Bayesian Nonparametric Analysis of Longitudinal Data with Ordinal Outcomes and Non-Monotone Non-Ignorable Missingness

• Activity Number: 122 - Clinical Trial Design and Missing Data
• Type: Contributed
• Date/Time: Monday, July 30, 2018 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Yu Cao* and Nitai Mukhopadhyay
• Companies: Virginia Commonwealth University and Virginia Commonwealth University
• Keywords: Missing not at random; Ordinal outcomes; Bayesian nonparametric analysis ; Non-monotone missingness

Abstract

Missing outcomes are common in longitudinal studies with repeated measures. Ignoring the missingness may lead to biased estimates, especially when it is missing not at random (MNAR). Likelihood-based imputation methods are often used to handle this type of missingness. Although research has been done when the missingness is monotone (i.e. a subject is not measured again once he/she leaves the study), little work has been done when the missingness is non-monotone. Previously, we proposed a method that imputes the missing continuous outcomes using the estimated likelihood of the observed data and evaluated it using pattern-mixture models in the nonparametric Bayesian framework. This method used a latent class analysis to group the patterns of missingness, and extrapolated the likelihood of unobserved data conditioning on the latent classes of the observed data. In this study, we extended this method to handle ordinal outcomes using a normal ogive model that transforms the ordinal data into latent continuous scores. The results showed improved accuracy in dealing with ordinal outcomes with missing values.