Mix and match. A simulation study on the impact of mixed-treatment comparison methods on health-economic outcomes.

Abstract

BACKGROUND:Decision-analytic cost-effectiveness (CE) models combine many parameters, often obtained after meta-analysis. AIM:We compared different methods of mixed-treatment comparison (MTC) to combine transition and event probabilities derived from several trials, especially with respect to health-economic (HE) outcomes like (quality adjusted) life years and costs. METHODS:Trials were drawn from a simulated reference population, comparing two of four fictitious interventions. The goal was to estimate the CE between two of these. The amount of heterogeneity between trials was varied in scenarios. Parameter estimates were combined using direct comparison, MTC methods proposed by Song and Puhan, and Bayesian generalized linear fixed effects (GLMFE) and random effects models (GLMRE). Parameters were entered into a Markov model. Parameters and HE outcomes were compared with the reference population using coverage, statistical power, bias and mean absolute deviation (MAD) as performance indicators. Each analytical step was repeated 1,000 times. RESULTS:The direct comparison was outperformed by the MTC methods on all indicators, Song's method yielded low bias and MAD, but uncertainty was overestimated. Puhan's method had low bias and MAD and did not overestimate uncertainty. GLMFE generally had the lowest bias and MAD, regardless of the amount of heterogeneity, but uncertainty was overestimated. GLMRE showed large bias and MAD and overestimated uncertainty. Song's and Puhan's methods lead to the least amount of uncertainty, reflected in the shape of the CE acceptability curve. GLMFE showed slightly more uncertainty. CONCLUSIONS:Combining direct and indirect evidence is superior to using only direct evidence. Puhan's method and GLMFE are preferred.