A Hidden Markov Model for Nonignorable Nonmonotone Missing Longitudinal Data for Medical Studies of Quality of Life

• Activity Number: 20
• Type: Topic Contributed
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Andrea Troxel*+ and Kaijun Liao
• Companies: Univ of Pennsylvania School of Medicine and University of Pennsylvania School of Medicine
• Keywords: pseudolikelihood ; hidden Markov model ; Backward-forward ; selection model ; shared parameter model

Abstract

We present a latent process approach for the analysis of longitudinal data with non-ignorable and non-monotone missingness. Hidden Markov models are widely used for applications in speech recognition, handwriting, bioinformatics, and gene finding and profiling. Multi-state Markov models are widely used to model disease progression and cancer screening. The hidden Markov model is a powerful extension of the multi-state Markov model for longitudinal studies that assumes the states are unobserved. Incorporating this approach with selection models and shared parameter models, we can identify differences among disease processes with incomplete data simultaneously in both the state-dependent model and missingness mechanism model. We propose the models in a generalized linear model and generalized linear mixed model framework, using a backward-forward algorithm to provide efficient parameter estimation in the general situation of non-ignorable non-monotone longitudinal missing data. A two stage pseudo-likelihood method is used to reduce the parameter space to make the model more attractive. We provide an example of cognitive functioning in patients with brain cancer.