Skip to main content
Effective Health Care Program

Findings of Bayesian Mixed Treatment Comparison Meta-Analyses: Comparison and Exploration Using Real-World Trial Data and Simulation

Research Report

People using assistive technology may not be able to fully access information in these files. For additional assistance, please contact us.

Structured Abstract

Background

Specific objectives were to examine the following: (1a) how results of Bayesian mixed treatment comparison (MTC) methods compare with several commonly considered frequentist indirect methods; (1b) how Bayesian MTC methods perform for different evidence network patterns; (2) how meta-regression can be used with Bayesian MTC meta-analysis to explore heterogeneity; and (3) how findings of Bayesian MTC meta-analyses compare for different numbers of studies and different network pattern assumptions. For objectives 1 and 2, we aimed to conduct case studies using data from two recent comparative effectiveness reviews (CERs). For objective 3, we aimed to use simulated data.

Methods

For objectives 1 and 2, we used data from CERs that examined second-generation antidepressants (SGAs) and biologic disease-modifying antirheumatic drugs (DMARDs) for rheumatoid arthritis (RA). For objective 1, we compared results of Bayesian MTC methods with those of three frequentist indirect methods: meta-regression, the Bucher method, and logistic regression for dichotomous and continuous outcomes. For objective 2, we conducted two types of meta-regression. One explored subgroup effects with a binary covariate to assess whether efficacy of SGAs differs between older adults (≥55 years) and adults of any age. The other explored a continuous covariate to assess whether treatment efficacy varies by disease duration of RA. For objective 3, we used simulated data to examine the Bayesian MTC method's ability to produce valid results for two data scenarios when varying numbers of studies were available for each comparison for various network patterns.

Results

Bayesian MTC methods permitted the calculation of results for more comparisons of interest than frequentist meta-regression or the Bucher method (when applied as they would typically be used). When comparisons were calculated, the findings generally agreed but differed for a small proportion (less than 10%) of comparisons. Regarding precision, logistic regression produced the most precise estimates, followed by the Bayesian MTC method.

Our meta-regressions found a trend toward lesser efficacy for SGAs in older adults and a trend toward greater efficacy of biologic DMARDs for those with greater mean disease duration.

Our simulations supported the validity of Bayesian MTC methods for star and ladder network patterns but raised some concerns about one closed loop (and possibly loop) network patterns. Simulations generally found similar probabilities for which drug was the best treatment for scenarios when only 1 study was available for each comparison and those when more studies (2, 3, 5, or 10) were available; precision increased as the number of available studies increased.

Conclusions

Bayesian MTC methods offer several advantages over frequentist indirect methods, including the ability to produce results for all comparisons of interest in a single analysis. Results of Bayesian MTC methods and those of frequentist indirect methods may differ for a small proportion of comparisons, which could lead to differences in conclusions when using different methods. Our findings raise some concerns about the validity of the results of Bayesian MTC methods for certain network patterns. Further research is needed to explore additional real-world datasets and simulated data to determine if our findings are reproducible or generalizable and to better understand the validity of Bayesian MTC methods for various scenarios.