Interpreting Parameter Estimates That Distinguish Multilevel Effects in Oral Health Research

• Activity Number: 253
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): David M Thompson*+ and Julie A Stoner and Jeffrey B Payne
• Companies: University of Oklahoma Health Sciences Center and University of Oklahoma and University of Nebraska
• Keywords: correlated data ; dental data ; generalized estimating equations ; longitudinal data ; multilevel data ; nested data

Abstract

In analyzing longitudinal data, careful parameterization permits a linear model to separate cross-sectional (cohort) and longitudinal (within-subject) effects. Models that do not separate the effects assume they are equal and that, graphically, regression functions estimated for "early" and "late" subject cohorts are collinear. To separately estimate the effects, a useful parameterization subtracts a subject-specific baseline value from each within-subject covariate. This shifts the origin of the metric for the within-subject covariate. A related parameterization centers within-subject effects on subject-specific means. The "centered" parameterization can be extended to multilevel data, including longitudinal and hierarchical data structures. This study explores an extended parameterization, accounting for cluster, subcluster and longitudinal effects in data from a trial of periodontal disease that includes observations across five time points at 56 sites within a subject's mouth. Graphical approaches illuminate how the extended parameterization's estimates can be interpreted in complex multilevel designs. (Supported by NIDCR grants R01DE12872, R03DE019805-01A1)