Choosing cutoff values for correlated continuous diagnostic data to estimate sensitivity and specificity
This dissertation includes two projects and an R package for the 1st project, focusing on choosing cutoff values to estimate the sensitivity and specificity for continuous diagnostic tests. A continuous diagnostic test needs to be dichotomized to generate positive and negative test outcomes by choosing a cutoff value. The choice of the cutoff value depends on the estimates of the sensitivity and specificity of the dichotomized test with this cutoff. There are two challenges during this process: 1) a typical experiment to validate a new diagnostic test usually involves multiple observations from the same subjects, resulting in correlated data values. This correlation increases the complexity of calculating the confidence intervals of the sensitivity and specificity given the true statuses. 2) In many diagnostic trials, the true statuses are unknown, which make it difficult even to calculate the point estimates of the sensitivity and specificity. In the 1st project (Chapter 2), we propose a method to calculate the confidence intervals of sensitivity and specificity with the true statuses of subjects given based on a parametric model for the correlated continuous diagnostic test data. In the second project (Chapter 3), we focus on the challenge of unknown statuses. Due to the lack of the statuses, only the model-based method can be used to estimate the sensitivity and specificity. However, the estimations of model parameters are difficult because of the unknown statuses. We propose a Bayesian model with latent variables to model the unknown statuses. In Chapter 4, I create an R package for the 1st project for future uses.