Missing data and censored data are often present in observational studies, such as Customer Value Analysis (CVA), and designed experiments, such as in lifetime testing for quality improvement. Analyzing missing and/or censored data is challenging because, in addition to the problems that arise in analyzing complete data, it requires a model of missing-data mechanism and new software for fitting models and analyzing results. The latter often creates a tremendous programming burden, because analyzing incomplete data typically requires iterative methods, such as the EM algorithm for maximum likelihood estimation and MCMC methods for Bayesian estimation.

Our research program in this area takes two main approaches.

1. Direct Modelling: Such examples include

   • the Bayesian model for CVA project (Clark et al, 1998); and

   • analysis of censored data from fractionated experiments using covariance adjustments (Liu and Sun, 1998, Technical Report in preparation).

2. Multiple Imputation: The technique of Multiple Imputation of Rubin (1987) provides a simple way of analyzing incomplete data. For example,

   • Lorraine Denby is currently using this technique to understand and model the relationship among various attributes/questions in the CVA project; and

   • Andrew Gelman, Gary King, and Chuanhai Liu (1998) used this technique for analyzing the data from multiple sample surveys in a study of pre-election public opinion polls, in which not all the questions of interests are asked in all the surveys.