I have a question about trying to achieve balance on variables and their
interactions as we did on problem set 7, 1d - It seems intuitive to me
that if you have only a few covariates and can achieve balance on all of
the interactions between those you might well create two subsamples via
matching that are more similar to one another than balancing only on the
individual covariates without their interactions. However, if you put in
all the interactions (treat~x1*x2*x3*x4*x5), then calculating standardized
biases and this kind of thing becomes ridiculous - so what is recommended?
Do we put in less, so that the balance can be more easily assessed, or do
we put in more but still only assess the balance on the covariates and
interesting interactions?