Interval estimation for TPRs and FPRs of two diagnostic tests with unverified negatives.

Date

2011-05

Authors

Stock, Eileen Marie.

Access rights

Worldwide access.
Access changed 6/26/13.

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

In the clinical setting, the performance of a diagnostic or screening test is often summarized using the test's true positive rate (TPR) and false positive rate (FPR). However, estimation of the TPR and FPR for a diagnostic or screening test requires individuals to be identified as diseased or non-diseased using a gold standard, which is seldomly available. In cases for when the gold standard test is available, it may be too costly or time-consuming to implement the gold standard procedure. Furthermore, this verification procedure is frequently limited to individuals with one or more positive test results, thus yielding unverified negatives and possibly verification bias. We present two interval estimators for the estimation of the differences in the TPRs and FPRs of two dichotomous diagnostic tests applied to members of a population stratified into two distinct groups. The two groups have varying disease prevalences with unverified negatives. We obtain maximum likelihood estimates using the EM algorithm to devise a Wald interval for the differences in the TPRs and FPRs and compare its performance to a Bayesian credible interval for a spectrum of different TPRs, FPRs, and sample sizes. We further present a hierarchical Bayesian logit model when incorporating different locations into the model. In particular, we compare the accuracy of two low-cost procedures for screening of high-grade cervical intra-epithelial neoplasia (CIN2+) using a cross-sectional multi-center study. We model the prevalence of CIN2+ for two different age groups assumed to have varying disease prevalences with unverified negatives. Furthermore, we include random effects to account for heterogeneity among the locations. In the final chapter, we perform a power analysis for multivariate normality in the presence of a monotone missing data pattern. We assess the efficiency of several imputation procedures and multivariate normality tests through a simulation study where we compare the resulting powers using an alternative distribution. In particular, we consider different sample sizes and proportions of missingness for various parameter values of a multivariate skewed t-distribution. We calculate power using both the median and mean test statistics for each imputation-test combination.

Description

Keywords

Sensitivity and specificity., Hierarchical Bayesian model., Diagnostic or screening tests., Unverified negatives., Monotone missing data., Multivariate normality., Multiple imputation., Wald confidence interval.

Citation