OBJECTIVES: Estimates of real-world cost-effectiveness are mostly based on observational data with non-random treatment assignments. Several methods exist to address the resulting confounding-by-indication, including regression and methods based on propensity scores (PS). This study examined the performance of these methods in the context of cost-effectiveness analysis. The PS-methods were: PS matching (kernel and one-to-one), covariate adjustment using PS, inverse probability-of-treatment weighting (IPTW) and double robustness, each with several specifications. METHODS: Thirty-eight adjustment approaches were compared using Monte Carlo simulations. In each simulation, four differently confounded samples (n=2000) were drawn from a synthesized population. Incremental survival time and costs were calculated using the results of Weibull survival and generalized linear regression. These regressions – with treatment as sole covariate or fully specified with all confounders - were performed directly or after applying a PS-method. Each approach was assessed on bias (systematic deviation from the true effect), accuracy (proportion of simulated results within acceptable distance from the true values) and reliability (width of bootstrapped confidence intervals) of estimates of incremental effects, costs and cost-effectiveness ratios. RESULTS: In estimates of the average treatment effect in the treated (ATT), kernel and 1-to-1 PS matching had similar bias and accuracy results, but the reliability of kernel was better. Regarding average treatment effects for the sample as a whole (ATE), double robustness and IPTW had the least bias and the best accuracy and reliability. Combining PS methods with fully specified regression models was most likely to lead to good results. PS covariate adjustment and regression without a PS method scored worst. CONCLUSIONS: PS methods are preferable to conventional regression for use in observational cost-effectiveness studies. Combining a PS method with fully specified regression should be considered for the analysis. Since no method is always superior, it is advised that sensitivity analyses with different techniques be performed.