OBJECTIVES: Cardinality matching (CM) uses advancements in optimization algorithms and may outperform propensity score matching (PSM) in post-match residual covariate balance, sample size retention and generalizability. Estimating the average treatment effect on the controls (ATC) may be necessary in studies to understand the impact of expanding medical treatments to previously unserved populations. CM and PSM may differ in the generalizability of the post-match sample to the targeted estimand. The present study compared the performance of PSM and CM when targeting the ATC estimand in a retrospective study of patients undergoing minimally-invasive surgery (MIS) vs. open surgery (OS) for thoracic segmentectomy.

METHODS: Patients aged ≥18 years undergoing thoracic segmentectomy between 1-1-2010 to 6-30-2017 were identified from the Premier Healthcare Database (index=first admission) and classified as undergoing OS or MIS. Targeting the ATC estimand, we conducted PSM through 1:1 nearest-neighbor matching (caliper=0.15) and CM through 1:1 matching permitting a maximum standardized mean difference (SMD) of 0.05 between comparison groups and the ATC. We evaluated PSM and CM on post-match sample size, covariate balance, and balance between groups and the ATC estimand. An SMD>0.10 was considered imbalanced.

RESULTS: We identified 5,552 patients (MIS: 3,524, OS: 2,028) meeting the study criteria. Matching was conducted across 37 covariates (74 unique levels). Total post-match sample sizes were similar for PSM=3,604 and CM=3,600. However, CM resulted in better balance between the MIS and OS groups (PSM: mean SMD=0.03, range=0.0—0.19, imbalanced covariates=5; CM: mean=0.03, range=0.0—0.08) and matched groups which were more similar to the ATC estimand (PSM OS: SMD mean=0.02, range=0.0—0.13, imbalanced covariates=1; PSM MIS: mean=0.04, range=0.0—0.18, imbalanced covariates=7; CM OS: mean=0.0, range=0.0—0.03; CM MIS: mean=0.03, range=0.0—0.05).

CONCLUSIONS: In this applied comparison, CM outperformed PSM on balance between comparison groups and generalizability in targeting the ATC estimand, while post-match sample sizes were similar between methods.