Statistics Colloquium: Anne Presanis
Systematic conflict diagnostics in evidence synthesis
Bayesian evidence synthesis, where multiple independent data sources contribute to the likelihood, is becoming increasingly employed in various fields, including medical decision-making and epidemiology. Evidence synthesis methods are most useful for estimating quantities which can't be directly observed, but for which indirect evidence is available: for example, a treatment comparison for two treatments which have not been directly compared in the same clinical trial; or prevalence of undiagnosed HIV infection. However, the use of multiple sources informing common parameters leads to potential for different datasets to provide conflicting or inconsistent inferences about the common parameters. Cross-validatory posterior predictive methods ("node-splitting") have previously been proposed to detect and measure such conflict, as a crucial step in the model criticism process. Such conflict assessment could be targeted to specific parts of a model, or could be systematic, testing for conflict throughout a model. In the latter case, the multiple testing problem arises. A framework for systematic conflict assessment that accounts for the multiple tests is presented and illustrated through a network meta-analysis and a synthesis to estimate HIV prevalence in Poland.
|+44 1524 593606|