When Does It Pay to Break the Matches for Analysis of a Matched-Pairs Design?

Abstract

SUMMARYTwo methods of analyses are compared to estimate the treatment effect of a comparative study where each treated individual is matched with a single control at the design stage. The usual matched pairs analysis accounts for the pairing directly in its model, whereas regression adjustment ignores the matching but instead models the pairing using a set of covariates. For a normal linear model, the estimated treatment effect from the matched pairs analysis (paired t-test) is more efficient. For a Bernoulli logistic model, matched pairs analysis performs better when the sample size was small, but is inferior to logistic regression for large sample sizes.

A Multicomponent Intervention to Prevent Delirium in Hospitalized Older Patients

“…40,41 Subsequently, unmatched analyses by means of traditional logistic regression for new cases of delirium during the hospital stay and Cox proportional-hazards analysis for the risk of delirium per hospital day, with adjustment for the matching factors, were carried out to provide comparisons and alternatives to the matched analyses, as advocated by previous investigators. 42 Kaplan-Meier analysis and the log-rank test were used to compare the cumulative incidence of delirium, defined as the proba-*Plus-minus values are means ±SD. There were no significant differences in any of these characteristics between the intervention and control groups in matched or unmatched analyses.…”

Impact of childhood cancer on the mental health of parents

“…It could be argued that matched analyses might be preferable because each parent in a matched pair has a child of the same age and sex. If a normal linear model is used, the estimators of the childhood cancer effect should be unbiased for both matched and unmatched analyses, but the variances would differ [26]. In our particular situation, the correlations between the two groups (matched pairs) for the GHQ-12 score and the ratings for health, social life, mood, and coping ability were all extremely low.…”

The Econometrics of Randomized Experiments

“…See Lynn and McCulloch (1992) for a similar argument in the context of paired randomized experiments.…”