This is an old version, view current version.

5.2 Latent discrete parameterization

One way to parameterize a mixture model is with a latent categorical variable indicating which mixture component was responsible for the outcome. For example, consider \(K\) normal distributions with locations \(\mu_k \in \mathbb{R}\) and scales \(\sigma_k \in (0,\infty)\). Now consider mixing them in proportion \(\lambda\), where \(\lambda_k \geq 0\) and \(\sum_{k=1}^K \lambda_k = 1\) (i.e., \(\lambda\) lies in the unit \(K\)-simplex). For each outcome \(y_n\) there is a latent variable \(z_n\) in \(\{ 1,\dotsc,K \}\) with a categorical distribution parameterized by \(\lambda\), \[ z_n \sim \textsf{categorical}(\lambda). \]

The variable \(y_n\) is distributed according to the parameters of the mixture component \(z_n\), \[ y_n \sim \textsf{normal}(\mu_{z[n]},\sigma_{z[n]}). \]

This model is not directly supported by Stan because it involves discrete parameters \(z_n\), but Stan can sample \(\mu\) and \(\sigma\) by summing out the \(z\) parameter as described in the next section.