On reparameterization of random effects in linear mixed models

Abstract

The Empirical Best Linear Unbiased Predictor of random effects in linear mixed models may be non-unique. For fixed effects two approaches are used to derive unique solutions – one is based on using estimable linear combinations of parameters and the other one uses reparametrization constraints. It is shown that both approaches can be applied in a similar manner to derive unique prediction results for random effects.