Standards and analytic techniques used in evaluation of diagnostic tests

In the evaluation of accuracy for a diagnostic test, unbiased estimators for accuracy measures are preferred. However, some bias in measuring diagnostic accuracy of the test may occur. The most common biases among encountered are verification bias and imperfect gold standard bias. In this thesis, proposed approaches for correcting these biases were examined and alternative approaches were proposed for the situations involving problems for the current approaches.Verification bias arises when not all patients in the study are taken the gold standard test. When the diagnostic test is binary, depending on the missing mechanism, Begg&Greenes, Upper and Lower Bound by Zhou, Logistic Regression or Multiple Imputation approaches are used for correcting this bias. In this thesis, a new approach, called Artificial Neural Networks, was alternatively proposed and compared with other approaches in different comparison schemes using simulated and real data sets.Imperfect gold standard bias arises when using an imperfect gold standard test instead of gold standard test. Depending on the measurement scale of new diagnostic test, Discrepant Resolution, Composite Reference Standard, Non Parametric AUROC Estimation by Zhou and Bayesian Approaches are used for correcting this bias. In this thesis, a new approach, called Latent Class Analysis, was alternatively proposed and compared with other approaches in different comparison schemes using simulated and real data sets.In the simulations, in order to measure the correct performance of the approaches and to determine deficiencies involving, data were generated with a variety of experimental conditions; changing sample size, disease prevalence, and diagnostic accuracy of new test.When the missing mechanism is MAR, Begg&Greenes approach and when the missing mechanism is NMAR, Logistic Regression or Artificial Neural Networks approaches are preferable for correcting verification bias.When the measurement scale of the new test is binary and disease prevalence and diagnostic accuracy values are well-defined, Latent Class Analysis approach; when the measurement scale is ordinal and disease prevalence and the number of test and test categories are appropriate, Latent Class Analysis approach; and finally when the measurement scale is interval or ratio and with large sample, high AUROC and without confounding variable, Bayesian Approach and for other situations Latent Class Analysis approach are the best approaches for correcting imperfect gold standard bias.