|
Activity Number:
|
527
- Diagnostic Tests: Student Papers and Correlated Data
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 1, 2018 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Medical Devices and Diagnostics
|
|
Abstract #330482
|
Presentation
|
|
Title:
|
A New Half-Marginal Approach for Analyzing Cross-Correlated Binary Data from Multi-Reader Studies of Diagnostic Accuracy
|
|
Author(s):
|
Andriy Bandos* and Yuvika Paliwal
|
|
Companies:
|
Teva Pharmaceuticals and University of Pittsburgh
|
|
Keywords:
|
GLMM;
binary;
cross-correlated;
diagnostic
|
|
Abstract:
|
A standard approach for analyzing cross-correlated data is based on the Generalized Linear Mixed Models (GLMM) with crossed random effects. For models with non-linear link this implies a "subject-specific" type of the coefficients, which are not the primary targets for typical multireader studies of diagnostic accuracy and do not agree with the empirical estimates. We propose a half-marginal GLMM which offers a more natural parametrization for multireader studies. The model's parameters can be estimated using the pseudo-likelihood (PL) approach implemented in some GLMM packages. However, the half-marginal PL estimates for binary data are questionable and often considered non-probabilistic. To support the model's validity we have developed a new semiparametric approach for consistently estimating half-marginal coefficients based on probability models for individual readers. The half-marginal model's performance is investigated in a simulation study and illustrated on example of a diagnostic-accuracy study in radiology. Both fitting approaches offer accurate estimates which are naturally interpretable and agree with the empirical summaries for multireader diagnostic accuracy studies
|
Authors who are presenting talks have a * after their name.
