LIKELIHOOD IMPUTATION

The method of likelihood imputation is devised under the framework of latent structure models where the observation is a statistic of the co mplete data which can only be specified on a latent basis. The imputed data set is chosen to differ least from the observed one in their inf ormation contents - a concept with general implications for the analys is of incomplete-data. In contrast to the standard conditional-mean si ngle imputation, our procedure depends on an entire likelihood region instead of any single point in it, and yields consistent parameter est imators nevertheless. We explain its implementations and illustrate wi th data from panel surveys and linear regression with censorship. We a lso discuss its potentials in sensitivity analysis.