ABSTRACT:Cross-correlated data occur in multi-sample studies with a fully crossed design. An important type of binary cross-correlated data results from multi-reader diagnostic imaging studies where each of several readers independently evaluates the same sample of subjects for the presence or absence of a specific condition (e.g. disease). Such studies are used during development, optimization and regulatory approval of medical technologies, which can eventually be implemented for diagnostics and screening in clinical practice. The analysis of the fully crossed studies can be challenging because of the need to address both reader and subject variability and the related correlation structure. Generalized Linear Mixed Models (GLMM) are implemented in standard statistical software and offer a natural tool for the analysis of the cross-correlated data in the presence of covariates. However, performance of GLMMs in cross-correlated binary data from typical multi-reader studies is generally unknown and is questionable due to the specifics of the available estimation approaches.In the first part of the dissertation we investigate the standard built-in GLMM methods for cross-correlated binary data with and without covariates and explore simple combinations of built-in estimation techniques to overcome existing deficiencies. In the second part, we propose a half-marginal GLMM approach which offers superior interpretation in the context of multi-reader studies of diagnostic accuracy. Our investigation of this model demonstrates good quality of statistical inferences in typical scenarios, but indicates possible large-sample problems stemming from the pseudo-likelihood estimation approach. In the third part of the dissertation we develop an explicit approach for estimating half-marginal model parameters without using pseudo-likelihood. The consistent fixed-effect estimator and its variance are evaluated in a simulation study. The proposed approach can be implemented using non-iterative combination of results from robust GEE models and, for simple scenarios, provides estimates that are equivalent to the empirical estimates.
Last Updated On Tuesday, April 04, 2017 by Valenti, Renee Nerozzi
Created On Tuesday, March 14, 2017
Get a look at recent events at Pitt Public Health.
Update Your Profile