Doing so you will get samples of these two variables following the prior distribution. Since they have nothing to do with the model logp (besides adding some additional constant), they have no effect to NUTS or MH sampler.
You can also following the for loop like you did, but create a deterministic using tt.stack. I guess it depends on how many skips in your real case. If there are lots of subject with only one measurements (ie, you are skipping a lot), using a for loop then tt.stack might be more efficient.