2004 年 25 巻 2 号 p. 89-116
Missing data is a prevalent complication in the analysis of data from longitudinal studies, and remains an active area of research for biostatisticians and other quantitative methodologists. This paper reviews several statistical methods that are used to address outcome-related drop-out. We begin with a review of important concepts such as missing data patterns, missing data mechanisms, ignorability and likelihood-based inference, which were originally proposed by Rubin (1976, Biometrika 63, 581-592). Secondly, we review the simple analysis methods for handling drop-outs such as a complete-case analysis, an available data analysis and a last observation carried forward analysis, and their limitations are given. Thirdly, we review the more sophisticated approaches for handling drop-outs, which take account of the missing data mechanisms in the analysis. Inverse probability weighted methods and multiple imputation methods, which represent two distinct paradigms for handling missing data, are reviewed. The analysis methods for non-ignorable drop-outs are also reviewed. Three approaches, selection models, pattern mixture models and latent variable models are presented. We illustrate the analysis techniques using the longitudinal clinical trial of contracepting women reported by Machine et al (1988, Contraception 38, 165-179). We briefly review the analysis methods in the presence of missing covariates. Finally, we give some notice in the analysis of missing data.