The FMM Procedure (Experimental)

Since the FMM procedure fits finite mixtures of generalized linear models, it can also fit standard forms of these models in which the distribution of the data does not follow a mixture. This enables you to use the FMM procedure to estimate parameters in models that can be fit with the CATMOD, LOGISTIC, GENMOD, or GLIMMIX procedures. However, the FMM procedure does not fit models for multinomial data or models with random effects.

The FMM procedure has limited postprocessing capabilities compared to some other statistical procedures that are based on linear models. Concepts that are well understood and commonplace in linear models, such as (linear) estimable functions, estimability, and least squares means, do not apply to mixture models in the same way. For example, even the computation of a predicted value is not without ambiguity. You can estimate the means in the component distributions in addition to the overall mean of the mixture.

The FMM procedure provides a limited number of built-in distributions and link functions. User-defined distributions or link functions are not supported. Mixture models with component distributions that are not supported by the FMM procedure can be fit with the NLMIXED procedure.

For Bayesian estimation, the FMM procedure implements a small number of highly specialized sampling algorithms. These algorithms are very efficient and specifically designed for generalized linear models and their mixtures. This limits, for example, the allowable specifications for prior distributions of the model parameters. Models that do not fit the targeted algorithms of the FMM procedure can be fit with the MCMC procedure.