Methods for the Analysis of Longitudinal Data

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Longitudinal studies play a prominent role in all areas of medicine and public health. They are important in public health in understanding the development of chronic illness, and in studying the factors that alter the course of disease development. In addition, longitudinal data on markers of diseases progression and quality of life are an increasingly important component of clinical trials of therapies for cancer and AIDS. In longitudinal studies in the health sciences, missing data are the rule, not the exception. Longitudinal studies frequently suffer from dropouts and/or intermittent non-response. The primary goal of this project is to address the statistical problems that arise when non-response depends on the specific value that in principle should have been obtained, i.e., when non-response is non-ignorable. Much of the proposed work is based on a class of joint models for the longitudinal outcomes and the non-response indicators known as pattern mixture models. Pattern-mixture models offer many important advantages, e.g., they can be easily implemented and it is straightforward to characterize model assumptions. However, pattern-mixture models have one very important drawback that has, so far, limited their usefulness in many areas of application: the natural regression parameters of usual scientific interest are not available. With the goal of overcoming this acknowledged shortcoming of pattern-mixture models, the focus of Specific Aims 1-3 is on developing and implementing new methods for handling non-ignorable dropout and intermittent missingness by adopting "marginalized" parameterizations for the conditional distribution of the repeated measures given non-response patterns. Specific Aim 4 focuses on developing pattern-mixture models for non-ignorable non-response when it depends on extraneous covariates rather than directly upon unobserved responses. Finally, Specific Aim 5 focuses on developing approximate, but computationally simple, methods for handling ignorable non-response in the analysis of repeated binary data. The statistical methods to be developed will be applied to clinical trial data for schizophrenia and contraceptive therapies, and to data from a longitudinal epidemiological study of obesity in children; however, the methods will be equally useful in a much broader range of applications where longitudinal data commonly arise.