Approaches to Modeling Self-Rated Health in Longitudinal Studies: Best Practices and Recommendations for Multilevel Models

Isaac Sasson, University of Texas at Austin

Self-rated health (SRH) is a key measure in the study of population health with proven external validity in predicting mortality. Nevertheless, failing to address the measure’s ordinal scale in statistical analyses poses a potential threat to internal validity. Despite the advent of rich panel data, sociologists have generally been slow in adopting longitudinal methods for ordinal outcomes, and many discard valuable information in favor of simpler methods. This paper reviews and contrasts several approaches to modeling SRH in longitudinal studies under the generalized linear mixed model framework. Model performance is compared (e.g., linear versus nonlinear, conditional versus marginal) using simulation and data from the Health and Retirement Study. Findings suggest that conditional cumulative-logit models provide more statistical power than their linear counterparts, but result in similar substantive conclusions. By contrast, dichotomizing SRH significantly reduces power and is ill-advised. The paper concludes with recommendations for modeling ordinal outcomes in longitudinal studies.