Principal Stratification Designs To Estimate Outcomes Missing Due to Death

• Activity Number: 320
• Type: Topic Contributed
• Date/Time: Tuesday, August 8, 2006 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Constantine Frangakis*+ and Donald Rubin and Ming-Wen An
• Companies: Johns Hopkins Bloomberg School of Public Health and Harvard University and Johns Hopkins University
• Keywords: causal inference ; censoring by death ; missing data ; potential outcomes ; principal stratification

Abstract

We address studies of individuals after a critical event, such as injury, to measure outcomes about the period preceding the event. These studies aim to characterize the outcomes and predict the critical events based on these outcomes. The outcomes, however, are missing for the individuals who die after the critical event. Standard methods cannot estimate such missing outcomes accurately before death if these outcomes differ from observed strata. To address these aims, we propose a design using data on externally controlled factors that could prevent deaths. Based on this design, we develop a method that allows the missing outcomes to differ from any observed strata and still estimate the missing outcomes using the framework of principal stratification. Using injury data, we show that our design and method can uncover dramatically different results from those of standard methods.