Missing Data in clinical studies

摘要

This is a comprehensive treatment of practical methods for handling missing data from a non-Bayesian perspective, written by authors who have made substantial contributions to the field over the last 10 years or so. Their central viewpoint is to employ likelihood methods to handle missing value problems, and to shun methods such as complete-case analysis, last-observation-carried-forward, and standard imputation. All these methods are covered, and some of their pitfalls revealed, in comparison to their favoured methods. Likelihood analysis focuses on the actual observations, not on pseudo-observations generated, say, by imputation. Furthermore, the authors demonstrate how a likelihood approach is valid under quite light assumptions such as data missing at random (MAR).