Addressing between-study heterogeneity and inconsistency in mixed treatment comparisons: Application to stroke prevention treatments in individuals with non-rheumatic atrial fibrillation.

Authors: Cooper NJ (1) , Sutton AJ (1) , Morris D (2) , Ades AE (3) , Welton NJ (3)
(1) Department of Health Sciences, University of Leicester (2) Section of Epidemiology, Institute of Cancer Research, Sutton (3) Academic Unit of Primary Care, Department of Community Based Medicine, University of Bristol
Source: Stat Med. 2009 Jun 30;28(14):1861-81
DOI: 10.1002/sim.3594. Publication date: June 30, 2009 E-Publication date: April 27, 2009 Availability: abstract Copyright: © 2009 John Wiley & Sons, Ltd.
Language: English Countries: Not specified Location: Not specified Correspondence address: Nicola J. Cooper :
Centre for Biostatistics and Genetic Epidemiology, Department of Health Sciences, University of Leicester, Adrian Building, Leicester LE1 7RH, U.K.
Email :


Article abstract

Mixed treatment comparison models extend meta-analysis methods to enable comparisons to be made between all relevant comparators in the clinical area of interest. In such modelling it is imperative that potential sources of variability are explored to explain both heterogeneity (variation in treatment effects between trials within pairwise contrasts) and inconsistency (variation in treatment effects between pairwise contrasts) to ensure the validity of the analysis.The objective of this paper is to extend the mixed treatment comparison framework to allow for the incorporation of study-level covariates in an attempt to explain between-study heterogeneity and reduce inconsistency. Three possible model specifications assuming different assumptions are described and applied to a 17-treatment network for stroke prevention treatments in individuals with non-rheumatic atrial fibrillation.The paper demonstrates the feasibility of incorporating covariates within a mixed treatment comparison framework and using model fit statistics to choose between alternative model specifications. Although such an approach may adjust for inconsistencies in networks, as for standard meta-regression, the analysis will suffer from low power if the number of trials is small compared with the number of treatment comparators.

Find it online