Nonparametric Spatial Models for Periodontal Disease Data with Spatially Varying Non-Random Missingness

• Activity Number: 597
• Type: Topic Contributed
• Date/Time: Thursday, August 2, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Dipankar Bandyopadhyay*+ and Brian Reich and Howard Bondell
• Companies: and North Carolina State University and North Carolina State University
• Keywords: non-stationary ; non-random missingness ; spatial association ; skewed ; Bayesian

Abstract

Clinical attachment level (CAL) is measured at 6 locations per tooth throughout the mouth (excluding the molars), and is the most important marker for periodontal disease. However, analysis of CAL data comes with several interesting challenges. CAL progression is hypothesized to be spatially-associated, i.e. a diseased site might influence the decay of a group of neighbors. Besides, it can be non-stationary, usually with a symmetric distribution with small variance in the anterior (front) of the buccal cavity, and a skewed distribution with large variance in the posterior (back). Also, absent (missing) teeth are common and cannot be considered 'missing at random' since periodontal disease is the leading cause of adult tooth loss. In this talk, we propose a flexible Bayesian modeling framework which allows for a varying, non-Gaussian CAL distribution at each location. Rather than a simple regression on some one-number summary, we exploit the spatial referencing to improve upon the fixed-effects estimates and jointly model the observed CAL and location of an absent tooth allowing for complex missing data mechanisms that can vary in various regions of the mouth.