Abstract:
|
Many statistical analyses involving multivariate multilevel data are non-trivial because of uncontrollable missing data. When missing data exist only for the response variable, standard procedures (e.g., HLM, PROC MIXED) can be employed, as they allow for imbalance or missing data. These procedures, however, do not accommodate truly multivariate responses and missing covariates, and operate under standard error assumptions of hierarchical linear models. This work presents variety of computational techniques for model fitting and creating multiple imputations of covariates and responses in multilevel data applications. These techniques, based on iterative algorithms, implement appropriate data augmentation(DA) scheme depending on the application. Our parameter estimation algorithms, for example, adapt a strategy that augments observed data by missing data only, whereas algorithms for creating multiple imputations follow a DA scheme that involves missing data and random effects-covariances as auxiliary variables. Our algorithms operate under a flexible class of mixed-effects models that preserve random variation in means and covariances across clusters. Data examples will be provided.
|