Pair matching in other contexts

I want to be clear that my dissing of pair matching yesterday was in the context of non-experimental evaluations relying on unconfoundedness assumptions. In that context, the alternatives are other procedures for conditioning on observed covariates, such as inverse propensity weighting or parametric linear regression.

In contrast, pair matching, treated as a special binary case of blocking prior to random assignment, is a very useful idea indeed, as described in a post from a blog at the Wild Bank. In the experimental context, pair matching can improve power relative to simple random assignment, especially in cases where the units being randomly assigned are relatively heterogeneous. The idea is to do random assignment within matched pairs (or, more generally, blocks of similar observations).

Hat tip: Jess Hoel