Examining Assumptions About Missing Data Using Prior Distributions

• Activity Number: 87
• Type: Invited
• Date/Time: Monday, August 12, 2002 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section*
• Author(s): Daniel Scharfstein*+
• Affiliation(s): Johns Hopkins University

Abstract

We examine sensitivity to various assumptions about the missing data through formulation of prior distributions on the unidentified parameters. We build priors for both the complete data distribution and for the selection bias parameters and examine the interplay of these priors on final inferences about the parameters of interest. A strong prior on the complete data distribution alludes to the setting of parametric selection models (Diggle and Kenward, 1994) and a strong prior on the selection bias parameters with a weak prior on the complete data distribution alludes to the work of Scharfstein et al. (1999). This approach will allow us to jointly examine the sensitivity of inferences to both distributional assumptions on the complete data and the missing data mechanism explicitly. We illustrate our approach using data from an AIDS clinical trial.