A New Way to Multiply Impute Nonignorable Missing Outcomes

• Activity Number: 283
• Type: Topic Contributed
• Date/Time: Tuesday, August 5, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Shahab Jolani*+ and Stef Van Buuren
• Companies: Utrecht University and Utrecht University
• Keywords: Incomplete data ; Missing not at random ; Propensity score ; Selection function

Abstract

Models for dealing with missing outcomes are necessarily based on restrictive assumptions when the missing data are nonignorable. In order to avoid the often unrealistic normality assumption for the hypothetically complete outcomes, and to avoid choosing arbitrary sensitivity parameters, we adopt a pragmatic Bayesian methodology for estimating regression parameters using multiple imputation. This method is based on fully conditional specification that imputes the missing outcomes and remodels the missingness mechanism in an alternate fashion. The proposed method requires correct specification of the form of the missingness mechanism, up to an unknown parameter that is estimated from the data. The simulation shows that the method is insensitive to failure of the normality assumption, and clearly improves upon the selection model and multiple imputation under missing at random for the cases investigated.