Methods for Studying Variability as a Predictor of Health Status

As longitudinal data have become widely available in the past 20 years, methods have been developed to analyze mean outcomes or profiles over time (Breslow and Clayton 1993, Laird and Ware 1982, Zeger and Liang 1986). Thus epidemiologic cohort studies often measure both risk factors and disease outcomes repeatedly over time. To evaluate the effect of risk factors on the subsequent development of disease, it is often necessary to calculate summary measures of these factors to capture features of the risk profile. These methods typically consider correlations among repeated measurements on subjects as nuisance parameters. However, there is evidence that the residual variability in subject measures after accounting for subject profiles may sometimes predict future health outcomes of interest (Carroll 2003). The goal of this research is to develop novel methods to decompose within-subject variability into short-term and long-term variance measures, and to combine variance structures with mean structures such as mean longitudinal profiles to better understand the relationships between risk factors and health outcomes.