> Do you imply that I should use 3dcalc or something like that to compute the average coefficient map
> for each twin pair, for each condition; and then feed these average coefficient maps into 3dMVM?

No, just directly provide input files for each pair to 3dMVM: two rows per condition.

> for each pair, there would be 3 rows in the data table (one for each condition)?

6 rows per pair.

> By averaging the twins, am I not losing degrees of freedom / power?

Not really. You have enough number of twins, so I don't think that DF is an issue.

> That is, this approach would allow me the same degrees of freedom as splitting the sample in half
> (such that no twin pair is in the same group), and then running the same analyses on each cohort, correct?

Right.

> my incentive with looking into a LME approach here was to attempt to maximize power while
> accounting for non-independence in the data.

I'd just go with 3dMVM, which is more straightforward.

Gang