|
Title
|
|
Global Sensitivity Analysis in Missing Data -- Some Examples
|
|
Date / Time / Room
|
Sponsor
|
Type
|
08/12/2002
10:30 AM -
12:20 PM
Room: H-Concourse B
|
Biometrics Section*, Biopharmaceutical Section*, ENAR
|
Invited
|
|
Organizer:
|
Huiman X. Barnhart, Rollins School of Public Health of Emory University
|
|
Chair:
|
M. Elizabeth Halloran, Emory University
|
|
Discussant:
|
11:50 AM - Andrea Rotnitzky, Harvard University
|
|
Floor Discussion
|
12:05 PM
|
Description
Missing data are ubiquitous in statistical analysis. In the past decade, there have been a lot of activities in modeling missing data while accounting for the missing data mechanism. However, most of the methodology for missing data relies on some assumption in the missing data mechanism to achieve identifiability, e.g., missing at random or non-ignorable with some distributional form. The assumptions may be very plausible but ultimately are not testable or checkable. Lack of a priori knowledge about the true missing data mechanism is often recognized in the current literature and is frequently handled using sensitivity analyses where several plausible models are contrasted. However, the current methodology for sensitivity analysis is often implemented in an ad-hoc fashion. The most common approach to sensitivity analysis is to check how the parameter of interest changes as the selection parameter in the assumed selection model for missing data changes over a plausible range of values. This type of sensitivity analysis can be considered as a local sensitivity analysis because it only assesses the sensitivity of the inference around the assumed identifiable model. This session will emphasize the importance of sensitivity analysis in missing data problem and to further develop a framework and concept of global sensitivity analysis with presentations that illustrate systematic sensitivity analyses.
|
