Latent Variables via Data Augmentation

In order to fit finite Bayesian mixture models, the FMM procedure treats the mixture model as a missing data problem and introduces an assignment variable as in Dempster, Laird, and Rubin (1977). Since is not observable, it is frequently referred to as a latent variable. The unobservable variable assigns an observation to a component in the mixture model. The number of states, , might be unknown, but it is known to be finite. Conditioning on the latent variable , the component memberships of each observation is assumed to be known, and Bayesian estimation is straightforward for each component in the finite mixture model. That is, conditional on , the distribution of the response is now assumed to be . In other words, each distinct state of the random variable leads to a distinct set of parameters. The parameters in each component individually are then updated using a conjugate Gibbs sampler (where available) or a Metropolis-Hastings sampling algorithm.

The FMM procedure assumes that the random variable has a discrete multinomial distribution with probability of belonging to a component ; it can occupy one of states. The distribution for the latent variable is


where denotes a conditional probability density. The parameters in the density denote the probability that takes on state .

The FMM procedure assumes a conjugate Dirichlet prior distribution on the mixture proportions written as:


where indicates a prior distribution.

Using Bayes’ theorem, the likelihood function and prior distributions determine a conditionally conjugate posterior distribution of and from the multinominomial distribution and Dirichlet distribution, respectively.