Modeling of the ARMA random effects covariance matrix in logistic random effects models

Logistic random effects models (LREMs) have been frequently used to analyze longitudinal binary data. When a random effects covariance matrix is used to make proper inferences on covariate effects, the random effects in the models account for both within-subject association and between-subject variation, but the covariance matix is difficult to estimate because it is high-dimensional and should be positive definite. To overcome these limitations, two Cholesky decomposition approaches were proposed for precision matrix and covariance matrix: modified Cholesky decomposition and moving average Cholesky decomposition, respectively. However, the two approaches may not work when there are non-trivial and complicated correlations of repeated outcomes. In this paper, we combined the two decomposition approaches to model the random effects covariance matrix in the LREMs, thereby capturing a wider class of sophisticated dependence structures while achieving parsimony in parametrization. We then used our proposed model to analyze lung cancer data.