Primary Submission Category: Matching, Weighting
Balancing Act: Comparing Coarsened Exact Matching and Entropy Balancing in Cigna’s 2025 Value of Integration Study
Authors: Aran Canes, Kamala Swayampakala, Lukas Halim, Robert Wojewoda,
Presenting Author: Aran Canes*
Background: This study represents the 2025 edition of Cigna’s Value of Integration (VOI) study, which evaluates the impact of integrated benefits on healthcare costs.
Methods: Using a large observational dataset with approximately 1.8 million treated customers and 95,642 controls, we compared Propensity Score (PS) stratification, Coarsened Exact Matching (CEM) and Entropy Balancing (EB) for ATT estimation. PS stratification was explored using 5, 10 and 20 strata but failed to achieve balance on key confounders at any stratification level.
Results: Pre-adjustment estimates indicated substantial confounding. CEM retained 83% of treated customers but excluded a subset with markedly higher outcomes leading to an ATT of $241. EB retained the full treated population at the cost of reduced ESS (83%) for the controls and produced an ATT of $501, closer to the naive difference. No evidence of dominance by individual controls was observed.
Conclusions: Divergent ATT estimates from CEM and EB reflected differences in overlap handling and estimand definition rather than estimator instability. EB preserved the full treated distribution and efficiently absorbed overlap constraints, while CEM produced a more conservative estimate applicable to an overlap-restricted treated subpopulation. Cigna elected to publish the CEM estimate since it reflects outcomes with strong covariate overlap. These results underscore the importance of aligning estimator choice with the target estimand.