Simple and Effective Measures to Quantify Nonignorable Selection Bias

• Activity Number: 187
• Type: Contributed
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Hui Xie* and Donald Hedeker and Robin Mermelstein
• Companies: and The University of Chicago and University of Illinois at Chicago
• Keywords: Missing data ; sensitivity index ; clinical trial ; EMA ; survey

Abstract

Evidence-based health studies and policy making depend heavily on relevant and proper data. However, despite the best efforts to collect representative samples to infer about underlying populations, the available data are often plaqued by various selection issues such that a subset of ideal data is systematically excluded due to particular attributes (e.g., missing data or sample truncation). In particular, selection due to unobserved values renders the resulting observed data unrepresentative of the underlying populations in an unknown way, which threatens the validity of standard analyses designed for random samples by introducing potentially significant selection bias into these regular analyses. There is an increasing interest in developing simple and principled measures to quantify such bias. We develop simple measures to achieve this goal. As compared with exiting index methods, our methods provides adequate quantification of selection bias in much broader situations and applications. We illustrate new measures in both clinical trials and survey studies using modern real-time data capturing methods.