We consider the problem of estimating an unknown distribution function F in the presence of censoring under the conditions that a parametric model is believed to hold approximately. We use a Bayesian approach, in which the prior on F is a mixture of Dirichlet distributions. A hyperparameter of the prior determines the extent to which this prior concentrates its mass around the parametric family. A Gibbs sampling algorithm to estimate the posterior distributions of the parameters of interest is reviewed. An importance sampling scheme enables us to use the output of the Gibbs sampler to very quickly recalculate the posterior when we change the hyperparameters of the prior. The calculations can be done sufficiently fast to enable the dynamic display of the changing posterior as the prior hyperparameters are varied.
If you want the software, literate program, data files, and everything, you can get it as a tarred and gzipped file. It contains everything so will enable you to reproduce our work. In addition, you get a package of utilities for Lisp-Stat programming that you can use freely. You can get it here (you should probably save the link, since this file is big, about 25MB). If you want a glimpse of what the program can do, here's the README file.
The software uses routines written in C for speed. It works without modification on Windows 95/NT and Unix, except that in the latter case, a configure/make sequence tailors the stuff to the particular flavor of Unix. We don't have a version for the Mac because we don't know how to make a library for dynamic loading on the Mac.
Last modified: Fri Jun 8 14:34:39 EDT 2001