Using Gaussian mixtures with unknown number of components for mixed model estimation


Laurence Watier, Sylvia Richardson & Peter J. Green
Hierarchical mixed models are used to account for dependence between correlated data, in particular dependence created by a group structure within the sample. In such models, the correlation between observations is modelled by including, in the regression model, group-indexed parameters regarded as random variables, so called random effects. Gaussian distributions are commonly used for the random effects. However, this choice places a strong constraint on the shape of the random parameter distribution. In this presentation, we focus on misspecification in mixed model with random intercept, a commonly used model in epidemiology. We propose to model the prior distribution of the random intercept by gaussian mixtures with an unknown number of components in a Bayesian framework. This methodology has recently been developed by Richardson and Green (1997) to analyse heterogeneous data. Another use of gaussian mixtures with unknown number of components is that of density estimation.
Some keywords: Bayesian estimation; Gaussian mixtures; Misspecification; Mixed models.
Back to Peter Green's research page