A mean-constrained finite mixture of normals model

A simple constructive approach to imposing a mean constraint in a finite mixture of multivariate Gaussian densities is proposed. All parameters in the model except for one have closed-form full conditional distributions and are fit through a simple Markov chain Monte Carlo algorithm. For illustration, the mean-constrained finite mixture is implemented in a linear mixed model. Simulations reveal that the mean-constrained model is able to precisely estimate the regression coefficients and mean-constrained random effects distribution simultaneously. An analysis of the Framingham cholesterol data shows that, with relatively simple structure, the model has competitive predictive power with earlier approaches.