Background: Prior research has shown large-scale cardinality matching (CM) to be associated with superior patient retention and comparable systematic bias as compared to large-scale propensity score matching (PSM). However, large-scale methods may not be applicable in all settings.
Objectives: The current study compared non-large-scale PSM and CM in terms of post-match patient retention, covariate balance and residual confounding at progressively smaller sample sizes.
Methods: We identified new users of β-blockers versus angiotensin-converting enzyme inhibitors (ACEI) monotherapy between 10-01-2014 and 01-01-2017 from the IBM® MarketScan® Commercial Claims and Encounters database (index=first drug exposure). PSM was performed through nearest-neighbor matching (1:1, caliper=0.15). CM was performed at four prespecified matching criterion (1:1; max SMD=0.00, 0.01, 0.05 and 0.10 of matching covariates between study groups). Matching covariates (n=37) included patient demographic and select clinical characteristics. Observed covariates included patient demographics, and all observed prior conditions, drug exposures, and procedures. Expected absolute systematic error (EASE), a summary measure of expected systematic error, was assessed using 105 negative control outcome experiments perceived as unassociated with the treatment. All analyses were conducted within a 10% and 0.25% sample group consisting of 5 and 200 subsample draws, respectively.
Results: A total of 186,233 (β-blocker: 56,871; ACEI: 129,362) patients meeting the study criteria were identified. CM achieved superior patient retention and matching covariate balance in both sample groups. After PSM, an average of 1.6% and 15.3% of matching covariates were imbalanced in the 10% and 0.25% sample groups, respectively. There was no evidence of a significant differences in observed covariate balance between PSM and CM; however, CM with more stringent balance criteria (e.g., SMD < 0.01) was associated with a significant reduction in EASE as compared to PSM.
Conclusions: We recommend CM with more stringent balance criteria be considered as an alternative to PSM when using a small set of matching covariates.
