Some further insight into these challenges may be gained by realizing that pair-matching is most effective when each of the matched pairs constructed correspond to distinct levels of baseline risk. Although several published trials have enrolled more than 50 matched pairs, the ability to actually construct such a large number of distinct matches is likely to be very challenging in practice. This is because there is often only limited knowledge available on the factors likely to affect outcome. However, even if such knowledge exists, it may not be possible to secure matches for all eligible clusters.
Thus, rather than attempting to construct, for example, 52 matched pairs, it may be more practical to adopt a stratified design with, say, 28 strata enrolling four clusters each. The resulting assignment of two clusters to each of the intervention and control groups within each stratum also has important analytic advantages. These accrue because the assignment of multiple clusters to each stratum allows the intracluster correlation coefficient to be directly computed using routine methods (e.g., Donner and Klar, 2000, Section 6.4). This is not possible in the matched-pair design since the lack of cluster-level replication implies that the natural variation between two matched clusters is totally confounded with the effect of intervention. Without a direct measure of such between-cluster variation, additional assumptions and a fairly large number of matched pairs are needed to estimate ρ (Klar and Donner, 1997).
Flash is not available on mobile devices. Please view the Flash Description.