Conditioning in a Missing Data Problem

摘要

Observations are recorded on variables x and y but a mechanism, which may depend on the observed x values, causes some of the y values to be missing. For three parametric examples, exact or approximate ancillary statistics are constructed. Conditioning on these ancillaries enables the missing data mechanism to be ignored under certain conditions. A correspondence is shown between these conditional procedures and the use of the observed information matrix in measuring the dispersion of the maximum likelihood estimator.