Abstract
Ignoring or inappropriately treating missing data can lead to inefficiency and biased estimation. Three classes of missing data are defined and an overview of various methods for accommodating missing data is presented. A model-based approach to missing data is advanced, and tools for finding maximum- likelihood estimates in this context-in particular, the Expectation-Maximization (EM) algorithm - are described. Strengths and limitations of the EM algorithm are specified. Application of techniques to linear and loglinear models are described, and other possible applications are cited. Implementation is illustrated using epidemiological accident data.
KEY WORDS: