I will describe a comprehensive R package we have developed for fitting
finite mixture distributions to grouped data. The component distributions
may be normal, lognormal, gamma, Weibull, binomial, negative binomial,
Poisson, or Gumbel. The grouped data may have missing or censored bins and
may include conditional as well as mixed data. Maximum likelihood estimates
are calculated by a combination of EM and quasi-Newton algorithms.
Ill-conditioned problems can be resolved by constraints on the parameters.
Graphical methods include the rootogram as a diagnostic tool. This talk will
illustrate the logic and functionality of Rmix with a number of examples
from different areas of application.
This work is joint with Juan Du and Ruochu Gao.
About the Speaker
Peter Macdonald is professor of statistics at McMaster University and
co-ordinator of the Graduate Program in Statistics. His major research
interests are in applied statistics, biological applications of statistics
and analysis of mixture distributions. Prof. Macdonald did his doctoral
work in Biomathematics in Oxford under the supervision of M.S. Bartlett.
Since returning to Canada in 1971 and coming to McMaster, Prof. Macdonald
has been very active in the statistical community including being president
of the Statisitcal Society of Canada in 1990/91.
References
See Dr. Macdonald's
home page for the MIX software to find out more
about mixture models and the software package described in the talk
including a bibliography.