from another group in the sense that they performed the same task with a slight variation. We are interested in modelling the paired difference between them as well as estimate subject effects. I don’t think using a paired t-test as in a typical within-subject design is appropriate. What will be the best way to model this?
Each subject from one group is paired with a subject from another group in the sense that they performed the same task
with a slight variation. We are interested in modelling the paired difference between them as well as estimate subject effects.
How exactly is the pairing? There are equal number of subjects between the two groups: subject A1 is paired with B1, A2 is paired with B2, so on and so forth? If it’s something different, could you elaborate a little bit more?
Thank you, Gang.
Yes, there are exactly same number of subjects in each group like you understood; A1 paired with B1, A2 with B2 and so on…
A paired t-test is the only way I can think of in this case.
Clarification: the two groups have different sets of subjects, i.e, “paired” subjects A1 and B1 are not the same subject. You still think paired t-test is the only way?
I don’t know the nature of the pairing. Ideally such a paired difference would be better designed as a within-subject effect. However, with the data you currently have, paired t-test is the way to go and I can’t think of a better alternative.
We are using 3dMEMA to run a paired t-test. There is not much documentation on how to do a paired t-test in 3dMEMA, but I looked at the documentation of 3dttest++. It mentioned two ways for doing a paired t-test: I can either provide the paired subjects as separate sets, or I can provide a single set in which each input is the beta difference of the pair.
I think I have to use the second way in the 3dMEMA. I can get the beta difference within each pair by using 3dcalc (as suggested in 3dttest++ documentation). Should I do the same thing for corresponding tstats? This doesn’t seem right. Or do you recommend running fixed effect analysis on each pair and use their outputs in the 3dMEMA?
To use 3dMEMA for paired testing, you need to provide the contrast estimate as well as the t-stat for the contrast as input. In other words, you have to perform the contrast testing at the individual level through, for example, 3dREMLfit.
I am not sure if I can do that, as my pairing is not between different conditions but it’s between different groups. For example, pairing between pre and post treatment groups.
Chirag, in that case just use 3dttest++.