Missing Data Methods for Non-Random Attrition in Longitudinal Studies

点击翻译按钮获取中文摘要

This project will develop new methods for multivariate longitudinal data with nonrandomly missing data such as arise with attrition from the sample. Maximum likelihood and Bayesian inference will be derived for a variety of pattern- mixture models, which stratify the data by the pattern of missing data and identify parameters by exploiting assumptions about the missing-data mechanism. Methods will be developed for general patterns of missing data, normal repeated measures models, multivariate t models that accommodate outlying values, and data involving mixtures of continuous and categorical variables. The methods will be compared with existing methods, including those based on stochastic censoring models. Tests for the type of missing-data mechanism will also be developed. Many important scientific studies involve repeated measures of subjects over time. A common problem in analyzing such studies is that some subjects are missing some of their measurements, because they miss visits or do not stay in the study to the end. Discarding such incomplete cases is wasteful, and can lead to erroneous conclusions when the cases that are completely observed differ systematically from those that are incomplete. This grant studies methods of analyses that include all the data and incorporate a variety of reasons for why values are missing.