You would introduce a bias between subjects. I would not recommend it.
There seems to be some slight differences between the sources of the cortical and mixed models (see images attached).
The source space is different, the maps are therefore going to be different.
Quantifying the differences between different source spaces is complicated, saying which one is better is impossible with real recordings (you need to run simulations where you know exactly the truth).
We only have a small group of five subjects and this issue might not be anything significant, but we were wondering if this is something possible to consider and not too big a task?
We would like to use the constrained model as it would give suitable results in relation to literature.
It would be simpler to run your analysis both with a fully unconstrained model (volume source model using a group grid) and with a constrained model (location + orientation constraints).
Ideally, your observations with both approaches should converge.