OBJECTIVES: Uncertainty in the estimation of the incremental cost-effectiveness ratio (ICER) is usually described by bootstrap confidence intervals. When bootstrap replicates cover more than one quadrant of the cost-effectiveness plane angular transformation of the ICER is appropriate. In a simulation study the performance of the most common bootstrap methods (normal, percentile, bias corrected with and without acceleration correction) is investigated. METHODS: Cost data were generated according to a lognormal distribution, effect data were drawn from normal distributions. The simulation frame work covers 180 settings differing in correlation between costs and benefits ( 0.9, 0.45, 0, 0.45, 0.9), in standardized mean difference of benefit and cost (each 0.2, 1, 5) and in sample size (50, 100, 200, 400). For each setting 10,000 samples were generated. The bootstrap procedures were based on 1000 replications. The performance of each procedure was described by the observed coverage probability of the 95%-confidence intervals. RESULTS: Overall, the percentile method demonstrates the highest robustness with miscoverage frequencies between 4% and 7%. The normal approach shows comparable results except in case of small effects and costs. While for negatively correlated data up to 15% miscoverage is observed , a strong conservative behaviour is found in case of positive correlation. With increasing differences of cost and/or effect both bias corrected methods show coverage frequencies of less than 50% falling below 10% for larger sample sizes. CONCLUSIONS: The bias corrected methods suffer on violated parametric assumptions of the bootstrap distribution and result in seemingly short confidence intervals. Recently developed methods as bootstrapping the bootstrap or (weighted) jackknife after bootstrap, which do not depend on these assumptions, may provide a promising alternative. However, in order to avoid computationally intensive procedures the bootstrap percentile method is generally recommended.