Purpose: To demonstrate the use of a bivariate random effects meta-analysis (BRMA) to synthesise two outcomes from a systematic review of randomised controlled trials (RCTs) investigating the use of continuous positive airways pressure (CPAP) for the treatment of obstructive sleep apnoea syndrome (OSAS).

Methods: The available RCT data assessing the clinical effectiveness of CPAP for OSAS described a range of relevant outcome data. Two outcomes were determined to be of particular interest for the cost-effectiveness analysis: mean difference in Epworth Sleepiness Scale (ESS) score and mean difference in blood pressure (BP). A BRMA was undertaken to jointly synthesise the available evidence on each outcome and to incorporate information on the within-study correlation and the observed between-study correlation. This paper describes the specification and results of the BRMA in comparison to separate univariate random effects meta-analyses (URMAs) for each outcome.

Results: ESS was reported in the majority of studies and the pooled estimate differed little between analyses. BP was infrequently reported and the pooled estimate varied between the URMA and BRMA according to the ability of the between-study correlation to inform the missing data where trials reported only ESS. When the synthesis was limited to trials that used ambulatory blood pressure measurement (ABPM) only four studies provided information on both outcomes. When office measurements were included seven studies reported both outcomes and the estimate of between-study correlation was more precise. An informative prior for the within-study correlation had little impact on the pooled estimates.

Conclusions: In general systematic reviews may focus on more than one outcome of interest. The model described here is generalisable to any number of outcome measures and takes advantage of information provided by the between-study correlation. Where trials report different sub-sets of outcome measures a multivariate meta-analysis can be used to impute missing outcomes on the basis of observed characteristics. In contrast conducting separate URMAs for each outcome assumes that any missing outcomes are missing completely at random and that there is no correlation between alternative outcome measures. The difference between the two approaches may have important consequences when the results are used to inform a cost-effectiveness analysis, particularly in terms of characterising uncertainty and estimating the value of further research.