Premium
Bayesian model selection for meta‐analysis of diagnostic test accuracy data: Application to Ddimer for deep vein thrombosis
Author(s) -
Novielli Nicola,
Cooper Nicola J.,
Sutton Alexander J.,
Abrams Keith R.
Publication year - 2010
Publication title -
research synthesis methods
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 3.376
H-Index - 35
eISSN - 1759-2887
pISSN - 1759-2879
DOI - 10.1002/jrsm.15
Subject(s) - model selection , deviance information criterion , deviance (statistics) , covariate , computer science , bivariate analysis , bayesian probability , meta analysis , random effects model , context (archaeology) , missing data , statistical model , statistic , sensitivity (control systems) , statistics , selection (genetic algorithm) , econometrics , data mining , machine learning , artificial intelligence , bayesian inference , mathematics , medicine , paleontology , electronic engineering , engineering , biology
A number of statistical models have been developed for meta‐analysis (MA) of diagnostic test (DT) accuracy data. Here we consider these alternative MA models, explore the relationships between them, and consider the use of the deviance information criteria (DIC) to decide which is the most appropriate model for a given dataset. A Bayesian statistical approach is adopted throughout. The alternative MA models are applied to a dataset of 198 assays of Ddimer to diagnose deep vein thrombosis. In this example, based on the DIC, a bivariate random effects model for sensitivity and specificity fitted the data best. When considering the inclusion of study level covariates, allowing sensitivity to vary by study setting further improved the fit of the model. The model fitting approach is then repeated for a subset of the data, which highlights the less decisive results obtained when using the DIC with a more limited dataset. Formal approaches to model selection are often overlooked in an MA context; however, they offer sensible rationale to the analysis, particularly for complex models such as those proposed for DT accuracy. Specifically, the use of the DIC statistic appears to be well suited for deciding between potentially complex mixed‐effect MA models, possibly including covariates. Copyright © 2010 John Wiley & Sons, Ltd.