The MBC Procedure

The traditional goal of clustering has been to identify groups of observations that are similar by some measurement. Standard clustering methods achieve this primarily through heuristic methods, by using a measurement such as the distance to a center or a boundary. These methods appeal to intuition, but they cannot answer fundamental questions about the number of clusters, the suitability of certain cluster structures, or the handling of outliers. In addition, standard methods do not have well-defined statistical properties and so do not support formal inference.

In contrast, model-based clustering (MBC) models the observations by using a mixture of specific distributions. In this framework, the qualities of these clusters and cluster membership are parameters to be estimated by formal statistical methods.

The MBC procedure implements model-based clustering by using mixtures of multivariate Gaussian distributions. PROC MBC allows for a noise component and automatic model selection through the use of information criteria. In addition, it provides posterior scoring of new input data.