Diagnostic test meta-analysis by empirical likelihood under a Copas-like selection model

The validation of diagnostic test meta-analysis is often threatened by publication bias, which can be commonly characterized by the Copas selection model. Under this model, conventional approaches to diagnostic meta-analysis are based on conditional likelihood. Since they may have efficiency loss, we propose a full likelihood diagnostic meta-analysis method by integrating the usual conditional likelihood and a marginal semi-parametric empirical likelihood. We show that the resulting maximum likelihood estimators (MLEs) have a jointly normal limiting distribution, and the resulting likelihood ratio follows a central chisquare limiting distribution. Our numerical studies indicate that the proposed MLEs often have smaller mean square errors than the conditional likelihood MLEs. The full likelihood ratio interval estimators generally have more accurate coverage probabilities than the conditional-likelihood-based Wald intervals. We re-study two real meta analyses on influenza and mental health respectively for illustration.