A NEW NON-PARAMETRIC MATCHING METHOD (GENETIC MATCHING) FOR BIAS ADJUSTMENT

   Purpose: In cost-effectiveness analyses (CEA) that use observational data, the key concern is how to adjust for imbalances in baseline covariates. Current analytical methods for dealing with baseline imbalance are model dependent and inadequate. This paper demonstrates that a new non-parametric matching method, Genetic Matching, reduces bias in CEA compared to standard methods.

   Methods: Genetic Matching is a generalization of propensity score and Mahalanobis distance matching that automates the search for covariate balance (Sekhon forthcoming). We demonstrate Genetic Matching in a CEA of a clinical intervention, Pulmonary Artery Catheterization [PAC]. The method uses a genetic algorithm to maximize the balance of observed potential confounding factors across treatment and control groups. Genetic Matching uses non-parametric measures of balance that examine the entire QQ plot, unlike the usual practice of only examining mean differences.

   We compare the estimates of incremental effectiveness and cost-effectiveness after applying propensity score and Genetic Matching to the observational data, with estimates from the corresponding RCT.

   Results: Propensity score matching resulted in poor balance in important covariates such as the baseline probability of death (p=0.03). Genetic Matching achieved excellent balance (e.g., baseline probability of death, p=0.87). Following propensity score matching, PAC was associated with increased mortality (Odds Ratio=1.22, p=0.03) and a negative incremental net monetary benefit (INB) (-£27,000 95% CI from -£40,000 to -£14,000). After Genetic Matching, PAC was not associated with increased mortality (OR=1.09, p=0.35) and the odds ratio was similar to that observed in the RCT (1.13, p=0.36). The INB associated with PAC was negative, but not statistically significant following both Genetic Matching (-£12,000, 95% CI from -£25,000 to £834) and the RCT (-£3,000, 95% CI from -£22,000 to 12,000). Monte Carlo simulations used data from the RCT to examine the properties of Genetic Matching compared to propensity score matching. The propensity score is almost always unknown (as in the PAC study), and the simulations demonstrate that when the propensity score is misspecified, propensity score matching may actually increase bias. By contrast Genetic Matching gave estimates that were close to the true estimate and substantially reduced bias and RMSE compared to propensity score matching.

   Conclusions: Genetic Matching achieves excellent covariate balance in a CEA based on observational data, and unlike the extant literature replicates the results from the corresponding RCT.