Discrete mixtures in Bayesian networks with hidden variables: a latent time budget example

The existing methods of analysis applicable to time budget data are summarised. Latent budget models, a subclass of general reduced rank models for two-way contingency tables, are most appropriate as they view each of the observed conditional distributions of interest as a mixture of a small number of conditional distributions involving a hidden variable. However, they suffer from unusually complex unidentifiability problems which can cause standard estimation methods to perform badly and/or be misleading. Recent advances in estimation methods for this type of mixture model which address the unidentifiability issues are reported and demonstrated. (C) 2002 Published by Elsevier Science B.V.