Optimum Covariate Space in Random Change Point Models

• Activity Number: 421
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Kumar Rajan* and Denis A. Evans
• Companies: and Rush University Medical Center
• Keywords: Change Point Models ; Covariate Space ; Optimum Regions ; Risk Factors ; Random Parameters

Abstract

Covariate effects are often assumed to have a linear form either in continuum or discrete states, and may affect disease progression differently during various stages. Random change point models are used to model longitudinal disease progression mostly defined by two preclinical stages. Linear covariate effects in random change point models have been established previously, but threshold-specific covariate effects are difficult to establish due to the estimation process. Covariate thresholds may also be susceptible to subject-specific variability. We propose a Bayesian approach using Gibbs sampler under non-informative multivariate prior distribution to identify thresholds of systolic blood pressure on cognitive decline preceding Alzheimer's disease (AD) in a population-based sample of 2,149 older adults. Our preliminary findings suggest that for every 10 mm Hg increase in systolic BP above 135 mm Hg, cognitive decline increased by 5% (95% CI= 2-8%) before 6-year change point, but showed no association after change point. Identifying optimum covariate space in nonlinear longitudinal change can provide important information on risk factors and potential intervention components.