Daniel F. Heitjan, Columbia University

Ignorability and the Coarse-Data Model

When a fraction of the data are missing, it is common practice to act as though there had never been any intention of observing the units that were missed. This does not spoil inferences, provided the missing data are "ignorable". In many situations--for example, censored survival data--we do not know the precise values of the observations, but they are not exactly missing either, so it is not clear precisely how to express the idea of ignorability. I will describe the coarse-data model, a general model for parametric inference that permits us to extend ignorability to a broader range of incomplete-data problems. If time permits, I will describe the relationship of ignorability to the concepts of sufficient and ancillary statistics.