Quantifying how diagnostic test accuracy depends on threshold in a meta-analysis




Short oral session 2: Considerations for meta-analyses


Wednesday 13 September 2017 - 11:00 to 12:30


All authors in correct order:

Jones HE1, Gatsonis CA2, Trikalinos TA3, Ades AE1
1 School of Social and Community Medicine, University of Bristol, United Kingdom
2 Center for Statistical Sciences, School of Public Health, Brown University, USA
3 Center for Evidence Synthesis in Health, School of Public Health, Brown University, USA
Presenting author and contact person

Presenting author:

Hayley Jones

Contact person:

Abstract text
Background: Many tests for disease produce an explicit continuous measure, e.g. the concentration of a biomarker in a blood sample. This is dichotomised at some threshold to call the result positive or negative. In a meta-analysis of diagnostic test accuracy, the threshold used often varies across studies. To explain some of the heterogeneity in the meta-analysis and – more importantly – to identify the optimum threshold for clinical practice, it is intuitive to include reported threshold values as a covariate. However, guidance warns that this involves implicit assumptions that might not hold in practice.

Objectives: 1) To examine the assumptions involved in including threshold as a covariate in a meta-analysis of sensitivity and specificity; and, 2) to develop a more flexible model, requiring fewer assumptions.

Methods and Results: The implicit assumptions when including threshold directly as a covariate in the meta-analysis model are strong and not widely plausible. However, these can be relaxed by using additional data that are often available. In particular, it is common for some studies to report sensitivity and specificity estimates at multiple thresholds. Although this is widely regarded as problematic (due to the additional complexities involved in data synthesis), these extra data allow much greater flexibility in modelling. We describe a new model for the effect of threshold on sensitivity and specificity, which makes use of these additional data and can be considered a generalised version of that recently described by Steinhauser et al. We fit the model using the WinBUGS software and demonstrate its utility with 2 case studies.

Conclusions: Using more data, where available, allows the effect of threshold on sensitivity and specificity to be modelled flexibly, requiring minimal assumptions. This increases the potential clinical utility of the meta-analysis results.

Steinhauser S, Schumacher M. Rücker G, 2016. Modelling multiple thresholds in meta-analysis of diagnostic test accuracy studies. BMC Medical Research Methodology, 16: 97.