Bayesian Meta-Analyses for Comparative Effectiveness and Informing Coverage Decisions

Berry, Scott M. PhD*; Ishak, K. Jack PhD†; Luce, Bryan R. PhD‡; Berry, Donald A. PhD*§

doi: 10.1097/MLR.0b013e3181e24563
Comparative Effectiveness

Background: Evidence-based medicine is increasingly expected in health care decision-making. The Centers for Medicare and Medicaid have initiated efforts to understand the applicability of Bayesian techniques for synthesizing evidence. As a case study, a Bayesian analysis of clinical trials of implantable cardioverter defibrillators was undertaken using patient-level data not typically available for analysis.

Purpose: Conduct Bayesian meta-analyses of the defibrillator trials using published results to demonstrate a Bayesian approach useful to policy makers.

Data Sources, Study Selection, Data Extraction: We reconsidered trials in a 2007 systematic review by Ezekowitz et al (Ann Intern Med. 2007;147:251–262) and extracted information from the original published articles. Employing a Bayesian hierarchical approach, we developed a base model and 2 variants, and modeled hazard ratios separately within each year of follow-up. We considered sequential meta-analyses over time and found the predictive distribution of the results of the next trial, given its sample size.

Data Synthesis: For the most robust of 3 models, the probability that the mean defibrillator effect (in the population of trials) is beneficial is greater than 0.999. In that model, about 5% of trials in the population of trials would have a detrimental effect. Despite the moderate amount of heterogeneity across the trials, there was stability of conclusions after the first 3 of the 12 total trials had been conducted. This stability enabled reasonable predictions for the results of future trials.

Limitations: Inability to assess treatment effects within subsets of patients.

Conclusions: Bayesian meta-analyses based on literature surveys can effectively inform coverage decisions. Bayesian modeling for endpoints such as mortality can elucidate treatment effects over time. The Bayesian approach used in a sequential manner over time can predict results and help assess the utility of future clinical trials.

From the *Berry Consultants, College Station, TX; †United BioSource Corporation, Montreal, Québec, Canada; ‡United BioSource Corporation, Bethesda, MD; and §Division of Quantitative Sciences, MD Anderson Cancer Center, The University of Texas, Houston, TX.

Supported by Boston Scientific Corp. Boston Scientific Corp reviewed a draft of this manuscript, but made no substantial comments and had no material input into the manuscript. KJI and BRL received salary from United BioSource Corp that had a contract with Boston Scientific Corporation.

S.M.B. and D.A.B. are co-owners of Berry Consultants, LLC, which had a subcontract with United BioSource Corp. Berry Consultants also has contracts with Medtronic Corp.

Reprints: Scott M. Berry, PhD, Berry Consultants, 3145 Chaco Canyon Drive, College Station, TX 77845. E-mail: scott@berryconsultants.com.

© 2010 Lippincott Williams & Wilkins, Inc.