matchedpairs

Matched pairs, or paired design, is a statistical approach where subjects are grouped into pairs based on similarity on observed characteristics and then assigned to different treatments within each pair. This structure creates within-pair comparisons that help control for confounding factors that are similar within the pair, increasing the efficiency of the analysis.

Pairs can be formed through exact matching on key covariates, propensity score matching, or other matching algorithms. In some contexts, naturally occurring pairings exist, such as twins or before-and-after measurements on the same subject, which naturally yield paired data.

The primary analysis uses the pairwise differences. For each pair i, compute di = Yi1 − Yi0, where Yi1 and Yi0 are outcomes under the two conditions. A paired t-test tests whether the mean difference is zero, using t = d̄ / (s_d / √n), where d̄ is the average difference and s_d is the standard deviation of differences. If the differences are not approximately normal, a nonparametric alternative, the Wilcoxon signed-rank test, can be used. Effect size can be reported as the mean difference or as a standardized measure such as the mean difference divided by the standard deviation of differences.

Matched-pairs designs are common in clinical and observational studies, including before-and-after experiments and twin studies, because they reduce variability due to between-subject differences. Limitations include reliance on the quality of matching, potential residual confounding, and inability to assess effects of variables used for matching without additional methods.