General comparison of pairwise and larger-strata designs under attrition

Characterize the general trade-off between tighter matching in pairwise randomization and the greater retention of within-stratum variation in larger-strata designs under attrition, and determine when pairwise designs analyzed with optimally weighted randomization inference outperform larger-strata designs analyzed with standard methods.

Background

The paper establishes dominance of its optimally weighted randomization-inference procedure over two standard procedures within pairwise designs. It does not establish whether pairwise designs analyzed with that procedure outperform designs using larger strata, such as strata of four units, analyzed with conventional methods.

A simulation provides examples in which pairwise designs perform better, but the authors emphasize that the comparison depends on attrition, matching quality, and the ability of larger strata to retain within-stratum variation. A general rule or formal characterization of this design trade-off remains unresolved.