Matching and other observational methods are commonly used in health economic cost-effectiveness studies. Many turn to observational data because of the rarity of long-term randomized trials that compare many treatment options and that collect cost data. Cost-effectiveness analysis is of growing interest to U.S. policy makers. It is a component of health reform in the U.S., and it is already used by policy makers elsewhere, such as the U.K.
However, the reliance on observational methods makes inferences unreliable. Aside from the issue of unobserved confounders, a key concern is how to adjust for imbalances in observed confounders due to non-random treatment assignment. Traditional methods of covariate adjustment such as regression depend on correct model specification. Alternatives such as propensity score matching depend on covariate balance being achieved. We demonstrate a non-parametric matching method, Genetic Matching, which uses a search algorithm to optimize covariate balance. Genetic Matching is a generalization of propensity score and Mahalanobis distance matching. We apply Genetic Matching to an economic evaluation of a clinical intervention, Pulmonary Artery Catheterization for which an experimental benchmark is available. Placebo tests show that the experiment provides a valid benchmark for the observational study. We find that Genetic Matching achieves better covariate balance than previous studies that used propensity score matching. Unlike previous studies, we recover the experimental benchmark for clinical outcomes, but like the extant literature, the cost estimates from the observational study differ sharply from the experiment. Statistical and policy implications are explored.
