Incorporation of meta-analyses of diagnostic test accuracy studies into a clinical/economic decision analytic framework

An accurate diagnosis is a crucial part of an effective treatment. Diagnostic errors cause unwanted side effects for healthy individuals and witheld treatments for diseased patients. Meta-analysis techniques allow the accuracy of diagnostic tests to be estimated using all the available sources of evidence. The most common measures of diagnostic accuracy are sensitivity (true positive rate) and specificity (true negative rate). As part of this thesis, current methods developed for synthesising data from diagnostic test studies are reviewed and critiqued, and then applied to estimate the accuracy of the Ddimer test for diagnosing Deep Vein Thrombosis (DVT). The fit of the different models is assessed via the Deviance Information Criterion and the Residual Deviance and the most complex synthesis models are found to provide the best fit to the data. When covariates are added to these models, only the incorporation of study setting sensitivity is found to improve the fit of the model. Diagnostic tests are rarely used in isolation and consideration of multiple tests in combination may also require evaluation. In this thesis, a multiple equations with shared parameters approach is proposed which estimated the accuracy of a combination of tests in two stages: i) estimate the conditional accuracy of the tests; and ii) estimate the accuracy of possible combinations of tests as functions of the conditional accuracies. Such a modeling approach allows the inclusion of different sources of evidence to be used simultaneously. The final part of the thesis evaluated the cost-effectiveness of different strategies for diagnosing DVT by incorporating the results from the aforementioned evidence synthesis models into an economic decision analytic model. In conclusion, the assumption of conditional independence can affect the analyses of the effectiveness and the cost-effectiveness of combinations of diagnostic tests, thus leading to potentially wrong decisions if the dependence is not explicitly modelled.