|
Statistical inference with missing data is a very important problem since missing values are frequently encountered in practice. In fact, most statistical problems can be considered incomplete because not all variables are observed for each unit (or possible unit) in a study. For example, randomization in a clinical trial generates missing values since the outcome that would have been observed had a subject been randomized to a different treatment group is not observed. Missing values can be both planned and unplanned. Unplanned missing data can arise when study subjects fail to report to a clinic for monthly evaluations, when respondents refuse to answer certain questions on a questionnaire, or when data is lost. On the other hand, data can be missing by design in a randomized clinical trial or in a Latin square experimental design. Although the problems associated with incomplete data are well-known, they are often ignored, and the analysis is restricted to those observations with complete data. This method of analysis is still the default method in most software packages despite the development of statistical methods that handle missing data more appropriately. In particular, likelihood-based methods, multiple imputation, methods based on weighted estimating equations, and fully Bayesian methods have gained increasing popularity since they have become more computationally feasible in recent years. In this short course, we examine each of these methods in some detail, and compare and contrast them under various settings. In particular, we will examine missing covariate and response data in generalized linear models, random effects models, and survival models. Ignorable missingness as well as nonignorable missingness will be presented for theses models, as well as frequentist and Bayesian methods for analysis. The newly developed statistical package XMISS (Cytel Software) will be used and demonstrated for several real data examples. In addition, live demos of the XMISS software and data analysis using the various models will be given using an LCD projector.
The course presents a balance between theory and applications, and for each class of methods and models discussed, detailed examples and analyses from case studies are presented whenever possible. The applications are all essentially from the health sciences including cancer, AIDS, epidemiology, and the environment. Overall, this course will be applied in nature and will focus on the applications of frequentist and Bayesian methods for research problems arising in the medical sciences. Live demo real data examples will be given using the XMISS software.
|
= Applied Session,
= Theme Session,
= Presenter