Correcting for verification bias in ROC estimation with covariates
Abstract
The ROC curve is a commonly used summary measure of the accuracy of a diagnostic test. It is a plot of the true positive fraction, sensitivity, against the false positive fraction, one minus specificity, for increasingly stringent positivity criterion. Extensive methodology for estimating the ROC curve exists in the situation that patients who receive disease status verification represent a random sample of those who receive the diagnostic test. Bias can occur in the estimation of ROC curves if only some of the patients given the diagnostic test are selected for disease verification. If the probability of selection for verification depends on the diagnostic test result and/or covariates, then the verified sample is not a simple random sample. The bias that possibly arises from using only the verified subjects in the analysis is known as "verification bias". In this thesis, methods for deriving the maximum likelihood estimate for an ROC curve adjusted for variables affecting the likelihood of being verified, is given. Procedures which adjust for a verification mechanism which depends on the diagnostic test have been developed. Our methodology extends these procedures to adjust for "verification bias" when verification depends not only on the subject's diagnostic test result but also on additional covariates. Implementation of these methods are carried out using interactive software which we have developed. Our methods and software are illustrated on data from a two-phase study of dementia disorders; where selection for verification depends on the diagnostic test result, age and site.
Degree
Ph.D.
Advisors
McCabe, Purdue University.
Subject Area
Biostatistics|Mental health
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.