gms | German Medical Science

Meta-analysis of diagnostic test accuracy studies is often done by jointly analyzing pairs of sensitivity and specificity in a generalized linear mixed model. Bivariate random effects account for correlation between these two measures. Mainly likelihood approaches are used for model fitting. However, the estimation approach can be very unstable or produce poor estimates of the between-study correlation, especially for a small number of studies. Bayesian versions avoid such problems. Up to now, Bayesian hierarchical models generally implied the use of Markov Chain Monte Carlo methods, which can be very time-consuming and require diagnostic checks to ensure convergence of the simulated samples. Recently, Rue et al. (2009) proposed a new Bayesian deterministic inference approach for latent Gaussian models using integrated nested Laplace approximations (INLA). With this approach MCMC sampling becomes redundant as the posterior marginal distributions are directly and accurately approximated. In this talk I will discuss the use of INLA for bivariate meta-analysis. In a case study, the telomerase meta-analysis for the diagnosis of bladder cancer provided by Glas et al. (2003) are re-analyzed. In particular, the influence of different prior distributions on the results is investigated. A comparison with the likelihood procedure SAS PROC NLMIXED in a simulation study indicates that INLA is more stable and gives generally better coverage probabilities for the pooled estimates and less biased estimates of variance parameters.