BACKGROUND: Decision-analytic cost-effectiveness (CE) models combine many different parameters like transition probabilities, event probabilities, utilities and costs, which are often obtained after meta-analysis. The method of meta-analysis may affect the CE estimate. AIM: Our aim was to perform a simulation study that compares the performance of different methods of meta-analysis, especially with respect to model-based health economic (HE) outcomes. METHODS: A reference patient population of 50,000 was simulated from which sets of samples were drawn. Each sample drawn represented a clinical trial comparing two fictitious interventions. In several scenarios, the heterogeneity between these trials was varied, by drawing one or more of the trials from predefined subpopulations. Parameter estimates from these trials were combined using frequentist fixed (FFE) and random effects (FRE), and Bayesian fixed (BFE) and random effects (BRE) meta-analysis. The pooled parameter estimates were entered into a probabilistic cost-effectiveness Markov model. The four methods of meta-analysis resulted in different parameter estimates and HE outcomes, which were compared with the true values in the reference population. Performance statistics were: (1) the percentage of repetitions that the confidence interval of the probabilistic sensitivity analysis covers the true value (coverage), (2) the difference between the estimated and true value (bias), (3) the mean absolute value of the bias (MAD) and (4) the percentage of repetitions that result in a statistically significant difference between the two interventions (statistical power). As the differences between methods could be due to chance, we repeated every step of the analysis 1,000 times to study whether differences were systematic. RESULTS: FFE, FRE and BFE lead to different parameter estimates, but, when entered into the model, they do not lead to large differences in the point estimates of the HE outcomes, even in scenarios where we built in heterogeneity. Random effects methods do not necessarily reduce bias when heterogeneity is added to the trials, and may even increase bias in certain situations. BRE tends to overestimate uncertainty reflected in the CE acceptability curve. CONCLUSION: FFE, FRE and BFE lead to comparable HE outcomes. BRE tends to overestimate uncertainty. Based on this study, we recommend FRE as the preferred method of meta-analysis.
BACKGROUND: Decision-analytic cost-effectiveness (CE) models combine many different parameters like transition probabilities, event probabilities, utilities and costs, which are often obtained after meta-analysis. The method of meta-analysis may affect the CE estimate. AIM: Our aim was to perform a simulation study that compares the performance of different methods of meta-analysis, especially with respect to model-based health economic (HE) outcomes. METHODS: A reference patient population of 50,000 was simulated from which sets of samples were drawn. Each sample drawn represented a clinical trial comparing two fictitious interventions. In several scenarios, the heterogeneity between these trials was varied, by drawing one or more of the trials from predefined subpopulations. Parameter estimates from these trials were combined using frequentist fixed (FFE) and random effects (FRE), and Bayesian fixed (BFE) and random effects (BRE) meta-analysis. The pooled parameter estimates were entered into a probabilistic cost-effectiveness Markov model. The four methods of meta-analysis resulted in different parameter estimates and HE outcomes, which were compared with the true values in the reference population. Performance statistics were: (1) the percentage of repetitions that the confidence interval of the probabilistic sensitivity analysis covers the true value (coverage), (2) the difference between the estimated and true value (bias), (3) the mean absolute value of the bias (MAD) and (4) the percentage of repetitions that result in a statistically significant difference between the two interventions (statistical power). As the differences between methods could be due to chance, we repeated every step of the analysis 1,000 times to study whether differences were systematic. RESULTS: FFE, FRE and BFE lead to different parameter estimates, but, when entered into the model, they do not lead to large differences in the point estimates of the HE outcomes, even in scenarios where we built in heterogeneity. Random effects methods do not necessarily reduce bias when heterogeneity is added to the trials, and may even increase bias in certain situations. BRE tends to overestimate uncertainty reflected in the CE acceptability curve. CONCLUSION: FFE, FRE and BFE lead to comparable HE outcomes. BRE tends to overestimate uncertainty. Based on this study, we recommend FRE as the preferred method of meta-analysis.
Authors: A E Ades; Mark Sculpher; Alex Sutton; Keith Abrams; Nicola Cooper; Nicky Welton; Guobing Lu Journal: Pharmacoeconomics Date: 2006 Impact factor: 4.981