Analysis of Longitudinal Smoking Outcomes with Missing Data Under a Simple Nonignorable Multiple Imputation Model

• Activity Number: 279
• Type: Invited
• Date/Time: Tuesday, August 2, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Donald Hedeker*+ and Hakan Demirtas and Robin J. Mermelstein
• Companies: University of Illinois at Chicago and University of Illinois at Chicago and University of Illinois at Chicago
• Keywords: missing data ; binary outcomes ; multiple imputation ; longitudinal data

Abstract

We consider the problem of missing smoking outcomes in a longitudinal two-group design. In this situation, some researchers deterministically recode the missing data to smoking, or assume that the missing data are equal to the last available smoking outcome for a subject (i.e., last observation carried forward, or LOCF). Building on our work for data at a single timepoint (Hedeker, Mermelstein, & Demirtas, 2007, Addiction), we describe how these deterministic assumptions can be relaxed by allowing missingness to be imperfectly related to the smoking outcome, and stratified on past patterns of the smoking outcome. Thus, one can examine the robustness of study findings to the assumed strength of the relationship between missingness and smoking. Our approach uses multiple imputation to take into account the uncertainty inherent in the imputed data. We illustrate the methods using data from a smoking cessation study, and describe computer syntax to perform the analyses.