A traditional Gaussian hierarchical model assumes a nested multilevel structure for the mean and a constant variance at each level. We propose a Bayesian multivariate multilevel factor model that assumes a multilevel structure for both the mean and the covariance matrix. That is, in addition to a multilevel structure for the mean we also assume that the covariance matrix depends on covariates and random effects. This allows to explore whether the covariance structure depends on the values of the higher levels and as such models heterogeneity in the variances and correlation structure of the multivariate outcome across the higher level values. The approach is applied to the three-dimensional vector of burnout measurements collected on nurses in a large European study to answer the research question whether the covariance matrix of the outcomes depends on recorded system-level features in the organization of nursing care, but also on not-recorded factors that vary with countries, hospitals, and nursing units. Simulations illustrate the performance of our modeling approach.

, , , ,
doi.org/10.1002/sim.6062, hdl.handle.net/1765/72229
Statistics in Medicine
Department of Biostatistics

Li, B., Bruyneel, L., & Lesaffre, E. (2014). A multivariate multilevel Gaussian model with a mixed effects structure in the mean and covariance part. Statistics in Medicine, 33(11), 1877–1899. doi:10.1002/sim.6062