|
Abstract:
|
Generalized Linear Mixed Models (GLMMs) provide a powerful built-in tool for analyzing cross-correlated data, especially in presence of covariates. However, conventional GLMMs for binary data are notorious for possible biases. Yet, the quality of statistical inferences are generally unknown for data typically encountered in multireader studies of diagnostic accuracy (where each of several readers evaluates the same sample of subjects for presence/absence of a specific condition). Furthermore, the structure of these studies favors consideration of much less known half-marginal models that substantially reduce the number of random effects by marginalizing over subjects (e.g., for estimating sensitivity). We investigated GLMMs for cross-correlated binary data without covariates, with binary covariates, and with a continuous covariate. Using simulations we evaluated estimation bias as well as the coverage and length of the confidence intervals (CI) for the fixed effects. We demonstrated that unlike subject-specific models, which often resulted in bias and degraded CI coverage, the better-interpretable half-marginal models showed low bias, adequate coverage and shorter length of the CIs
|
